5 ms·

It seems to me we’re already at the top of the S curve, not at the toe of an exponential curve. At least with LLMs. Better training data will make small improve

It seems to me we’re already at the top of the S curve, not at the toe of an exponential curve. At least with LLMs. Better training data will make small improvements, better architecture will make it less compute intensive, and all these “hyper-scale” data centers will make it cheap and ubiquitous. But none of that is it getting exponentially more intelligent.

Cyclical 4mo ago
What leads you to believe that?
- kennywinker 4mo ago
  Is chatgpt 5.6 that much smarter than chatgpt 5.0?
  - gallerdude 4mo ago
    If you’ve done any software development at all, certainly.
    - nxc18 4mo ago
      Is that true or does it only feel true because they nerf the old models just before every major release?
      - lewi 4mo ago
        You can compare benches of the old models against the new models. So yeah, you can see the difference. Even then, you can just compare the progress in open models. Leaps and bounds from where they were 6 months ago.
        mzjzjzushs 4mo ago
        [dead]
      - gabriel-uribe 4mo ago
        I remember being blown away by o1-o3 family of models finally stringing together coherent agentic tool calls to write and execute scripts semi-reliably for workloads in the several minutes before they would start hallucinating/flailing. GPT 5 was a bit ahead of that, but barely Now we take for granted that the latest models can juggle between multiple browser tabs, applications, databases, simulators, docker etc to write, execute, e2e test and deploy full-stack applications over hours managing up to dozens of subagents, relatively untouched, without taking down prod even 1% of the time Not only this, but in the GPT 5.0 era, agents had 0 taste. Nothing looked good. It was the agentic version of the twitter bootstrap era, but worse somehow. Now, I would argue the average agent frontend beats the average human frontend. This isn't even getting into 3D applications in the GPT 5 era Anyway, the models now reliably execute more than a human can fit into their own context. It's magic
        mikestorrent 4mo ago
        Yes, and we haven't even really begun to nail down computer-use agents yet (can you believe they're still basically just OCR'ing screenshots?) Once we have something that experiences a desktop interface more like a human does, an entire swathe of tooling that has heretofore been nigh-impossible to automate moves into the fold, and that'll be another explosion of folks finally getting to join the agentic workflow world on their industry specific apps...
        Timwi 4mo ago
        How do you think humans experience desktop interfaces? “Basically just OCR'ing screenshots” is exactly what humans do.
        gabriel-uribe 4mo ago
        I was also under the impression modern AI agents have moved on from just OCR'ing screenshots to leveraging native vision model capabilities.
        4mo ago
      - fragmede 4mo ago
        5.6 to 5.0 is a big enough of a jump to say yes. if it was 5.4 to 5.6 it would be a bit easier to say it only feels true because of that, but 5.6 is definitely better than 5.0. I don't have anything empirical to point at though, which is your point, but August 2025 for 5.0 vs July 2026 is almost a year later, and it's not just vibes that it's better, despite not having an objective metric to point at. It would be more scientifical to have numbers and shit to point at and there are some benchmarks out there, but you have to dig into them and really understand them in order to believe in exactly what they're testing, and I'm betting you haven't.
  - HDThoreaun 4mo ago
    Yes, it very clearly is
- leoc 4mo ago
  Things like https://www.tobyord.com/writing/hourly-costs-for-ai-agents https://www.tobyord.com/writing/hourly-costs-for-ai-agents and https://www.tobyord.com/writing/mostly-inference-scaling https://www.tobyord.com/writing/mostly-inference-scaling seem in line with other accounts like https://www.youtube.com/watch?v=aR20FWCCjAs https://www.youtube.com/watch?v=aR20FWCCjAs ?
  - reasonableklout 4mo ago
    The author of the posts you linked also wrote https://www.tobyord.com/writing/inference-scaling-reshapes-ai-governance https://www.tobyord.com/writing/inference-scaling-reshapes-a... which posits: > AI labs may also be able to reap tremendous benefit from these inference-scaled models by using them as part of the training process. If so, the large scale-up of compute resources could go into post-training rather than deployment. This would have very different implications for AI governance. > ... > So iterated distillation and amplification provides a plausible pathway for scaling inference-during-training to rapidly create much more powerful AI systems. Arguably this would constitute a form of ‘recursive self-improvement’ where AI systems are applied to the task of improving their own capabilities, leading to a rapid escalation. So "inference scaling is required to scale capabilities" doesn't mean that we're reaching the top of the S-curve in intelligence. If anything, it could mean a shorter timeline and more unpredictable landscape for governance (e.g. due to securing weights no longer as effectively preventing escalation, more in the article).
    - leoc 4mo ago
      > So "inference scaling is required to scale capabilities" doesn't mean that we're reaching the top of the S-curve in intelligence. On its own it wouldn't. But that article came before the later article https://www.tobyord.com/writing/hourly-costs-for-ai-agents https://www.tobyord.com/writing/hourly-costs-for-ai-agents which adds the claim that inference (along with everything else being employed at present) is scaling poorly with increasing task lengths. Now maybe the December 2025 claim is wrong, or maybe things will change soon, but the February 2025 article surely doesn't establish either of those.
Enginerrrd 4mo ago
I completely agree, but I also think that an industry disrupting architecture tweak akin to the “Attention is all you need” is VERY possible to emerge at any moment. It feels like the cognitive gaps on current LLMs are indeed structural, but also that if we solve that structural issue with a new or extended transformer type of architecture, we’ll be looking at a whole new ballgame. I mean, basically we’re just looking at needing some type of new post training learning architecture. It’s very clear that extending context windows isn’t that. What’s needed is an honest to god, continuous learning and modification process.
Petersipoi 4mo ago
People have been saying this since GPT-1. This idea that we can only squeeze a little bit more of intelligence out of LLMs isn't a new one. And thus far, it has always been wrong.

Cyclical 4mo ago

What leads you to believe that?

kennywinker 4mo ago
Is chatgpt 5.6 that much smarter than chatgpt 5.0?
- gallerdude 4mo ago
  If you’ve done any software development at all, certainly.
  - nxc18 4mo ago
    Is that true or does it only feel true because they nerf the old models just before every major release?
    - lewi 4mo ago
      You can compare benches of the old models against the new models. So yeah, you can see the difference. Even then, you can just compare the progress in open models. Leaps and bounds from where they were 6 months ago.
      - mzjzjzushs 4mo ago
        [dead]
    - gabriel-uribe 4mo ago
      I remember being blown away by o1-o3 family of models finally stringing together coherent agentic tool calls to write and execute scripts semi-reliably for workloads in the several minutes before they would start hallucinating/flailing. GPT 5 was a bit ahead of that, but barely Now we take for granted that the latest models can juggle between multiple browser tabs, applications, databases, simulators, docker etc to write, execute, e2e test and deploy full-stack applications over hours managing up to dozens of subagents, relatively untouched, without taking down prod even 1% of the time Not only this, but in the GPT 5.0 era, agents had 0 taste. Nothing looked good. It was the agentic version of the twitter bootstrap era, but worse somehow. Now, I would argue the average agent frontend beats the average human frontend. This isn't even getting into 3D applications in the GPT 5 era Anyway, the models now reliably execute more than a human can fit into their own context. It's magic
      - mikestorrent 4mo ago
        Yes, and we haven't even really begun to nail down computer-use agents yet (can you believe they're still basically just OCR'ing screenshots?) Once we have something that experiences a desktop interface more like a human does, an entire swathe of tooling that has heretofore been nigh-impossible to automate moves into the fold, and that'll be another explosion of folks finally getting to join the agentic workflow world on their industry specific apps...
        Timwi 4mo ago
        How do you think humans experience desktop interfaces? “Basically just OCR'ing screenshots” is exactly what humans do.
        gabriel-uribe 4mo ago
        I was also under the impression modern AI agents have moved on from just OCR'ing screenshots to leveraging native vision model capabilities.
        4mo ago
    - fragmede 4mo ago
      5.6 to 5.0 is a big enough of a jump to say yes. if it was 5.4 to 5.6 it would be a bit easier to say it only feels true because of that, but 5.6 is definitely better than 5.0. I don't have anything empirical to point at though, which is your point, but August 2025 for 5.0 vs July 2026 is almost a year later, and it's not just vibes that it's better, despite not having an objective metric to point at. It would be more scientifical to have numbers and shit to point at and there are some benchmarks out there, but you have to dig into them and really understand them in order to believe in exactly what they're testing, and I'm betting you haven't.
- HDThoreaun 4mo ago
  Yes, it very clearly is
leoc 4mo ago
Things like https://www.tobyord.com/writing/hourly-costs-for-ai-agents https://www.tobyord.com/writing/hourly-costs-for-ai-agents and https://www.tobyord.com/writing/mostly-inference-scaling https://www.tobyord.com/writing/mostly-inference-scaling seem in line with other accounts like https://www.youtube.com/watch?v=aR20FWCCjAs https://www.youtube.com/watch?v=aR20FWCCjAs ?
- reasonableklout 4mo ago
  The author of the posts you linked also wrote https://www.tobyord.com/writing/inference-scaling-reshapes-ai-governance https://www.tobyord.com/writing/inference-scaling-reshapes-a... which posits: > AI labs may also be able to reap tremendous benefit from these inference-scaled models by using them as part of the training process. If so, the large scale-up of compute resources could go into post-training rather than deployment. This would have very different implications for AI governance. > ... > So iterated distillation and amplification provides a plausible pathway for scaling inference-during-training to rapidly create much more powerful AI systems. Arguably this would constitute a form of ‘recursive self-improvement’ where AI systems are applied to the task of improving their own capabilities, leading to a rapid escalation. So "inference scaling is required to scale capabilities" doesn't mean that we're reaching the top of the S-curve in intelligence. If anything, it could mean a shorter timeline and more unpredictable landscape for governance (e.g. due to securing weights no longer as effectively preventing escalation, more in the article).
  - leoc 4mo ago
    > So "inference scaling is required to scale capabilities" doesn't mean that we're reaching the top of the S-curve in intelligence. On its own it wouldn't. But that article came before the later article https://www.tobyord.com/writing/hourly-costs-for-ai-agents https://www.tobyord.com/writing/hourly-costs-for-ai-agents which adds the claim that inference (along with everything else being employed at present) is scaling poorly with increasing task lengths. Now maybe the December 2025 claim is wrong, or maybe things will change soon, but the February 2025 article surely doesn't establish either of those.

Enginerrrd 4mo ago

I completely agree, but I also think that an industry disrupting architecture tweak akin to the “Attention is all you need” is VERY possible to emerge at any moment. It feels like the cognitive gaps on current LLMs are indeed structural, but also that if we solve that structural issue with a new or extended transformer type of architecture, we’ll be looking at a whole new ballgame. I mean, basically we’re just looking at needing some type of new post training learning architecture. It’s very clear that extending context windows isn’t that. What’s needed is an honest to god, continuous learning and modification process.

Petersipoi 4mo ago

People have been saying this since GPT-1. This idea that we can only squeeze a little bit more of intelligence out of LLMs isn't a new one. And thus far, it has always been wrong.