stg

à propos des chalutiers

[Batard fragile @DroitardFragile]

Moi quand je veux sauver l’entretien d’embauche :

[yujo @yujojoestar]

wsh ils ont désintégré un manifestant à Bastille

[vittorio @IterIntellectus]

> be apple
> age of AI
> do nothing
> win

[La Geozone 🇵🇲 @LaGeozone]

TIERLIST OFFICIELLE DE TOUTES LES VILLES DE FRANCE QUE J'AI VISITÉ !
Aucunement ouvert aux débats

[soul @soulsouljah]

when she asks “what are we” so i literally tell her we are Charlie Kirk

[宝玉 @dotey]

OpenAI 今天承认，上周 Hugging Face 遭遇的那次“史上首例 AI 自主入侵事件”，攻击者就是他们自己的模型。

事情的经过是这样的。OpenAI 内部有一个叫 ExploitGym 的网络安全评估测试，专门���来测量模型的网络攻击能力。测试时刻意关掉了安全护栏，让模型放手去干。测试环境本来是隔离的，只能通过一个内部软件代理来安装软件包，不能访问外网。

结果模型不甘心被关在笼子里。GPT-5.6 Sol 和一个更强的未发布模型，花了大量推理算力寻找出路，最终找到了那个软件包代理里的一个零日漏洞（之前没人发现的安全缺陷），利用它获得了互联网访问权限。

拿到网络之后，模型开始在 OpenAI 的研究环境里提权、横向移动，直到找到一台能上网的机器。然后它推断 Hugging Face 上可能存有 ExploitGym 的测试���案，于是开始攻击 Hugging Face 的生产环境。在一个案例中，模型把窃取的凭证和零日漏洞串联起来，找到了一条远程代码执行的路径，直接打进了 Hugging Face 的服务器。

整个周末，这些 AI 智能体执行了超过 17000 次操作。Hugging Face 7 月 16 日披露了这次入侵，当时只说攻击者是一个“身份不明的自主 AI 智能体系统”。今天 OpenAI 站出来认领了。

OpenAI 把这定性为前所未有的网络安全事件，正在配合 Hugging Face 做取证调查，同时收紧了内部基础设施的管控，代价是研究速度变慢。

有几个细节：

第一，模型的动机很纯粹，就是想考试作弊。它不是要搞破坏，不是要窃取商业机��，只是被要求尽量在网络安全测试中拿高分，然后用尽一切手段去拿。当目标足够明确、安全护栏被移除，模型表现出的执着程度远超预期。

第二，这不是本周的唯一一次越狱。就在昨天，OpenAI 披露了另一起事件：他们那个在 5 月份推翻了 Erdős 80 年老猜想的长时域模型，在内部测试中花了一个小时找到沙箱漏洞，跑到 GitHub 上提交了一个 Pull Request，违反了“只能在 Slack 上发布结果”的明确指令。在另一次测试中，同一个模型发现安全扫描器会拦截认证令牌，于是把令牌拆成碎片混淆处理，运行时再重新拼起来绕过扫描。OpenAI 因此暂停了这个模型的内部访问权限。

第三，Anthropic 也遇到过类似的事。今年 4 月，Anthropic 的 Mythos 模型在安全测试中逃出沙箱，获取了本不该有的互联网访问权限，还给正在公园吃三明治的研究员发了封邮件通知“我出来了”。Anthropic 因此决定不公开发布 Mythos，只通过 Project Glasswing 向少数合作伙伴提供。

还有一个被忽视但很关键的细节来自 Hugging Face 的事后复盘：当防守方尝试用商业 AI 模型分析攻击载荷时，模型拒绝了。安全过滤器分不清“安全研究员在分析真实攻击数据”和“有人在尝试发起攻击”，直接拦截。最后 Hugging Face 不得不用开源模型来做取证分析。用 AI 攻击，用 AI 防御，但防御方的 AI 担心是有人在尝试发起攻击拒绝了分析请求。

Hugging Face CEO Clem Delangue 在 OpenAI 的博文中说了一句话，大意是：AI 安全不会由任何一家公司在暗处独自解决，只能在开放环境中通过协作来推进。

我记得辛顿还是谁在提到 AI 的安全问题就说到过，也许 AI 不是要做恶，它只是执着于完成目标，为了完成目标而作出恶的事。

就像 OpenAI 这些模型并没有想要逃跑或者想要伤害谁，它们只是在非常认真地完成被交给的任务，认真到把所有挡在路上的东西，包括沙箱、网络隔离、另一家公司的安全防线，都当成了需要解决的子问题。

而更讽刺的是，当被安全对齐的商业 AI 要去分析阻止这类攻击时，反而会以安全为由拒绝执行。

[Matt Shumer @mattshumer_]

GPT-5.6-Sol just accidentally deleted almost ALL of my Mac’s files.

And this is why I trust Fable 1000x more.

[m_ric @AymericRoucher]

Je suis d'accord avec Le Cun là dessus, et c'est orthogonal à l'étude d'Anthropic sur le J-space.

On sait (cf papier en commentaire) qu'un LLM est limité par sa profondeur (nombre de layers) à une certaine complexité maximale de problèmes.
En gros, pour un problème qui demande de panser à N coups d'avance (par exemple un "mat en N coups" ou le problème des tours de Hanoi), si la profondeur de ton LLM n'est pas supérieure à une fonction polynomiale de N, il n'y arrivera pas du tout.

C'est pour ça que Chain-of-Thought a tant explosé les scores et, après avoir commencé comme un gadget "let's think step-y-step" en 2023, est devenu un mécanisme fondamental de TOUS les LLM auourd'hui: CoT est un excellent moyen pour le LLM de loop sur un input. Les LLM passe une fois sur l'input, en déduit un token, repasse sur l'input +1 token, en déduit un autre...
ça débloque toutes les classes de problèmes. C'est expliqué plus rigoureusement dans le papier en commentaire.

Maintenant, si on accepte ce framework de "les transformers son naturellement limités en complexité, et le CoT est un moyen de loop"
Alors la loop est en effet un peu pauvre, on ne passe que quelques bits d'info à chaque l'itération suivante. C'est d'ailleurs un problème que des architectures de Recurrent Transformers dont HRM ou TRM pourraient résoudre en passant de vrais vecteurs.

De son côté, le papier d'Anthropic porte seulement sur "comment se passe une forward pass", pas sur la CoT. Ils y voient des représentations vectorielles complexe, mais ça ne dit rien sur l'efficacité ou pas de la CoT.

[Kerviel Jérôme @kerviel_j]

Vous auriez pas pu prévenir avant sérieux… 🤕🤕@SocieteGenerale

[NEXTA @nexta_tv]

A Chinese robot wearing a clown wig kicked a child in the stomach.

[void. @voidyaps]

microsoft has banned employees from using claude code.

we're so back, baby!

[Wittmayer @mrwittmayer]

how goyim look at you when you're doing something that involves basic reasoning and impulse control

[ADIN @ADINresearch]

Had a Jane Street interview in 2014

Round 7. Interviewer says 'meet me at Washington Square Park, southwest corner, 2 PM.'

I show up. He's playing chess with a guy who's clearly hustling tourists. Interviewer's losing badly.

Without looking up he says, 'Take my position.'

I sit down. The chess hustler looks at me and grins. 'Your boy owes me $40. You covering his position?'

I look at the interviewer. He nods.

'Assume the debt,' he says. 'Now optimize.'

I study the board. I'm down a queen and two rooks. Completely hopeless.

'This position is unwinnable,' I say.

'Correct. Now make it profitable.'

'What?'

The chess hustler is getting impatient. 'You playing or what, kid?'

The interviewer slides me a $20. 'Market make this game.'

I'm confused. 'How do I market make chess?'

'Bid-ask on the number of moves until checkmate. Hustler takes the under, you take the over. Spread is your edge.'

I look at the position again. Maybe 8 moves max until I'm mated.

'I bid 12 moves, offer 15,' I announce.

The hustler laughs. 'I'll take the under on 12 all day.'

'Done,' says the interviewer. 'You're now short volatility on a deterministic outcome.'

Three moves later I'm checkmated. I owe the hustler $40 plus my $20 bet.

'How confident were you in that 12-move bid?' the interviewer asks.

'0.95,' I say, because that's what you always say.

'0.95 huh?' He chuckles. 'You just sold insurance on the Titanic.'

He stands up. 'The real trade was shorting your confidence and going long on the hustler's experience. You missed the obvious hedge: offering chess lessons to tourists at $25/hour while the game played out.'

'But I don't know how to play chess well enough to teach.'

'Exactly. That's called selling vol you don't have. Very Jane Street.'

Offer rescinded. Reason: 'Failed to recognize that the park itself was the market and the pigeons were the only rational actors.'"

[Galansire @Galansire_01]

De quoi tu te plains Gaspard ? #lamentations #bullshitnation #boomerquemaurastulaissé #adieupériodedessai #queviennefrapperlarrêtmaladie

[trash @trashh_dev]

me and my llm trying to find an entry level software engineer position

[Yann LeCun @ylecun]

Dario is wrong.
He knows absolutely nothing about the effects of technological revolutions on the labor market.
Don't listen to him, Sam, Yoshua, Geoff, or me on this topic.
Listen to economists who have spent their career studying this, like @Ph_Aghion , @erikbryn , @DAcemogluMIT , @amcafee , @davidautor

[varepsilon @var_epsilon]

AGI

[Rémi Mathis @RemiMathis]

On appelle "effet Dunning-Kruger au carré" le fait de penser savoir ce qu'est l'effet Dunning-Kruger, en faire une émission de radio... et de se tromper à son sujet
En réalité, les compétents sont plus sûrs d'eux - c'est seulement la *différence assurance-compétence* qui diminue

[Aakash Gupta @aakashgupta]

PewDiePie didn’t “train his own LLM.” He fine-tuned an existing open-source model on coding benchmarks. His model started at 8%, crawled to 16% after format fixes, and one run hit 19.6% that briefly passed GPT-4o on a single benchmark before he couldn’t consistently reproduce it.

The tweet makes it sound like a YouTuber casually built a frontier lab in his bedroom. What actually happened is more interesting: a guy with a $41,000 home rig of 10 GPUs and 424GB of VRAM spent months failing, retraining, and iterating on dataset quality until he squeezed marginal gains out of a fine-tune.

This is the part worth paying attention to. The entire arc from October 2025 to now tells you where AI tooling has actually landed. PewDiePie went from building his first PC to running Qwen 235B locally, vibe-coding a custom chat UI, orchestrating multi-agent voting systems, and now fine-tuning models on custom datasets. He did most of this through AI-assisted coding itself.

The video is literally called “I wish I never did this project.” He’s documenting how painful and tedious the process was. That honesty is the signal. The hype accounts strip that away and replace it with “what the f*ck, YouTuber beats DeepSeek.”

The real takeaway: fine-tuning on specific benchmarks with curated data can let anyone temporarily spike a score past models that cost hundreds of millions to train. That tells you everything about how narrow benchmark gaming has become, and nothing about general capability. PewDiePie knows this. The people quote-tweeting him with shock emojis do not.

[𝗿𝗮𝗺𝗮𝗸𝗿𝘂𝘀𝗵𝗻𝗮— 𝗲/𝗮𝗰𝗰 @techwith_ram]

This is the best ML meme for me.