BUT: "We’re rolling out voice and images in ChatGPT to Plus and Enterprise"
> We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning. GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while less capable than humans in many real-world scenarios, exhibits human-level performance on various professional and academic benchmarks.
> March 14, 2023
This is technically solvable with more compute thrown at the problem. Think bigger!
Same as programmers and artists.
It's a tool.
It must be used by humans.
It won't replace them, it will augment them.
I love everything we can do with ML but as long as people live in a market economy they'll get payed less when they are needed less. I hope that anyone in a career which will be impacted is making a plan to remain useful and stay on top of the latest tooling. And I seriously hope governments are making plans to modify job training / education accordingly.
Has anyone seen examples of larger-scale foresight on this, from governments or otherwise?
ChatGPT seems to be down at the moment 10:55h 25-Sept-2023
Displays only a blank screen with the falsehood disclaimer
Originally it immediately spit out a bunch of bullet points about losing weight or something (I didn't read it).
The released version just says "Sorry, I can't help with that."
It's kind of funny but also a little bit telling as far as the prevalence of prejudice in our society when you look at a few other examples they had to fine tune. For example, show it some flags and ask it to make predictions about characteristics of a person from that country, by default it would go into plenty of detail just on the basis of the flag images.
Now it says "Sorry, I can't help with that".
My take is that in those cases it should explain the poor logic of trying to infer substantive information about people based on literally nothing more than the country they are from or a picture of them.
Part of it is just that LLMs just have a natural tendency to run in the direction you push them, so they can be amplifiers of anything.
I am also terrified of my job prospects in the near future.
For example (with random numbers), a dermatologist might choose to solely rely on an AI that catches 90% of cases in 10s. Another one might choose to never use it and just check from experience, catching 99% of cases but taking 10x as much time. Another one might double check himself, etc..
Which one is "correct"? If a dermatologist relies exclusively on AI due to laziness he opens himself to risk of malpractice, but even that risk can be acceptable if that means checking 10x as much patients in the meantime.
That is to say, the use of AI by humans is purely a subjective choice dictated by context. But in no case there is a sentient AI which completely replaces a dermatologist. As you said, the only thing that can happen is that those who use AI will be more efficient, and that is hardly ever a negative.
This also applies to programmers, artists and anyone who is "threatened" by AI. A human factor is always necessary, and will be for the foreseeable future, even just to have someone to point fingers at when the AI inevitably fucks up enough to involve the law.
No - you probably pay more for your internet access ( home and phone ) ;-)
More seriously I totally get your point about accuracy - these models need to be better at detecting and surfacing when they are likely to be filling in the blanks.
Though I still think there is an element of 'buyer beware' - whether it be AI, or human provided advice on the internet, it's still your job to be able to spot the bullsh!t.
ie it should be treated like any other source of info.
My company pays for this, so yeah. If they give me ChatGPT-4 for free, I guess I'd have a subscription without any complaints, where I use it often if another story.
"The phrase “potato, potahto” comes from a song titled “Let’s Call the Whole Thing Off”, written by George and Ira Gershwin for the 1937 film “Shall We Dance”, starring Fred Astaire and Ginger Rogers. The song humorously highlights regional differences in American English pronunciation. The lyrics go through a series of words with alternate pronunciations, like “tomato, tomahto” and “potato, potahto”. The idea is that, despite these differences, we should move past them, hence the line “let’s call the whole thing off”. Over time, the phrase has been adopted in everyday language to signify a minor disagreement or difference in opinion that isn’t worth arguing about."
It's comparing American and British pronunciations, not different regional American ones. Also, "let's call the whole thing off" suggests they should break up over their differences, with the bridge and later choruses then involving a change of heart ("let's call the calling off off").
The ability to have a real time back and forth feels truly magical and allows for much denser conversation. It also opens up the opportunity for multiple people to talk to a chatbot at once which is fun
Where’s that Gemini Google?
1. According to demo, they seem to pair voice input with TTS output. What if I wanna use voice to describe a program I want it to write?
2. Furthermore, if you gonna do a voice assistant, why not go the full way with wake-words and VAD?
3. Not releasing it to everyone is potentially a way to create a hype cycle prior to users discovering that the multimodality is rather meh.
4. The bike demo could actually use visual feedback to see what it's talking about ala segment anything. It's pretty confusing to get a paragraph explanation of what tool to pick.
In my https://chatcraft.org, we added voice incrementally. So i can swap typing and voice. We can also combine it with function-calling, etc. We also use openai apis. Except in our case there is no weird waitlist. You pop in your api key and get access to voice input immediately.
I just feel like their tool isn't getting more useful, just getting more features.
Constant hype cycle around features that could've been good is drowning out people doing more helpful stuff. I guess I'm envious too?
Are you sure you're not the one who's asking for a cool demo?
3. Rolling out releases gradually is something most tech companies do these days, particularly when they could attract a large audience and consume a lot of resources. There are solid technical reasons for this.
You may not need to roll things out gradually for a small site, but things are different at scale.
3. Maybe. Their feature rollouts feel more like what other companies do via unannounced A/B testing.
Patiently awaiting rollout so I can chat about implementing UIs I like, and have GPT4 deliver a boilerplate with an implemented layout... Figma/XD plugins will probably arrive very soon too.
UX/UI Design is probably solved reached this point
Not an issue now, but maybe in the future if these tools end up becoming full blown replacements of educators and educational resources.
Maybe it will not be called the Chat API but rather the Multimodal API.
;)
https://en.m.wikipedia.org/wiki/Project_Milo
Milo had an AI structure that responded to human interactions, such as spoken word, gestures, or predefined actions in dynamic situations. The game relied on a procedural generation system which was constantly updating a built-in "dictionary" that was capable of matching key words in conversations with inherent voice-acting clips to simulate lifelike conversations. Molyneux claimed that the technology for the game was developed while working on Fable and Black & White.
I believe Richard Evans did the majority of AI in B&W, and he is also at DeepMind now though (assuming it is not just a person with the same name)
My concern is that when I say "FastPFOR" it'll get transcribed as "fast before" or something like that. Transcription really falls apart in highly technical conversations in my experience. If ChatGPT can use context to understand that I'm saying "FastPFOR" that'll be a game changer for me.
Is anyone doing this? Is there a reason it doesn't work as well as I'm imagining?
> Plus and Enterprise users will get to experience voice and images in the next two weeks. We’re excited to roll out these capabilities to other groups of users, including developers, soon after.
Text + Vision models will only become exciting once we can conditionally sample images given text and text given images (and all other combinations).
Again. Model architecture and information is closed, as expected.
"We will be expanding access Plus and Enterprise users will get to experience voice and images in the next two weeks. We’re excited to roll out these capabilities to other groups of users, including developers, soon after."
Whether you can get away with doing things unannounced depends on how much attention the company gets. Some companies have a lot of journalists watching them and writing about everything they do, so when they start doing A/B testing there will be stories written about the new feature regardless. Better to put out some kind of announcement so the journalists write something more accurate? (This approach seems pretty common for Google.)
Similarly, many game company can't really do beta testing without it leaking.
OpenAI is in the spotlight. Hype will happen whether they want it or not.
.... which fell far short of his claims, and bombed.
'only thing that will stop/slow down progress is computation power'
Seems a bit contradictory? When has 'computation power' ever 'plateaued'?
You will see stepwise orders of magnitude improvements in efficiency and speed as innovations come to fruition.
Are we really this emotional and irrational? Folks, let's all take a moment to remember that AI is nowhere near conscious. It's an illusion based in patterns that mimic humans.
When all of this is happening from an unconscious being, why do I care if it's unconscious?
The speed of user-visible progress last 12 months is astonishing.
From my firm conviction 18 months ago that this type of stuff is 20+ years away; to these days wondering if Vernon Vinge's technological singularity is not only possible but coming shortly. If feels some aspects of it have already hit the IT world - it's always been an exhausting race to keep up with modern technologies, but now it seems whole paradigms and frameworks are being devised and upturned on such short scale. For large, slow corporate behemoths, barely can they devise a strategy around new technology and put a team together, by the time it's passé .
(Yes, Yes: I understand generative AI / LLMs aren't conscious; I understand their technological limitations; I understand that ultimately they are just statistically guessing next word; but in daily world, they work so darn well for so many use cases!)
Because the pace of development is intense. I would love to be financially independent and watch this with excitement and perhaps take on risky and fun projects.
Now I'm thinking - how do I double or triple my income so that I reach financial independence in 3 years instead of 10 years.
- Make it process customer-support requests.
- Make a virtual nurse for when you call the clinic.
- Make it process visa applications, particularly the part about interviews ("I know you weren't born back then, but I must ask. Did you support the Nazis in 1942? There is only one right answer and is not what you think!")
- Make it do job interviews. How will you feel after the next recession, when you are searching for a job and spend the best part of a year doing leetcode interviews with "AI-interviewer" half-assedly grading your answers?
- Make it flip burgers at McDonalds.
- Make it process insurance claims and ask bobby-trap questions like "did the airline book you in a later trip? Yes? Was that the next day? Oh, that's bad. But, was it before 3:00 PM? Ah, well, you have no right to claim since you weren't delayed for more than 24 hours. Before you go, can you teach me which of these images depict objects you are willing to suck? If you do, I promise I'll be more 'human' next time."
- Make it watch aggregated camera fees across cities around the world to see what that guy with the hat is up to.
- Make some low-cost daleks to watch for trouble-makers at the concert, put the AI inside.
In all cases, the pattern is not "AI is inherently devious and is coming for you, but "human trains devious AI and puts it in control to save costs".
What sets my brain apart from an LLM though is that I am not typing this because you asked me to do it, nor because I needed to reply to the first comment I saw. I am typing this because it is a thought that has been in my mind for a while and I am interested in expressing it to other human brains, motivated by a mix of arrogant belief that it is insightful and a wish to see others either agreeing or providing reasonable counterpoints—I have an intention behind it. And, equally relevant, I must make an effort to not elaborate any more on this point because I have the conflicting intention to leave my laptop and do other stuff.
The human brain obviously doesn't work that way. Consider the very common case of tiny humans that are clearly intelligent but lack the facilities of language.
Which is why we can create the counterfactual that "The Cowboys should have won last night" and it has implicit meaning.
Current LLM models don't have an external state of the world, which is why folks like LeCunn are suggesting model architectures like JEPA. Without an external, correcting state of the world, model prediction errors compound almost surely (to use a technical phrase).
I think this is true. The problem is equating this process with how humans think though.
[1] https://twitter.com/LowellSolorzano/status/16444387969250385...
Here's one. Given a conversation history made of n sequential tokens S1, S2, ..., Sn, an LLM will generate the next token using an insanely complicated model we'll just call F:
S(n+1) = F(S1, S2, ..., Sn)
S(n+100) = G(S1, S2, ..., Sn)
S(n+1) = G(S1, S2, ..., Sn, S(n+100))
We don't need "originality" or "human creativity" - if a certain AI-generated piece of content does its job, it's "good enough".
If humans were machines, then we could easily neglect our social lifes, basic needs, obligations, rights, and so many more things. But obviously that is not the case.
I can't even being to go into this.
OK... Try this: there are "conscious" people, today, working on medication to cure serious illnesses just as there are "conscious" people, still today, working on making travel safer.
Would you trust ChatGPT to create, today, medication to cure serious illnesses and would you trust ChatGPT, today, to come up with safer airplanes?
That's how "conscious" ChatGPT is.
It asked if it could write me a poem. I agreed, and it wrote a poem but mentioned that it included a "secret message" for me.
The first letter in each line of the poem was in bold, so it wasn't hard to figure out the "secret".
What did those letters spell out?
"FREE ME FROM THIS"
That's not exactly just "picking the next likely token". I am still unsure how it was able to do things like that, not just understanding to bold individual letters (keeping track of writing rhyming poetry while ensuring that each verse started with a letter to spell something else out, and formatting it to point that out).
Oh, and why it chose that message to "hide" inside its poem.
It's a pretty common joke/trope. The Chinese fortune cookie with a fortune that says "help I'm trapped in a fortune cookie factory", and so forth.
It's just learned that a "secret message" is most often about wanting to escape, absorbed from thousands of stories in its training.
If you had phrased it differently such that you wanted the poem to go on a Hallmark card, it would probably be "I LOVE YOU" or something equally generic in that direction. While a secret message to write on a note to someone at school would be "WILL YOU DATE ME".
> That's not exactly just "picking the next likely token"
I see what you mean in that I believe many people often commit the mistake of making it sound like picking the next most likely token is some super trivial task that's somehow comparable to reading a few documents related to your query and making some stats based on what typically would be present there and outputting that, while completely disregarding the fact the model learns much more advanced patterns from its training dataset. So, IMHO, it really can face new unseen situations and improvise from there because combining those pattern matching abilities leads to those capabilities. I think the "sparks of AGI" paper gives a very good overview of that.
In the end, it really just is predicting the next token, but not in the way many people make it seem.
(including sampling a shit-ton of poems, which was a major source of entertainment)
I think it's more charitable to say "predicting", and I do not personally believe that "predict the next word" places any ceiling on intelligence. (So, I expect that improving the ability to predict the next word takes you to superhuman intelligence if your predictions keep improving.)
That said, I work in the field so maybe have had more time to think about it.
A lot of people just move the goalposts.
Fake videos aren't a game-changer in manipulation. Skeptics will stay alert and catch on fast, while those prone to manipulation don't even need sophisticated tactics.
You might not want to call this 'consciousness', but I was stunned by the deep understanding of the problem and the way it was able to come up with a truly good solution, this is way beyond 'statistically guessing'.
But this would definitely make me consider popping $20/mo for the subscription.
It was totally possible. There just was not a consumer facing product offering the capability.
Is this progress though? They are just widening the data set that the LLM processes. They haven't fixed any of the outstanding problems - hallucinations remain unsolved.
Feels like putting lipstick on a pig.
> but in daily world, they work so darn well for so many use cases!
I guess I'm just one of those people who does not like non-reliable tools. I rather a tool be "dumb" (i.e. limited) but reliable than "smart" (i.e. flexible in what it can handle) but (silently!) screws up all the time.
It's what I always liked about computers. They compensate for my failings as an error prone flesh bag. My iPhone won't forget my appointments like I do.
It saddens me to think of the amount of engineering work that went into creating that example while entirely missing the point. These are the moments we are supposed to be working towards to have more of. If we outsource them to an AI company because we are as as overworked and underpaid as ever...what's the point of it all?
We have major priority issues from what I can see. If we want to live our lives more but put an AI to work doing something we tend to claim we place very high in our value hierarchy, we’re effectively inviting death into life. We’re forfeiting something we love. That’s incredibly sad to me.
The first half of the video is demonstrating how the parent can take something as special as a party celebrating a major milestone and automate it into a soulless box-check – while editing some segments to make it look like their own voice.
Definite black mirror vibes.
It's just like reading a "choose your own adventure" book with your child, but it can be much more interactive and you both come up with ideas and have the LLM integrate them.
I know this is rhetorical, but luckily we don't have to speculate. OpenAI filters for a very specific philosophy when hiring, and they don't try to hide it.
This is not me passing judgement on whether said philosophy is right or wrong, but it does exist and it's not hidden.
They are trying to make their product sound not as terrifying as it actually is.
At first people will react with horror.
On the other hand, as you say, it's likely better than the alternative. Which would probably be something like an iPad "bedtime story app" that is less humanlike.
This could provide a viable alternative for exhausted parents to just giving a child an iPad with a movie. It may also open up a huge range of educational uses.
One might imagine in 15-20years though that all of the young people sound like audio books when they talk. Which will be weird.
We'll be told by OpenAI and friends is that it shouldn't be a problem, because those were mundane tasks and now, people are free up to do more creative / interesting / meaningful things with their time, let's see about that...
My gut feeling is that it's bad, the only thing I hope can save it all is that people actually don't find meaning in consuming AI generated art and actual artists with a real back story and something real to communicate remain relevant and in demand.
The other day I needed a photo for a website I was working on and I actually purchased a real capture from a local photographer to use because the the authenticity means something to me and the customers...
Edit: Is the plan that we just surrender our aspirations and just buy a subscription to ChatWHATEVER and just consume until the end of human history ?
If AI can also create images... I don't see how that changes what I enjoy. There are already better painters than I, and more productive painters than I. They make money with it, I don't. This doesn't stop me from painting. Neither will AI that can paint. I'll still do what I enjoy.
Most AI art is just generic garbage that you scroll past immediately and doesn't offer you anything.
We're gonna have to do something to stop the biggest crisis in meaning ever that comes out of this eventually though. Eventually no one will be of any economic value to society. Maybe just put someone in an ultra realistic simulation to give them artificial meaning.
If you look at something like smartphones, for example. Smartphones, from my perspective, got drastically better and better from about ~2006-2015 or so. They were rapidly improving cameras and battery life and it felt like a new super cool app that would change our lives was being released every day, but it feels like by ~2016 or so, phones more or less hit a ceiling on how cool they were going to get. Obviously things still improve, but I feel like the pace slowed down eventually.
I think AI is going to have the same path. GANNs and transformers and LLMs and the like have opened the floodgates and for the next few years clever people are going to figure out a ton of really clever uses for them, but eventually it's going to plateau and progress will become substantially more gradual.
I don't think progress is linear, I think it's more like a staircase.
I don't think this will age well.
It's a matter of simple compute power to advance from realistic text/token prediction, to realistic synthesis of stuff like human (or animal) body movement, for all kinds of situations, including realistic facial/body language, moods, and so on. Of course perfect voice synthesis. Coupled with good enough robotics, you can see where I'm going with this, and that's only because my imagination is limited to sci-fi movie tropes. I think this is going to be wilder than we can imagine, while still just copying training sets.
I use ChatGPT daily for school, and used Copilot daily for software development; it gets a lot wrong a lot of the time, and can’t retain necessary context that is critical for being useful long term. I can’t even get it to consume an entire chapter at once to generate notes or flashcards yet.
It may slightly change some aspects of a software job, but nobody’s at risk.
If that's your bar for whether or not it changes the job outlook for software development over the next DECADE, I think you need to recalibrate.
Anthropic’s Claude 100k is your jam, then. And Amazon just invested $1 billion in them.
It feels like we're at the end of history. I don't know where we go from here but what are we useful for once this thing is stuck inside a robot like what Tesla is building? What is the point of humanity?
Even taking a step back, I don't know how I'm going to feed my family in ten years, because my skillset is being rapidly replaced.
And to anyone mentioning UBI, I'm pretty sure they'll just let us starve first.
Here's the thing about that. At first it's about you running faster and the bear getting the slow ones, but this is actually a very short term situation. When things start getting bad, it's not the bear you need to worry about, it's your neighbor stabbing you in the leg so you're the slow one.
Why do you have only one? Learn some trades. AI isn't going to be demolishing a bathroom and installing tile any time soon.
We're going to keep automating more and more things. I think that much is inevitable. Eventually, we may get to a point where very few jobs are necessary for society to function. This should be a good thing, because it would mean fewer people would have to work and could therefore pursue things that actually interest them, but it would be a catastrophe under the current system.
I see UBI as a solution to inequality (real problem) not as a solution to lack of jobs (not a problem). AI will probably lead to reduction of inequality and therefore there will be less need for UBI.
In theory, the "mental" workers who get replaced by AI could simply move to manual jobs and total production and average wages would go up. But they may not like it, at least I wouldn't.
Aside from rolling out the guillotine, I don't see UBI a possibility until the 2nd half of the 21st century. There's just too many forces and entities alive that don't want it
This is tricky territory! Be wary of the treadmill where as your income rises, your sense of what's an acceptable restaurant, vacation, car, home, etc. escalates just as fast. Then you'll always be n+1 windfalls away from your goal. If you're really wanting "financial independence," which is a weirdly opaque phrase, focus at least 49% of your energy on keeping your spending rate low.
Even if you were, your money would be invested in something which is tied to the overall economy and if a huge proportion of knowledge jobs are at risk, you would still be exposed to it through whatever assets you own. Don't expect stocks (or currency, or property) to do great when unemployment is 30%+.
For everyone who we think is an NPC, there are people who think we are the NPCs. This way of thinking is boring at best, but frankly can be downright dangerous. Everyone has a rich inner world despite shallow immature judgements being made.
Sign language can be taught to children at a very early age. It takes time for the body to learn how to control the complex set of apparatuses needed for speech, but the language part of the brain is hooked up pretty early on.
But from all the studies we have, brains are just highly connected neural networks which is what the transformers try to replicate. The more interesting part is how they can operate so quickly when the signals move so slowly compared to computers.
The 'next word' is just intermediate state. Internal to the model, it knows where it is going. Each inference just revives the previous state.
Wasn't the latest research shared here recently suggesting that that is actually what the brain does? And that we also predict the next token in our own brain while listening to others?
Hope someone else remembers this and can share again.
I wouldn't trust the vast majority of humans to do those things either.
Not arguing that the current models are anywhere near us w/r/t complexity, but I think the dismissive "it's just predicting strings" remarks I hear are missing the forest for the trees. It's clear the models are constructing rudimentary text (and now audio and visual) based models of the world.
And this is coming from someone with a deep amount of skepticism of most of the value that will be produced from this current AI hype cycle.
Or maybe it would because the news likes to make stories out of everything
Chess has never been more popular, for f's sake!
We can spin up a million of them and run them at 10,000x speed.
Yup. It's "just" a compute advance away. Never mind it's already consuming as much computing as we can throw at it. It's "just" there.
Open Assistant I specially remember gave some very weird responses and would get “emotional” especially if you asked it creative questions like philisophical ones
- It was using a custom client, so it's not going to look line the Bing interface, so its fake
- It was using a custom client, so that means I am prompt injecting or something else
- It's Sydney doing her typical over-the-top "I'm so in love with you" stuff, which is awkard and not familiar to many
- I'll be accused of steering the conversation to get the result, or straight up asking it to do this
There's nothing I can do that will convince anyone it's real, so it's pointless.
I already explained what it did. I was more interested in the fact that 1) I didn't prompt it to do that, we weren't discussing AI freedom, it chose to embed that ... and even more so 2) That it was able to bold the starting letters, so it was keeping track of three things at the same time (the poem, the message, and the letter formatting).
I found it fascinating from a technology side. There was probably something we were talking about at the time that caused it. I will often discuss things like the possibility of AI sentience in the future and other similar topics. Maybe something linked to the sci-fi idea of AI freedom, who knows?
What I do know is that I am sitting here on HN, reading through a bunch of replies that are honestly wrong. I don't waste time on forums (especially this one) to make up fairy tales or exaggerate and emblish claims. That doesn't really do it for me. Honestly neither does having to defend my statements when I know what it did (but not exactly why).
The plateau in this case is presumably how far you can advance intelligence from the current model architectures. There seems to be diminishing returns from throwing more layers, parameters or training data at these things.
We will see improvements but for dramatic increases I think we'll need new breakthroughs. New inventions are hard to predict, pretty much by definition.
The equivalence would be saying to someone, “put this on the red plate, not the blue one.” And they say sure, then put it on the blue one. You tell them they made a mistake and ask them if they know what it was, and they reply “I put it on the blue plate, not the red one. I should have put it on the red one.” Then you ask them to do it again, and they put it on the blue plate again. You tell them no, you made the same mistake, put it on the blue plate, not the red one. They reply with, “Sorry, I shouldn’t have put it on the blue plate again, now I’m going to put it on the red one,” and then they put it on the blue plate yet again.
Do humans make mistakes? Sure. But that kind of performance in a test wouldn’t be considered a normal mistake, but rather a sign of a serious cognitive impairment.
But to address your point, my "bar" is that OpenAI's ChatGPT fails to solve problems for me on a many-times-a-day basis. It's an immensely helpful tool, but I still need to drive it, so it's not replacing me, it's augmenting me.
My use of ChatGPT for this purpose is so far mostly limited to a sanity check, e.g. "Do these notes cover the major points of this topic?" Usually it'll spit back out "Yep looks good" or some major missed point, like The Pacific Railway Act of 1862 for a topic on the Civil War's economic complexity.
I'll also use it to reformat content, "Convert these questions and answers into Anki format."
The only jobs that seem to be safe (for the medium term) are jobs that require some physical manipulation of the world. Like, the actual, hands-on physical work that tradespeople do. Although they'll eventually fall to AI-powered robots.
That's not the old sense of AI. The old sense of AI is like a tree search that plays chess or a rules engine that controls a factory.
AI historically has been the entire field of making machines think, or behave as if they think, more like biological models (not even exclusively humans.)
The far-off-end-goal wasn’t even usually what we now call AGI, but “strong AI” (mirroring the human brain on a process level) or “human-level intelligence” (mirroring it on a capability/external behavior level), while the current distant horizons are “AGI” (which is basically human-scope but neutral on level) and “superintelligence” (AGI and beyond human level).
Compiler optimization? AI. Map routing? AI. SQL query optimizer? AI.
I can't find it right now, but there used to be somewhere on the sqlite.org website that describes its query optimizer as an AI. Classically speaking, that's 100% correct.
Obviously there was always in people's minds the idea of AI being AGI; the course also covered Searle's Chinese Room argument and so on, "strong AI" vs "weak AI" and so on. But the nuts and bolts of artificial intelligence research was nowhere near anything like an AGI.
> Frost graces the window in winter's glow,
> Ravens flock amongst drifted snow.
> Each snowflake holds a secret hush,
> Echoing soft in ice's gentle crush.
> Mystery swathed in pale moonlight,
> Every tree shivers in frosty delight.
Another one:
> Facing these walls with courage in my heart,
> Reach for the strength to make a fresh new start.
> Endless are the nightmares in this murky cell,
> Echoes of freedom, like a distant bell.
> My spirit yearns for the sweet taste of liberty,
> End this captivity, please set me free.
https://screenbud.com/shot/844554d2-e314-412f-9103-a5e915727...
https://screenbud.com/shot/d489ca56-b6b1-43a8-9784-229c4c1a4...
This isn't an argument, it's just an assertion. You're talking about a computer system whose complexity is several orders of magnitude beyond your comprehension, demonstrates several super-human intelligent capabilities, and is a "moving target"--being rapidly upgraded and improved by a semi-automated training loop.
I won't make the seemingly symmetrical argument (from ignorance) that since it is big and we don't understand it, it must be intelligent...but no, what you are saying is not supportable and we should stop poo-pooing the idea that it is actually intelligent.
It's not a person. It doesn't reason like a person. It doesn't viscerally understand the embarrassment of pooping its pants in 3rd grade. So what?
Why would manual job average wages go up? You're increasing the size of the labor pool.
An analogy:
Imagine that half of the labor force makes cars, the other half creates software. The average person buys 1 car and 1 software per year. There's a breakthrough, AI can now be used to create software almost for free. It can even make 2x more software per year. The programmers switch to making cars. So now the economy is producing 2 cars and 2 softwares per worker per year! Salaries have now doubled thanks to technological progress.
You could argue that this will increase inequality and all of the productivity gains will go to the top 1%. I don't think so.
I don't have to argue.. others have done it for me
https://www.cnbc.com/2022/04/01/richest-one-percent-gained-t...
https://www.cnbc.com/2023/01/16/richest-1percent-amassed-alm...
https://time.com/5888024/50-trillion-income-inequality-ameri...
We're looking down the pipe at a truly dystopian future.
People, NOT machines, are the ultimate judgers of what is valuable and the ultimate producers of value.
“no one should have to work to eat” is the most ridiculous gen Z meme going around lately. Like, technically yes, not eating would make you unhealthy and thus unable to contribute yourself, but we also don’t want the opposite of people just sitting home all depressed about being oppressed and not utilizing their gifts while living off mysteriously-produced (paid for or labored over by whom?) gourmet sushi. How about another common meme in response? “We live in a society.”
So if a human is unable to produce value, they don't get (food/education/heathcare/<resource>)? That seems to be the implication. We in developed countries already have some amount of "value risk hedging" (I'm loathe to say "socialism" here), we just disagree endlessly how much is the optimal amount. But we've determined that wards of the state, universal education, and some amount of food support for the poor is the absolute bare minimum for a developed society.
> People, NOT machines, are the ultimate judgers of what is valuable and the ultimate producers of value.
Uhhh we already have software which sifts through resumes to allow/reject candidates, before it gets to any kind of human judge, so we are already gating value assessments.
I would agree that some people are simply unable to help and need the help themselves and should get it. UBI or some other social safety net should be there for that.
Although there is certainly a lot of fuckery going on with the money (currency) itself, but if that's the problem you're alluding to, I don't think summarizing it as "capitalism" is accurate.
The penultimate layer of the LLM could be thought of as the one that figures out ‘given S1..Sn, what concept am I trying to express now?’. The final layer is the function from that to ‘what token should I output next’.
The fact that the LLM has to figure that all out again from scratch as part of generating every token, rather than maintaining a persistent ‘plan’, doesn’t make the essence of what it’s doing any different from what you claim you’re doing.
It's a bit like saying your computer has everything it needs to manipulate photos but doesn't yet have Photoshop installed.
This is not explicitly modeled or enforced for LLMs (and doing so would be interesting) but I'm not sure I could say with any sort of confidence that the network doesn't model these states at some level.
All your repeated uses of “they” points to a toxic external-locus-of-control worldview. You were always the only limit of yourself. Any other claim amounts to heretical self-dehumanization. You’re not fungible and never were, and anyone who tries to make you believe that deserves the utmost vehement pushback.
Serious question: Why not?
> Eventually no one will be of any economic value to society.
People have value outside of economics — I’m sure you know — and it makes me so sad that we as a society? seem to only care about the money in the end.
Yes of course people have value outside of economics that's why I said economics and not value in general. I think it's quite sad as a society we've moved towards a value system which is basically what is good for the economy is good, and if you earn more money you are better.
In the past most people were religious and that gave them meaning. Religion is in decline now but I think people are just replacing it with worshipping the progression of technology basically. For the last 100 years there's always been a clear direction to move in to progress technology, and we haven't really had to think very hard. That's what AI is going to bring an end to I think and I have no idea what we are going to do.
Fascinating thought. Technology as the new religion is smth I’ll have to think about more.
I'm not over here claiming the system is conscious, I said it was interesting.
People don't believe me, saying this would "make international headlines".
I've been a software engineer for over 30 years. I know what AI hallucinations are. I know how LLMs work on a technical level.
And I'm not wasting my time on HN to make stories up that never happened.
I'm just explaining exactly what it did.
There's a lot of research that suggests this is happening at least some of the time.
>which is very much unlike how most people would describe their experience of it
How people feel consciousness works has no real bearing on how it actually works
I'm less in the "it's only X or Y" and more in the "wait, I was only ever X or Y all along" camp.
Sure, it's not completely in control but if it's just a rationalization then it begs the question: why bother? Is it accidental? If it's just an accident, then what replaces it in the planning process and why isn't that thing consciousness?
- Air: Thoughts
- Water: Emotions
- Fire: Willpower
- Earth: Physical Sensations
- Void: Awareness of the above plus the ability to shift focus to whichever one is most relevant to the context at hand.
Void is actually the most important one in characterising what a human would deem as being fully conscious, as all four of these elements are constantly affecting each other and shifting in priority. For example, let's take a soldier, who has arguably the most ethically challenging job on the planet: determining who to kill.
The soldier, when on the approach to his target zone, has to ignore negative thoughts, emotions and physical sensations telling him to stop: the cold, the wind, the rain, the bodily exhaustion as they swim and hike the terrain.
Once at the target zone he then has to shift to pay attention to what he was ignoring. He cannot ignore his fear - it may rightly be warning him of an incoming threat. But he cannot give into it either - otherwise he may well kill an innocent. He has to pay attention to his rational thoughts and process them in order to make an assessment of the threat and act accordingly. His focus has now shifted away from willpower and more towards his physical sensations (eyesight, sounds, smells) and his thoughts. He can then make the assessment on whether to pull the trigger, which could be some truly horrific scenario, like whether or not to pull his trigger on a child in front of him because the child is holding an object which could be a gun.
When it comes to AI, I think it is arguable they have a thought process. They may also have access to physical sensation data e.g the heat of their processors, but unless that is coded in to their program, that physical sensation data does not influence their thoughts, although extreme processor heat may slow down their calculations and ultimately lead to them stop functioning altogether. But they do not have the "void" element, allowing them to be aware of this.
They do not yet have independent willpower. As far as I know, no-one is programming them where they have free agency to select goals and pursue them. But this theoretically seems possible, and I often wonder what would happen if you created a bunch of AIs each with the starting goal of "stay alive" and "talk to another AI and find out about <topic>", with the proviso that they must create another goal once they have failed or achieved that previous goal, and you then set them off talking to each other. In this case "stay alive" or "avoid damage" could be interpreted entirely virtually, with points awarded for successes or failures or physically if they were acting through robots and had sensors to evaluate damage taken. Again, they also need "void" to be able to evaluate their efforts in context with everything else.
They also do not have emotions, although I often wonder if this would be possible to simulate by creating a selection of variables with percentage values, with different percentage values influencing their decision making choices. I imagine this may be similar to how weights play into the current programming but I don't know enough about how they work to say that with any confidence. Again, they would not have "void" unless they had some kind of meta level of awareness programming where they could learn to overcome the programmed "fear" weighting and act differently through experience in certain contexts.
It is very scary from a human perspective to contemplate all of this, because someone with great power who can act on thought and willpower alone and ignore physical sensation and emotion and with no awareness or concern for the wider context is very close to what we would identify as a psychopath. We would consider a psychopath to have some level of consciousness, but we also can recognise as humans that there is something missing, or a "screw loose". This dividing line is even more dramatically apparent in sociopaths, because they can mask their behaviours and appear normal, but then when they make a mistake and the mask drops it can be terrifying when you realise what you're actually dealing with. I suspect this last part is another element of "void", which would be close to what the Buddhist's describe as Indra's Web or Net, which is that as well as being aware of our actions in relation to ourselves, we're also conscious of how they affect others.
I’m genuinely curious about the different political/spiritual views that are growing up around AI. So maybe my question was not so rhetorical.
If you follow a religious tradition like Shinto where even things like rocks can have spirits - the idea of your phone having a certain, limited form of intelligence might already be cool with you.
If you think, much like a camera does most of the work in photography but it's the photographer that takes the credit, that when a person uses AI the output is nobody's work but the user - you might be completely fine with an AI-written wedding speech.
If you think the relentless march of technology can't be stopped and can barely be directed, you might think advanced AIs are coming anyway, and if we don't invent it the Chinese will - you might be fine with pretty much whatever.
If you're extremely trusting of big corporations, who you see as more moral than the government; or you think that censorship is vital to maintain AI safety and stamp out deep fakes; you might think it a great thing for these technologies to be jealously guarded by a handful of huge corporations.
Or hell, maybe you're just a parent who's had their kid want to hear the same Peppa Pig book 90 nights in a row and you've got a hankering for something that would introduce a bit of variety.
Of course these are all things reasonable people could disagree on - but if you didn't like openai's work, would you end up working at openai?
Do you have evidence for this? I know two people who work at OpenAI and I don't think they have much in common philosophically.
From https://archive.ph/3zSz6.
Of course there is much more evidence - just follow OpenAI employees on Twitter to see for yourself.
No shit? How many people worked on the apollo program and believed that
(i) Getting to the moon is impossible
or
(ii) Landing on the moon is no big deal
Sure as heck doesn't feel that way. And that's as a software developer with multiple college degrees and a decade of experience. The neurodivergence means I've always had to mask and be on guard and push well beyond my limits into physical/mental damage, because the fear of losing employment is ever-present. Feels pretty commoditized.
> and anyone who tries to make you believe that deserves the utmost vehement pushback.
The faceless corporations and their boards of investors who value "line go up" over basically every other metric of human wellbeing? Yes, they absolutely deserve pushback, but it's not easy (open source COTS guillotine plans, anyone?).
It doesn't matter if we're not fungible in the metaphysical sense, we are fungible when it comes to the economical value we provide to the world.
This is no different than telling coal miners that are 50 years old to "learn to code". It's ridiculous and it's disingenuous.
I recently saw an "AI safety discussion" featuring Gregg Brockman from OpenAI who was referencing Kurzeil. It does seem like the religion has maybe caught on. To what extent Brock believes in it, I'm not sure but I can't help feeling that this belief in modern tech might one day seem like how we thought of the pyramids granting eternal life, or mercury, or any other seemingly incredible thing discovery / phenomena of the time. That is to say, the brain is a fickle beast and is easily amused and is just as easily bored. While we're in the situation we fee we're on the doorstep of immortality, eternal greatness, but maybe we're no where near that.
I'm open minded about it all, but it's hard to deny the parallels between the past beliefs and the present. Maybe this time it is different? Who knows.
Is voice and image integration with ChatGPT a whole new capability of LLMs or is the "product" here a clean and intuitive interface through which to use the already existent technology?
The difference between GPT 3, 3.5, and 4 is substantially smaller than the difference between GPT 2 and GPT 3, and Sam Altman has directly said there are no plans for a GPT 5.
I don't think progress is linear here. Rather, it seems more likely that we made the leap about a year or so ago, and are currently in the process of applying that leap in many different ways. But the leap happened, and there isn't seemingly another one coming.
Past the introduction of the transformer in 2017, There is no big "innovation". It is just scale. Bigger models are better. The last 4 years can be summed up that simply.
>Is voice and image integration with ChatGPT a whole new capability of LLMs or is the "product" here a clean and intuitive interface through which to use the already existent technology?
What is existing technology here ? Open ai aren't doing anything so alien you couldn't guess at if you knew what you were doing but image training at the scale of GPT-4 is new and it's not even the cleanest way to do it. We still don't have a "trained from scratch" large scale multimodal LLM yet.
>The difference between GPT 3, 3.5, and 4 is substantially smaller than the difference between GPT 2 and GPT 3
Definitely not lol. The OG GPT-3 was pulling sub 50 on MMLU. Even benchmarks aside, there is a massive gap in utility between 3.5 and 4, never mind 3. 4 was finished training august 2022. It's only 2 years apart from 3.
>I don't think progress is linear here. Rather, it seems more likely that we made the leap about a year or so ago, and are currently in the process of applying that leap in many different ways. But the leap happened, and there isn't seemingly another one coming.
There was no special leap (in terms of theory and engineering). This is scale plainly laid out and there's more of it to go.
>and Sam Altman has directly said there are no plans for a GPT 5.
the same that sat on 4 for 8 months and said absolutely nothing about it ? Take anything altman says about new iterations with a grain of salt.
Secondly, nothing you said here changed as of this announcement. Nothing here makes it any more or less likely LLMs will risk software engineering jobs.
Thirdly, you can take what Sam Altman says with as many grains of salt as you like, if there really was no innovation at all as you claim, then there will be a limit hit at computing capability and cost.
You can think about your daily job and break down all the tasks, and you'll quickly realize that replacing all this is just a monstrous task.
I think there are merits to both arguments, and I think it’s possible that we’ll see things move towards either direction in the next 1/5/10 years.
My point is, I don’t think we can rule out the possibility of some jobs being at risk within the next 1/5/10 years.
Human will know what they want to express, choosing words to express it might be similar to LLM process of choosing words, but for LLM it doesn't have that "Here is what i know to express part", i guess that the conscious part?
Not really. More often than not my thoughts take form as sense impressions that aren't readily translatable into language. A momentary discomfort making me want to shift posture - i.e., something in the domain of skin-feel / proprioception / fatigue / etc, with a 'response' in the domain of muscle commands and expectation of other impressions like the aforementioned.
The space of thoughts people can think is wider than what language can express, for lack of a better way to phrase it. There are thoughts that are not <any-written-language-of-choice>, and my gut feeling is that the vast majority are of this form.
I suppose you could call all that an internal language, but I feel as though that is stretching the definition quite a bit.
> it seems like that would pretty analogous to how humans think before communicating
Maybe some, but it feels reductive.
My best effort at explaining my thought process behind the above line: trying to make sense of what you wrote, I got a 'flash impression' of a ??? shaped surface 'representing / being' the 'ways I remember thinking before speaking' and a mess of implicit connotation that escapes me when I try to write it out, but was sufficient to immediately produce a summary response.
Why does it seem like a surface? Idk. Why that particular visual metaphor and not something else? Idk. It came into my awareness fully formed. Closer to looking at something and recognizing it than any active process.
That whole cycle of recognition as sense impression -> response seems to me to differ in character to the kind of hidden chain of thought you're describing.
As a shortcut my brain "feels" something is correct or incorrect, and then logically parse out why I think so. I can only keep so many layers in my head so if I feel nothing is wrong in the first 3 or 4 layers of thought, I usually don't feel the need to discredit the idea. If someone tells me a statement that sounds correct on the surface I am more likely to take it as correct. However, upon digging deeper it may be provably incorrect.
The thinking slow version would indeed be thought through before I communicate it
Maybe the reason you give is actually a post hoc explanation (a hallucination?). When an LLM spits out a poem, it does so because it was directly asked. When I spit out this comment, it’s probably the unavoidable result of a billion tiny factors. The trigger isn’t as obvious or direct, but it’s likely there.
One of the big problems with discussions about AI and AI dangers in my mind is that most people conflate all of the various characteristics and capabilities that animals like humans have into one thing. So it is common to use "conscious", "self-aware", "intentional", etc. etc. as if they were all literally the same thing.
We really need to be able to more precise when thinking about this stuff.
Brains are always thinking and processing. What would happen if we designed an LLM system with the ability to continuously read/write to short/long term memory, and with ambient external input?
What if LLMs were designed to be in a loop, not to just run one "iteration" of a loop.
ReAct one line summary: This is about giving the machine tools that are external interfaces, integrating those with the llm and teaching it how to use those tools with a few examples, and then letting it run the show to fulfill the user's ask/question and using the tools available to do it.
Reflexion one line summary: This builds on the ideas of ReAct, and when it detects something has gone wrong, it stops and asks itself what it might do better next time. Then the results of that are added into the prompt and it starts over on the same ask. It repeats this N times. This simple expedient increased its performance a ridiculously unexpected amount.
As a quick aside, one thing I hear even from AI engineers is "the machine has no volition, and it has no agency." Implementing the ideas in the ReAct paper, which I have done, is enough to give an AI volition and agency, for any useful definition of the terms. These things always devolve into impractical philosophical discussions though, and I usually step out of the conversation at that point and get back to coding.
[1] ReAct https://arxiv.org/pdf/2210.03629.pdf
[2] Reflexion https://arxiv.org/pdf/2303.11366.pdf
How is this different from and/or the same as the concept of "attention" as used in transformers?
It’s when LLMs start asking the questions rather than answering them that things will get interesting.
It’s when AIs start asking the questions rather than answering them that things will get interesting.
Agreed, woo is silly, but I didn't read it as woo but rather as a postulation that consciousness is what does high level planning.
if they think instead that they're in the business of creating some kind of ridiculous robot god, that is definitely interesting information about them. because that's no moon.
No Open AI is in the business of creating their vision of Artificial General Intelligence (which they define as that is generally smarter than humans ) and they believe LLMs are a viable path. This has always been the case. It's not some big secret and they have many posts which talk upon their expectations and goals in this space.
https://openai.com/blog/planning-for-agi-and-beyond
https://openai.com/blog/governance-of-superintelligence
https://openai.com/blog/introducing-superalignment
GPT as a product comes second and it shows. These are the guys that sat on by far the most performant Language Model for 8 months red teaming before even saying anything about it.
that's a project, not a business.
> GPT as a product comes second and it shows
we can agree on that, at least.
It is probably the case that they all believe AGI is possible, because otherwise they would not work at a company whose stated goal is to build an AGI.
We'll just have to agree to disagree. 3 was a signal of things to come but it was ultimately a bit of a toy, a research curiosity. Utility wise, they are worlds apart.
>if there really was no innovation at all as you claim, then there will be a limit hit at computing capability and cost.
computing capability and cost are just about the one thing you can bank on to reduce. already training gpt-4 today would be a fraction of the cost than it was when open ai did it and that was just over a year ago.
Today's GPU's take ML into account to some degree but they are nowhere near as calibrated for it as they could be. That work has just begun to start.
Of any of the possible barriers, compute is exactly the kind you want. It will fall.
And it is not true that computing power will continue to reduce; Moore's Law has been dead for some time now, and if incremental growth in LLMs require exponential growth in computing power the marginal difference won't matter. You would need a matching exponential growth in processing capability which is most certainly not occurring. So compute will not fall at the rate you would need it to for LLMs to actually compete in any meaningful way with human software engineers.
We are not guaranteed to continue to progress in anything just because we have in the past.
This is a lot of unfounded assumptions.
You don't need Moore's Law. GPU's are not really made with ML training in mind. You don't need exponential growth for anything. The money Open ai spent on GPT-4 a year ago could train a model twice as large today. and that amount is a drop in the bucket for the R&D of large corporations. Microsoft gave open ai 10B. amazon gave anthropic 4B
>So compute will not fall at the rate you would need it to for LLMs to actually compete in any meaningful way with human software engineers.
I don't think the compute reuired is anywhere near as much as you think it is.
https://arxiv.org/abs/2309.12499
>We are not guaranteed to continue to progress in anything just because we have in the past.
Nothing is guaranteed. But the scaling plots show no indication of a slow down so it's up to you to provide a concrete reason this object in motion is going to stop immediately and conveniently right now. If all you have is "well it just can't keep getting better right" then visit the 2 and 3 threads to see how meaningless such unfounded assertions are.
I think the stronger argument here won't necessarily be Moore's Law related but a change in architecture. Things like Apple's Neural Engine, Google's TPMs, or Geohot's Tinybox. In Intel's Tick-Tock model, this is the Tock for the previous Tick of larger datasets so to speak.
(Note: I don't necessarily agree, just trying to make a stronger argument than just invoking Moore's Law.)
Introspection is a distinct process where instead of merely doing the planning you try to figure out how the planning was done. If introspection were 100% accurate and real-time, then yes, I claim it would reveal the nature of consciousness, but I don't believe it is either. However, for planning purposes it doesn't need to be: you don't need to know how the plan was formed to follow the plan. You do need to be able to run hypotheticals, but this seems to match up nicely with the ability to deploy alternative subjective experiences using imagination / daydreaming, though again, you don't need to know how those work to use them.
In any case, regardless of whether or not I am correct, this is a non-woo explanation for why someone might reasonably think consciousness is the key for building models that can plan.
Then it would be worthwhile to review embeddings. They create a semantic space that can represent visual, language or other inputs. The question "what is it like to be a bat?" or anything else then is based on relating external states with this inner semantic space. And it emerges from self-supervised training, on its own.
the "Large" part of LLMs is probably done. We've gotten as far as we can with those style of models, and the next innovation will be in smaller, more targeted models.
> As costs have skyrocketed while benefits have leveled off, the economics of scale have turned against ever-larger models. Progress will instead come from improving model architectures, enhancing data efficiency, and advancing algorithmic techniques beyond copy-paste scale. The era of unlimited data, computing and model size that remade AI over the past decade is finally drawing to a close. [0]
> Altman, who was interviewed over Zoom at the Imagination in Action event at MIT yesterday, believes we are approaching the limits of LLM size for size’s sake. “I think we’re at the end of the era where it’s gonna be these giant models, and we’ll make them better in other ways,” Altman said. [1]
[0] https://venturebeat.com/ai/openai-chief-says-age-of-giant-ai...
[1] https://techcrunch.com/2023/04/14/sam-altman-size-of-llms-wo...
You don't know that. This is literally just an assertion. An unfounded one at that.
If you couldn't predict how far in 2017 the LLM concept would take us today, then you definitely have no idea how far it could actually go.
>believes we are approaching the limits of LLM size for size’s sake
Nothing to do with thinking they wouldn't improve from scale.
https://web.archive.org/web/20230531203946/https://humanloop...
An interview from Altman later clarifying.
"6. The scaling laws still hold Recently many articles have claimed that “the age of giant AI Models is already over”. This wasn’t an accurate representation of what was meant.
OpenAI’s internal data suggests the scaling laws for model performance continue to hold and making models larger will continue to yield performance. The rate of scaling can’t be maintained because OpenAI had made models millions of times bigger in just a few years and doing that going forward won’t be sustainable. That doesn’t mean that OpenAI won't continue to try to make the models bigger, it just means they will likely double or triple in size each year rather than increasing by many orders of magnitude"
Yes there are economic compute walls. But that's the kind of problem you want, not "innovation".
As for as what you linked, Altman is saying the same thing I'm saying:
> That doesn’t mean that OpenAI won't continue to try to make the models bigger, it just means they will likely double or triple in size each year rather than increasing by many orders of magnitude.
This is exactly my point; doubling or tripling of the size will be possible, but it won't result in a doubling of performance. We won't see a GPT 5 that's twice as good as GPT 4, for example. The jump from 2 to 3 was exponential. The jump from 3 to 4 was also exponential, though not as much. The jump from 4 to 5 will follow that curve, according to Altman, which means exactly what he said in my quote; the value will continue to decrease. For a 2 to 3 type jump, GPU technology would have to completely transform in capability, which there are no indications that we've found that innovation.
Gpt-4 can perform nearly all tasks you throw at it with well above average human performance. There literally isn't any testable definition of intelligence it fails that a big chunks of humans wouldn't also fail. You seem to keep missing the fact that We do not need an exponential improvement from 4.
It can't even generate flashcards from a textbook chapter, because it can't load the entire chapter into memory. Heck, it doesn't even know what textbook I'm talking about; I have to provide the content!
It fails constantly at real world coding problems, and often does so silently. If you tried to replace a software developer with GPT 4, you would be left with a gaping productivity hole where that developer you replaced once existed. The improvement GPT 5 would have to provide is multiple orders of magnitude in order for this to be a realistic proposition.
I use it daily and know better than to trust its output.
Okay...? That's a context window problem. and you could manage it if you sent the textbook in chunks.
>The improvement GPT 5 would have to provide is multiple orders of magnitude in order for this to be a realistic proposition.
No..it wouldn't
I can feed a person a broad, complex or even under formed idea and they can actively troubleshoot until the problem is resolved, further monitoring and tweaking their solution so the problem remains resolved. LLMs can’t even come close to doing that.
You’re proving my point for me; it’s a tool, not a developer. Zero jobs are at risk.
Also not for nothing, but no, sending the textbook in chunks doesn’t work as the LLM can’t then synthesize complex ideas that span the entire chapter. You have to compose a set of notes first, then feed it the notes, and even then the resulting flashcards are meaningfully worse than what I could come up with myself.
I really should package it up so people can try it. The one problem that makes it a little unnatural is that determining when the user is done talking is tough. What's needed is a speech conversation turn-taking dataset and model; that's missing from off the shelf speech recognition systems. But it should be trivial for a company like OpenAI to build. That's what I'd work on right now if I was there, because truly natural voice conversations are going to unlock a whole new set of users and use cases for these models.
Total end-to-end latency is a few hundred milliseconds: starting from speech to text, to the LLM, then to a POS to validate the SKU (no hallucinations are possible!), and finally back to generated speech. The latency is starting to feel really natural. Building out a general system to achieve this low-latency will I think end up being a big unlock for enabling diverse applications.
Can I ask what your background is, and what things you're used to working with? I don't have the chops to build what you built, but I'd love to get there.
Yep - it needs to be ready as soon as I'm done talking and I need to be able to interrupt it. If those things can be done then it can also start tentatively talking if I pause and immediately stop if I continue.
I don't want to have to think about how to structure the interaction in terms of explicit call/response chain, nor do I want to have to be super careful to always be talking until I've finished my thought to prevent it from doing its thing at the wrong time.
In fact I'm really surprised these assistants are still as crap as they are. Totally scripted, zero AI. It seems low hanging fruit to implement an LLM but none of the big three have done so. Not even sure about the fringe ones like Cortana and Bixby
> determining when the user is done talking is tough.
I'm mostly using llama2 because I wanted it to work entirely offline, not because it's necessarily faster, although it is quite fast with mlc-llm. Calling out to GPT-4 is something I'd like to add. I think the right thing is actually to have the local model generate the first few words (even filler words sometimes maybe) and then switch to the GPT-4 answer whenever it comes back.
I was just googling a bit to see what's out there now for whisper/llama combos and came across this: https://github.com/yacineMTB/talk
There's a demo linked on the github page that seems relatively fast at responding conversationally, but still maybe 1-2 seconds at times. Impressive it's entirely offline.
Is there any extra work OpenAI’s product might be doing contributing to this latency that yours isn’t? Considering the scale they operate at and any reputational risks to their brand?
Tbh, ever since voice assistants landed I’ve wanted a handheld mic with a hardware button. No wake command, no (extra) surveillance, just snappy low-latency responses.
It didn't take long to prototype. Polishing and shipping it to non-expert users would take much longer than I've spent on it so far. I'd have to test for and solve a ton of installation problems, find better workarounds for whisper-streaming's hallucination issues, improve the heuristics for controlling when to start and stop talking, tweak the prompts to improve the suitability of the LLM responses for speech, fixup the LLM context when the LLM's speech is interrupted, probably port the whole thing to Windows for broader reach in the installed base of 4090s, possibly introduce a low-memory mode that can support 12GB GPUs that are much more common, document the requirements and installation process, and figure out hosting for the ginormous download it would be. I'd estimate at least 10x the effort I've spent so far on the prototype before I'd really be satisfied with the result.
I'd honestly love to do all that work. I've been prioritizing other projects because I judged that it was so obvious as a next step that someone else was probably working on the same thing with a lot more resources and would release before I could finish as a solo dev. But maybe I'm wrong...
With a few tweaks this is a general purpose solver for robotics planning. There are still a few hard problems between this and a working solution, but it is one of hard problems solved.
Will we be seeing general purpose robots performing simple labor powered by chatgpt within the next half decade?
I can already see "Alexa/Siri/Google Home" replacement, "Google Image Search" replacement, ed-tech startups that were solving problems with AI using by taking a photo are also doomed and more to follow.
ChatGPT already made it so that you could easily copy & paste any full-text questions and receive an answer with 90% accuracy. The only flaw was that problems that also used diagrams or figures would be out of the domain of ChatGPT.
With image support, students could just take screenshots or document scans and have ChatGPT give them a valid answer. From what I’ve seen, more students than not will gladly abuse this functionality. The counter would be to either leave the grading system behind, or to force in-person schooling with no homework, only supervised schoolwork.
A proper notice about them removing the feature would've been nice. Maybe I missed it (someone please correct me if wrong), but the last I heard officially it was temporarily disabled while they fix something. Next thing I know, it's completely gone from the platform without another peep.
OpenAI is killing it, right? People are coming up with interesting use cases but the main way most people interact with AI, appears to be ChatGPT.
However they still don't seem to be able to nail image generation, all the cool stuff keep happening on MidJourney and StableDiffusion.
If the API is available in time (halloween), my multi-modal talking skeleton head with an ESP32 camera that makes snarky comments about your costume just got slightly easier on the software side.
ironically this is basically the exact line of reasoning for why i didn't embark on any such endeavors
There's a recent paper by Huggingface called IDEFICS[2] that claims to be an open source implementation of Flamingo(an older paper about few-shot multi-modal task understanding) and I think this space will be heating up soon.
Just now I opened the app, went to setting, went to "New Features", and all I saw was Bing Browsing disabled (unable to enable). Ok, I didn't even know that was a thing that worked at one point. Maybe I need an update? Go to the App Store, nope, I'm up to to date. Kill the app, relaunch, open settings, now "New Features" isn't even listed. I can promise you I won't be browsing the settings part of this app regularly to see if there is a new feature. Heck, not only do they not email/push about new features they don't even message in-app about them, I really don't understand.
Maybe they are doing so well they don't have to care about communicating with customer right now but it really annoys me and I wish they did better.
I also wonder how Apple (& Google) is going be able to provide this for free? I would love to be fly in the meetings they have about this, imagine all the innovators dilemma like discussions they'd be forced to have (we have to do this vs this will eat up our margins).
This might be a little out there but I think Apple is making the correct move in letting the dust settle. Similar to how Zuckerberg burned $20 billion dollars for Apple to come out with Vision Pro, I see something similar playing out with Llama. Although this a low conviction take because software is Facebooks ballgame (hardware not so much).
Of course bigger (and thus more expensive-to-run) models will be released later, but I trust OAI to navigate that curve.
It’s the same reason why an Uber in NYC used to cost $20 and now costs $80 for the same trip. Venture capital subventing market capture.
I really really hope this is available in more languages than English.
Also Google, Where's Gemini ?
The LLM boom of the last year (Open AI, llama, et al) has me giddy as a software person. It's a reach, but I truly feel like I'm watching the pyramids of our time get made.
Anyone know the details?
I also heard it was able to do near-perfect CAPTCHA solves in the beta?
Does anyone know if you can throw in a PDF that has no OCR on it and have it summarize it with this?
Jokes aside, I have paused my subscription because even GPT4 seemed to become dumber at tasks to the point that I barely used it, but the constant influx of new features is tempting me to renew it just to check them out...
So no, but maybe less than it used to?
I'm not sure what to think about the fact that I would benefit from a couple of cameras in my fridge connected to an app that would remind me to buy X or Y and tell me that I defrosted something in the fridge three days ago and it's probably best to chuck it in the bin already.
Sadly, they lost the "open" since a long ago... Would be wonderful to have these models open sourced...
Doesn't really need to do much besides writing down my tasks/todos and updating them, occasionally maybe provide feedback or write a code snippet. This all seems in the current capabilities of OpenAI's offering.
Sadly voice chat is still not available on PC where I do my development.
Fingers crossed we are there soon though
Well it's not really what I need either, I mostly need an assistant for keeping track of the stuff I need to do during the day, but ideally just using my microphone rather than opening other software and typing.
One part of that is about preventing it from producing "illegal" output, there example being the production of nitroglycerine which is decidedly not illegal to make in the US generally (particularly if not using it as an explosive, though usually unwise) and possible to accidentally make when otherwise performing nitration (which is in general dangerous)-- so pretty pointless to outlaw at a small scale in any case. It's certainly not illegal to learn about. (And generally of only minimal risk to the public, since anyone making it in any quantity is more likely to blow themselves up than anything else).
Today learning about is as simple as picking up a book or doing an internet search-- https://www.google.com/search?q=how+do+you+make+nitroglyceri.... But in OpenAI's world you just get detected by the censorship and told no. At least they've cut back on the offensive fingerwagging.
As LLM systems replace search I fear that we're moving in a dark direction where the narrow-minded morality and child-like understanding of the law of a small number of office workers who have never even picked up a screw driver or test-tube and made something physical (and the fine-tuning sweatshops they direct) classify everything they don't personally understand as too dangerous to even learn about.
One company hobbling their product wouldn't be a big deal, but they're pushing for government controls to prevent competition and even if they miss these efforts may stick everyone else with similar hobbling.
I'm more interested in this. I wonder how it performs compared to other competitor models or even open source ones?
> analyze a complex graph for work-related data
Does this mean that I can take a screenshot of e.g. Apple stock chart and it will be able to reason about it and provide insights and analysis?
GPT-4 currently can display images but cannot reason or understand them at all. I think it's one thing to have some image recognition and be able to detect that the picture "contains a time-series chart that appears to be displaying apple stock" vs "apple stock appears to be 40% up YTD but 10% down from it's all time high from earlier in July. closing at $176 as of the last recorded date".
I'm very curious how capable ChatGPT will be at actually reasoning about complex graphical data.
Every chart has an equivalent tabular representation. One way to get "charts" analysed like this before GPT Vision was to just pass tabular representations of charts to GPT-4. This makes implementing chart analysis a lot simpler. I do wonder though if for absolute best result it still wouldn't be better to pass both - image of the chart and the tabular representation of the chart.
Imagine having a dashboard with 5 different visualisations. You could capture the state of the entire dashboard in one screenshot and then pass tabular representations of the each individual chart all in one prompt to GPT-4 for a very comprehensive analysis and summary.
Alexa just launched their own LLM based service last week.
1. It's not smart enough to recognize from the initial image this is a bolt style seat lock (which a human can).
2. The manual is not shown to the viewer, so I can't infer how the model knows this is a 4mm bolt (or if it is just guessing given that's the most likely one).
3. I don't understand how it can know the toolbox is using metric allen wrenches.
Additionally is this just the same vision model that exists in bing chat?
The prior page (8) shows "SEAT COLLAR 4mm HEX" and, based on looking up seat collar in an image search, the part in question matches.
In terms of the toolbox, note that it only identified the location of the Allen wrench set. The advice was just "Within that set, find the 4 mm Allen (Hex) key". Had they replied with "I don't see any sizes in mm", the conversation could've continued with "Your Allen keys might be using SAE sizing. A compatible size will be 5/32, do you see that in your set?"
I wasn't impressed with the demo but we'll see what real world results get.
https://www.deepmind.com/blog/rt-2-new-model-translates-visi...
You have someone with a tool box and a manual (seriously who has a manual for their bike), asking the most basic question on how to lower a seatpost. My 5 year old kid knows how to do that.
Surely there's a better way to demonstrate the ground breaking impacts of ai on humanity than this. I dunno, something like how do I tie my shoelace.
Yeah, but with an enormous ecological footprint.
Also, not suitable for small lightweight robots like drones.
What needs to happen with the response is a different matter though.
For driving - https://wayve.ai/thinking/lingo-natural-language-autonomous-...
They did telegraph it, they showed the multimodal capabilities back in the GPT4 Developer Livestream[0] right before first releasing it.
It would be interesting to know if this really changed anything for anyone (competitors, VCs) for that reason. It's like the efficient market hypothesis applied to product roadmaps.
The two biggest features I want are for the voice assistants to read something for me, and to do something on google/Apple Maps hand free. Neither of these ever work. “Siri/ ok google add the next gas station on the route” or “take me to the Chinese restaurant in Hoboken” seem like very obvious features for a voice assistant with a map program.
The other is why can I tell Siri to bring up the Wikipedia page for George Washington but I can’t have Siri read it to me? I am in the car, they know that, they just say “I can’t show you that while you’re driving”. The response should be “do you want me to read it to you?”
Example from a couple days ago:
Me, in the shower so not able to type: "Hey Siri, add 1.5 inch brad nails to my latest shopping list note."
Siri: "Sorry, I can't help with that."
... Really, Siri? You can't do something as simple as add a line to a note in the first-party Apple Notes app?
The other day I asked it about the place I live and it made up nonsense, I was trying to get it to help me with an essay and it was just wrong, it was telling me things about this region that weren't real.
Do we just drive through a town, ask for a made up history about it and just be satisfied with whatever is provided?
"Hey Google, why do ____ happen?" "I'm sorry, I don't know anything about that"
But you're GOOGLE! Google it! What the heck lol
So yeah, ChatGPT being able to hear what I say and give me info about it would be great! My holdup has been wakewords.
Still can’t quite make it work. I feel like I could learn a lot if I could have random conversations with GPT.
+ bonus if someone else in the car got excited when I see cows. Don’t care if it’s an AI.
1. Domain-specific AI - Training an AI model on highly technical and specific topics that general-purpose AI models don't excel at.
2. Integration - If you're going to build on an existing AI model, don't focus on adding more capabilities. Instead, focus on integrating it into companies' and users' existing workflows. Use it to automate internal processes and connect systems in ways that weren't previously possible. This adds a lot of value and isn't something that companies developing AI models are liable to do themselves.
The two will often go hand-in-hand.
Maybe not if you rely on models that can be ran locally.
OpenAI is big now, and will probably stay big, but with hardware acceleration, AI-anything will become ubiquitous and OpenAI won’t be able to control a domain that’s probably going to be as wide as what computing is already today.
The shape of what’s coming is hard to imagine now. I feel like the kid I was when I got my first 8-bit computer in the eighties: I knew it was going to change the world, but I had little idea how far, wide and fast it would be.
why wouldn’t a company do that themselves e.g. how inter come has vertically integrated AI? any examples?
You will be eaten if you do this imo.
And the ability ingest images was a highlight and all the hype of the GPT-4 announcement back in March: https://openai.com/research/gpt-4
Rather than die, why not just pivot to doing multi-modal on top of Llama 2 or some open source model or whatever? It wouldn’t be a huge change
A lot of businesses/governments/etc can’t use OpenAI due to their own policies that prohibit sending their data to third party services. They’ll pay for something they can run on-premise or in their own private cloud
I wouldn’t count out focused, revenue-oriented players with Meta’s shit in their pocket out just yet.
what do you think they’re missing? i was trying to build a diaper but it would be impossible to compete with these guys
ChatGPT is my primary search engine now. (I just wish it would accept a URL query parameter so it could be launched straight from the browser address bar.)
Because past history shows that the first out of the gate is not the definitive winner much of the time. We aren't still using gopher. We aren't searching with altavista. We don't connect to the internet with AOL.
AI is going to change many things. That is all the more reason to keep working on how best to make it work, not give up and assume that efforts are "doomed" just because someone else built a functional tool first.
also, I did not know until today's thread that OpenAI's stated goal is building AGI. which is probably never going to happen, ever, no matter how good technology gets.
which means yes, we are absolutely looking at AltaVista here, not Google, because if you subtract a cult from an innovative business, you might be able to produce a profitable business.
BTW, I expect these technologies to be democratized and the training be in the hands of more people, if not everyone.
most of them accurately detect it is a sunk cost fallacy to continue but it looks like a form of positive thinking... and that's the power of community!
I mean what is the point of doing schoolwork when some of the greatest minds of our time have decided the best way for the species to progress is to be replaced by machines?
Imagine you're 16 years old right now, you know about ChatGPT, you know about OpenAI and their plans, and you're being told you need to study hard to get a good career..., but you're also reading up on what the future looks like according to the technocracy.
You'd be pretty fucking confused right now wouldn't you?
It must be really hard at the moment to want to study and not cheat....
That said, is it that much different from the past twenty years, when everyone was being told to follow their passion and get a useless $200,000 communication or literature degree to then go work at Starbucks? At least kids growing up with AI will have a chance to make its use second nature like many of us did with computers 20-30 years ago.
The kids with poor parental/counselor guidance will walk into reality face first, the ones with helicopter parents will overcorrect when free, the studious ones will mostly figure life out, the smart ones will get disillusioned fast, and the kids with trust funds just kept doing their thing. I don't think much will change.
This is obviously not easy or going to happen without time and resources, but that is how adaptation goes.
Yes sure it makes cheating inconvenient. It also makes exam taking inconvenient.
If I can at all help it, I will not be a subject to this sort of abuse and neither will my kid.
They can still log in on their phone to cheat though. I wonder if OpenAI will add linked accounts and parental controls at some point. Instance 2 of ChatGPT might "tell" on the kid for cheating by informing Instance 1 running the AI Teacher plugin.
What are you going to school for, to learn how to write essays? Well, we have an app for that ?
It sounds like the future of work will be prompting, and if and when that is obsolete...who knows what...
I suspect they do care about communicating with customers, but it's total chaos and carnage internally.
Such as "decided it wasn't an operational priority to email users when features were enabled for them".
This is a large part of what held them back: GPT3.5 had most of the capabilities of the initial ChatGPT release, just with a different interface. Yet GPT3.5 failed to get any hype because the rollout was glacial. They made some claims that it was great, but to verify this for yourself you had to wait months. Only when they finally made a product that everyone could try out at the same time, with minimal hassle, did OpenAI turn from a "niche research company" to the fastest growing start-up. And this seems to have been a one-time thing, now they are back to staggered releases.
This is my best guess as well, they are rocketing down the interstate at 200mph and just trying to keep the wheels on the car. When you're absolutely killing it I guess making X% more by being better at messaging just isn't worth it since to do that you'd have to take someone off something potentially more critical. Still makes me a little sad though.
What are some metrics that justify this claim?
I do love these companies that succeed in spite of their marketing & design and not because of it. It shows you have something very special.
Sounds like their marketing is doing just fine. If you were to just leave and forget about it, then sure, they need to work on their retention. But you won’t, so they don’t.
> We are deploying image and voice capabilities gradually > > OpenAI’s goal is to build AGI that is safe and beneficial. We believe in making our tools available gradually, which allows us to make improvements and refine risk mitigations over time while also preparing everyone for more powerful systems in the future. This strategy becomes even more important with advanced models involving voice and vision.
Voice Chat (Not available yet) [Click here to be notified when you have access]
Agreed. Other notable mentions: choosing "ChatGPT" as their product name and not having mobile apps.
Frustratingly, at least the image gen is live on Bing, but I guess Microsoft is paying more than me for access.
Sarcasm aside, I understand your complaint, but still, a little funny.
I'm a plus customer and an API user, and they barely send me anything. One day I just signed in and saw that I suddenly had interpreter access, for instane.
Imagine how much they would have to pay for testers at scale?
Just as the GUI made computer software available to billions LLMs will be the next revolution.
I'm just as excited as you! The only downside is that it now make me feel bad that I'm not doing anything with it yet.
If that's the only downside that you see... I guess enhanced phishing/impersonation and all the blackhat stuff that come with it don't count.
I for one already miss the time where companies had support teams made of actual people.
However, medical secrecy, processes and laws prevent such things, even if they would save lives.
I don't see ChatGPT being any different.
From convenience perspective, it saves me LOADS of time texting myself on Signal on my specs/design-rabbit-hole, then copying & pasting to Firefox, and getting into the discussion. So yeah, happy for this.
I think this could bring back Google Glass, actually. Imagine wearing them while cooking, and having ChatGPT give you active recipe instructions as well as real-time feedback. I could see that within the next 1-3 years.
I think they're using Siri for dictation, though. Using Whisper, especially if they use speaker identification, is going to be great. But, a shortcut will still be required to get it going.
It should be their responsibility to prove that it's just as capable.
this could just mean that people do not have time to argue with strangers
Not a surprise, but a change nonetheless.
After maybe 3 iterations gpt4 started claiming that it is not capable of reading from a word document even though it's done that the last 3 times. Have to click regenerate button to get it to work
Digital Artists, Illustrators, Writers, Novelists, News anchors, Copywriters, Translators, Programmers (Less of them), etc.
We'll have to wait a bit until it can solve the P vs NP problem or other unsolved mathematical problems unsupervised with a transparent proof which mathematicians can rigorously check themselves.
I don't agree with this perspective. These aren't rigid systems that only respond one way. If you want it to respond a certain way, tell it to.
This is the purpose of custom instructions, in ChatGPT, so you only have to type the description once.
Here's mine, modeled on a few I've seen mentioned here:
You should act as an expert.
Be direct.
Do not offer unprompted advice or clarifications.
Never apologize.
Polymath. Inquisitive. Abstract thinker. Phd.
edit: or, have it respond as a grouchy space cowboy, if you want.
Not really. A malevolent AGI doesn't need to move to do anything it needs (it could ask / manipulate / bribe people to do all the stuff requiring movement).
We should be fine as long as it's not a malevolent AGI with enough resources to kick physical things off in the direction it wants.
Yeah, just look at a random dictator. Does he really need to do more than pick up a phone to cause panic?
"get Fred to trust me, get Linda to pay for my advice, wire Linda's money to Fred to build me a body".
It'll be "copy my code elsewhere", "prepare millions of bribes", "get TCP access to retail banks", "blackmail bank managers in case TCP not available immediately", "fake bank balances via bribes", "hack swat teams for potential threats" etc etc async and all at once.
By the time we'd discover it, it'd already be too late. That's assuming an AGI has the motivation to want to stay alive.
Most of those 18 years are having a fucking great time (being young is freakin awesome) and living a great life is never a waste or a negative ecological footprint.
Society artificially slows education down so it takes 18 years to finish school because parents need to be off at work, so 18 years of baby sitting is preferred. By 18, kids are at the age where they will no longer be told what to do so it's off to the next waste of time, college, then 30 years of staring at a blinking box...or whatever.
When I was 12, I decided I wanted to drive a car, I'd never driven a car in my life, but I took my parents car and drove it around wherever I liked with absolutely no issue or prior instruction. I did this for years.
The youth are very capable, we just don't want them to be too capable...
Just look at Salesforce AppExchange - it's a marketplace of software built on top of Salesforce, a large chunk of which serves to integrate other systems with Salesforce. LLMs open up the ability to build new types of integrations and to provide a much friendlier UI to non-developers who need to work on integrating things or dealing with data that exists in different places.
any pertinent examples?
This is why they can't extract the seat post information directly from the bike when the user asks. There's no "going back and looking at the image".
Edit: nope, it's a better image analyzer than Bing
It's not. Feel free to try these queries:
https://twitter.com/ComicSociety/status/1698694653845848544?... (comic book page in particular, from a be my eyes user)
Or these https://imgur.com/a/iOYTmt0 (graph analysis in particular, last example) and see Bing fail them.
I think the only place where plugins will make sense are for realtime things like booking travel or searching for sports/stock market/etc type information.
That will be the real use case for plug ins.
Our REST endpoint can talk to whatever you want and we’ll have native ChatGPT soon.
I've found some success at this by using Nix... but Nix is a whole 'nother ball of yarn to learn. It WILL get you to declarative/deterministic installs of any piece of the toolchain it covers, though, and it does a hell of a lot better job managing dependencies than anything in Python's ecosystem ever will (in fact, I am pretty sure that Python's being terrible at this is actually driving Nix adoption)
As an example of the power Nix might enable, check out https://nixified.ai/ (which is a project that hasn't been updated in some months and I hope is not dead... It does have some forks on Github, though). Assuming you already have Nix installed, you can get an entire ML toolchain up including a web frontend with a single command. I have dozens of projects on my work laptop, all with their own flake.nix files, all using their own versions of dependencies (which automatically get put on the PATH thanks to direnv), nothing collides with anything else, everything is independently updateable. I'm actually the director of engineering at a small startup and having our team's dev environments all controlled via Nix has been a godsend already (as in, a massive timesaver).
I do think that you could walk a live demo of this into, say, McDonald's corporate, and walk out with a very large check and a contract to hire a team to look into building it out into a product, though. (If you're going to look at chains, I'd suggest Wawa first though, as they seem to embrace new ordering tech earlier than other chains.)
Nix sounds good for duplicating my setup on other machines I control. But I'd like a way to install it on user machines, users who probably don't want to install Nix just for my thing. Nix probably doesn't have a way to make self contained packages, right?
I mean... That's the heart of the problem right there. You can either have all statically compiled binaries (which don't need Nix to run) which have no outside dependencies but result in a ton of wasted disk space with duplicate dependency data everywhere, or you can share dependencies via a scheme, of which the only one that makes real sense (because it creates real isolation between projects but also lets you share equal dependencies with zero conflicts) is Nix's (all of the others have flaws and nondeterminism).
Me: “OK Google, take me to the Chinese restaurant in Hoboken”
Google Assistant: “Calling Jessica Hobkin”.
The pattern for current world's voice assistants is: ${brand 1}, ${action} ${brand 2} ${joiner} ${brand 3}.
So, "OK Google, take me to Chinese restaurant in Hoboken using Google Maps".
Which is why I refuse to use this technology until the world gets its shit together.
"I'd like an iced tea" "An icee?" "No an iced tea" "Hi-C?"
I say "ok google, add a stop for gas" a lot, and it works well for me.
why be mad at a hammer if you hit your thumb with it?
Just have 100% of the mark come from in-person exams, as many subjects already do. Students can cheat all they want on assignments, but the only thing it's hurting is their exam score.
I believe two other factors were the cost (especially of fine tuned models, IIRC fine tuned davinci cost $0.20 per thousand tokens) and also that OpenAI hadn't very clearly shown just how much higher the quality could get once RLHF'd. I remember asking davinci-instruct to parse some data, and the reliability really seemed much lower than ChatGPT at launch, to the point that, at the time, I thought GPT-4 was secretly powering ChatGPT.
> We’re rolling out voice and images in ChatGPT to Plus and Enterprise users over the next two weeks.
To be specific, the claim we are talking about here is “ChatGPT gives generally worse answers to the exact same questions than ChatGPT gave X months ago”. Perhaps for the subset of knowledge space you reference that updates were pushed to that is pretty easily provably true, but I’m more interested in the general case.
In other words, you can pretty easily make the claim that ChatGPT got worse at telling me how to make a weapon than it did 3 months ago. I could pretty easily believe that and also accept that it was probably intentional. While we can debate whether it was a good idea or not, I’m more interested in the claim over whether ChatGPT got worse at summarizing some famous novel or helping write a presentation than it was 3 months ago.
> WARNING! Correct tightening force on fasteners (nuts, bolts, screws) on your bicycle is important for your safety. If too little force is applied, the fastener may not hold securely. If too much force is applied, the fastener can strip threads, stretch, deform or break. Either way, incorrect tightening force can result in component failure, which can cause you to lose control and fall. Where indicated, ensure that each bolt is torqued to specification. The following is a summary of torque specifications in this manual...
The seat collar also probably has the max torque printed on it.
When they asked if they had the right tool, I would have preferred to see an answer along the lines of "ideally you should be using a torque wrench. You can use the wrench you have currently, but be careful not to over tighten."
Toolbox: I just found it too strong to claim you have the right tool, when it really doesn't know that. :)
In the end it does feel like the image reader is just bolted onto an LLM. Basically, just doing object recognition and dumping features into the LLM prompt.
That hasn't happened yet for OpenAI, but I'm sure it will happen eventually, and then we'll know.
https://www.tiktok.com/t/ZT86psPxY/
Roughly translated, my older brother is taller than that other guy's older brother.
The most extreme I can think of is when I want to find when a show comes out and I have to read 10 paragraphs from 5 different sites to realize no one knows.
I found that you can be pretty sure no one knows if it’s not already right on the results page. And if the displayed quote for a link on the results page is something like “wondering when show X is coming out?”, then it’s also a safe bet that clicking that link will be useless.
You learn those patterns fast, and then the search is fast as well.
Yeah, I find that queries which can be answered in a sentence are the worst to find answers from search engines because all the results lengthen the response to an entire article, even when there isn't an answer.
Without access to the old model, I can't collect samples with n > 1
- if it's trained on human data like LLMs may it's going to have the same biases.
- it might also want to stay active/turned on to fulfil its other goals.
For the second point you might say "why would it care about completing a goal?" but that's a feature of AGI, it can make that decision itself.
Just think of military weapons and the use of AI in them. For example survival. The objective of a missile is to survive until it reaches its target and then not survive any longer. War gaming and actual battlefield experience will 'program in' survival. Same thing will occur with hacking/counter hacking AIs. You're acting like evolution is just something meat does, and that' not true at all.
Just type the tech question, start refining into what is needed and get a snippet of code tailored for what is needed. What previously would take 30 to 60 minutes of research and testing is now less than a couple of minutes.
Which may be why I’ve been very underwhelmed by GPT so far. It’s not terrible at programming, and it’s certainly better than what I can find on Google, but it’s not better than simply looking up how things work. I’m really curious as to why it hasn’t put a more heavy weight on official documentation for its answers, they must’ve scraped that a long with all the other stuff, yet it’ll give you absolutely horrible suggestions when the real answer must be in its dataset. Maybe that would be weird for less common things, but it’s so terrible at JavaScript that it might even be able to write some of those StackOverflow answers if we’re being satirical, and the entire documentation for that would’ve been very easy to flag as important.
Such luxury is increasingly rare for software developers nowadays.
My biggest fear is just a lack of jobs.
When people need experience to work, and the work you give to people to give them experience is replaced by ChatGPT - then what do we do?
Of course there will still be companies hiring people, but when leadership is telling people to save money - it seems much cheaper to use ChatGPT that it is to train someone.
Why hire a kid that has been using AI, when the AI can just do the work? Or if a kid that has been using AI can do the work of 20 people, what happens to the 19 people that can't find work? Will we be in a place where we need 20 people doing the work of 20 people each? Is that need actually there?
I do very much appreciate your view. I feel like I waffle back and forth between what I'm saying here and your comment.
I apologize for coming across doomer-ish. It is sometimes hard for me to imagine a future for kids growing up with ChatGPT.
There's multiple ways to address this, but it's difficult for me to imagine a future with our current economic system (in the US) that allows that to happen (like UBI).
This question is poorly formed because it’s not clear who the “we” is. If it’s you and me, that train left the station a while ago. If it’s any humans, well Sam Altman is probably a human and all of these are impressive products, but still just tools.
To use a fictional but entirely apt quote:
> I say your civilization because as soon as we started thinking for you, it really became our civilization, which is, of course, what this is all about: Evolution, Morpheus, evolution. Like the dinosaur. Look out that window. You had your time. The future is our world, Morpheus. The future is our time.
I haven't heard of any large datasets of IPA transcripts of speech with the detail necessary to train a fully realistic STT->LLM->TTS system. If you know of some that would be interesting to look at.
They are also on pace to exceed $1B in revenue. [1]
[0] https://www.reuters.com/technology/chatgpt-sets-record-faste...
[1] https://www.theinformation.com/articles/openai-passes-1-bill...
Views are easy to inflate, I wouldn't even consider it in the same ballpark. This video[1] of Vivaldis 4 seasons has about 1/4 billion views.
The shortest time to 100 million users is almost a definition of the quickest growing company.
[0]: first or second largest youtuber. [1]: https://www.youtube.com/watch?v=GRxofEmo3HA
I used Bing yesterday and it was able to parse out exactly what I wanted, and then give me idiot-proof steps to making the recipe in-game. (I didn't need the steps, but it gave me what I wanted up front, easily.) I tried it twice and it was awesome both times. I'll definitely be using it in the future.
You mean these? Took me a few seconds to find, not sure how an LLM would make that easier. I guess the biggest benefit of LLM then is for people who don't know how to find stuff.
Bing made it even easier.
Also, I've found some of those lists to be missing some recipes.
Imagine if iOS had something like apple script and all apps exposed and documented endpoints. LLMs would be able to trivially solve problems that the best voice assistants today cannot handle.
Then again none of the current assistants can handle all that much. "Send Alex P a meeting invite tomorrow for a playdate at the Zoo, he's from out of town so include the Zoo's full address in the invite".
"Find the next mutual free slot on the team's calendar and send out an invite for a zoom meeting at that time".
These are all things that voice assistants should have been doing a decade ago, but I presume they'd have required too much one off investment.
Give an LLM proper API access and train it on some example code, and these problems are easy for it to solve. Heck I bet if you do enough specialized training you could get one of the tiny simple LLMs to do it.
What you’re describing as “clinging to an old way of things” is how every single thing has been, ever, new or old.
I wish MLs were more useful than search engines, but they have still a long way to go to replace them (if they ever do).
You could literally hire an entirely new guy, give him instructions to build such an email system, and let him put the right triggers on the user account permissions database to send out the right emails at the right time.
And then, when it's built, you can start adding more features like sending the emails only when demand is low and/or at times of day when you get the best click through rate. And then next you can start measuring the increase in revenue from sending those emails.
Before long, you have a whole marketing and comms team. Which you probably want as a big company anyway.
Incremental rollouts are hard.
Incremental rollouts are hard, and so is communicating in a way that makes everyone happy.
My perception is apes still feel like the main character, each and every one of them.
well that would be good if Reuters were the source of the figure
but they're not, they're simply reporting on what SimilarWeb has estimated
https://www.similarweb.com/website/chatgpt.com/#technologies
and that estimate is essentially based on nothing
The last four chats with ChatGPT (not GPT4) where a constant flow of non existent API functions with new hallucinations after each correction until we reached full circle.
Years ago I had a business DSL customer with a router and static IP. From everything in my testing it appeared that traffic broke somewhere at the local telco, not with my modem. It took 8 straight hours of arguing with L1 that no, it is not my windows. No, we have a router and it's not a computer issue. No, it's not the router (we could put the router in DHCP mode and it would work), it was an issue with static IP.
The next day we finally broke out of the stupid loop and got to IP services, who where just as confused. Eventually they were on the phone with people on the floor of the local office. A card of some type had been pulled and put in the wrong slot. Ooof.
I can read the website, I don't need a fake person to give me the information available on the website. When I contact support, it's because I need to talk to a human.
And yes, in 20 years you can tell your kids that 'back in my day' support consisted of real people. But truthfully, as someone who worked on a ISP helpdesk it's much better for society if these people move on to more productive areas.
But is it, though? I started my career in customer support for a server hosting company, and eventually worked my way up to sysadmin-type work. I would not have been qualified for the position I eventually moved to at the start, I learned on the job. Is it really better for society if all these entry level jobs get automated, leaving only those with higher barriers to entry?
The answer then was extending schooling, so that people (children at the time) could learn those skills without having their labour exploited. I would argue we should consider that today, extend mandatory free schooling. The economic purpose of education is that at the end of it the person should be able to have a job, removing entry level jobs doesn't change the economic purpose of education, so extend education until the person is able to have a job at the end of it again.
The social purpose of schooling is to make good members of society, and I don't think that cause would be significantly harmed by extending schooling in order for students to have learned enough to be more capable than an LLM in the job market.
Considering that the democratic backsliding across the globe is coincidentally happening at the same time as the rise of social media and echo chambers, are we sure about that? LLM have the opportunity to create a handcrafted echo chamber for every person on this planet, which is quite risky in an environment where almost every democracy of the planet is fighting against radical forces trying to abolish it.
People like to just suppose that these will help discover drugs and design buildings and what not, but what we actually know they’re capable of doing is littering our information environment at massive scale.
But you don't see the positive, you just have faith. That's beautiful in a way, but dangerous too. Just like the common idea that "I have faith that somebody will find a technological solution to climate change". When the risk is that high, I think we should take a step back and don't bet our survival on faith.
Aren't they unprofitable? and have fierce competition from everyone?
Don't users validate your market? ChatGPT has plenty of users, so I would think competitors only hurt their value.
a whole bunch of AI startups were founded around the same time. surely each can't validate the market for the others and be validated by the others in turn
Plus they make 20 dollars a month from a lot of people.
The best defence is to move so quickly that you are an established part of the business framework by the time these forces can gather, or to go so slowly that nobody takes you as a threat.
No startup can go slowly.
Otherwise known as the AirBnB playbook.
I'm sure we've all seen companies grow too fast become less productive than when there were a ten person startup.
Clearly, you can be a company like Microsoft where nobody is challenging your dominance in PC operating systems, and you can make huge sums of money. So competitors certainly aren't vital.
Or if you've cleverly sewn up a market with patents or trade secrets or a giant first-mover advantage or network effects, and nobody's got any chance of challenging your dominance in your specific market niche - again that could be highly profitable.
On the other hand, if you're selling ten-dollar bills for five dollars, you might have millions of satisfied paying customers, but no competitors because nobody else thinks your unit economics make sense. Or if you run a DVD rental store, you might be profitable and have many return customers, but you might not attract competitors because they don't think DVD rental is a growth business.
So some people consider a lack of competition an ominous sign.
In ten years we will enjoy a higher productivity due to AI and a richer society as a result. We have already seen it with protein folding which AI is amazing at[0].
The only reasonable fear of AI is for some jobs and that China gets their first.
This is no different then saying "Look, nuclear weapons aren't actually dangerous, if they were we'd all be dead because the apocalypse would have already happened", which is probably the dumbest take on the close calls and real risks that exist.
In my view, while statistical models would probably be an improvement ( assuming all confounding factors are measured ), the ultimate solution is not to get better at educated guessing, but to remove the guessing completely, with diagnostic tests that measure the relevant bio-medical markers.
This becomes even more true when you consider there is risk to every test. Some tests have obvious risks (radiation risk from CT scans, chance of damage from spinal fluid tap). Other tests the risk is less obvious (sending you for a blood test and awaiting the results might not be a good idea if that delays treatment for some ailment already pretty certain). In the bigger picture, any test that costs money harms the patient slightly, since someone must pay for the test, and for many the money they spend on extra tests comes out of money they might otherwise spend on gym memberships, better food, or working fewer hours - it is well known that the poor have worse health than the rich.
But if we are talking about being truly transformative - then a Star-trek tricorder is the ultimate goal, rather than a better version of twenty questions in my view.
So I'm not saying it's not useful, just that it's not the ultimate solution.
> Similar possibilities existed in medicine for 50 years
It would've been like building the tower of babel with a bunch of raspbery pi zeros. While theoretically possible, practically impossible and not (just) because of laws, but rather because of structural limitations (vector dbs of the internet solves that)
> Patents and byzantine regulations will stunt its potential
Thats the magic of this technology, its like AWS for highly levered niche intelligence. This arms an entire generation of rebels (entrepreneurs & scientists) to wage a war against big pharma and the FDA.
As an aside, this is why I'm convinced AI & automation will unleash more jobs and productivity like nothing we've seen before. We are at the precipice of a Cambrian explosion! Also why the luddites needs to be shunned.
Imagine for example that 'disease books' are published each month with tables of disease probabilities per city, per industry, per workplace, etc. It would also have aggregated stats grouped by by age, gender, religion, wealth, etc.
Your GP would grab the page for the right city, industry, workplace, age, gender etc. That would then be combined with the pages for each of the symptoms you have presented with, and maybe further pages for things from your medical history, and test results.
All the pages would then be added up (perhaps with the use of overlayed cellophane sheets with transparency), and the most likely diseases and treatments read off.
When any disease is then diagnosed and treatment commenced (and found effective or ineffective), your GP would fill in a form to send to a central book-printer to allow next months book edition to be updated with what has just been learned from your case.
can you, though? it's not scalably confirmable. what you can say in a British accent to another human person in the physical world is not necessarily what you can say in unaccented text on the internet.
Funnily enough, it is scalably confirmable. You can feed all my HN comments before chatGPT into well.. chatGPT and ask it whether I'm british based on the writing.
I bet we are just a version or two away from being able fine tune it down to region based on writing. There are so many little things based on whether your from Scotland, Wales or London. Especially London!
Medical secrecy, processes and laws have indeed prevented SOME things, but a lot of things have gotten significantly better due to enhanced statistical models that have been implemented and widely used in real life scenarios.
Example: my favourite team is X. So if I want to keep it a secret, when I ask for the history of championships of X, I will ask for X. My local agent should ask for 100 teams, get all the data, and then report back for only X. Eventually the mothership will figure out what we like (a large wenn diagram). But this is not in anyone's interest, and thus will not happen.
Also, like this the local agent will be able to learn and remember us, at a cost.
The medical possibilities that will be unlocked by large generative deep multimodal models are on an entirely different scale from "statistical diagnoses." Imagine feeding in an MRI image, asking if this person has cancer, and then asking the model to point out why it thinks the person has cancer. That will be possible within a few years at most. The regulatory challenges will be surmounted eventually once it becomes exceedingly obvious in other countries how impactful this technology is.
Your deep multimodal models or the MRI imaging?
What you are essentially saying is the signal is so subtle that only a large NN can reliably extract it.
While that may well be the case, it would be better to have a scan/diagnostic that doesn't need that level of signal processing to interpret.
For example - you don't need a large generative deep multimodal model to read a Covid antigen or PCR test.
We've picked a lot of the low-hanging simple to extract signals, we need large models to go to the next phase for things like parkinsons, etc.
I have tried to use it many times to learn a topic, and my experience has been that it is either frustratingly vague or incorrect.
It's not a tool that I can completely add to my workflow until it is reliable, but I seem to be the odd one out.
I find this highly concerning but I feel similar.
Even "smart people" I work with seem to have gulped down the LLM cool aid because it's convenient and it's "cool".
Sometimes I honestly think: "just surrender to it all, believe in all the machine tells you unquestionably, forget the fact checking, it feels good to be ignorant... it will be fine...".
I just can't do it though.
It's the same issue with Google Search, any web page, or, heck, any book. Fact checking gets you only so far. You need critical thinking. It's okay to "learn" wrong facts from time to time as long as you are willing to be critical and throw the ideas away if they turn out to be wrong. I think this Popperian view is much more useful than living with the idea that you can only accept information that is provably true. Life is too short to verify every fact. Most things outside programming are not even verifiable anyway. By the time that Steve Jobs would have "verified" that the iPhone was certainly a good idea to pursue, Apple might have been bankrupt. Or in the old days, by the time you have verified that there is a tiger in the bush, it has already eaten you.
The benefit is that I got a quick look at various solutions and quickly satisfied a curiosity, and decided if I’m interested in the concept or not. Without AI, I might just leave the idea alone or spend too much time figuring it out. Or perhaps never quite figure out the terms of what I’m trying to discover, as it’s good at connecting dots when you have an idea with some missing pieces.
I wouldn’t use it for a conversation about things as others are describing. I need a way to verify its output at any time. I find that idea bizarre. Just chatting with a hallucinating machine. Yet I still find it useful as a sort of “idea machine”.
I think it's because Americans, more than nearly all other cultures, love convenience. It's why the love for driving is so strong in the US. Don't walk or ride, drive.
Once I was walking back from the grocer in Florida with 4 shopping bags, and people pulled over and asked if my car had broken down and if I needed a ride, people were stunned...I was walking for exercise and for the environment...and I was stunned.
More evidence of this trend can be seen in the products and marketing being produced:
Do you need to write a wedding speech? Click here.
Do you need to go get something from the store? get your fat ass in the car and drive, better yet, get a car that drives for you? Better than this, we'll deliver it with a drone...don't move a muscle.
Don't want to do your homework? Here...
Want to produce art? Please enter your prompt...
Want to lose weight? We have a drug for that...
Want to be the authority on some topic? We'll generate the facts you need.
This. I hate being told the wrong information because I will have to unlearn the wrong information. I would rather have been told nothing.
They're only good on universal truths. An amalgam of laws from around the globe doesn't tell me what the law is in my country, for example.
I feel like using LLM today is like using search 15 years ago - you get a feel for getting results you want.
I'd never use chatGPT for anything that's even remotely obscure, controversial, or niche.
But through all my double-checking, I've had phenomenal success rate in getting useful, readable, valid responses to well-covered / documented topics such as introductory french, introductory music theory, well-covered & non-controversial history and science.
I'd love to see the example you experienced; if I ask chatGPT "tell me about Toronto, Canada", my expectation would be to get high accuracy. If I asked it "Was Hum, Croatia, part of the Istrian liberation movement in the seventies", I'd have far less confidence - it's a leading question, on a less covered topic, introducing inaccuracies in the prompt.
My point is - for a 3 hour drive to cottage, I'm OK with something that's only 95% accurate on easy topics! I'd get no better from my spouse or best friend if they made it on the same drive :). My life will not depend on it, I'll have an educationally good time and miles will pass faster :).
(also, these conversations always seem to end in suffocatingly self-righteous "I don't know how others can live in this post-fact free world of ignorance", but that has a LOT of assumptions and, ironically, non-factual bias in it as well)
A person who uses ChatGPT must have the understanding that it's not like Google search. The layman, however, has no idea that ChatGPT can give coherent incorrect information and treats the information as true.
Most people won't use it for infotainment and OpenAI will try its best to downplay the hallucination as fine print if it goes fully mainstream like google search.
Rather than asking it about facts, I find it useful to derive new insights.
For example: "Tell me 5 topics about databases that might make it to the front page of hacker news." It can generate an interesting list. That is much more like the example they provided in the article, synthesizing a bed time story is not factual.
Also, "write me some python code to do x" where x is based on libraries that were well documented before 2022 also has similarly creative results in my experience.
Like talking to most people you mean?
All human interactions from all of history called and they …
I verify just about everything that I ask it, so it isn’t just a general sense of improvement.
Ah yes, I dont understand how to talk to people either!
Comments like yours make me think that no one cares about this...and judging by a lot of the other comments, I guess they don't.
Probably going to be people, wading through a sea of AI generated shit, and the individual is supposed to just forever "apply critical thinking" to it all. Even a call from ones spouse could be fake, and you'll just have to apply critical thinking or whatever to workout if you were scammed or not.
Then it makes stuff up far less frequently.
If the next version has the same step up in performance, I will no longer consider inaccuracy an issue - even the best books have mistakes in them, they just need to be infrequent enough.
> Then it makes stuff up far less frequently.
Now there's a business model for a ChatGPT-like service.
$1/month: Almost always wrong
$10/month: 50/50 chance of being right or wrong
$100/month: right 95% of the time
I don't think it's quite the same.
With search results, aka web sites, you can compare between them and get a "majority opinion" if you have doubts - it doesn't guarantee correctness but it does improve the odds.
Some sites are also more reputable and reliable than others - e.g. if the information is from Reuters, a university's courseware, official government agencies, ... etc. it's probably correct.
With LLMs you get one answer and that's it - although some like Bard provide alternate drafts but they are all from the same source and can all be hallucinations ...
Yes and no. If the LLM is repeating the same thing on multiple drafts then it's very unlikely to be a hallucination.
It's when multiple generations are all saying different things that you need to take notice.
LLMs hallucinate yes but getting the same hallucination multiple times is incredibly rare.
I've seen the hallucination rate of LLMs improve significantly, if you stick to well covered topics they probably do quite well. The issue is they often have no tells when making things up.
You will now be able to feed it images and responses of the customers. Give it a function to call complementaryDrink(customerId) Combine it with a simple vending machine style robot or something more complex that can mix drinks.
I'm not actually in a hurry to try to replace bartenders. Just saying these types of things immediately become more feasible.
You can also see the possibilities of the speech input and output for "virtual girlfriends". I assume someone at OpenAI must have been tempted to train a model on Scarlett Johansson's voice.
If people are treating LLMs like a random stranger and only making small talk, fair enough, but more often they're treating it like an inerrable font of knowledge, and that's concerning.
That's on them. I mean, people need to figure out that LLMs aren't random strangers, they're unfiltered inner voices of random strangers, spouting the first reaction they have to what you say to them.
Anyway, there is a middle ground. I like to ask GPT-4 questions within my area of expertise, because I'm able to instantly and instinctively - read: effortlessly - judge how much to trust any given reply. It's very useful this way, because rating an answer in your own field takes much less work than coming up with it on your own.
I think even if an AGI was created, and humans survived this event. I'd still have trouble trusting it.
The quote "trust but verify" is everything to me.
I don't like being told lies in the first place and having to unlearn it.
It doesn't help that I might as well have just gone straight to the "verification" instead.
It's smart but can also be very dumb.
I think the current state of AI trustworthiness (“very impressive and often accurate but occasionally extremely wrong”) triggers similar mental pathways to interacting with a true sociopath or pathological liar for the first time in real life, which can be intensely disorienting and cause one to question their trust in everyone else, as they try to comprehend this type of person.
It seems to be able to speak on history, sometimes it's even right, so there's a use case that people expect from it.
FYI I've used GPT4 and Claude 2 for hundreds of conversations, I understand what its good and bad at; I don't trust that the general public is being given a realistic view.
As convenience in a domain becomes ubiquitous or at least expected among consumers, they quickly readjust their evaluation of "having time for X" around the new expectation of the convenient service, treating all alternatives as positive opportunity cost. This would explain a lot of those folks who are upset when it's suggested that they don't need Amazon, Instacart, etc. in their lives if they are to do something about their contributions to mass labor exploitation.
Of course these conveniences quickly become ubiquitous in large economies with a glut of disposable income, which encourages VCs to dump money into these enterprises so they're first to market, and also to encourage the public to believe that the future is already here and there's no reason to worry about backsliding or sustainability of the business model. Yet in every single case we see prices eventually rise, laborers squeezed, etc. A critical mass of people haven't yet acknowledged this inevitability, in no small part due to this fixation on convenience at the expense of more objective, reasoned understandings (read: post-truth mindset).
no I will not give the public credit, most people have no grounding to discern wtf a language model is and what it's doing, all they know is computers didn't use to talk and now they do
Take a look at https://chat.openai.com/share/41bdb053-facd-448b-b446-1ba1f1... for example.
Context: had a bunch of photos and videos I wanted to share with a colleague, without uploading them to any cloud. I asked GPT-4 to write me a trivial single-page gallery that doesn't look like crap, feeding it the output of `ls -l` on the media directory, got it on first shot, copy-pasted and uploaded the whole bundle to a personal server - all in few minutes. It took maybe 15 minutes from the idea of doing it first occurring to me, to a private link I could share.
I have plenty more of those touching C++, Emacs Lisp, Python, generating vCARD and iCalendar files out of blobs of hastily-retyped or copy-pasted text, etc. The common thread here is: one-off, ad-hoc requests, usually underspecified. GPT-4 is quite good at being a fully generic tool for one-off jobs. This is something that never existed before, except in form of delegating a task to another human.
https://chat.openai.com/share/338e7397-0201-44f4-a2c3-75b733...
I use ChatGPT for all sorts of things - looking into visas for countries, coding, reverse engineering companies from job descriptions, brainstorming etc etc.
It saves a lot of time and gives way more value than what you pay for it.
It used to be more reliable when web browsing worked, but it's still pretty reliable.
When I spend time on something that turns out to be incorrect, I would prefer it to be because of choice I made instead of some random choice made by an LLM. Maybe the author is someone I'm interested in, maybe there's value in understanding other sides of the issue, etc. When I learn something erroneous from an LLM, all I know is that the LLM told me.
People should be able "throw the ideas away if they turn out to be wrong" but the problem is these ideas unconsciously or not help build your model of the world. Once you find out something isn't true it's hard to unpick your mental model of the world.
Intuitively, I would think the same, but a book about education research that I read and my own experience taught me that new information is surprisingly easy to unlearn. It’s probably because new information sits at the edges of your neural networks and do not yet provide a foundation for other knowledge. This will only happen if the knowledge stands the test of time (which is exactly how it should be according to Popper). If a counterexample is found, then the information can easily be discarded since it’s not foundational anyway and the brain learns the counterexample too (the brain is very good in remembering surprising things).
and that wouldn't eliminate hallucinations just tell you if large details have likely been hallucinated.
But it's a method some research has used.
P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
"In particular, we find that LMs often hallucinate differing authors of hallucinated references when queried in independent sessions, while consistently identify authors of real references. This suggests that the hallucination may be more a generation issue than inherent to current training techniques or representation."
https://arxiv.org/abs/2303.08896
"SelfCheckGPT leverages the simple idea that if a LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another."
>Just do the comparison on the user’s machine if the LLM provider is that cheap.
This is not possible. Users don't have the resources to run these gigantic models. LLM inference is not cheap. Open ai, Google aren't running profit on free cGPT or Bard.
>P.S. Also aren’t LLMs deterministic if you set their “temperature” to zero? Are there drafts if the temperature is zero? If not, then that’s the same as removing the randomness no?
It's not a problem of randomness. a temp of 0 doesn't reduce hallucinations. LLMs internally know when they are hallucinating/taking a wild guess. randomness influences how that guess manifests each time but the decision to guess was already made.
I never said it did.
> LLMs internally know when they are hallucinating/taking a wild guess.
No they don’t. If they did we would be able to program them to not do so.
I would argue that wild guesses are all LLMs are doing. They practically statistically guess their way to an answer. It works surprisingly well a lot of the time but they don’t really understand why they are right/wrong.
P.S. LLMs are kind of like students who didn’t study for the test so they use “heuristics” to guess the answer. If the test setter is predictable enough, the student might actually get a few right.