Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

OpenAI Progress

progress.openai.com

Post reply on HN
Print view View original

291–300 of 372 posts

Re: OpenAI Progress

#291

Post by sarchertech »

pram wrote: post #101
mkozlows wrote: post #81

Earlier quoted context omitted.

Modern ChatGPT will (typically on its own; always if you instruct it to) provide inline links to back up its answers. You can click on those if it seems dubious or if it's important, or trust it if it seems reasonably true and/or doesn't matter much. The fact that it provides those relevant links is what allows it to replace Google for a lot of purposes.

It does citations (Grok and Claude etc do too) but I've found when I read the source on some stuff (GitHub discussions and so on) it sometimes actually has nothing to do with what the LLM said. I've actually wasted a lot of time trying to find the actual spot in a threaded conversation where the example was supposedly stated.

Same experience with Google search AI. The links frequently don’t support the assertions, they’ll just say something that might show up in a google search for the assertion.

For example if I’m asking about whether a feature exists in some library, the AI says yes it does and links to a forum where someone is asking the same question I did, but no one answered (this has happened multiple times).

Re: OpenAI Progress

#292

Post by barrell »

iammrpayments wrote: post #29
simianwords wrote: post #7

My interpretation of the progress. 3.5 to 4 was the most major leap. It went from being a party trick to legitimately useful sometimes. It did hallucinate a lot but I was still able to get some use out of it. I wouldn't count on it for most things however. It could answer simple questions and get it right mostly but never one or two levels deep. I clearly remember 4o was also a decent leap - the accuracy increased su…

I must be crazy, because I clearly remember chatgpt 4 being downgraded before they released 4o, and I felt it was a worse model with a different label, I even choose the old chatgpt 4 when they would give me the option. I canceled my subscription around that time.

Nah that rings a bell. 4o for me was the beginning of the end - a lot faster, but very useless for my purposes [1][2]. 4 was a very rocky model even before 4o, but shortly after the 4o launch it updated to be so much worse, and I cancelled my subscription.

[1] I’m not saying it was a useless model for everyone, just for me.

[2] I primarily used LLMs as divergent thinking machines for programming. In my experience, they all start out great at this, then eventually get overtrained and are terrible at this. Grok 3 when it came out had this same magic; it’s long gone now.

Re: OpenAI Progress

#293

Post by brookst »

svantana wrote: post #283
furyofantares wrote: post #97

Earlier quoted context omitted.

I have a theory about why it's so easy to underestimate long-term progress and overestimate short-term progress. Before a technology hits a threshold of "becoming useful", it may have a long history of progress behind it. But that progress is only visible and felt to researchers. In practical terms, there is no progress being made as long as the thing is going from not-useful to still not-useful. So then it goes from…

> why it's so easy to underestimate long-term progress and overestimate short-term progress I dunno, I think that's mostly post-hoc rationalization. There are equally many cases where long-term progress has been overestimated after some early breakthroughs: think space travel after the moon landing, supersonic flight after the concorde, fusion energy after the H-bomb, and AI after the ENIAC. Turing himself guesstimat…

With the exception of h-bomb/fusion and ENIAC/AI, I think all of those examples reflect a change in priority and investment more than anything. There was a trajectory of high investment / rapid progress, then market and social and political drivers changed and space travel / supersonic flight just became less important.

Re: OpenAI Progress

#294

Post by brookst »

robrenaud wrote: post #222
muzani wrote: post #195

Earlier quoted context omitted.

I don't have a source for this (there's probably no sources from anything back then) but anecdotally, someone at an AI/ML talk said they just added more data and quality went up. Doubling the data doubled the quality. With other breakthroughs, people saw diminishing gains. It's sort of why Sam back then tweeted that he expected the amount of intelligence to double every N years. I have the feeling they kept on this u…

The input size to output quality mapping is not linear. This is why we are in the regime of "build nuclear power plants to power datacenters". Fixed size improvements in loss require exponential increases in parameters/compute/data.

Most of the reason we are re-commissioning a nuclear power plant is demand for quantity, not quality. If demand for compute had scaled this fast in the 1970’s, the sudden need for billions of CPUs would not have disproven Moore’s law.

It is also true that mere doubling of training data quantity does not double output quality, but that’s orthogonal to power demand at inference time. Even if output quality doubled in that case, it would just mean that much more demand and therefore power needs.

Re: OpenAI Progress

#296

Post by ACCount37 »

stavros wrote: post #141
simianwords wrote: post #7

My interpretation of the progress. 3.5 to 4 was the most major leap. It went from being a party trick to legitimately useful sometimes. It did hallucinate a lot but I was still able to get some use out of it. I wouldn't count on it for most things however. It could answer simple questions and get it right mostly but never one or two levels deep. I clearly remember 4o was also a decent leap - the accuracy increased su…

All the replies are spectacularly wrong, and biased by hindsight. GPT-1 to GPT-2 is where we went from "yes, I've seen Markov chains before, what about them?" to "holy shit this is actually kind of understanding what I'm saying!" Before GPT-2, we had plain old machine learning. After GPT-2, we had "I never thought I would see this in my lifetime or the next two".

GPT-2 was the first wake-up call - one that a lot of people slept through.

Even within ML circles, there was a lot of skepticism or dismissive attitudes about GPT-2 - despite it being quite good at NLP/NLU.

I applaud those who had the foresight to call it out as a breakthrough back in 2019.

Re: OpenAI Progress

#297

Post by dgfitz »

Joeri wrote: post #250
jibal wrote: post #209

Earlier quoted context omitted.

> I'm not brave enough to draw a public conclusion about what this could mean. I'm brave enough to be honest: it means nothing. LLMs execute a very sophisticated algorithm that pattern matches against a vast amount of data drawn from human utterances. LLMs have no mental states, minds, thoughts, feelings, concerns, desires, goals, etc. If the training data were instead drawn from a billion monkeys banging on typewrit…

LLMs are not people, but they are still minds, and to deny even that seems willfully luddite. While they are generating tokens they have a state, and that state is recursively fed back through the network, and what is being fed back operates not just at the level of snippets of text but also of semantic concepts. So while it occurs in brief flashes I would argue they have mental state and they have thoughts. If we bu…

> Consider this thought exercise: if we were to ever do an upload of a human mind, and it was executing on silicon, would they not be experiencing feelings because their thoughts are provably a deterministic calculation?

You just said “consider this impossibility” as if there is any possibility of it happening. You might as well have said “consider traveling faster than the speed of light” which sure, fun to think about.

We don’t even know how most of the human brain even works. We throw pills at people to change their mental state in hopes that they become “less X” or “more Y” with a whole list of caveats like “if taking pill reduce X makes you _more_ X, stop taking it” because we have no idea what we’re doing. Pretending we can use statistical models to create a model that is capable of truly unique thought… stop drinking the kool-aid. Stop making LLMs something they’re not. Appreciate them for what they are, a neat tool. A really neat tool, even.

This is not a valid thought experiment. Your entire point hinges on “I don’t believe in souls” which is fine, no problem there, but it does not a valid point make.

Re: OpenAI Progress

#298

Post by verelo »

simianwords wrote: post #7

My interpretation of the progress. 3.5 to 4 was the most major leap. It went from being a party trick to legitimately useful sometimes. It did hallucinate a lot but I was still able to get some use out of it. I wouldn't count on it for most things however. It could answer simple questions and get it right mostly but never one or two levels deep. I clearly remember 4o was also a decent leap - the accuracy increased su…

Everyone talks about 4o so positively but I’ve never consistently relied on it in a production environment. I’ve found it to be inconsistent in json generation and often it’s writing and following of the system prompt was very poor. In fact it was a huge part of what got me looking closer at anthropics models.

I’m really curious what people did with it because while it’s cool it didn’t compare well in my real world use cases.

Re: OpenAI Progress

#299

Post by svantana »

brookst wrote: post #293
svantana wrote: post #283

Earlier quoted context omitted.

> why it's so easy to underestimate long-term progress and overestimate short-term progress I dunno, I think that's mostly post-hoc rationalization. There are equally many cases where long-term progress has been overestimated after some early breakthroughs: think space travel after the moon landing, supersonic flight after the concorde, fusion energy after the H-bomb, and AI after the ENIAC. Turing himself guesstimat…

With the exception of h-bomb/fusion and ENIAC/AI, I think all of those examples reflect a change in priority and investment more than anything. There was a trajectory of high investment / rapid progress, then market and social and political drivers changed and space travel / supersonic flight just became less important.

That's the conceit for the tv show For All Mankind - what if the space race didn't end? But I don't buy it, IMO the space race ended for material reasons rather than political. Space is just too hard and there is not much of value "out there". But regardless, it's a futile excuse, markets and politics should be part of any serious prognostication.

Re: OpenAI Progress

#300

Post by weird-eye-issue »

vineyardmike wrote: post #259
muzani wrote: post #198

Earlier quoted context omitted.

ChatGPT was a proper product, but as an engine, GPT-3 (davinci-001) has been my favorite all the way until 4.1 or so. It's absolutely raw and they didn't even guardrail it. 3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions. Both of these had an IQ of around 70 or so, so the customer service trainin…

> 3.5 was like Jenny from customer service. davinci-001 was like Jenny the dreamer trying to make ends meet by scriptwriting, who was constantly flagged for racist opinions. How do you use the product to get this experience? All my questions warrant answers with no personality.

https://platform.openai.com/docs/models/davinci-002
Post reply on HN