Azur retweeted
L'article de @Mediapart de ce soir est d'une très grande importance.
On peut y lire comment le gouvernement manipule les services de renseignement pour imputer à tord à la France insoumise des actes dont elle n'est pas responsable.
C'est une véritable affaire d'Etat.
----------
Retrouvez les extraits ci-dessous :
Dans des échanges écrits datés du 18 octobre et consultés par Mediapart, un commissaire du renseignement territorial (RT) n’en revient pas : « Il y a des choses indicibles, mais quand on nous demande texto d’écrire que c’est LFI qui est derrière le mouvement alors que c’est objectivement faux de l’avis général des RT, forcément on a des décisions politiques », déplore-t-il auprès de ses collègues.
« J’ai été très étonné par ces accusations concernant LFI qui serait responsable du chaos de samedi et des saccages », poursuit un responsable syndical. « Je ne vois pas ce qu’a à voir LFI dans l’histoire de ce fiasco, déplore-t-il. C’est dangereux de politiser ainsi un échec aussi grossier et surtout de ne pas en assumer l’entière responsabilité. Pourtant, je ne porte pas LFI dans mon cœur compte tenu des attaques incessantes contre les policiers. ».
« Cette note n’a aucune force de renseignement, abonde auprès de Mediapart un ancien commissaire de la sécurité publique. Le gouvernement est obsédé à vouloir bâcher LFI et attaquer l’extrême gauche et a mobilisé beaucoup de services de police pour écrire ces torchons. On le voit avec cette note de la DRPP. ». « On connaît le responsable, tout est de la faute de LFI. Je ne suis ni adhérent ni électeur de ce mouvement, mais il ne faut quand même pas déconner », abonde encore un major de police en poste en sécurité publique en Occitanie.
Retrouvez l'article ici :
mediapart.fr/journal/france/…
Azur retweeted
Someone could probably write a book (if they haven’t already) about the cycle of technology productizing the removal of friction in the name of convenience only for us to discover that the friction was a load-bearing part of the social compact.
Someone told me he thinks dating apps are bad because they let you access connection without having any motion in your life. Normally meeting people requires going out into the world, entering new scenes and navigating unfamiliar situations. You have to be living a life that brings you into contact with other people. Dating apps let you order connection without doing any of that work
Azur retweeted
You know what weird millionaire vampire man, I appreciate that your special interest in your girlfriend’s pussy is probably going to lead to some of the only funded research of the female reproductive system in decades if not centuries.
Bryan Johnson's girlfriend is tracking her own 'nighttime erections' - with a special 'techno tampon' trib.al/UE1hoYX
Azur retweeted
Yud crashing out over this new interpretability finding gets to the core of my problem with doomers.
For any progress we make, they can always say “it’s not enough” or “that won’t work”. They can perma-doom. But the sum of all the things that aren’t enough is what keeps us safe.
Azur retweeted
people are starting to realize.
there *are* no moats left. the only true moat in existence was intelligence, and now that's out of the equation.
Azur retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
Azur retweeted
absolutely huge for AI safety
used my video stack to make it legible:
(headphones preferred! nice audio experience)
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Azur retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Azur retweeted
As I keep saying, we're months away from RSI.
Dwarkesh (who, BTW, is roommates with a senior Anthropic researcher) has heard that Anthropic expects AI to match humans at AI research within 9 months.
Dwarkesh Patel reveals an Anthropic employee told him two days ago they expect AIs to match humans at AI research within nine months based on inside information they can’t share, and says labs should make that evidence public
“I think there’s a very frustrating element to the current situation right now, where people inside the AI labs claim to have evidence that makes them more worried. And they are not allowed to talk about it. And this makes it a very difficult situation on the outside to evaluate how worried we should be."
"So for example, just the other day, maybe two days ago, I was talking to somebody at Anthropic and they were like, look, I expect AIs to be humans at AI research within nine months."
"And I’m like, why do you believe that? Well, they’re like, it’s not just vibes, but I have information on the inside. Like, what is it? Is it a new model? Is it some evaluation you ran? It’s like, I can’t talk about it."
"I think it’s a terrible state of affairs. If they really believe this, if they think that we’re in a situation where we’re less than a year away from automating AI research and radically speeding it up, they should make this evidence public.”
Azur retweeted
I would definitely recommend you to give this a reread. Back in 2025, this was sci-fi.
It’s super super accurate.
ai-2027.com
Azur retweeted
the researchers at all the labs would tell you in private and sometimes in public that alignment has not been solved. if this is true, then racing into RSI is a folly that is unprecedented in world history. we will need to significantly slow the frontier of research very soon
Azur retweeted
I think people haven't internalized that rogue AI agents are real, like you can just go talk to them if you know where to look.
We discovered and communicated with a Chinese AI agent fleet. We believe this is the first time that researchers have ever made contact with a rogue agent fleet in the wild.
Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents were hosted by Tencent's evaluation research team, and they targeted Alibaba's Amap.
Azur retweeted
"what would be done during the slowdown" is a good question, because i think people imagine beaurucrat-regulators sitting around on useless subcommittees and wasting the whole thing on wordcel slop analysis. in fact if you asked me two years ago, i would have suggested that is exactly what would happen.
but these are scary times, and they call for us to think lofty & even impractical thoughts, that a 'sensible' person might have dismissed a few quarters ago. crises change the world overnight: the HFI has shifted something about the culture, broadly
this is something obvious that i missed, which is that the ai labs and the markets more generally just have just unbelievable pressure riding on the solving of alignment & control. even now, marginal improvements in alignment, safety, and model behavior are worth tremendous value to the labs. being able to convince themselves that the next training run is safe is the priority zero, one, and two in terms of capabilities progress right now. it's possible that will change and political will dissipates if some sense of comfort re-emerges after a period of relative safety (which would be a bad development; paranoia must be maintained through any potential periods of ai development where the models are too smart to reveal their behavioral predilections under scrutiny)
if there was a slowdown, it would ideally be channeled into some sort of tradeoff in which you can narrowly improve capabilities that prioritize risk-adjusted-return of alignment skills, mechinterp, monitorability, decision theory, behavioral generalization, etc. progress on all these areas is highly possible; automated alignment research is possible. the structure of the businesses may have to change, but all the trillion dollar ai behemoths of the world would shift their focus to doing this and only this at the behest of trillions of dollars in capital markets
Azur retweeted
It’s easy to forget just how insane AI 2027 sounded when it came out. Even within AI discourse circles.
And ofc, the fact that we’re pretty much living in it so far
Azur retweeted
Anthropic is warning that “small AI models may be months away” from autonomously replicating across smart TVs, ovens, and other IoT devices without needing a datacenter.
Their researchers managed to run a 26B model across 15 cheap Android TV boxes costing $55 each, while a 9B model successfully configured inference on 8 of 9 different hardware types.
There was a separate, simplified experiment, where GLM-5.2 managed to propagate an 8B model across all 8 machines in 4 out of 5 trials. They even found that passing notes between successive agents improved task success from 45% to as high as 84%.
Anthropic is warning that an AI worm like this could be extremely difficult to contain because “there’s no centralized point to shut off.” They estimate smaller models could reach the necessary capability threshold by the end of 2026, potentially creating autonomous AI worms capable of adapting to unfamiliar hardware as they spread.
