Ingénieur de formation. J'enseigne depuis plus de 30 ans. J'utilise le machine learning depuis 2015 dans mes projets artistiques personnels et professionnels
France
Joined July 2023
- Tweets1.7K
- Following488
- Followers156
- Likes907
Olivier Heckendorn retweeted
Plus j'avance sur mon projet d'IA juridique, plus je vois le problème.
Beaucoup de juges ont un raisonnement totalement défaillant.
Ils ignorent les JP majeures, interprètent n'importe comment les faits, font des liens de causalité erronés, se satisfont de conditions non remplies pour rentrer au chausse-pied dans les cases. Un mathématicien qui regarde ces démonstrations devient fou.
Les tribunaux de 1ère instance et d'appel ont facilement 30 ou 50% de décisions qui ne tiendraient pas en cassation, sans mêle parler des erreurs d'interprétation des faits. Et la cour de cassation invente du droit comme ça l'arrange.
L'IA a du mal à suivre, car elle essaie de prévoir avec rigueur un comportement qui est finalement aléatoire. Donnez le même dossier à 3 juges, il vous sortira trois jugements complètement différents.
Olivier Heckendorn retweeted
C’est surtout très stupide de jouer à ça. Un IA n’est pas un humain et ce genre de mise en scène va causer plus de soucis qu’elle ne va résoudre de problèmes.
Upping my p(doom) again just because one more time humans do thing without any thinking of the consequences.
Olivier Heckendorn retweeted
It still makes really stupid writing in any expert subject. LLM are still really bad writer and incredible engineers 🤪
Olivier Heckendorn retweeted
Je pense vraiment qu’à se concentrer sur la perte de contrôle ( ou les agents vont se heurter rapidement au mur de la réalité des ressources) on oublie que le vrai danger est la démocratisation des moyens d’attaque cyber ou biologique. Les humains restent à ce jours bien plus mal intentionnés que les IA « autonomes «
Olivier Heckendorn retweeted
If you think of LLMs as computer programs, you have the wrong mental model. They are nondeterministic, their behavior isn't even programmed.
Think of them as computer people
Olivier Heckendorn retweeted
Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead.
You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence. You’ve also claimed the lead is widening because of recursive self-improvement.
I don’t see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible.
But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier.
Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want.
Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sanders’ “shut it all down.” China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well.
So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it.
If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know this was just another bid for regulatory capture — or an election-season psyop.
Olivier Heckendorn retweeted
It is now getting very hard to track where your work and conversations are in Claude and ChatGPT: Local computer? Cloud? Which computer? Phone? The computer connected via Dispatch or Remote? The one you are talking to in Tag? Is that in a project? The current UX is little help.
Olivier Heckendorn retweeted
It’s revealing to see so many medical imaging experts being dismissive towards Midjourney (although of course they all preface their dismissive comments by saying that they aren’t doing that… but no one is fooled), calling them naive, and giving all the reasons why it won’t work and won’t change anything.
It sort of helps to explain why imaging is so insanely overpriced and rare in elective, prophylactic, discretionary contexts.
Some of the arguments are so absurd. “Oh, water was used in the 50s, we walked away from that. That’s going backwards.”
Really? They had sub-millimeter resolution back then, too? Oh. So maybe that’s totally irrelevant.
It can’t see through bones? So take images from all around the bone. Sure, you won’t be able to see inside the bones, but you don’t need to for it to still be very useful.
The criticisms all seem to make a common analytical mistake that I used to see all the time in the hedge fund business, which is ignoring key aspects of a disruptive change when applying an argument.
If the cost of the scan is 99.9% lower, and it takes a minute instead of an hour, and there is no risk from radiation, then you can realistically do LOTS of scans much more frequently.
That means you have a whole new modality to work with: time. You can look at changes over time. Duh. That’s a really big difference!
This one aspect of the innovation dramatically changes the calculus, and suddenly most of the “fatal flaws” identified by the experts melt away into obvious irrelevance.
Also, some of the flaws they identify have obvious next steps that could address them, like adding additional modalities (like light) to help differentiate tissue types.
Anyway, it’s all a bit maddening to see this tripe, and also seeing all the people gleefully reposting and echoing it as if it’s some kind of “gotcha.”
In 5 years, it will all seem as laughable as the video of that European space executive dismissing SpaceX and reusable rockets as a pipe dream 8 years ago.
One of the great things about the free market system is that smart people like @DavidSHolz can just do this stuff and the old guard can’t stop them.
Under a Soviet central-planning system, David would be dismissed as an ignorant dilettante with dangerous, naive, wrong ideas.
He would be laughed out of the room, kind of like how the Russians laughed at Musk when he tried to buy some of their rockets before deciding to build his own.
Olivier Heckendorn retweeted
Matthieu, 22 ans, a réalisé la place Stanislas de Nancy en Lego. Si son modèle atteint 10.000 votes, la marque en fera l'une de ses références comme jouet officiel. Un petit vote peut faire la différence : urls.fr/QE507T
Merci pour lui et pour Nancy ! 🫶 ❤️🤍
Olivier Heckendorn retweeted
Here’s my definition: To understand a phenomenon is to have a world model some part of which has an isomorphism to the key parts of the phenomenon. And for that isomorphism to have been there even if those parts of the phenomenon had been different.
Olivier Heckendorn retweeted
🔴 À l'unanimité, les sénateurs viennent de voter une loi qui interdit de fait tout entrainement de modèle d'IA en France.
Voire même toute commercialisation d'une IA.
À l'unanimité.
Olivier Heckendorn retweeted
We're doing a bunch of redesigns on our site and we're putting them up on alpha.midjourney.com for you to try! The first new design is a click-and-scroll style creator, second is a rework of all our dropdown settings panels into a universal sidebar. Tell us whatcha think <3
