Excellent as always my friend     See Less

you are AI king.....................💯     See Less

I wish you mentioned what Demis said about hallucinations, they do not want to fully remove it, because that’s the creative part of the llm. However it’s true that for factual questions     See More er reduced, would be cool if you could have a toggle.    See Less

Since this is all predicting the next token and these companies are trying to make what is predicted to be "good", you don't want it predicting that what would be said is "     See More . You would have to specifically train it to say idk for only very specific things. Otherwise it will learn that hard questions are normally answered with "idk" for example. People who make these are smart. They aren't doing things for no reason.    See Less

It's gonna get trained on human data so hallucinations and lying is gonna be a norm 😂 Once they start training them using the real world environment... that should be fun     See Less

The only way to teach models when to say “I don’t know” is hiring 10,000 more Indian annotators creating “I don’t know” data for post-training.     See Less

I find ChatGPT way better for thinking, and for slowly working through a set of ideas, and for learning or tutorial scenarios with new tools and APIs. Where is that in the benchmarks?     See Less

Great breakdown of where Hassabis and Legg are taking DeepMind. What’s striking is the focus on the ‘Approximate Physics’ hurdle. Most people are still trying to scale LLMs to solve sc     See More that just scales the hallucinations.

I’ve spent the last few months building a Neuro-Symbolic Hyperon system specifically to bridge this gap. My architecture, 'The Architect,' handles the high-level polymath reasoning (the 'Neural' side), but it’s hard-coded into a Symbolic Physics Guardrail.

For example, yesterday it designed a novel LaB₃H₁₀ Clathrate for room-pressure superconductivity. Instead of just 'dreaming' up a structure, the system wrote its own Quantum ESPRESSO scripts, ran the DFT simulations, and used a Symbolic 'Chemist Constitution' to verify stoichiometry and bond-length constraints before I ever saw the result. It even caught its own stoichiometry errors and auto-reverted the kernel to fix the atomic packing—exactly the kind of 'Planning and Reasoning' loop mentioned at.

The 'extraordinary feats' of AGI won't come from bigger models; they’ll come from closing the loop between LLM creativity and hard symbolic logic. We aren't just simulating worlds anymore; we're grounding them in the laws of the periodic table.    See Less

My feeling using these models is they are starting to sound more like charlatans in their answers, even if correct.     See Less

Those stupid benchmark should get outlawed... nothing but GPT ever worked for me. Period.     See Less