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

GPT-5

openai.com

Post reply on HN
Print view View original

701–710 of 1001 posts

Re: GPT-5

#701

Post by lisperforlife »

highfrequency wrote: post #331

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie the…

I don't think models are fundamentally getting better. What is happening is that we are increasing the training set, so when users use it, they are essentially testing on the training set and find that it fits their data and expectations really well. However, the moat is primarily the training data, and that is very hard to protect as the same data can be synthesized with these models. There is more innovation surrounding serving strategies and infrastructure than in the fundamental model architectures.

Re: GPT-5

#702

Post by wouldbecouldbe »

highfrequency wrote: post #331

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie the…

It's all based on the theory of singularity. Where the AI can start trainig & relearning itself. But it looks like that's not possible with the current techniques.

Re: GPT-5

#703

Post by throwaway_2898 »

henriquegodoy wrote: post #432

That SWE-bench chart with the mismatched bars (52.8% somehow appearing larger than 69.1%) was emblematic of the entire presentation - rushed and underwhelming. It's the kind of error that would get flagged in any internal review, yet here it is in a billion-dollar product launch. Combined with the Bernoulli effect demo confidently explaining how airplane wings work incorrectly (the equal transit time fallacy that NAS…

TBH Claude Code max pro's performance on coding has been abhorrent(bad at best). The core of the issue is that the plan produced will more often than not use humans as verifiers(correctness, optimality and quality control). This is a fundamentally bad way to build systems that need to figure out if their plan will work correctly, because an AI system needs to test many plans quickly in a principled manner(it should be optimal and cost efficient).

So you might get that initial MvP out the door quickly, but when the complexity grows even just a little bit, you will be forced to stop and look at the plan and try to get it to develop it saying things like: "use Design agent to ultrathink about the dependencies of the current code change on other APIs and use TDD agent to make sure tests are correct in accordance with the requirements I stated" and then one finds that even the all the thinking there are bugs that you will have to fix.

Source: I just tried max pro on two client python projects and it was horrible after week 2.

Re: GPT-5

#705

Post by kmacdough »

m4x wrote: post #678
highfrequency wrote: post #331

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie the…

It's quite possible that the models from different companies are clustering together now because we're at a plateau point in model development, and won't see much in terms in further advances until we make the next significant breakthrough. I don't think this has anything to do with AGI. We aren't at AGI yet. We may be close or we may be a very long way away from AGI. Either way, current models are at a plateau and a…

What does AGI mean to you, specifically?

As is, AI is quite intelligent, in that it can process large quantities of diverse unstructured information and build meaningful insights. And that intelligence applies across an incredibly broad set of problems and contexts. Enough that I have a hard time not calling it general. Sure, it has major flaws that are obvious to us and it's much worse at many things we care about. But that's doesn't make it not intelligent or general. If we want to set human intelligence as the baseline, we already have a word for that: superintelligence.

Re: GPT-5

#706

Post by samsullivan »

henriquegodoy wrote: post #432

That SWE-bench chart with the mismatched bars (52.8% somehow appearing larger than 69.1%) was emblematic of the entire presentation - rushed and underwhelming. It's the kind of error that would get flagged in any internal review, yet here it is in a billion-dollar product launch. Combined with the Bernoulli effect demo confidently explaining how airplane wings work incorrectly (the equal transit time fallacy that NAS…

Thats why they need to pay 300k for a slide designer https://openai.com/careers/creative-lead-presentation-design...

Re: GPT-5

#707

Post by onlyrealcuzzo »

beeflet wrote: post #361
highfrequency wrote: post #331

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie the…

Perhaps it is not possible to simulate higher-level intelligence using a stochastic model for predicting text. I am not an AI researcher, but I have friends who do work in the field, and they are not worried about LLM-based AGI because of the diminishing returns on results vs amount of training data required. Maybe this is the bottleneck. Human intelligence is markedly different from LLMs: it requires far fewer examp…

> Human intelligence is markedly different from LLMs: it requires far fewer examples to train on, and generalizes way better.

Aren't we the summation of intelligence from quintillions of beings over hundreds of millions of years?

Have LLMs really had more data?

Re: GPT-5

#708

Post by mmcconnell1618 »

hnlmorg wrote: post #633
highfrequency wrote: post #331

It is frequently suggested that once one of the AI companies reaches an AGI threshold, they will take off ahead of the rest. It's interesting to note that at least so far, the trend has been the opposite: as time goes on and the models get better, the performance of the different company's gets clustered closer together. Right now GPT-5, Claude Opus, Grok 4, Gemini 2.5 Pro all seem quite good across the board (ie the…

The reason AGI would create a singularity is because of its ability to self learn. Presently we are still a long way from that. In my opinion we at least are as far away from AGI as 1970s mainframes were from LLMs. I really don’t expect to see AGI in my lifetime.

Self-learning opens new training opportunities but not at the scale or speed of current training. The world only operates at 1x speed. Today's models have been trained on written and visual content created by billions of humans over thousands of years.

You can only experience the world in one place in real time. Even if you networked a bunch of "experiencers" together to gather real time data from many places at the same time, you would need a way to learn and train on that data in real time that could incorporate all the simultaneous inputs. I don't see that capability happening anytime soon.

Re: GPT-5

#709

Post by metzpapa »

zone411 wrote: post #693

GPT-5 set a new record on my Confabulations on Provided Texts benchmark: https://github.com/lechmazur/confabulations/

For how much I’ve seen it pushed that this model has lower hallucination rates, it’s quite odd that every actual test I’ve seen says the opposite.

Re: GPT-5

#710

Post by samsullivan »

punee94 wrote: post #577

I ran the below prompt to both Kimi2 and GPT5. how many rs in cranberry? -- GPT5's response: The word cranberry has two “r”s. One in cran and one in berry. Kimi2's response: There are three letter rs in the word "cranberry".

answering correctly is completely dependent on the attention blocks to somehow capture the single letter nuance given word tokenization constraints. does the attention block in kimi have a more receptive architecture to this?
Post reply on HN