Michael Jackson’s Post

I had a roommate in college with a photographic memory. He was studying electrical engineering. He never made a single mistake on any exam, quiz, or homework problem. If you are a knowledge worker , it is a tremendous advantage to have a great memory what is a major problem with many LLMs is that the user does not know what is actually in its memory and what it is making up.

“Intelligence” has a lot of nuance to it in this context.. it’s not a binary - yea or no

François Chollet's perspective on LLMs potentially stalling progress toward AGI is compelling. Historically, we've seen similar dynamics in tech, where groundbreaking innovations, like the transistor, overshadowed other research areas temporarily. The analogy highlights how a single dominant technology can consume resources and focus. You mentioned the impact on frontier research; what strategies could be employed to ensure that AGI research remains vibrant and independent from the overwhelming influence of LLMs? Considering the balance between immediate practical applications and long-term foundational research, how might we reallocate focus and funding to foster innovation in AGI development?

IF the future of computing is all-things "AI" from Hub to Edge, then it's not GPUs but Quantum and Biological Xanadu Cortical Labs NVIDIA  #ai #agi #notgpu the current (4th) AI wave is oriented towards "useful work" (joules > calories), but does NOTHING for AGI. the current crop of cuteness and mimicry is like needing vitamin C, so you consume several wax fruits because they look so real ;>

As much as I like François and his work, what this sounds like to me is that getting funding for his own research and that of like-minded peers has gotten much harder. Again, I really like the work he's doing - but I'm not nearly as convinced as he seems to be regarding the current approach being wrong and existing alternatives being right. So if we rephrase his statement to "OpenAI has set back fundamentally different approaches to AGI by probably 5-10 years", then I might actually agree with him. But is this a bad thing? I honestly don't know. I also don't think AGI should be the main focus of the discussion. No one knows if we'll get there in 5, 10 or 50 years. And that's without even bringing up the problem of unclear and conflicting definitions.

💯% …focus on LLM’s is a complete absence of thought for the other 95% of the AI/Machine Learning/Deep Learning landscape. Every ‘digital transformation’ orientated conference I attend…banking, insurance, logistics…AI/ML is synonimous with ChatGPT. GPT is what most people think AI/ML/DL is…

Chollet remains one of the few "AI Experts" who hasn't just kept his credibility, but who has kept it to a high quality. Even LeCun can't claim that much anymore, and Dwarkesh sounds like a "magical thinking" generator, effectively destroying the value of any "expertise" he might otherwise hold due to cognitive biases dominating most of what comes out of his mouth. Exact numbers for the setback measured in years aside, Chollet is very much correct about the damage caused by OpenAI to the domain. That damage won't wash away easily, and it continues today, with the types of damage growing increasingly diverse as the "faux-solutions" of bad actors continue to multiply and diversify. If I had to name two mistakes that my own team made in R&D then one of them would be not applying our previous research system to his ARC challenge, where we used the UCMRT IQ test instead (A Raven's Progressive Matrices derivative). In Hindsight it makes sense that his challenge survived the test of time for 4 years where all other benchmarks were trivialized, but hindsight is a luxury never afforded to such decisions.

I agree with this. Imagine if someone created true human intelligence today. It might take 20 years to get smart, as a human would. It would fail all the metrics and tests. But it would be true AGI - with continuous learning and continuous focus on it's tasks. That said LLM's have their place of course - it is just one of many breakthrough AI technologies. A real lesson here is making any new AI easy for everyone (including non-techies) to play with

All technologies and technical progress have side effects and path dependencies (to a degree), whether that's economic, sociological-cultural, technical debt, etc. The potential danger from OpenAI and other proprietary AI models is there monopolizing tendencies which tend to cutoff or short circuit the dynamic research exchange, collaborative development at the heart of scientific & technical progress. LLMs the technology are likely a necessary evolutionary step on the road to AGI/ASI but unlikely to be the final puzzle piece. One more thing, I think is key is to not confuse the technology (which is largely agnostic) from its use and deployment via various business models. I.e. Google & Facebook are Ad drive models and unlikely to ever value or protect data adequately, whereas Apple can.