[DanMcInerney]

I'm really hoping GPT5 is a larger jump in metrics than the last several releases we've seen like Claude3.5 - Claude4 or o3-mini-high to o3-pro. Although I will preface that with the fact I've been building agents for about a year now and despite the benchmarks only showing slight improvement, I have seen that each new generation feels actively better at exactly the same tasks I gave the previous generation.

It would be interesting if there was a model that was specifically trained on task-oriented data. It's my understanding they're trained on all data available, but I wonder if it can be fine-tuned or given some kind of reinforcement learning on breaking down general tasks to specific implementations. Essentially an agent-specific model.

[codingwagie]

I'm seeing big advances that arent shown in the benchmarks, I can simply build software now that I couldnt build before. The level of complexity that I can manage and deliver is higher.

[IanCal]

A really important thing is the distinction between performance and utility.

Performance can improve linearly and utility can be massively jumpy. For some people/tasks performance can have improved but it'll have been "interesting but pointless" until it hits some threshold and then suddenly you can do things with it.

[shmoogy]

Yeah I kind of feel like I'm not moving as fast as I did, because the complexity and features grow - constant scope creep due to moving faster.

[protocolture]

I am finding that my ability to use it to code, aligns almost perfectly with increasing token memory.

[kevinqi]

yeah, the benchmarks are just a proxy. o3 was a step change where I started to really be able to build stuff I couldn't before

[alightsoul]

mind telling examples?

[motorest]

Not OP, but a couple of days ago I managed to vibecode my way through a small app that pulled data from a few services and did a few validation checks. By itself its not very impressive, but my input was literally "this is how the responses from endpoint A,B and C look like. This field included somewhere in A must be somewhere in the response from B, and the response from C must feature this and that from response A and B. If the responses include links, check that they exist". To my surprise, it generated everything in one go. No retry nor Agent mode churn needed. In the not so distant past this would require progressing through smaller steps, and I had to fill in tests to nudge Agent mode to not mess up. Not today.

[corysama]

I’m wrapping up doing literally the same thing. I did it step-by-step. But, for me there was also a process of figuring out how it should work.

[alightsoul]

what tools did you use?

[motorest]

> what tools did you use?

Nothing fancy. Visual Studio Code + Copilot, agent mode, a couple prompt files, and that's it.

[munksbeer]

Do you mind me asking which language and if you have any esoteric constraints in the apps you build? We use a java in a monorepo, and have a full custom rolled framework on top of which we build our apps. Do you find vibe coding works ok with those sort of constraints, or do you just end up with a generic app?

[iLoveOncall]

Okay but this has all to do with the tooling and nothing to do with the models.

[efunnekol]

I mostly disagree with this.

I have been using 'aider' as my go to coding tool for over a year. It basically works the same way that it always has: you specify all the context and give it a request and that goes to the model without much massaging.

I can see a massive improvement in results with each new model that arrives. I can do so much more with Gemini 2.5 or Claude 4 than I could do with earlier models and the tool has not really changed at all.

I will agree that for the casual user, the tools make a big difference. But if you took the tool of today and paired it with a model from last year, it would go in circles

[mofeien]

Can you explain why?

[iLoveOncall]

You can write projects with LLMs thanks to tools that can analyze your local project's context, which didn't exist a year ago.

You could use Cursor, Windsurf, Q CLI, Claude Code, whatever else with Claude 3 or even an older model and you'd still get usable results.

It's not the models which have enabled "vibe coding", it's the tools.

An additional proof of that is that the new models focus more and more on coding in their releases, and other fields have not benefited at all from the supposed model improvements. That wouldn't be the case if improvements were really due to the models and not the tooling.

[eru]

You need a certain quality of model to make 'vibe coding' work. For example, I think even with the best tooling in the world, you'd be hard pressed to make GPT 2 useful for vibe coding.

[iLoveOncall]

I'm not claiming otherwise. I'm just saying that people say "look what we can do with the new models" when they're completely ignoring the fact that the tooling has improved a hundred fold (or rather, there was no tooling at all and now there is).

[broast]

Chatgpt itself has gotten much better at producing and reading code since a year ago, in my experience

[pyman]

They're using a specific model for that, and since they can't access private GitHub repos like MS, they rely on code shared by devs, which keeps growing every month.

[energy123]

That would require AIME 2024 going above 100%.

There was always going to be diminishing returns in these benchmarks. It's by construction. It's mathematically impossible for that not to happen. But it doesn't mean the models are getting better at a slower pace.

Benchmark space is just a proxy for what we care about, but don't confuse it for the actual destination.

If you want, you can choose to look at a different set of benchmarks like ARC-AGI-2 or Epoch and observe greater than linear improvements, and forget that these easier benchmarks exist.

[croddin]

There is still plenty of room for growth on the ARC-AGI benchmarks. ARC-AGI 2 is still <5% for o3-pro and ARC-AGI 1 is only at 59% for o3-pro-high:

"ARC-AGI-1: * Low: 44%, $1.64/task * Medium: 57%, $3.18/task * High: 59%, $4.16/task

ARC-AGI-2: * All reasoning efforts: <5%, $4-7/task

Takeaways: * o3-pro in line with o3 performance * o3's new price sets the ARC-AGI-1 Frontier"

- https://x.com/arcprize/status/1932535378080395332

[saberience]

I’m not sure the arcagi are interesting benchmarks, for one they are image based and for two most people I show them too have issues understanding them, and in fact I had issues understanding them.

Given the models don’t even see the versions we get to see it doesn’t surprise me they have issues we these. It’s not hard to make benchmarks that are so hard that humans and Lims can’t do.

[nipah]

"most people I show them too have issues understanding them, and in fact I had issues understanding them" ??? those benchmarks are so extremely simple they have basically 100% human approval rates, unless you are saying "I could not grasp it immediately but later I was able to after understanding the point" I think you and your friends should see a neurologist. And I'm not mocking you, I mean seriously, those are tasks extremely basic for any human brain and even some other mammals to do.

[viraptor]

> so extremely simple they have basically 100% human approval rates

Are you thinking of a different set? Arc-agi-2 has average 60% success for a single person and questions require only 2 out of 9 correct answers to be accepted. https://docs.google.com/presentation/d/1hQrGh5YI6MK3PalQYSQs...

> and even some other mammals to do.

No, that's not the case.

[nipah]

No, I think I saw the graphs on someone's channel, but maybe I misinterpreted the results. But to be fair, my point never depended on 100% of the participants being right 100% of the questions, there are innumerous factors that could affect your performance on those tests, including the pressure. The AI also had access to lenient conventions, so it should be "fair" in this sense.

Either way, there's something fishy about this presentation, it says: "ARC-AGI-1 WAS EASILY BRUTE-FORCIBLE", but when o3 initially "solved" most of it the co-founder or ARC-PRIZE said: "Despite the significant cost per task, these numbers aren't just the result of applying brute force compute to the benchmark. OpenAI's new o3 model represents a significant leap forward in AI's ability to adapt to novel tasks. This is not merely incremental improvement, but a genuine breakthrough, marking a qualitative shift in AI capabilities compared to the prior limitations of LLMs. o3 is a system capable of adapting to tasks it has never encountered before, arguably approaching human-level performance in the ARC-AGI domain.", he was saying confidently that it would not be a result of brute-forcing the problems. And it was not the first time, "ARC-AGI-1 consists of 800 puzzle-like tasks, designed as grid-based visual reasoning problems. These tasks, trivial for humans but challenging for machines, typically provide only a small number of example input-output pairs (usually around three). This requires the test taker (human or AI) to deduce underlying rules through abstraction, inference, and prior knowledge rather than brute-force or extensive training."

Now they are saying ARC-AGI-2 is not bruteforcible, what is happening there? They didn't provided any reasoning for why one was bruteforcible and the other not, nor how they are so sure about that. They "recognized" that it could be brute-forced before, but in a way less expressive manner, by explicitly stating it would need "unlimited resources and time" to solve. And they are using the non-bruteforceability in this presentation as a point for it.

--- Also, I mentioned mammals because those problems are of an order that mammals and even other animals would need to solve in reality for a diversity of cases. I'm not saying that they would literally be able to take the test and solve it, nor to understand this is a test, but that they would need to solve problems of similar nature in reality. Naturally this point has it's own limits, but it's not easily discarded as you tried to do.

[saberience]

lol 100% approval rates? No they don’t.

Also mammals? What mammals could even understand we were giving it a test?

Have you seen them or shown them to average people? I’m sure the people who write them understand them but if you show these problems to average people in the street they are completely clueless.

This is a classic case of some phd ai guys making a benchmark and not really considering what average people are capable of.

Look, these insanely capable ai systems can’t do these problems but the boys in the lab can do them, what a good benchmark.

[nipah]

quoting my own previous response: > Also, I mentioned mammals because those problems are of an order that mammals and even other animals would need to solve in reality for a diversity of cases. I'm not saying that they would literally be able to take the test and solve it, nor to understand this is a test, but that they would need to solve problems of similar nature in reality. Naturally this point has it's own limits, but it's not easily discarded as you tried to do.

---

> Have you seen them or shown them to average people? I’m sure the people who write them understand them but if you show these problems to average people in the street they are completely clueless.

I can show them to people on my family, I'll do it today and come back with the answer, it's the best way of testing that out.

[clbrmbr]

You may be above average intelligence. Those challenges are like classic IQ tests and I bet have a significant distribution among humans.

[achierius]

No, they've done testing against samples from the general population.

[HDThoreaun]

arc agi is the closest any widely used benchmark is coming to an IQ test, its straight logic/reasoning. Looking at the problem set its hard for me to choose a better benchmark for "when this is better than humans we have agi"

[saberience]

There are humans who cannot do arc agi though so how does an LLM not doing it mean that LLMs don’t have general intelligence?

LLMs have obviously reached the point where they are smarter than almost every person alive, better at maths, physics, biology, English, foreign languages, etc.

But because they can’t solve this honestly weird visual/spatial reasoning test they aren’t intelligent?

That must mean most humans on this planet aren’t generally intelligent too.

[HDThoreaun]

> LLMs have obviously reached the point where they are smarter than almost every person alive, better at maths, physics, biology, English, foreign languages, etc.

I dont think memorizing stuff is the same as being smart. https://en.wikipedia.org/wiki/Chinese_room

> But because they can’t solve this honestly weird visual/spatial reasoning test they aren’t intelligent?

Yes. Being intelligent is about recognizing patterns and thats what arc agi tests. It tests ability to learn. A lot of people are not very smart.

[XCSme]

I remember the saying that from 90% to 99% is a 10x increase in accuracy, but 99% to 99.999% is a 1000x increase in accuracy.

Even though it's a large10% increase first then only a 0.999% increase.

[zmgsabst]

Sometimes it’s nice to frame it the other way, eg:

90% -> 1 error per 10

99% -> 1 error per 100

99.99% -> 1 error per 10,000

That can help to see the growth in accuracy, when the numbers start getting small (and why clocks are framed as 1 second lost per…).

[XCSme]

Still, for the human mind it doesn't make intuitive sense.

I guess it's the same problem with the mind not intuitively grasping the concept of exponential growth and how fast it grows.

[pixl97]

The lily pad example of the lake being half full on the 29th day out of 30 is also a good one.

[XCSme]

ChatGPT quick explanation:

Humans struggle with understanding exponential growth due to a cognitive bias known as *Exponential Growth Bias (EGB)*—the tendency to underestimate how quickly quantities grow over time. Studies like Wagenaar & Timmers (1979) and Stango & Zinman (2009) show that even educated individuals often misjudge scenarios involving doubling, such as compound interest or viral spread. This is because our brains are wired to think linearly, not exponentially, a mismatch rooted in evolutionary pressures where linear approximations were sufficient for survival.

Further research by Tversky & Kahneman (1974) explains that people rely on mental shortcuts (heuristics) when dealing with complex concepts. These heuristics simplify thinking but often lead to systematic errors, especially with probabilistic or nonlinear processes. As a result, exponential trends—such as pandemics, technological growth, or financial compounding—often catch people by surprise, even when the math is straightforward.

[bobbylarrybobby]

I think the proper way to compare probabilities/proportions is by odds ratios. 99:1 vs 99999:1. (So a little more than 1000x.) This also lets you talk about “doubling likelihood”, where twice as likely as 1/2=1:1 is 2:1=2/3, and twice as likely again is 4:1=4/5.

[jsjohnst]

The saying goes:

From 90% to 99% is a 10x reduction in error rate, but 99% to 99.999% is a 1000x decrease in error rates.

[AtlasBarfed]

What's the required computation power for those extra 9s? Is it linear, poly, or exponential?

Imo we got to the current state by harnessing GPUs for a 10-20x boost over CPUs. Well, and cloud parallelization, which is ?100x?

ASIC is probably another 10x.

But the training data may need to vastly expand, and that data isn't going to 10x. It's probably going to degrade.

[littlestymaar]

> I'm really hoping GPT5 is a larger jump in metrics than the last several releases we've seen like Claude3.5 - Claude4 or o3-mini-high to o3-pro.

This kind of expectations explains why there hasn't been a GPT-5 so far, and why we get a dumb numbering scheme instead for no reason.

At least Claude eventually decided not to care anymore and release Claude 4 even if the jump from 3.7 isn't particularly spectacular. We're well into the diminishing returns at this point, so it doesn't really make sense to postpone the major version bump, it's not like they're going to make a big leap again anytime soon.

[indigo945]

I have tried Claude 4.0 for agentic programming tasks, and it really outperforms Claude 3.7 by quite a bit. I don't follow the benchmarks - I find them a bit pointless - but anecdotally, Claude 4.0 can help me in a lot of situations where 3.7 would just flounder, completely misunderstand the problem and eventually waste more of my time than it saves.

Besides, I do think that Google Gemini 2.0 and its massively increased token memory was another "big leap". And that was released earlier this year, so I see no sign of development slowing down yet.

[sailingparrot]

> We're well into the diminishing returns at this point

Scaling laws, by definition have always had diminishing returns because it's a power law relationship with compute/params/data, but I am assuming you mean diminishing beyond what the scaling laws predict.

Unless you know the scale of e.g. o3-pro vs GPT-4, you can't definitively say that.

Because of that power law relationship, it requires adding a lot of compute/params/data to see a big jump, rule of thumb is you have to 10x your model size to see a jump in capabilities. I think OpenAI has stuck with the trend of using major numbers to denote when they more than 10x the training scale of the previous model.

* GPT-1 was 117M parameters.

* GPT-2 was 1.5B params (~10x).

* GPT-3 was 175B params (~100x GPT-2 and exactly 10x Turing-NLG, the biggest previous model).

After that it becomes more blurry as we switched to MoEs (and stopped publishing), scaling laws for parameters applies to a monolithic models, not really to MoEs.

But looking at compute we know GPT-3 was trained on ~10k V100, while GPT-4 was trained on a ~25k A100 cluster, I don't know about training time, but we are looking at close to 10x compute.

So to train a GPT-5-like model, we would expect ~250k A100, or ~150k B200 chips, assuming same training time. No one has a cluster of that size yet, but all the big players are currently building it.

So OpenAI might just be reserving GPT-5 name for this 10x-GPT-4 model.

[littlestymaar]

> but I am assuming you mean diminishing beyond what the scaling laws predict.

You're assuming wrong, in fact focusing on scaling law underestimate the rate of progress as there is also a steady stream algorithmic improvements.

But still, even though hardware and software progress, we are facing diminishing returns and that means that there's no reason to believe that we will see another leap as big as GPT-3.5 to GPT-4 in a single release. At least until we stumble upon radically new algorithms that reset the game.

I don't think it make any economic sense to wait until you have your “10x model” when you can release 2 or 3 incremental models in the meantime, at which point your “x10” becomes an incremental improvement in itself.

[avereveard]

There's a new set of metrics that capture advances better than MMLU or it's pro version but nothing yet as standardized and specifically very few have a hidden set of tests to keep advancements from been from directional fine tuning.

[jstummbillig]

It's hard to be 100% certain, but I am 90% certain that the benchmarks leveling off, at this point, should tell us that we are really quite dumb and simply not good very good at either using or evaluating the technology (yet?).

[motorest]

> (...) at this point, should tell us that we are really quite dumb and simply not good very good at either using or evaluating the technology (yet?).

I don't know about that. I think it's mainly because nowadays LLMs can output very inconsistent results. In some applications they can generate surprisingly good code, but during the same session they can also do missteps and shit the bed while following a prompt to small changes. For example, sometimes I still get prompt responses that outright delete critical code. I'm talking about things like asking "extract this section of your helper method into a new methid" and in response the LLM deletes the app's main function. This doesn't happen all the time, or even in the same session for the same command. How does one verify these things?

[alightsoul]

either that or the improvements aren't as large as before.