The pace of advancement is still going strong, especially in math and “reasoning”. For example the frontier math benchmark is showing big strides for the frontier models. A year ago they were only getting <10% of questions, now sol from openai is scoring 89% and they’re reporting the unreleased astra is at 98% .
They even had to create a new frontier math benchmark called erdos that all previous models scored 0% on and astra got 3% , which may not look like much but is a huge relative jump.
I’m aware of what the pace is. You might want to check up on the current controversy surrounding OpenAI’s math “achievements”.
By my reckoning, the difference between Mythos and Opus is smaller than the difference between Opus and Sonnet. Same with the difference between GPT 5.5 to 5.6 is smaller than the difference between GPT 4 to GPT 5.
The size of improvements over time is diminishing. We’re not in “big bang” territory anymore and we’re about two years into the “incremental refinement” period. We’re about to enter the next AI Winter unless somebody comes up with a new architectural component as revolutionary as transformers have been for ML models.
The core problem is that LLMs do not create. Full stop. All creativity is borne by the human inputs. Until that changes - until the model gains the capability to truly create new information, we’ve hit the limits in raw capability.
The size of improvements over time is diminishing. We’re not in “big bang” territory anymore and we’re about two years into the “incremental refinement” period. We’re about to enter the next AI Winter unless somebody comes up with a new architectural component as revolutionary as transformers have been for ML models.
The models are also getting extremely big as well, since the big improvement currently seems to largely be stuffing the model with more parameters, and making that work.
I can only imagine that the training cost has also been skyrocketing.
The core problem is that LLMs do not create. Full stop. All creativity is borne by the human inputs. Until that changes - until the model gains the capability to truly create new information, we’ve hit the limits in raw capability.
The models suppress outliers by design. Statistically speaking, the most novel thing is a garbled mess of random words, but random noise is useless, so it ends up being suppressed. You can see by fiddling the samplers, or increasing the temperature.
The Library of Babel is the most creative thing in the world, containing every possible combination of English words and letters. You can basically act like an LLM by trying to find a new coherent sentence in it, but also one that hasn’t been said before. It’s basically impossible.
But that is what an improvement is supposed to be. Compare that to finding a sentence that has been said, or something close to it.
You might want to check up on the current controversy surrounding OpenAI’s math “achievements”.
I’m aware of the controversies, this benchmark isn’t about making new proofs on previously unsolved problems, it’s whether it can answer complex math problems, which it’s getting better at.
By my reckoning, the difference between Mythos and Opus is smaller than the difference between Opus and Sonnet. Same with the difference between GPT 5.5 to 5.6 is smaller than the difference between GPT 4 to GPT 5.
Do you have any benchmarks or data to back this “reckoning”
Do you have any benchmarks or data to back this “reckoning”
I work with LLMs daily. I read papers as they hit arxiv. Also daily. You clearly don’t.
I’m not interested in convincing anyone, which is why I’m speaking non-technically.
The benchmarks being cited aren’t as interesting as you appear to believe they are. You’ve not fully grasped the fact that solving pre-made problems where the solutions are known or knowable isn’t anywhere close to the same thing as asking truly novel research questions independent of a human prompt. For OpenAI to also be embroiled in allegations of plagiarism only serves to underscore the gap between the two concepts.
fact that solving pre-made problems where the solutions are known or knowable isn’t anywhere close to the same thing as asking truly novel research questions independent of a human prompt.
I understand that, but the original statement was about the models stalling out in progress.
Would you say a student stalled out in progress if they could barely do 2 + 2 a couple years ago and is now able to consistently solve some of the most complex math problems known just because that student isn’t creating novel research?
We use AI as an umbrella term for so many things. Literally the type of machine learning you’re talking about has been in use in science and mathematics for quite some time and unrelated to all this buzz. It is decidedly not what OpenAI, Anthropic and Meta’s product are. Those are LLMs friend. The promised harbinger of AGI via agenetic AI. So far, all it’s been is an embarrassingly expensive and destructive waste of resources.
I would say that AI is becoming much more effective for the tokens produced. I have been running Qwen 3.8 Flash-Next on my new PC. This AI model weighs in at 120b parameters, with another 50b added via an Engram table that costs about an extra 20% in RAM for its parameter count. This is a major jump in world knowledge, and the model in question has 262k context length.
While experimental and with rough edges, this model has been able to use Lingua Gacha to undertake a game translation project - understanding and converting a Chinese game into English.
Agentic AI has become practical on local hardware over this last year, and there looks to be much more refinement to come.
00000000
What we are seeing with Anthropic and the other American companies, is a scramble to protect themselves from the consequences of badly planned data centers and poor practices in AI development. Their Chinese competition has produced much more efficient models, while delivering almost the same level of quality as American offerings. This means that local AI is becoming “good enough”, which makes data centres and AI subscriptions unnecessary for many businesses if they distrust the cloud.
The pace of advancement is still going strong, especially in math and “reasoning”. For example the frontier math benchmark is showing big strides for the frontier models. A year ago they were only getting <10% of questions, now sol from openai is scoring 89% and they’re reporting the unreleased astra is at 98% .
They even had to create a new frontier math benchmark called erdos that all previous models scored 0% on and astra got 3% , which may not look like much but is a huge relative jump.
I’m aware of what the pace is. You might want to check up on the current controversy surrounding OpenAI’s math “achievements”.
By my reckoning, the difference between Mythos and Opus is smaller than the difference between Opus and Sonnet. Same with the difference between GPT 5.5 to 5.6 is smaller than the difference between GPT 4 to GPT 5.
The size of improvements over time is diminishing. We’re not in “big bang” territory anymore and we’re about two years into the “incremental refinement” period. We’re about to enter the next AI Winter unless somebody comes up with a new architectural component as revolutionary as transformers have been for ML models.
The core problem is that LLMs do not create. Full stop. All creativity is borne by the human inputs. Until that changes - until the model gains the capability to truly create new information, we’ve hit the limits in raw capability.
The models are also getting extremely big as well, since the big improvement currently seems to largely be stuffing the model with more parameters, and making that work.
I can only imagine that the training cost has also been skyrocketing.
The models suppress outliers by design. Statistically speaking, the most novel thing is a garbled mess of random words, but random noise is useless, so it ends up being suppressed. You can see by fiddling the samplers, or increasing the temperature.
The Library of Babel is the most creative thing in the world, containing every possible combination of English words and letters. You can basically act like an LLM by trying to find a new coherent sentence in it, but also one that hasn’t been said before. It’s basically impossible.
But that is what an improvement is supposed to be. Compare that to finding a sentence that has been said, or something close to it.
I’m aware of the controversies, this benchmark isn’t about making new proofs on previously unsolved problems, it’s whether it can answer complex math problems, which it’s getting better at.
Do you have any benchmarks or data to back this “reckoning”
I work with LLMs daily. I read papers as they hit arxiv. Also daily. You clearly don’t.
I’m not interested in convincing anyone, which is why I’m speaking non-technically.
The benchmarks being cited aren’t as interesting as you appear to believe they are. You’ve not fully grasped the fact that solving pre-made problems where the solutions are known or knowable isn’t anywhere close to the same thing as asking truly novel research questions independent of a human prompt. For OpenAI to also be embroiled in allegations of plagiarism only serves to underscore the gap between the two concepts.
I understand that, but the original statement was about the models stalling out in progress.
Would you say a student stalled out in progress if they could barely do 2 + 2 a couple years ago and is now able to consistently solve some of the most complex math problems known just because that student isn’t creating novel research?
Isn’t it having the ability to create that would make it extremely potentially dangerous?
We use AI as an umbrella term for so many things. Literally the type of machine learning you’re talking about has been in use in science and mathematics for quite some time and unrelated to all this buzz. It is decidedly not what OpenAI, Anthropic and Meta’s product are. Those are LLMs friend. The promised harbinger of AGI via agenetic AI. So far, all it’s been is an embarrassingly expensive and destructive waste of resources.
I would say that AI is becoming much more effective for the tokens produced. I have been running Qwen 3.8 Flash-Next on my new PC. This AI model weighs in at 120b parameters, with another 50b added via an Engram table that costs about an extra 20% in RAM for its parameter count. This is a major jump in world knowledge, and the model in question has 262k context length.
While experimental and with rough edges, this model has been able to use Lingua Gacha to undertake a game translation project - understanding and converting a Chinese game into English.
Agentic AI has become practical on local hardware over this last year, and there looks to be much more refinement to come.
00000000
What we are seeing with Anthropic and the other American companies, is a scramble to protect themselves from the consequences of badly planned data centers and poor practices in AI development. Their Chinese competition has produced much more efficient models, while delivering almost the same level of quality as American offerings. This means that local AI is becoming “good enough”, which makes data centres and AI subscriptions unnecessary for many businesses if they distrust the cloud.