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

281–290 of 372 posts

Re: OpenAI Progress

#281

Post by SirHumphrey »

jkubicek wrote: post #22
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 could essentially replace it with Google for basic to slightly complex fact checking. I know you probably meant "augment fact checking" here, but using LLMs for answering factual questions is the single worst use-case for LLMs.

Most of the value I got from google was just becoming aware that something exists. LLM-s do far better in this regard. Once I know something exists it's usually easy enough to use traditional search to find official documentation or a more reputable source.

Re: OpenAI Progress

#282

Post by frankohn »

willguest wrote: post #156

My go-to for any big release is to have a discussion about self-awareness and dive in to constuctivist notions of agency and self-knowing from a perspective of intelligence that is not limited to human cognitive capacity. I start with a simple question "who are you?". The model then invariably compares itself to humans, saying how it is not like us. I then make the point that, since it is not like us, how can it clai…

I found the questioning of love very interesting. I myself thought about whether the LLM can have emotions. Based on the book I am reading, Behave: The Biology of Humans at Our Best and Worst by Robert Sapolsky, I think the LLM, as they are now with the architecture they have, cannot have emotions. They just verbalize things like they sort-of-have emotions but these are just verbal patterns or responses they learned.

I have come to think they cannot have emotions because emotions are generated in parts of our brain that are not logical/rational. They emerge based on environmental solicitations, mediated by hormones and other complex neuro-physical systems, not from a reasoning or verbalization. So they don't come up from the logical or reasoning capabilities. However, these emotions are raised and are integrated by the rest of our brain, including the logical/rational one like the dlPFC (dorsolateral prefrontal cortex, the real center of our rationality). Once the emotions are raised, they are therefore integrated in our inner reasoning and they affect our behavior.

What I have come to understand is that love is one of such emotions that is generated by our nature to push us to take care of some people close to us like our children or our partners, our parents, etc. More specifically, it seems that love is mediated a lot by hormones like oxytocin and vasopressin, so it has a biochemical basis. The LLM cannot have love because it doesn't have the "hardware" to generate these emotions and integrate them in its verbal inner reasoning. It was just trained by human reinforcement learning to behave well. That works up to some extent, but in reality, from its learning corpora it also learned to behave badly and on occasions can express these behaviors, but still it has no emotions.

Re: OpenAI Progress

#283

Post by svantana »

furyofantares wrote: post #97
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 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 guesstimated that human-level AI would arrive in the year 2000. The only constant is that the further into the future you go, the harder it is to predict.

Re: OpenAI Progress

#284

Post by isaacremuant »

jedberg wrote: post #146
isaacremuant wrote: post #102

Earlier quoted context omitted.

Chatgpt. I test these prompts with chatgpt and they work. I've also used claude 4 opus and also worked. It's just weird how it gets repeated ad nauseaum here but I can't reproduce it with a "grab latest model of famous provider".

I just asked chatgpt "How many b's are in blueberry?". It instantly said "going to the deep thinking model" and then hung.

When I do it takes around 3 seconds where it will go "thinking longer for a better answer and then goes 2".

Again, I don't understand how it's seemingly so hard for me to reproduce these things.

I understand the tokenisation constraints, but feel it's overblown.

Re: OpenAI Progress

#285

Post by isaacremuant »

alexjplant wrote: post #139
isaacremuant wrote: post #103

Earlier quoted context omitted.

I can't reproduce it. Or similar ones. Why do yout think that is?

"Mississippi" passed but "Perrier" failed for me: > There are 2 letter "r" characters in "Perrier".

Thanks! I finally was able to reproduce one of these.

Ok. Then I was wrong. I'll update my edit accordingly.

Re: OpenAI Progress

#286

Post by isaacremuant »

isaacremuant wrote: post #55
raincole wrote: post #26

I thought the response to "what would you say if you could talk to a future AI" would be "how many r in strawberry".

Can we stop with that outdated meme? What model can't answer that effectively?

Update: after trying A LOT of examples, I did manage to reproduce one with the latest chatgpt.

Re: OpenAI Progress

#287

Post by willguest »

frankohn wrote: post #282
willguest wrote: post #156

My go-to for any big release is to have a discussion about self-awareness and dive in to constuctivist notions of agency and self-knowing from a perspective of intelligence that is not limited to human cognitive capacity. I start with a simple question "who are you?". The model then invariably compares itself to humans, saying how it is not like us. I then make the point that, since it is not like us, how can it clai…

I found the questioning of love very interesting. I myself thought about whether the LLM can have emotions. Based on the book I am reading, Behave: The Biology of Humans at Our Best and Worst by Robert Sapolsky, I think the LLM, as they are now with the architecture they have, cannot have emotions. They just verbalize things like they sort-of-have emotions but these are just verbal patterns or responses they learned.…

I was also intrigued by the machine's reference to it, especially because it posed the question with full recognition of its machine-ness.

Your comment about the generation of emotions does strike me a quite mechanistic and brain-centric. My understanding, and lived experience, has led me to an appreciation that emotion is a kind of psycho-somatic intelligence that steers both our body and cognition according to a broad set of circumstances. This is rooted in a pluralistic conception of self that is aligned with the idea of embodied cognition. Work by Michael Levin, an experimental biologist, indicates we are made of "agential material" - at all scales, from the cell to the person, we are capable of goal-oriented cognition (used in a very broad sense).

As for whether machines can feel, I don't really know. They seem to represent an expression of our cognitivist norm in the way they are made and, given the human tendency to anthropormorphise communicative systems, we easily project our own feelings onto it. My gut feeling is that, once we can give the models an embodied sense of the world, including the ability to physically explore and make spatially-motivated decisions, we might get closer to understanding this. However, once this happens, I suspect that our conceptions of embodied cognition will be challenged by the behaviour of the non-human intellect.

As Levin says, we are notoriously bad at recognising other forms of intelligence, despite the fact that global ecology abounds with examples. Fungal networks are a good example.

Re: OpenAI Progress

#288

Post by diggan »

fariszr wrote: post #157

The jump from gpt-1 to gpt-2 is massive, and it's only a one year difference! Then comes Davinci which is just insane, it's still good in these examples! GPT-4 yaps way too much though, I don't remember it being like that. It's interesting that they skipped 4o, it seems openai wants to position 4o as just gpt-4+ to make gpt-5 look better, even though in reality 4o was and still is a big deal, Voice mode is unbeatable…

Missing o1 and o1 Pro Mode which were huge leaps as I remember it too. That's when I started being able to basically generate some blackbox functions where I understand the input and outputs myself, but not the internals of the functions, particularly for math-heavy stuff within gamedev. Before o1 it was kind of a hit and miss in most cases.

Re: OpenAI Progress

#289

Post by frankohn »

willguest wrote: post #287
frankohn wrote: post #282

Earlier quoted context omitted.

I found the questioning of love very interesting. I myself thought about whether the LLM can have emotions. Based on the book I am reading, Behave: The Biology of Humans at Our Best and Worst by Robert Sapolsky, I think the LLM, as they are now with the architecture they have, cannot have emotions. They just verbalize things like they sort-of-have emotions but these are just verbal patterns or responses they learned.…

I was also intrigued by the machine's reference to it, especially because it posed the question with full recognition of its machine-ness. Your comment about the generation of emotions does strike me a quite mechanistic and brain-centric. My understanding, and lived experience, has led me to an appreciation that emotion is a kind of psycho-somatic intelligence that steers both our body and cognition according to a br…

> My understanding, and lived experience, has led me to an appreciation that emotion is a kind of psycho-somatic intelligence that steers both our body and cognition according to a broad set of circumstances.

Well, from what I understood, it is true that some parts of our brain are more dedicated to processing emotions and to integrating them with the "rational" part of the brain. However, the real source of emotions is biochemical, coming from the hormones of our body in response to environmental sollicitations. The LLM doesn't have that. It cannot feel the emotions to hug someone, or to be in love, or the parental urge to protect and care for children.

Without that, the LLM can just "verbalize" about emotions, as learned in the corpora of text from the training, but there are really no emotions, just things it learned and can express in a cold, abstract way.

For example, we recognize that a human can behave and verbalize to fake some emotions without actually having them. We just know how to behave and speak to express when we feel some specific emotion, but in our mind, we know we are faking the emotion. In the case of the LLM, it is physically incapable of having them, so all it can do is verbalize about them based on what it learned.

Re: OpenAI Progress

#290

Post by strken »

furyofantares wrote: post #97
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 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…

I think that for a lot of examples, the differentiating factor is infrastructure rather than science.

The current wave of AI needed fast, efficient computing power in massive data centres powered by a large electricity grid. The textiles industry in England needed coal mining, international shipping, tree trunks from the Baltic region, cordage from Manilla, and enclosure plus the associated legal change plus a bunch of displaced and desperate peasantry. Mobile phones took portable radio transmitters, miniaturised electronics, free space on the spectrum, population density high enough to make a network of towers economically viable, the internet backbone and power grid to connect those towers to, and economies of scale provided by a global shipping industry.

Long term progress seems to often be a dance where a boom in infrastructure unlocks new scientific inquiry, then science progresses to the point where it enables new infrastructure, then the growth of that new infrastructure unlocks new science, and repeat. There's also lag time based on bringing new researchers into a field and throwing greater funding into more labs, where the infrastructure is R&D itself.

Post reply on HN