It doesn’t seem like AIs are accelerating at all if you ask me. We seem to be plateauing. Smaller and open weight models are catching up to the closed weight frontier models on benchmarks. If AI labs were able to use the smartest models to accelerate their development, the likes of OpenAI and Anthropic would be accelerating away from their competitors. But no such thing is happening. The focus has shifted from intelligence to cost and speed.
The breakthroughs in mathematics are impressive. But it doesn’t feel that different from what machine learning has done with Chess, Go and protein folding. They’re finding patterns in our systems and in nature. That’s what they’ve always been good at.
Mostly I'm seeing breakthroughs in the ways we use LLMs. Agentic harnesses, MCPs, etc - its the wild west still but we've come a long way from a basic chatbot. Gains are now coming from tools that make better use of the LLMs existing capabilities, and put guardrails on their worst tendencies.
I personally feel like we've plateaued in raw model intelligence (even regressed, I find sonnet 4.6 to perform better than Opus 5) but we've gifted them new skills that allow them to run for longer and explore the search space more thoroughly, making them more effective at the same level of "intelligence".
Take two smart people and a problem to solve. Give one person the tools, the other person nothing. The one with the best tools wins. At some point its more about abilities than raw intelligence. Watching a "frontier" model fumble with basic syntax is still common, but not if you give it treesitter.
The math and cybersecurity improvements this year don’t look like plateauing to me. They’re clearly improving?
But I think they’re becoming more specialized. Luna is working fine for me for ordinary web development, but I was impressed by Sol tracking down an OS-level bug that was causing my Playwright tests to be flaky. Previous models couldn’t figure it out.
A lot of the improvements might be on questions that ordinary users aren’t normally asking.
Agree. We’re going to hear a lot of buzz about recursive self-improvement in the near future, which I’d cynically say is meant to address the naked emperor you just called out.
Can labs get a boost augmenting more of their processes with automation? Sure. Will it result in a self-sustaining takeoff to infinity? … No. It will saturate too. Just my opinion.
Nothing wrong with a conservative take, especially given all the hype, but I often wonder if its hard to see the magnitude of the change at "day-to-day" speed.
And the emotions one gets from reading the "takes" on news sites, blogs, etc leaves a bad taste in ones mouth and in hoping there is no change, we make it even harder to see.
A few years ago, when this all started, we had major breakthroughs fairly often that would throw out all the existing tooling. Now, most of the breakthroughs have to do with improvements to the tools instead. All the gains I see personally have to do with agent upgrades, system prompt tuning, skill tuning, context optimization, etc. Didn’t the last Anthropic release come with some system prompt updates too? Plateauing of the LLMs themselves has long since been the case but that means they’re stable enough to build on which isn’t bad.
We are also using these things to like 1% of their ability. The amount of manual process in the world that could be documented once and then automated is staggering. Decoding and transitioning such process is time consuming, and change is slow until fast.
> So let’s go with the 5%. Or, in plain English: we pretty much don’t have a clue about how the world really works.
World is not same as the universe. We may not know about 95% of the stuff that makes up universe, but that doesn't mean we don't know much about our world. World is what matters to us as earthlings. We knew a lot about what's available to us - resources, information etc.
>> In other words, he and most of his contemporaries were convinced that science was pretty much done. This was of course before Einstein’s theories of relativity, quantum mechanics and today’s fruitless search for dark matter.
>> Science wasn’t done then and it’s not done today.
Ah, the good ol' "someone nearly 150 years ago made an incorrect assertion, which means that anybody that ever makes a similar assertion ever again is automatically wrong" argument.
- please rename your substack from signal and chaos to signal and noise but unfortunately noise is all i see
- i went through your list of posts and there is nothing but one opinion after the other with no hard facts to back any of the claims
- i see what you guys are doing there and even here on HN "riding the AI wave"
- honestly i would appreciate if someone could give far more balanced takes with actual facts to back them up
- your point "ai is accelerating....." explain to me how the same architecture transformers is accelerating? infact after the leading frontier labs pull all their gimmicks in this year and maybe the next, they are going to run out of gas unless they invent a new architecture
I kind of agree with vivzkestrel here, it would be nice to have more falsifiability in the post. I stopped reading at "AIs are outperforming humans in coding", please author put this sentence more towards the beginning of the post.
At least this does not seem AI-generated, if it was, kudos you have prompted for perfect human vague.
No one is calling out the bit where the guy goes full Mulder:
> A few days ago the US Department of War released the fifth tranche of records related to objects that appeared to have a faster mode of transportation. Anyone taking UFOs seriously used to be ridiculed. Not long ago the disclosures started. Will the implementation of “alien technology” be next?
It doesn’t seem like AIs are accelerating at all if you ask me. We seem to be plateauing. Smaller and open weight models are catching up to the closed weight frontier models on benchmarks. If AI labs were able to use the smartest models to accelerate their development, the likes of OpenAI and Anthropic would be accelerating away from their competitors. But no such thing is happening. The focus has shifted from intelligence to cost and speed.
The breakthroughs in mathematics are impressive. But it doesn’t feel that different from what machine learning has done with Chess, Go and protein folding. They’re finding patterns in our systems and in nature. That’s what they’ve always been good at.
Mostly I'm seeing breakthroughs in the ways we use LLMs. Agentic harnesses, MCPs, etc - its the wild west still but we've come a long way from a basic chatbot. Gains are now coming from tools that make better use of the LLMs existing capabilities, and put guardrails on their worst tendencies.
I personally feel like we've plateaued in raw model intelligence (even regressed, I find sonnet 4.6 to perform better than Opus 5) but we've gifted them new skills that allow them to run for longer and explore the search space more thoroughly, making them more effective at the same level of "intelligence".
Take two smart people and a problem to solve. Give one person the tools, the other person nothing. The one with the best tools wins. At some point its more about abilities than raw intelligence. Watching a "frontier" model fumble with basic syntax is still common, but not if you give it treesitter.
Real breakthrough happens with custom trained models
Having your grandpa create a todo app over the weekend is great, but churning out code faster does not really make human science progress.
Custom models trained on niches could perform better scientifically, for example solving protein folding problems.
The math and cybersecurity improvements this year don’t look like plateauing to me. They’re clearly improving?
But I think they’re becoming more specialized. Luna is working fine for me for ordinary web development, but I was impressed by Sol tracking down an OS-level bug that was causing my Playwright tests to be flaky. Previous models couldn’t figure it out.
A lot of the improvements might be on questions that ordinary users aren’t normally asking.
Agree. We’re going to hear a lot of buzz about recursive self-improvement in the near future, which I’d cynically say is meant to address the naked emperor you just called out.
Can labs get a boost augmenting more of their processes with automation? Sure. Will it result in a self-sustaining takeoff to infinity? … No. It will saturate too. Just my opinion.
Nothing wrong with a conservative take, especially given all the hype, but I often wonder if its hard to see the magnitude of the change at "day-to-day" speed.
And the emotions one gets from reading the "takes" on news sites, blogs, etc leaves a bad taste in ones mouth and in hoping there is no change, we make it even harder to see.
A few years ago, when this all started, we had major breakthroughs fairly often that would throw out all the existing tooling. Now, most of the breakthroughs have to do with improvements to the tools instead. All the gains I see personally have to do with agent upgrades, system prompt tuning, skill tuning, context optimization, etc. Didn’t the last Anthropic release come with some system prompt updates too? Plateauing of the LLMs themselves has long since been the case but that means they’re stable enough to build on which isn’t bad.
We are also using these things to like 1% of their ability. The amount of manual process in the world that could be documented once and then automated is staggering. Decoding and transitioning such process is time consuming, and change is slow until fast.
The frontier labs seem to be sandbagging the full capability of the models available in the name of "alignment", which is unfortunate.
> They’re finding patterns in our systems and in nature.
Finding patterns in data is a pretty succinct description of what intelligence does, no?
Calling an "LLM" "AI" does a disservice to AI before it even exists.
> So let’s go with the 5%. Or, in plain English: we pretty much don’t have a clue about how the world really works.
World is not same as the universe. We may not know about 95% of the stuff that makes up universe, but that doesn't mean we don't know much about our world. World is what matters to us as earthlings. We knew a lot about what's available to us - resources, information etc.
>> In other words, he and most of his contemporaries were convinced that science was pretty much done. This was of course before Einstein’s theories of relativity, quantum mechanics and today’s fruitless search for dark matter.
>> Science wasn’t done then and it’s not done today.
Ah, the good ol' "someone nearly 150 years ago made an incorrect assertion, which means that anybody that ever makes a similar assertion ever again is automatically wrong" argument.
The reality shift: why is there no food
Where did the water go
- please rename your substack from signal and chaos to signal and noise but unfortunately noise is all i see
- i went through your list of posts and there is nothing but one opinion after the other with no hard facts to back any of the claims
- i see what you guys are doing there and even here on HN "riding the AI wave"
- honestly i would appreciate if someone could give far more balanced takes with actual facts to back them up
- your point "ai is accelerating....." explain to me how the same architecture transformers is accelerating? infact after the leading frontier labs pull all their gimmicks in this year and maybe the next, they are going to run out of gas unless they invent a new architecture
I kind of agree with vivzkestrel here, it would be nice to have more falsifiability in the post. I stopped reading at "AIs are outperforming humans in coding", please author put this sentence more towards the beginning of the post. At least this does not seem AI-generated, if it was, kudos you have prompted for perfect human vague.
Once all reality is copmpressed i wonder what the end game is.
No one is calling out the bit where the guy goes full Mulder:
> A few days ago the US Department of War released the fifth tranche of records related to objects that appeared to have a faster mode of transportation. Anyone taking UFOs seriously used to be ridiculed. Not long ago the disclosures started. Will the implementation of “alien technology” be next?
Yep this is where I stopped reading.
[flagged]