Actualités IA Quotidiennes
jeudi 21 mai 2026
Today an OpenAI model disproved a conjecture that mathematicians had been working on for decades. Not solved — disproved. That distinction matters. The model did not find the answer humans expected. It found that the question was wrong. That is a qualitatively different kind of output from AI than writing code or summarizing documents.
At the same time, Anthropic is moving into Colossus2 with NVIDIA's most powerful GPUs, while a stealth startup called Hark just raised $700 million to build what it calls a universal AI interface. The compute buildout and the funding rounds are accelerating, not slowing. The companies betting against that pace keep losing.
The pattern to watch: capability gains are outrunning the frameworks people use to think about AI. The plagiarism debate trending today assumes AI training works like human copying. The geometry result suggests it does not. When a tool surprises you — not by doing what you asked better, but by doing something you did not think to ask — your mental model of the tool needs updating.
New AI tools, features, and services launching today
AI Agents That Actually Sell
StoreClaw is a new tool for online store owners that uses AI agents to help increase sales. The idea is simple: instead of just showing products to shoppers, the agents actively guide customers through the buying process — surfacing the right products at the right moment, handling objections, and pushing toward purchase.
The pitch is that most e-commerce stores leave money on the table because the shopping experience is passive. StoreClaw tries to make it active, using AI that understands how to sell rather than just display.
It launched today on Product Hunt and is generating strong interest from store owners who have tried other AI tools and found them too generic.
💡 Pourquoi ça compte
Most AI tools for e-commerce focus on writing product descriptions or generating images. StoreClaw is going after the harder problem: conversion. If AI agents can meaningfully increase the percentage of visitors who actually buy, that is a direct revenue impact that is easy to measure. Store owners have a clear reason to try it.
Email That Reaches the Inbox
mailX is a new product from mailwarm, a Y Combinator-backed startup that has been working on email since 2020. The product is a deliverability toolkit built for both human senders and AI agents that send email automatically.
As AI tools take over more outreach and communication tasks, email deliverability is becoming a new bottleneck. AI agents can send thousands of emails, but if those emails land in spam, the whole pipeline breaks. mailX addresses that gap directly.
Mailwarm has real history in this space, which gives mailX more credibility than a fresh launch. The YC pedigree and six years of focus on email infrastructure show up in the product's depth.
💡 Pourquoi ça compte
AI agents are taking on more outreach tasks, but inbox placement has not kept up. A message that goes to spam does nothing, no matter how good the AI that wrote it. mailX is one of the first tools built specifically for the new reality where AI agents — not just humans — are the ones hitting send. For any business using AI for email outreach, this is a practical problem that needs a practical fix.
Major business and policy developments shaping the AI industry
AI Breaks Real Math
An OpenAI model has disproved a conjecture in discrete geometry that mathematicians had been unable to crack for years. This is not a benchmark result or a demo — it is a genuine contribution to pure mathematics that changes what experts thought was true.
The model did not just find a proof. It found a counterexample, showing that the conjecture was false in the first place. That kind of result requires exploring areas of the problem space that human researchers had not thought to check.
Mathematicians online are calling this the math story of the year. The broader AI community is watching closely, because this is the clearest sign yet that AI can do original intellectual work — not just assist with it.
💡 Pourquoi ça compte
AI has beaten humans at chess, Go, and protein folding. Those were domains with clear rules and known solution paths. Pure mathematics is different. Disproving a conjecture means finding something no one knew existed. If this is repeatable, it puts AI in a new category: not a tool that helps researchers, but a research partner that can make discoveries on its own.
$700M for a Secret Interface
Hark, a startup that has kept its product almost entirely out of public view, just raised $700 million in a Series A round. The company says it is building a universal AI interface — a personal AI platform that works on top of your existing apps and services rather than replacing them.
Hark says its first multimodal models will launch this summer. Beyond that, details are thin. The company has shared little about what the product actually looks like or how it differs from existing AI assistants.
A $700 million Series A for a product that is not yet public is a striking signal. Investors are betting that whoever builds the layer connecting users to all their AI tools will end up in a powerful position.
💡 Pourquoi ça compte
The race to own the AI interface layer is heating up. Microsoft has Copilot, Apple has Apple Intelligence, and now Hark is making a large bet with backing to match. A universal interface that sits across all your tools — email, calendar, files, apps — could become the most used software on your computer. Seven hundred million dollars suggests some very serious investors think Hark has a real shot at that position.
The GPU Bet Behind Claude 4
Anthropic is reportedly expanding its compute infrastructure to Colossus2 and will use NVIDIA's GB200 chips — currently among the most powerful AI hardware available — according to a post from a senior researcher close to the company's training operations.
This is a significant infrastructure move. The GB200 is designed specifically for large AI training runs, and Colossus2 is one of the largest GPU clusters in existence. Gaining access to that level of compute puts Anthropic in a stronger position to train bigger, more capable models.
The post generated immediate discussion in the AI community today about what the company might be training next and what the jump in hardware capability signals about the scale of Anthropic's upcoming model work.
💡 Pourquoi ça compte
Compute is still one of the clearest signals of what an AI lab is planning. When a company moves to bigger, more expensive hardware, it is usually because it has models it wants to train that its current setup cannot handle. Moving to GB200s on Colossus2 suggests the next generation of Claude could be significantly more powerful than what is available today.
Notable GitHub projects and open-source releases
The Agent That Learns You
NousResearch has released Hermes Agent, an open-source AI agent designed to adapt to the person using it over time. Most AI tools treat every session as a fresh start. Hermes Agent remembers how you work, what you care about, and adjusts its behavior accordingly.
NousResearch has a strong reputation in the open-source AI world — they have previously released some of the most capable openly available language models. Hermes Agent brings that same philosophy to the agent layer: powerful tools that anyone can run, inspect, and modify.
The project is trending strongly in the developer community today, with builders excited about an agent that does not need to be re-configured every time.
💡 Pourquoi ça compte
Memory and personalization are the missing piece in most AI agents today. You set them up, they work well, and then tomorrow they have forgotten everything. Hermes Agent makes persistence a core feature rather than an afterthought. For developers building tools that people use repeatedly, this is a practical foundation worth examining.
Who Owns What AI Learned
A developer's essay arguing that AI training on copyrighted content is "unauthorised plagiarism at a bigger scale" is generating one of the most intense community discussions today. The author makes a direct case: when AI systems reproduce patterns, styles, and knowledge absorbed from copyrighted works without compensation to creators, that is not learning — it is taking.
The argument is not new, but the framing landed near the top of Hacker News and the comment thread that followed is dense with practitioners, lawyers, and creators debating where the line falls between training on data and copying it.
The legal cases around this question are still working through courts in multiple countries. But the community conversation has not slowed down while waiting for answers.
💡 Pourquoi ça compte
The courts will eventually decide whether AI training on copyrighted data is legal. But the social question of whether it is acceptable is being decided in discussions like this one — and the intensity of the debate tells you that neither side has convinced the other. For anyone building or using AI products, the outcome matters for what you build and on what terms.
⚡ En Bref
A developer community site called No Slop Grenade is gaining traction with a simple message: stop pasting raw AI output into team conversations. The argument is practical — long AI-generated walls of text slow reviews, bloat documentation, and erode the quality of communication. Edit before you send.
noslopgrenade.com →A developer published a detailed breakdown of how Google's "anti-gravity" demo from Google I/O was misleading — showing capabilities that do not yet exist in any product users can access. The post is reigniting questions about how honestly AI companies present their research at public events.
0xsid.com →Anthropic held its two-day "Code with Claude" developer event in London on May 19-20 — timed exactly against Google I/O. MIT Technology Review covered the event and described it as a picture of where software development is heading whether developers embrace it or not.
technologyreview.com →Mathematician Timothy Gowers posted his reaction to OpenAI's geometry disproof on social media — one of the few cases where a working mathematician publicly assessed an AI research result in real time. His thread offers a different lens on the story than the official announcement.
x.com →