@EricTopol A chicken-and-egg problem! The low risk groups - RG1 and RG2 (i.e., general population) show uninteresting Hazard ratios (< 1), so it is low-value, IMO, as a prognostic tool! More confirmatory (just a nice-to-have) than profoundly predictive.
@karpathy 0. Ideal = world models + LLMs.
1. We will buy reliability, not tokens.
2. Verifiers will become more valuable than generators.
3. The ultimate AI capability will be "forgetting."
4. The frontier model is no longer the product.
5. Distillation will swallow black-box monopolies.
4 biggest influencers of this upcoming direction - Alphabet (Google), Nvidia, Amazon, and Microsoft. As an investor, you should consider accummulating these in your portfolio.
The way we build, buy, and measure AI is still broken.
In the next 12 to 24 months, raw "intelligence" benchmarks will cease to matter. The future belongs to feedback-looped, cognitive control systems.
Here are the 5 core shifts that will rewrite the AI playbook:
๐งต1/n
5. Distillation will swallow black-box monopolies.
Expensive closed models will act as teachers. Frontier agents will map complex workflows into smaller, open-weight models to compress & mimic trajectories. Moat shifts from static weights to continuous post-training & telemetry.
@QBacquele@ScienceMagazine Brilliant; finally a great use of a UMAP! I'm sure you've also looked at frequency domain techniques such as wavelets for classification.
The world has made tremendous progress.
I believe we will overcome the major crises of our time and continue advancing humanity.
Be an agent for good.
via @toddrjones