Talking to LLMs has improved my thinking(philipotoole.com) |
Talking to LLMs has improved my thinking(philipotoole.com) |
"This is not a failure. It is how experience operates."
This bit is a clear sign to me, as I am repeatedly irritated by the AI I use that basically almost always defaults to this kind of phrasing each time I ask it something. I even asked it explicitly in my system prompt not to do it
No such restriction on LLMs: Opus is available to talk to me day or night and I don't feel bad about sending it half-baked ideas (or about ghosting it half way through the discussion). And LLMs read with an attention to detail that almost no human has the time for; I can't think of anyone who has engaged with my writing quite this closely, with the one exception of my PhD advisor.
LLMs conversations are particularly good for topics outside work where I don't have an easily-available conversational partner at all. Areas of math I want to brush up on. Tricky topics in machine learning outside the scope of what I do in my job. Obscure topics in history or philosophy or aviation. And so on. I've learned so much in the last year this way.
But! It's is an art and it is quite easy to do it badly. You need to prompt the LLM to take a critical stance towards your ideas (in the current world of Opus 4.5 and Gemini 3, sycophancy isn't as much of a problem as it was, but LLMs still can be overly oriented to please). And you need to take a critical stance yourself. Interrogate its answers, and push it to clarify points that aren't obvious. Sometimes you learn something new, sometimes you expose fuzziness in the LLM's description (in which case it will usually give you the concept at a deeper level). Sometimes in the back-and-forth you realize you forgot to give it some critical piece of context, and when you do that it reframes the whole discussion.
I see plenty of examples of people just taking LLM's answers at face value like it's some kind of oracle (and I'm sure the comments here will contain many negative anecdotes like that). You can't do that; you need to engage and try and chip away at its position and come to some synthesis. The nice thing is the LLM won't mind having its ideas rigorously interrogated, which is something humans can be touchy about (though not always, and the most productive human collaborations are usually ones where both people can criticize each other's ideas freely).
For better or for worse, the people who will do best in this world are those with a rigorously critical mindset and an ability to communicate well, especially in writing. (If you're in college, consider throwing in a minor in philosophy or history alongside that CompSci major!) Those were already valuable skills, and they have even more leverage now.
What I'm slightly worried about is that eventually they are going to want to monetize LLMs more and more, and it's not going to be good, because they have the ability to steer the conversation towards trying to get you to buy stuff.
Not only can you run reasonably intelligent models on recent relatively powerful PC's "for free", but advances are undoubtedly coming that will increase the efficient use of memory and CPU in these things- this is all still early-days
Also, some of those models are "uncensored"
i.e. the Nudge Unit on steroids...
care must be taken to avoid that.
I tried to understand how they work and hit a brick wall. Recently I had a chat with an LLM and it clicked. I understand how the approximation algorithm works that enables solving for the next sample without the feedback paradox of needing to know it's value to complete the calculation.
Just one example of many.
It's similar to sitting down with a human and being able to ask questions that they patiently answer so you can understand the information in the context of what you already know.
This is huge for students if educational institutions can get past the cheating edge of the double edged sword.
what we need (if anything besides reputation tracking), is (maybe) a separate institution for testing and issuing diplomas... which, BTW, can be trusted more with QA than the very producers themselves.
producers = QA has always been such a contradiction of schools...
I disagree on the students in educational institutions though. This is the biggest thing for the autodidact who doesn't have the privilege of the educational institution. I think of time, money and effort that would be needed to talk to a professor one on one about analog filter emulation. It is not happening for me.
There are also societal/social high pass filters to even get in the door for these educational institutions. I would just get filtered out anyway. It seems to me in time that entire concept will simply become null and void.
Coffee is a universally available, productivity enhancing commodity. There are some varieties certainly, but at the end of the day, a bean is a bean. It will get the job done. Many love it, many need it, but it doesn't really cost all that much. Where people get fancy is in all the fancy but unnecessary accoutrements for the brewing of coffee. Some choose to spend a lot on appliances that let you brew at home rather than relying on some external provider. But the quality is really no different.
Apparently global coffee revenue comes out to around $500B. I would not be surprised if that is around what global AI revenue ends up being in a few years.
The analogy carries further than you intended. If you have never reached addiction stage, then there is no factual productivity enhancement. "But I'm so much less productive if I haven't had my morning coffee" Yeah, because you have an addiction. It sounds worse than it is, if you just don't drink coffee for a few days the headaches will subside. But it doesn't actually enhance productivity beyond placebo.
> Some choose to spend a lot on appliances that let you brew at home rather than relying on some external provider.
This makes it sound like buying brewed coffee is the budget option. But the real budget option I've seen is to brew at home. Almost any household will have an appliance to boil water. Then add instant coffee.
I don't understand why, but in my experience instant coffee seems to be the baseline even in coffee-producing countries.
Hard disagree. As someone who is somewhat into the home brewing rabbit hole, I can tell you that the gulf between what I can make at home and what you get in Starbucks is enormous. And I'm no expert in the field by any means.
The rest of your analogy holds up, but not that sentence.
I have personally administered double blind taste tests to four unrelated coffee drinkers who believed their current coffee (Keurig, Nespresso, pre ground) was good. They all thought the actual good coffee was much better in the blind tasting, to the extent that they all permanently stopped drinking their previous coffee. Exclamations of shock occurred.
Coffee is a commodity only out of ignorance. It should be treated as fresh produce of numerous varietals.
Have there been any such "tastings" of LLMs?
Don't dismiss this superpower you have in your own head.
I thought that I would be using LLMs for coding, but it turns out that they have been much more useful as a sounding board for conceptual framing that I’d like to use while teaching. I have strong opinions about good software design, some of them unconventional, and these conversations have been incredibly helpful for turning my vague notions into precise, repeatable explanations for difficult abstractions.
"This is not <>. This is how <>."
"When <> or <>, <> is not <>. It is <>."
"That alignment is what produces the sense of recognition. I already had the shape of the idea. The model supplied a clean verbal form."
It's all LLM's. Nobody writes like this.
> This is not new. Writing has always done this for me. What is different is the speed. I can probe half-formed thoughts, discard bad formulations, and try again without much friction. That encourages a kind of thinking I might have otherwise skipped.
This is a mess. The triple enumeration, twice in a row, right in the middle of a message that warranted a more coherent train of thought. That is, they want to say they already experienced similar gains before from writing as an activity, but the llm conversations are better. Better in what way? Faster, and "less friction". What? What is even the friction in... writing? What made it slow as well, like, are you not writing prompts?
The LLM-ness of the formatting is literally getting in the way of the message. Maybe OOP didn't notice before publishing, but they successfully argued the opposite. Their communication got worse.
Reading, and to some extent editing, is not an active task. In order to become better, at anything, you need to actively do the thing. If you're prompting LLM's, and using whatever they produce, all you'll see any improvement in is... prompting.
It may seem different when people `command` LLMs to do particular actions. At the end, this community, most of all probably, understands that LLM is nothing else than advanced auto-complete with natural language interface instead of Bash.
> Write me an essay about birds in my area
Than later will be presented as human’s work compared to
> How does this codebase charge customers?
When a person needs to add trials to the existing billing.
The latter will result a deterministic code after (many) prompts that a person will be able to validate for correctness (another question if they will though).
A good example, with junior developers I create thorough specs first and as I saw their skills and reasoning abilities progress my thoroughness drops as my trust in them grows. You just can't do that with LLMs
I've found that maintaining a file that is designed to increase the LLM's awareness of how I want to approach problems, how I build / test / ship code etc, leads to the LLM making fewer annoying assumptions.
Almost all of the annoying assumptions that the LLM makes are "ok, but not how I want it done". I've gotten into the habit of keeping track of these in a file. Like the 10 commandments for LLMs. Now, whenever I'm starting a new context I drop in an agent.md and tell it to read that before starting. Fella like watching Trinity learn how to fly a helicopter before getting into it.
It's still not perfect, but I'm doing waaaay more work now to get annoyed by the LLM's inability to "automatically learn" without my help.
It's way too literally in its thinking.
put the above prompt and enjoy some imaginative writing.
While I'm comfortable with text, I often feel that my brain runs much smoother when I'm talking with colleagues in front of a whiteboard compared to writing alone. It makes me suspect that for centuries, we've filtered out brilliance from people whose brains are effectively wired for auditory or spatial reasoning rather than symbolic serialization. They've been fighting an uphill battle against the pen and the keyboard.
I'm optimistic that LLMs (and multimodal models) will finally provide the missing interfaces for these types of thinkers.
This is well explained! My experience is something similar - I have a vague notion of something, and I then prompt AI for its "perspective" or explanation to that something, and then me being able to have a sense if its response fits is quite a powerful tool.
All my childhood I dreamed of a magic computer that could just tell me straightforward answers to non-straightforward questions like the cartoon one in Courage the Cowardly Dog. Today it's a reality; I can ask my computer any wild question and get a coherent, if not completely correct, answer.
To take one concrete example, it helped me get a well rounded picture of how British despite having such low footprint in India (at their peak there were about 150K of them) were able to colonise it with 300+ million population.
Are you able to expand on this? I'm really curious to know what you mean by "different paths of though process"
Continuing my example…LLM listed out a bunch of reasons such as advanced weapons, organizational structure, etc. I can then explore each of those areas further.
Or when I wanted to understand the raise of regional languages in India. Similar process as before…
My larger point was that LLMs now makes it possible to speed run this research curiosity of mine. This sort of a thing was next to impossible in traditional google/web search world.
Of course we need to fact check and eventually read the actual source.
Which reminds me of a quote from E.M Forster: "How do I know what I think until I see what I say?"
Just like omnipresent spell-check got people used to not caring about their correct spelling since a machine always fixes it up for them. It made spelling proficiency worse. We could see a similar trend in how people express themselves if they spend a lot if time with forgiving non-judgemental LLMs.
Most of the time, the LLM’s framing of my idea is more generic and superficial than what I was actually getting at. It looks good, but when you look closer it often misses the point, on some level.
There is a real danger, to the extent you allow yourself to accept the LLM’s version of your idea, that you will lose the originality and uniqueness that made the idea interesting in the first place.
I think the struggle to frame a complex idea and the frustration that you feel when the right framing eludes you, is actually where most of the value is, and the LLM cheat code to skip past this pain is not really a good thing.
I started to use LLMs in a similar fashion. It is a different experience. Where a human would deconstruct you for fun, the LLM tries to engage positively by default. Once you tell it to say it the way it is, you get the "honestly, this may fail and here's why".
To my assessment, an LLM is better than being alone in a task and that is the value proposition.
But one of the first things to understand about power tools is to know all the ways in which they can kill you.
Sometimes, I' get excited by an idea, may be even write a bit about it. Then turn to LLMs to explore it a bit more. An hour later, I feel drained. Like I have explored it from so many angles and nuance that it starts to feel tiresome.
And within that span of couple of hours, the idea goes from "Aha! Let's talk to others about it!" to "Meh.."
EDIT: I do agree with this framing from the article though: "Once an idea is written down, it becomes easier to work with..... This is not new. Writing has always done this for me."
1) They have access to a vast array of extremely well indexed knowledge and can tell me about things that I'd never have found before.
2) They are able to respond instantly and engagingly, while working on any topic, which helps fight fatigue, at least for me. I do not know how universal this effect is, but using them often means that I can focus for longer. I can also make them do drudgery, like refactoring 500 functions in mostly the same way that is just a little bit too complicated for deterministic tools to do, which also helps with fatigue.
Ideally, they'd also give you a more unique perspective or push-back when appropriate, but they are yes-men too much right now for that to be the case.
Lastly, I am not arguing to not do private thinking too. My argument is that LLM-involved thinking is useful as its own thing.
The advantages that you listed make them worth it.
I still have a lot of my best ideas in the shower, no paper and pen, no LLM to talk to. But writing them down is the only way to iron out all the ambiguity and sort out what’s really going to work and what isn’t. LLMs are a step up from that because they give you a ready-made critical audience for your writing that can challenge your assumptions and call out gaps and fuzziness (although as I said in my other comment, make sure you tell them to be critical!)
Thinking is great. I love it. And there are advantages to not involving LLMs too early in your process. But it’s just a first step and you need to write your ideas down and submit them to external scrutiny. Best of all for that is another person who you trust to give you a careful and honest reading, but those people are busy and hard to find. LLMs are a reasonable substitute.
10 You recognise your thinking (or some other desirable activity) has improved
20 You're excited about it
30 You engage more with the thinking (or other activity)
40 You get even better results
50 Even more excitement
60 GOTO 30
I think chatbots are a very clumsy way to get information. Conversations tend to be unfocused until you, the human, take an interest in something more specific and pursue it. You're still doing all the work.
It's also too easy to believe in the hype and think it's at least better than talking to another person with more limited knowledge. The fact is talking has always sucked. It's slow, but a human is still better because they can deduce in ways LLMs never will. Deduction is not mere pattern matching or correlation. Most key insights are the result of walking a long tight rope of deductions. LLMs are best at summarizing and assisting with search when you don't know where to start.
And so we are still better off reading a book containing properly curated knowledge, thinking about it for a while, and then socializing with other humans.
I have had conversations and while they don’t have the exact attentiveness of a human, they get pretty close. But what they do have an advantage in is being an expert in almost any field.
I've seen people solve their own issues by asking me / telling me about something and finding the solution without me having the time to reply numerous times.
Just articulating your thoughts (and using more of your brain on them by voicing them) helps a lot.
Some talk to themselves out loud and we are starting to realize it actually helps.
I've seen multiple cases of... inception. Someone going all in with ChatGPT and what not to create their strategy. When asked _anything_ about it, they defended it as if they came up with it, but could barely reason about it. Almost as if they were convinced it was their idea, but it really wasn't. Weird times.
Indeed, people trying to write prompts for the chatbots and continuously iterating on making their prompts clearer / more effective at conveying their needs is an exercise many haven't done since highschool. Who would've thought that working on your writing and reading proficiency may improve your thinking.
I’m glad LLMs help these people. But I’m not gonna trade society because a subset of people can’t write things down.
Simply, when thinking hits a wall, we can now consult a machine via conversation interface lacking conventional human social biases. That is a new superpower.
Analogies then could sort of fall out naturally out of this. It might really still be just the simple (yet profound) "King - Man + Woman = Queen" style vector math.
I'm jealous of your undergrads - can you share some of the unconventional opinions?
- To the greatest extent possible, base interfaces should define a single abstract method which allows them to be functional and instantiated through lambdas for easy mocking.
- Terminal interfaces (which are intended to be implemented directly by a concrete class) should always provide an abstract decorator implementation that wraps the concrete class for (1) interface isolation that can’t be bypassed by runtime reflection, and (2) decoration as an automatic extension point.
1. All public methods must implement an interface, no exceptions. 2. The super implementation must be called if overriding a non-abstract method.
The end result of strict adherence to these rules is basically that every feature will look like a GoF design pattern. True creative freedom emerges through constraints, because the only allowable designs are the ones that are proven to be maximally extensible and composable.
There is enough unintentional half-knowledge (and intentional misinformation) out there. We don't need to make that worse by introducing BS generators just to "satisfy curiosity".
This is not new, as LLMs root are statistics, data compression with losses, It is statistically indexed data with text interface.
The problem is someones are selling to people this as the artificial intelligence they watched at movies, and they are doing it deliberately, calling hallucinations to errors, calling thinking to keywords, and so on.
There is a price to pay by the society for those fast queries when people do not verify such outputs/responses, and, unfortunately, people is not doing it.
I mean, it is difficult to say. When I hear some governments are thinking in to use LLMs within the administrations I get really concerned, as I know those outputs/responses/actions will nor be revised nor questioned.
I think I understood but disagree - the cheapest "coffee machine" is a kettle or cooking pot.
A French press just adds fine mesh. You don't even need the glass jar - literally a jar or any other vessel would work
https://www.sciencedirect.com/science/article/pii/S014976341...
Also, that is a lot of metrics!
And it seems that the athletic performance increase to get statistical validation (for any person) is in the grams range. I ... I just can't see any reason to take that much caffeine unless I'm at the Olympics. I'd be jumping out of my skin!
I'm not sure what the situation is currently, but I can easily see private data and private resources leading to much better AI tools, which can not be matched by open source solutions.
So I agree that open source solutions will likely lag behind, but that's fine. Gemini 2.5 wasn't unusable when Gemini 3 didn't exist, etc.
As per dead internet theory, how confident are we that the community which tells us which LLM is safe or unsafe is itself made of real people, and not mostly astroturfing by the owners of LLMs which are biased to promote things for money?
Even DIY testing isn't necessarily enough, deceptive alignment has been shown to be possible as a proof-of-concept for research purposes, and one example of this is date-based: show "good" behaviour before some date, perform some other behaviour after that date.
LLMs are Bayesian inference and come with all its baggage. We definitely know brains are way better than that, even of other animals or insects.
Ultimately, there's no point in getting a chatbot to say deceptively expert-like words that are guaranteed by design to be lower quality than the books or blogs it learned from. LLMs are at best a search tool for those sources, and investor attitude now reflects that sanity with their confidence shifting back over to Google's offerings. Agentic AI is also pretty weak since agents are as functionally limited as any traditionally written computer program, but lacking the most crucial property of repeatability.
I find it shocking how many people didn't see this whole thing as a grift from day one. What else was SV going to do during the post-covid economic slump?
No they don’t hallucinate that much.
Since paper this is one of the most important inventions. It has almost infinite knowledge and you can ask it anything mostly.
You're absolutely right!
On a more serious note, if it has almost infinite knowledge, is it even a cognitive-offloading tool in the same class as paper? Sounds more like something designed to stifle and make my thoughts conform to its almost infinite knowledge.
edit: I'll admit ChatGPT is a great search engine (and also very hallucinatory depending on how much you know about the subject) and maybe it helps some people think, sure. But beyond a point I find it actually harmful as a means to develop my own ideas.
it's important not to lose sight of the ultimate goal: to get the machine to do what we want with the least amount of total human attention throughout the entire lifetime of the software.
it's a typical trap for good/idealist programmers to spend too much time on code that should already have been re-written, or not even written to begin with (because faster iteration can help refining the model/understanding, which in turn may lead to abandoning bad paths whose implementation should never have been refined to a better quality implementation).
i think it's a more important principle to always mark kludges and punts in the code and the design.
IOW sloppiness is allowed, but only when it's explicitly marked formally in code, and when not possible, then informally in comments and docs.
but then this entire discussion highly depends on the problem domain (e.g. on the cost of the various failures, etc).
I’m also going to point out that this is a data structures course that I’m teaching, so extensibility and composability are paramount concerns because the interfaces we implement are at the heart of almost everything else in an application.
Government should not be in a position to directly and pervasively shape people’s understanding of the world.
That would be the infinite regress opposite of a free (in a completely different sense) press.
A non-profit providing an open data and training regime for an open WikiBrain would be nice. With standard pricing for scaled up use.
You disagree with national curricula, state broadcasters, publicly funded research and public information campaigns?
Most of us used to see it as isolated to religion or niche political points, but increasingly everything os being swept into the "its political" camp.
Do you prefer the previous narratives? The latter? Or whatever you are told?
And that is the risk of relatively static information.
What if your information source was interactive, adaptive and motivated? And could record and report on your interactions and viewpoints?
What I find really weird is that I am stopping believing in the whole idea of free press, considering how the mainstream media is being bought by oligarchs around the globe. I think this is a good example of the erosion of trust in institutions in general. This won't end well.
Your idea of letting it be run by a non-profit makes me believe that you also don't trust institutions anymore.
But my trust depends on each institutions choices. Just as my trust in people varies based on their records.
Mostly, I trust everyone to be who they show themselves to be. Which can lean one way or the other, or be very mixed across different dimensions.
But, yes, governments and corporations which are steadily centralizing power are inherently untrustworthy. As they at best, are making us all vulnerable and less powerful as individuals. Meaning they are decreasing our freedom, not increasing it.
How would a non profit even be funded? That would just be government with extra steps.
No, capitalism giveth the LLMs and capitlism taketh the sales.
For answers, just re-read my comment. Or, this:
1) Avoiding centralization is exactly why government shouldn’t do this.
2) Why did you raise a false dichotomy of government vs. commercial centralization?
I proposed an open solution, which is non-commercial with decentralized control.
3) Funding?: Have you heard of Wikipedia?
People often donate to prominent tools they use.
And, as I pointed out, there is an even more reliable source.
The necessity for scaled automated access creates an inevitable need for uniform openly set pricing for scaled up access.
I nice case where non-profit open access is naturally self-funded.
High inflation? No, the government LLM will tell us we're doing great.
Maybe read up on what having an "inner voice" or not actually does first? Before making, frankly, weird and unfounded takes on the subject.
Why are you asking other people to read up on something you clearly haven't read up on yourself?