2024-11-18T17:54:39.204534 | 🔗
The convergence of the biomolecular foundation models (see recent results on Chai and Boltz) reflect the convergence of foundation models onto their data (much like the GPT-4 class language models). This suggests two approach. Go back and get better data (not just more). Go forward and do post-training techniques to bias the accuracy of the foundation models up.
2024-11-18T17:47:31.810445 | 🔗
“He began with foundation stones of knowledge: time, space, motion. I do not define time, space, place, and motion, as being well known to all, he wrote in midlife—then a reclusive professor, recondite theologian and alchemist, seldom leaving his room in Trinity College, Cambridge.6 But he did mean to define these terms. He salvaged them from the haze of everyday language. He standardized them. In defining them, he married them, each to the others.He dipped his quill in an ink of oak galls and wrote a minuscule Latin script, crowding the words edge to edge: The common people conceive those quantities under no other notions but from the relation they bear to sensible objects. And thence arise certain prejudices.… ” Excerpt From Isaac Newton James Gleick https://itunes.apple.com/WebObjects/MZStore.woa/wa/viewBook?id=0 This material may be protected by copyright.
2024-11-18T17:46:23.442705 | 🔗
“At twenty-four, Newton believed he could marshal a complete science of motion, if only he could find the appropriate lexicon, if only he could set words in the correct order. Writing mathematics, he could invent his own symbols and form them into a mosaic. Writing in English, he was constrained by the language at hand.29 At times his frustration was palpable in the stream of words. Axiom 103: … as the body (a) is to the body (b) so must the power or efficacy vigor strength or virtue of the cause which begets the same quantity of velocity. …30Power efficacy vigor strength virtue—something was missing. But these were the laws of motion, in utero.” Excerpt From Isaac Newton James Gleick https://itunes.apple.com/WebObjects/MZStore.woa/wa/viewBook?id=0 This material may be protected by copyright.
2024-11-18T16:26:57.883993 | 🔗
What is the world’s tiniest bureaucracy?
2024-11-18T01:56:26.851419 | 🔗
Having a crazy person at the top will only drive the Deep State forward. I think it'll be a hint towards the ability, promise, and danger of controlling AI networks.
2024-11-16T18:48:18.868458 | 🔗
We are probably at the optimum already in terms of the mean intelligence of foundation models. Probably not possible to curate data too much more without losing generalizing power. Now the problem is post-training to sample as far along the distribution as possible.
2024-11-16T18:25:16.074574 | 🔗
I'm sitting in the Leverett House Library and thinking that the data curation here -- in this very room --- is probably better training data than any other dataset out there.
2024-11-16T15:35:36.969004 | 🔗
h/t https://x.com/chadnotchud/status/1857479296446771610?s=46. Trumpisms that are actually great: "many are saying this", "So true!", "Sad!", "Thank you X, very cool", "concepts of a plan", "big if true".
2024-11-16T15:13:33.128774 | 🔗
AI Agents screw things up by introducing duplicatable "people". Many frameworks are not able to handle "independent clones" or perfect anonymity.
2024-11-16T14:39:04.406163 | 🔗
"dolphins have these huge brains but they aren't after anything" they don't build cities, wage war, manufacture guns. maybe "intelligence" isn't really "intelligence" but rather a function of a cultural craze