michfrapp

Le constat est identique partout : la majorité des tokens IA produites le sont désormais en dehors des IA frontières américaines. Chine et on-premise ont pris une part conséquente en quelques mois. La conséquence est que Anthropic accélère soudainement son calendrier d'IPO et si ce graphique est exact, on peut comprendre pourquoi même certains pourraient observer que c'est peu éthique.

Current state of the stock market

🇺🇸 Anthropic is teaching Claude to fix its memories while it's idle. It's testing a process called "dreaming," where Claude goes back through old conversations, checks what it remembers and corrects mistakes it made earlier. And no, it's not dreaming like a human. It's basically cleaning up its memory so it can do a better job next time. In one test, Claude got stuck playing Pokémon Red because it had remembered a location incorrectly. After reviewing its old notes against the original gameplay, it spotted the mistake, fixed it and finally reached the next city. Anthropic also tested the idea on 350,000 company documents. Claude's performance on a set of questions improved from 13% to 48% after building a connected knowledge base. The idea is pretty interesting. Instead of just remembering more, AI could get smarter by learning which memories it got wrong. Source: Interesting Engineering / Writer: Daniyal

Starship flight maneuvers

Altman, Hassabis, and Amodei are the ones doing massive corporate lobbying at the moment. They are the ones who are attempting to perform a regulatory capture of the AI industry. You, Geoff, and Yoshua are giving ammunition to those who are lobbying for a ban on open AI R&D. If your fear-mongering campaigns succeed, they will *inevitably* result in what you and I would identify as a catastrophe: a small number of companies will control AI. The vast majority of our academic colleagues are massively in favor of open AI R&D. Very few believe in the doomsday scenarios you have promoted. You, Yoshua, Geoff, and Stuart are the singular-but-vocal exceptions. like many, I very much support open AI platforms because I believe in a combination of forces: people's creativity, democracy, market forces, and product regulations. I also know that producing AI systems that are safe and under our control is possible. I've made concrete proposals to that effect. This will all drive people to do the Right Thing. You write as if AI is just happening, as if it were some natural phenomenon beyond our control. But it's not. It's making progress because of individual people that you and I know. We, and they, have agency in building the Right Things. Asking for regulation of R&D (as opposed to product deployment) implicitly assumes that these people and the organization they work for are incompetent, reckless, self-destructive, or evil. They are not. I have made lots of arguments that the doomsday scenarios you are so afraid of are preposterous. I'm not going to repeat them here. But the main point is that if powerful AI systems are driven by objectives (which include guardrails) they will be safe and controllable because *e* set those guardrails and objectives. (Current Auto-Regressive LLMs are not driven by objectives, so let's not extrapolate from their current weaknesses). Now about open source: your campaign is going to have the exact opposite effect of what you seek. In a future where AI systems are poised to constitute the repository of all human knowledge and culture, we *need* the platforms to be open source and freely available so that everyone can contribute to them. Openness is the only way to make AI platforms reflect the entirety of human knowledge and culture. This requires that contributions to those platforms be crowd-sourced, a bit like Wikipedia. That won't work unless the platforms are open. The alternative, which will *inevitably* happen if open source AI is regulated out of existence, is that a small number of companies from the West Coast of the US and China will control AI platform and hence control people's entire digital diet. What does that mean for democracy? What does that mean for cultural diversity? *THIS* is what keeps me up at night.

Rewards for Justice is offering a reward of up to $20 million and possible relocation for information about Iran’s nuclear program, including its facilities, equipment, and associated personnel, such as those in the scientific community, or the movement of any of the above. Submit a tip: rewardsforjustice.net/reward…

🇵🇾 Millions of Paraguayans are struggling to access their savings after a massive bank collapse. Meanwhile, the president's daughter is in the United States posting gym selfies wearing a shirt that says “TAX THE POOR.” Constanza Peña, who studies at Columbia University, is facing a wave of backlash over the photo. Her father, Santiago Peña, has just resigned amid the Ueno Bank scandal, which has affected roughly 3 million customers. The shirt may have been intended as a joke. But try explaining that one to people standing outside a bank, wondering whether they'll get their money back. Writer: Oliver

États-Unis: sept morts et plus de 5.000 cas de rougeole confirmés, pire épidémie depuis 35 ans

Connaissez-vous les codes de vos ados ? Cette expérience glaçante fait découvrir à des parents la violence que peuvent cacher des émojis en apparence innocents. Une campagne d'Orange qui rappelle l’importance d’écouter, même lorsqu’on ne comprend pas encore. 👏

Red Bull heeft alsnog besloten tot een (complete) motorwissel voor Max Verstappen. Dit omdat er vrijdag wat problemen aan het licht kwamen en het risico anders te groot is. Verstappen start zondag dus achteraan in São Paulo, in een naar verwachting natte Grand Prix. #F1

Paraguay's President Santiago Peña impeached and arrested bbc.in/4839WBm

What are your thoughts 🤯

🇨🇳🇹🇼 China has quietly deployed more than 100 warships and coast guard vessels across the Western Pacific, including two of its three aircraft carriers, in a massive military buildup around Taiwan. Taiwanese officials reported on November 3 that the deployments were already in place, with Chinese forces operating across multiple regions and an aircraft carrier group moving through the Philippine Sea. China's newest aircraft carrier, the Fujian, was also spotted leaving its base on Hainan Island, while Taiwan tracked 49 Chinese military aircraft sorties on November 2 alone. But unlike previous major military exercises, Beijing has made no official announcement or even given the operation a name. The Institute for the Study of War warns that the unusually large deployment may be intended to demonstrate China's ability to cut Taiwan off from international air and sea access. The concern is that Beijing is rehearsing the military isolation of Taiwan, potentially laying the groundwork for a blockade without having to launch a full-scale invasion. Source: Taiwan News / Writer: Mhedi

نشرت روسيا عام 2020 مقاطع فيديو كانت سرية، لأضخم انفجار نووي شهدته الأرض، حين فجّر الاتحاد السوفيتي قنبلة "القيصر" في تجربة أجراها في أكتوبر 1961. بلغت قوتها التفجيرية 50 مليون طن، على ارتفاع نحو أربعة آلاف متر فوق أرخبيل "نوفايا زيمليا" قبل 66 عام ، فكيف اليوم !!!

Big Tech is making a trillion-dollar mistake. Alphabet alone has disclosed over $1 trillion in commitments. AI use will explode, and much of the work will move onto devices we already own. The companies betting their futures on keeping AI in giant data centers are toast. 1/5

Dario Amodei: A lot of AI generated videos and images are fake, addictive, and just slop. Anthropic isn't interested in competing in that market. “Google and OpenAI are fighting it out in consumer. It's existential to both of them.” “It's existential to OpenAI because that's their whole business. Existential to Google because they have the search business.” “For enterprise business, there's not really a demand for cats riding donkeys or whatever consumer video people want.” “If we ever need it, we can just contract a model from someone else.” “A lot of short-form video is fake, a lot of it's pretty addictive, a lot of it's slop.” “It's not a part of the market that I'm tripping over myself to get involved in.”

Yann LeCun was right & the AI industry is spending billions the other way. 109 million parameters beat 72 billion at matching a vision model. He has spent years saying scale isn't what makes a model understand the world. Meta bet on language models anyway. He left at the end of 2025 and started his own lab in Paris. Two AI models can end up with a similar picture of the world without ever seeing each other's data. Researchers named this the Platonic Representation Hypothesis in 2024, after Plato's cave. They saw vision models and language models grow more alike as they got bigger, but it stayed an observation. Randall Balestriero and Yann LeCun, from Brown, NYU, and AMI Labs, say they've proved it for a simplified world. The proof rests on one choice in training. The model learns from two separate descriptions of one scene, not from broken pieces of one sentence. Broken pieces teach a model the writer's wording, and the authors put next-word prediction, the method behind language models, in that group. Two descriptions keep a model on the scene. In one test, the scene explained 80% of what their model stored, against 61% for next-word prediction. With text from the same web collection, their method matched a vision model as well as next-word prediction did with 42 times fewer training samples, 21.6 million against 900 million. They only scored next-word prediction at 900 million, so the real gap could be smaller. Then they scaled it up. Their 109 million parameter text model never saw an image and averaged 0.184 across 3 runs on the benchmark from the original 2024 paper, where 1.0 would be a perfect match. A 4th run at a different setting scored 0.121, and all 4 averaged 0.168. Qwen2.5-72B, an open model from 2024 about 660 times larger, scored 0.154, and Llama-3-70B scored 0.151. Their model read 116 billion tokens, and those two read 15 to 18 trillion. They also built a photo captioner from parts trained separately. The only piece fitted on matching images and captions was one simple math map. The proof covers an idealized world, and this score measures agreement with a vision model, not chat or writing. An image captioner wrote two of every three training texts. In a smaller test on 2.24 million images, replacing those texts with paraphrases that used no image captioner dropped the score from 0.107 to 0.060. Both authors list AMI Labs, LeCun's $1 billion bet on world models over language models, so they have a stake here. If better training pairs can do this job for far less, what else are labs overpaying for with scale?

Bomb threats warning 'the clock is ticking' disrupt TV channels across Europe reut.rs/4Ak17iI