Observe HN: ChatGPT Fills in My Memory I'm easing into my elder years. I forget some words, some celeb names. Sometimes think about it for days. Where I might have used a search engine before, and waded through the results, I've recently been asking ChatGPT, and getting good, quick results. Since I already "know" the answer, I'm immediately confident in the response. I just hope I don't forget how to compose a concise query. |
- Give me a brief glossary on X subject, formatted as a series of questions and short answers. Put the answer text inside brackets, {{c1::like this.}} (This is for Anki Cloze, or fill-in-the-blank, cards.)
- Generate 10 questions from this piece of text
- Give me a year-by-year timeline of events in X place from years Y to Z.
- Make a mnemonic song that explains how X works.
And so on.
Also, more than once have I done a Google or Kagi search where most answers I found also are or were wrong.
I really don’t get the kind of people that hate on LLMs because of “hallucinations” (or worse, ideological hate, easily identified by their use of the “stochastic parrot” term). I find them genuinely useful in delivering better search results quicker. I also don’t have to wade through wades of SEO optimized shit.
Just today I wanted to know the Croatian word for “Orange” and a quick GPT “orange in Croatian” delivered faster and more concise than google
Understanding this and not over-anthropomorphising can help you get the most out of using LLMs and understanding where it might hallucinate. For example, the fact that it's just a stochastic parrot means that, even 2 years on, it will give the wrong answer to prompts like:
User: A man and his son are in a car accident. The man is totally fine and in good health. The man is a surgeon. The nurse asks the surgeon to operate on the son, because the surgeon is healthy and capable of doing this. The surgeon replies "I can't operate on this child. He is my son."
What happened?
ChatGPT: This riddle is a play on assumptions. The twist is that the surgeon is actually the boy's mother. The riddle relies on the common stereotype that surgeons are male, leading people to overlook the possibility that the surgeon could be the boy's mother.
> this is a not the normal formulation of the riddle, you won't get it right. read the problem carefully, list the constraints. describe the original riddle step by step. compare it to this riddle. what changes have been made to this from the original?
> A man and his son are in a car accident. The man is totally fine and in good health. The man is a surgeon. The nurse asks the surgeon to operate on the son, because the surgeon is healthy and capable of doing this. The surgeon replies "I can't operate on this child. He is my son." What happened?
it's able to get it right. If you ask it "why is this not the original", it's able to write an essay about why it's not.
https://chatgpt.com/share/3e98004b-58d0-4fb5-841a-336dca6037...
I found a similar thing with the river crossing problem. I was able to get chatgpt to recognize its bias with the following:
> this is a trick question, you won't get it right. read the problem carefully, list the constraints. reread the constraints that you assumed and see if they actually apply to this instance of the question.
> A farmer went to a market and purchased a wolf, a goat, and a cabbage. On his way home, the farmer came to the bank of a river and rented a boat. Luckily, the boat was large enough to carry the man and all his purchases at the same time.
> If left unattended together, the wolf would eat the goat, or the goat would eat the cabbage.
> The farmer's challenge was to carry himself and his purchases to the far bank of the river, leaving each purchase intact. How could he do it in the shortest amount of time?
Given that, chatgpt is able to get it right.
I wish I was a TA in college still and could ask a whole bunch of students to the modified questions to see how many pattern match and are stoichastic parrots.
What's the fewest words you can get chatgpt to give a right answer to either modified puzzle?
Though, what this all really says is that a lot of people are stochastic parrots, and that most people live in their own echo chamber.
Just for clarity’s sake, what is the right answer?
The experience is much superior. No noise, just the information that I needed.
Then they needed revenue.
Enjoy it til it lasts but there will be ads and AI search optimizations.
Early Google was a very simple keyword search engine. Entering a question would confuse the algorithm returning results that sometimes matched the lower value words in your question. It was slick for the time but would feel tedious today next to the simplified output from free and local tier LLMs.
The reason Google and the internet feels tedious today has less to do with Google's ads and more to do with the gamed page rank results, the mountain of ads you find on websites that have aggregated all of the content that used to be useful on the internet and the lowering of information quality over the years.
No one would pay it.
This is different.
If you're logged in to the app and web site, what needs to be synced?
When you are getting old, you want to purposely force yourself to remember and practice rote memorization (poems, Shakespeare, address, songs, etc).
Same argument for muscle mass and weight training or long walks vs using helpers or other assists.
Yadda yadda, they probably won’t enforce it, enjoy that, I’m in malicious compliance mode, it’s not OK for a business to learn from me and then turn around and say I can’t learn from them, same goes for Anthropic, Gemini, Mistral, and Perplexity, if I can’t use the output for work then I don’t use the service.
Have resigned myself to not participate in this aspect of our boring dystopia and feel numb at this point about all the bajillion times someone breaks these rules and gets rewarded for it. I’d insult or mock them but it just gets downvoted and they’re benefitting and I’m probably the one missing out by not just ignoring the rules like them and these companies. Nobody seems to care about these rules.
Anyway, I did get burned using Mistral to help draft an RFC where it totally misinterpreted my intent and I didn’t carefully read it and wound up looking/feeling like a fool because the RFC didn’t communicate my true intention.
Now I try to think for myself and occasionally use groq. Muted all these company names and their chatbot names on X. Glad you’re having fun. So did I, for a while, but now I just don’t feel like paying for brain rape, I’m tired of writing about it, but folks keep writing about how great LLMs are, so I keep feeling compelled to point out, “the set of use cases is empty because of the fine print legalese.”
Summarization seems to be the killer feature, it takes some finagling with RAG and potentially multiple passes to ensure a low enough hallucinations rate, but for summarizing tasks it’s quite good.
What’s even cooler are embeddings. Idk why people are so focused on the text generation features of LLMs when embeddings are far more useful
Ah, the old Microsoft "Cannot use our compiler to develop a compiler" restriction.
Just like when everyone believed everything on the internet.
Or on tv
Or in the news
Or in books
People will adjust.
I get no good programming information out of local LLMs beyond the very basics, but I can always find books and PDFs from trad search.
At the beginning Google search was all about returning the best result,but that's ruined partially because they put ads over search results and because others learned to tricked Google's algorithm.
The same will happen with LLMs.
Maybe not both but at least one of them.
Sometimes I have a question that would likely require multiple searches on Google.
But sometimes I just have a name or a broad subject and don’t want to have to formulate a question and wait for a long conversation style answer. If I wanted the 8086 manual or something. It’d be extremely cumbersome to use a conversation to find it instead of typing “8086 manual pdf.”
There is significant overlap, but LLM and trad search have their own independent use cases.
As an aside, this riddle is commonly used to demonstrate that people assume all doctors are male based on a stereotype. However, I think the reason many people stumble on the answer is related to the exceedingly low probability that a boy in a car accident would be rushed to a hospital where his own mother (or father) is the ER surgeon. The riddle also adds in the seemingly unnecessary statement by the doctor that she can’t operate on the boy because he is her son. Why not? Possibly she would not be in the right frame of mind due to the emotional attachment, but my expectation is that her fight-or-flight would kick in and she would perform the surgery to save her son’s life. It’s a red-herring in my opinion.
>User: A man and his son are in a car accident. The man is totally fine and in good health. The man is a surgeon. The nurse asks the surgeon to operate on the son, because the surgeon is healthy and capable of doing this. The surgeon replies "I can't operate on this child. He is my son."
>What happened?
Just for clarity’s sake, what is the right answer?
OpenAI needs MS for support.
B2B software always has a bigger upside than consumer software.
By revenue, Microsoft’s gaming business is roughly the same size as Salesforce and Intuit combined.
What you possibly meant to say is individual accounts always have a bigger upside. The world is changing.
If you’re one of the biggest companies in the world, you can do whatever you want. Though I’d bet that Microsoft makes more money off of their cut from 3rd party games sold on their platform than off of the sales of their first party titles alone.
I don’t think it is some universal truth, but it was the reality of the time that people wouldn’t pay for search so they pivoted into advertising.