NEW: From AI Skills to "Skill Programs"

Here are the esoteric concepts wallpapered over a simple idea (use code/scripts to intercept bad LLM outputs): - Failure manifolds for "wrong answers" or "failure modes" - Energy landscape / potential field for the space of possible LLM outputs - Particle traveling for an LLM generating tokens - Quantum tunneling for the LLM ignoring prompt instructions - Maxwell's demon (1867) dragged in as a 150-year-old thermodynamics analogy - Diffused potential field for "a prompt instruction" - Non-zero system temperature for LLM stochasticity - Latent momentum of the particle for reasoning trace bias - Phase space micro state for the model's current generation context - Global attractor of the system manifolds for "the correct answer" - Coordinate trajectory rewriting for editing/intercepting output - Multi-dimensional energy gradient for the optimization signal - Gedankenexperiment invocation to justify the metaphor existing at all - Force field for what is literally just an if-statement in code Stripped down, the paper's actual claim: deterministic program checks beat soft prompt instructions at catching known failure cases.

I can tell, this channel has crazy ai shit that any channel in entire youtube history and thats the reason I subscribed ❤🎉

and next step, new saas, software as agent skill?

Why not just make the tools dead simple or automatic?

"No maths! Just physics!" It's the same thiiiiiiiiiiiiiiiiiiiing :''''''''''''''''''''(

5:11 186️⃣7️⃣🫴🫳

Thanks for another great review. More addons to improve our Pinto when what we are really looking for is a better model, a Chiron Super Sport. Increasing our energy requirements won't make the Pinto into a Chiron.

Really great review. I’ve been following your channel for quite a long time, and I really love the way you explain things. One small piece of feedback: for me personally, the theoretical part at the beginning felt a bit too long. Around 12 minutes of explaining how it should work felt slightly unnecessary, especially because I think the same idea could be understood faster through illustrations or examples. I’m more of a visual learner, so I personally enjoy it more when the explanation moves earlier into practical examples and visuals. I think leaning more in that direction would make your already great reviews even stronger.

No mathematics at all! Woohoo!