Sometimes macroeconometrics by day, always maths by night (or viceversa) and algorithmics somewhere in between. Art lover. Tweets are my personal opinions.
Otro grande que se nos va. Si habre estudiado con sus libros, esp. "Parallel and Distributed Computation" con John Tsitsiklis y "Nonlinear Programming". @wsosaescudero
I've just learned that Prof. Dimitri Bertsekas @DBertsekas has passed away.
His contributions to optimization, dynamic programming, control, and reinforcement learning shaped generations of researchers and engineers. I am grateful for the knowledge and insight I gained from his books and work over the years.
My sincere condolences to his family, friends, colleagues, and students. His impact on our field will be felt for generations.
Testing, yet again, frontier models to execute meaningful analysis of rather complex codebases. Something that was bugging me. Thought "maybe I'm missing something here" and had a go. But, but...
It took at stance. I said "can't be, it does not make sense." Doubled down: "you are wrong, I am right." Put pen to paper: "this can't be, but let's try, maybe I'm missing something." Tested empirically: the proposed solution was wrong, as I predicted. The answer I got was "you are wrong, if it worked before it was by accident, and my fix stands" type of response.
Then I blew the gasket, on purpose. The answer coming back is telling.
But why? Frontier AI tend to get lost (fail to see the wider coding context) beyond, say, 200-300 lines of code at either side of whatever you're looking at. But also, and more importantly, you need swearing at them with full strength (see screen cap) to avoid post-rationalization and bullshitting. That's even more dangerous.
If you know what you're doing and why it gives you a foot to stand, but if you do not (as most vibe-coders, definitionally) you swallow whatever bullshit is thrown at you.
This is why vibe-coding and "the death of software engineers" is just gaslighting. That's why vibe coding stuff will invariably break things. The more vibe-coded, the more complex it will be to backtrack. That's also why it'll likely become a compounding problem.
Granted, frontier AI can help you with certain things, but you'll end into a deep and dark rabbit hole if you let it loose with even mildly complex stuff.
Receipts provided. Things to note:
- I did bully Claude to shake it out of a post-rationalization loop. That's always nearly necessary, but it does not guarantee you'll land in a better place.
- Bullying has side-effects: Frontier AI models tend to capitulate easily to avoid alienating the user (you) even when the user is wrong (without knowing).
- Consequently, if you really don't know what you're doing, you end in between a rock and a hard place. You need to swear at them to break their COT out of self-confirmatory loops, but if you don't know your stuff shouting will lead to appeasing (the seemingly default response when swearing at them; this is not uniformly the case though) which may end up potentially confirming your own biases.
- That's why, once again, vibe-coding stuff is potentially lethal in production environments.
@GaryMarcus@rohanpaul_ai@Scobleizer@rauchg@wsosaescudero
There is a deep why: LLMs and all traditional DL models cannot deal natively with information from unstructured environments. They need to form first a map, a landscape, encoding it through the model weights. Lacking any radical innovation in the underlying tech stack, this is the only hope (and a blunt mechanism at it) for traditional models to become relevant, e.g. for household robotic assistants. @Scobleizer@GaryMarcus
@Mente_Sabiaa No conoce nadie? Tom Ford o Brunello se compran en cualquier sitio. Ah, y te olvidaste de Berlutti. Kitton o Brioni? Publico muy distinto. Loro esta en toda ciudad importante. Todas marcas bien conocidas.
Por los guantes pense pilotaba un F1. Aparte de eso, esa aproximacion debajo de minimos por lo que sugiere el video (aunque a veces los videos y la posicion de la camara engañan) es temeraria, pero obvio hay que estar ahi sentado (tampoco se aprecia en que contexto se dio). Otro si digo: lo que se ve tenia en frente tampoco era para divertirse.