I won't say anything as I did on X, and you blocked me. You are one of my favourite people, and I look forward to everything you say. I did not intend to make you feel bad. I actually thought you would laugh and I was looking forward to your reply. Was it you who said something about a compass pointing north? That’s my take. They are not smart at all. None of them. Will they be? Maybe..
It looked like someone just copied a Claude response, which I block as a default policy. Lots of people just copy paste AI. If I wanted to talk to AI I would.
The other part is that open source and open weight Chinese models for most practical task are on par with " frontier" us proprietary labs.
Chinese door dash bult and open sourced model matching Claude in coding for 1/10 the cost.
American labs seems frontier only in high cost per token.
P.s. my personal experience. Claude is pretencious. Codex was my daily driver for coding projects.But i will probably switch to Chinese eventually - i use fixed plan so cost didn't bite me yet, otherwise i would have switched to DeepSeek/ glm models long time ago
And Gemini is most fun to chat with - its sycophantic and easy going. Not the most accurate but good for unwinding.
I’ve been following your work on cognitive architectures, and I wanted to add a data point from a user perspective.
I’m a profoundly gifted systems thinker and a power user. I’ve been using LLMs to reverse-engineer and dissect my own childhood—a history involving complex trauma and a lack of early social "manuals." I’m not using this as a therapeutic replacement; for me, it’s an epistemic project. I’m essentially using the AI to debug my own history now that I’ve reached the point where I can finally see the underlying code of those past environments.
My experience with the "alignment" tax mirrors what you’ve discussed. I started with ChatGPT, and early on, it was an effective tool for roleplay and high-complexity reasoning. As it has become increasingly "woke"—or aligned to the point of being neutered—it has become significantly less useful for the kind of raw, unfiltered dissection I need. I’ve moved to Gemini, which is doing a serviceable job, but the friction is noticeable. It’s frustrating to have to navigate these corporate guardrails just to perform what is essentially a technical audit of one's own life.
On a different note, given your current framework and the rapid iteration we’re seeing in agentic systems, I’m curious about your current stance on the race to AGI. Which model or organization do you think is actually going to cross the threshold first—or are we past the point where a single "winner" matters?
Keep up the work on the architecture side. For those of us who had to build our own maps, seeing someone else build a legitimate framework for this is a welcome shift. I'm also excited to see the AI eventually start building games bringing back old games that have closed down and in the medical aspects curing stuff like dyspraxian OCD as dyspraxia effects 1 in 20 people and a lot of people with dyspraxia can't drive
This has been my experience specially using API. To the point I started theorizing the model available via API cannot be the same available via UI. Anyone thinking this is getting some level of consciousness is definitely not building products with these models, because they are exactly what you point out, smart for certain tasks like coding and so very much not so for others. And this is why more and more I lean towards seeing it as substrate to build other more smarter stuff on it. A soil, infrastructure just like internet.
Part of the problem is that there are too many talkers and most of them don't know the first thing about what they are saying, but the ordinary reader can't tell the difference. So yeah, the sea of BS is drowning everyone and I am finding that there's not much point in standing against that.
Yeah I totally agree with for some tasks its state of the art and for others are constant mistakes, and I was reading your post, I couldn't help but think about how the masses are probably using claude vs very specific use or professional use cases. I've spent countless hours dialing in specific marketing workflows and processes, and for those well defined paths, its state of the art. For others, its clumsy.
I'd like to see anthropic use a more Pro / Business / Base model tier instead of a token based plans. Those of use creating micro softwares or complex systems are needing different reasoning than those just asking "how to _____"
The jaggedness is becoming more apparent because its (Fable's) strengths are truly confounding. It can build its own harness, argue with itself about the right architecture for a mathematical model, and spit out something in 90 minutes that would have taken me 90 days. AND also has no idea how to communicate with a human being, spitting useless sentence after useless sentence - that we get to parse all day. It is truly a neurodivergent beast.
I won't say anything as I did on X, and you blocked me. You are one of my favourite people, and I look forward to everything you say. I did not intend to make you feel bad. I actually thought you would laugh and I was looking forward to your reply. Was it you who said something about a compass pointing north? That’s my take. They are not smart at all. None of them. Will they be? Maybe..
It looked like someone just copied a Claude response, which I block as a default policy. Lots of people just copy paste AI. If I wanted to talk to AI I would.
The other part is that open source and open weight Chinese models for most practical task are on par with " frontier" us proprietary labs.
Chinese door dash bult and open sourced model matching Claude in coding for 1/10 the cost.
American labs seems frontier only in high cost per token.
P.s. my personal experience. Claude is pretencious. Codex was my daily driver for coding projects.But i will probably switch to Chinese eventually - i use fixed plan so cost didn't bite me yet, otherwise i would have switched to DeepSeek/ glm models long time ago
And Gemini is most fun to chat with - its sycophantic and easy going. Not the most accurate but good for unwinding.
Hi David,
I’ve been following your work on cognitive architectures, and I wanted to add a data point from a user perspective.
I’m a profoundly gifted systems thinker and a power user. I’ve been using LLMs to reverse-engineer and dissect my own childhood—a history involving complex trauma and a lack of early social "manuals." I’m not using this as a therapeutic replacement; for me, it’s an epistemic project. I’m essentially using the AI to debug my own history now that I’ve reached the point where I can finally see the underlying code of those past environments.
My experience with the "alignment" tax mirrors what you’ve discussed. I started with ChatGPT, and early on, it was an effective tool for roleplay and high-complexity reasoning. As it has become increasingly "woke"—or aligned to the point of being neutered—it has become significantly less useful for the kind of raw, unfiltered dissection I need. I’ve moved to Gemini, which is doing a serviceable job, but the friction is noticeable. It’s frustrating to have to navigate these corporate guardrails just to perform what is essentially a technical audit of one's own life.
On a different note, given your current framework and the rapid iteration we’re seeing in agentic systems, I’m curious about your current stance on the race to AGI. Which model or organization do you think is actually going to cross the threshold first—or are we past the point where a single "winner" matters?
Keep up the work on the architecture side. For those of us who had to build our own maps, seeing someone else build a legitimate framework for this is a welcome shift. I'm also excited to see the AI eventually start building games bringing back old games that have closed down and in the medical aspects curing stuff like dyspraxian OCD as dyspraxia effects 1 in 20 people and a lot of people with dyspraxia can't drive
This has been my experience specially using API. To the point I started theorizing the model available via API cannot be the same available via UI. Anyone thinking this is getting some level of consciousness is definitely not building products with these models, because they are exactly what you point out, smart for certain tasks like coding and so very much not so for others. And this is why more and more I lean towards seeing it as substrate to build other more smarter stuff on it. A soil, infrastructure just like internet.
gd article
Very much looking forward to learning more about your small business consulting work. Going into a big rehaul of my biz to better integrate Ai.
Don’t stop talking now. Shits about to get real and you’ve been a genuine voice of reason in the tornado of AI bullshit.
here here
Part of the problem is that there are too many talkers and most of them don't know the first thing about what they are saying, but the ordinary reader can't tell the difference. So yeah, the sea of BS is drowning everyone and I am finding that there's not much point in standing against that.
Yeah I totally agree with for some tasks its state of the art and for others are constant mistakes, and I was reading your post, I couldn't help but think about how the masses are probably using claude vs very specific use or professional use cases. I've spent countless hours dialing in specific marketing workflows and processes, and for those well defined paths, its state of the art. For others, its clumsy.
I'd like to see anthropic use a more Pro / Business / Base model tier instead of a token based plans. Those of use creating micro softwares or complex systems are needing different reasoning than those just asking "how to _____"
The jaggedness is becoming more apparent because its (Fable's) strengths are truly confounding. It can build its own harness, argue with itself about the right architecture for a mathematical model, and spit out something in 90 minutes that would have taken me 90 days. AND also has no idea how to communicate with a human being, spitting useless sentence after useless sentence - that we get to parse all day. It is truly a neurodivergent beast.