Since you wrote a book on LLM architecture I am surprised that you would write "I don’t know if Claude or other current LLMs have sentience or something analogous to consciousness". I am always surprised when people who understand LLM architecture suggest that this architecture would lead of something like consciousness.
I believe that LLMs alone are a dead end when it comes to something like AGI. Google has been adding reasoning algorithms to Gemini. I think that LLMs will be a front end to more powerful AI systems, but to me this is obviously an architecture that will not lead to complex reasoning.
On writing english vs. writing code: I don't have experience with Claude, but I have a lot of experience with the latest release of Gemini (Gemini 3.5). In both case the result can be mediocre. Especially when it comes to english. LLMs cannot resist hyperbole. The generate the average of their training set and the result is mediocre english.
Making my 90-hour video course and textbook on LLMs has definitely refined my thinking about how they work and what they are (and might be in the future) capable of. I don't know if frontier LLMs are capable of something like sentience, but I wouldn't dismiss it out of hand. We (humans) don't understand consciousness nearly well enough to be able to identify it in artificial systems -- and human consciousness might be a terrible analogy for AI sentience (if it exists). What is certainly true is that LLMs do not possess a human-like consciousness, and probably never will.
As for AGI: Honestly, it's a moving target. Current LLMs probably would have been considered "AGI" 20 years ago. I agree with you that the future of highly capable AI systems will be that -- systems that work together, with an LLM either at the executive top, or at least the one interfacing with humans. That is already the case, e.g., ChatGPT cannot produce images; it calls a separate image-diffusion model to create an image, then copies it to you.
As for coding capabilities: I was very unimpressed by code from any LLM until around March of this year, when Claude suddenly become quite a good coder. I've also had shit code from Gemini, though it was the free version; perhaps a higher tier of Gemini is a good coder. LLM coding abilities also depend heavily on how they are trained and what tools they are given (e.g., whether they smoke-test their own code in Python before giving it to you).
Since you wrote a book on LLM architecture I am surprised that you would write "I don’t know if Claude or other current LLMs have sentience or something analogous to consciousness". I am always surprised when people who understand LLM architecture suggest that this architecture would lead of something like consciousness.
I believe that LLMs alone are a dead end when it comes to something like AGI. Google has been adding reasoning algorithms to Gemini. I think that LLMs will be a front end to more powerful AI systems, but to me this is obviously an architecture that will not lead to complex reasoning.
On writing english vs. writing code: I don't have experience with Claude, but I have a lot of experience with the latest release of Gemini (Gemini 3.5). In both case the result can be mediocre. Especially when it comes to english. LLMs cannot resist hyperbole. The generate the average of their training set and the result is mediocre english.
Thanks for the comment, Ian.
Making my 90-hour video course and textbook on LLMs has definitely refined my thinking about how they work and what they are (and might be in the future) capable of. I don't know if frontier LLMs are capable of something like sentience, but I wouldn't dismiss it out of hand. We (humans) don't understand consciousness nearly well enough to be able to identify it in artificial systems -- and human consciousness might be a terrible analogy for AI sentience (if it exists). What is certainly true is that LLMs do not possess a human-like consciousness, and probably never will.
As for AGI: Honestly, it's a moving target. Current LLMs probably would have been considered "AGI" 20 years ago. I agree with you that the future of highly capable AI systems will be that -- systems that work together, with an LLM either at the executive top, or at least the one interfacing with humans. That is already the case, e.g., ChatGPT cannot produce images; it calls a separate image-diffusion model to create an image, then copies it to you.
As for coding capabilities: I was very unimpressed by code from any LLM until around March of this year, when Claude suddenly become quite a good coder. I've also had shit code from Gemini, though it was the free version; perhaps a higher tier of Gemini is a good coder. LLM coding abilities also depend heavily on how they are trained and what tools they are given (e.g., whether they smoke-test their own code in Python before giving it to you).