Today’s loudest claim is that OpenAI and Anthropic may have toppled a hypothesis that whole research careers were built on. Elsewhere, Oracle wants a meter running on top of your AI subscription, and Mistral is selling open weights as geopolitical insurance.
1. Oracle wants a meter running on top of your AI subscription
The key word in The Register’s Oracle story is “add.” Per the dek, Oracle says AI units will add to subscriptions, so this is a usage meter bolted onto existing contracts, not a clean swap from seats to consumption. The source here is only a headline and summary, so how an AI unit is defined and priced is unknown.
The context is a lot of money going out. The Register cites $90 billion-plus in capex, mounting debt and a $225 billion revenue goal for 2030. Our read, not Oracle’s stated reasoning, is that a flat per-seat fee is a poor way to recover costs that rise with every task an AI agent runs. A meter ties revenue to the work being done.
For customers, that moves risk. Seats were predictable, but task-based charges depend on how much automation you adopt and on how the vendor defines a task. Founders building on Oracle should ask for those definitions and rates before the pricing becomes a line item that grows faster than their revenue.
2. Two AI labs say they disproved ETH, and theorists are checking their own papers
Peter Jonsson has a paper due at a top theory conference in two weeks, and he says much of it is now “dead.” The cause is a claim from OpenAI and Anthropic this week that they disproved the exponential time hypothesis, a 27-year-old conjecture that a lot of complexity and algorithm-design work treats as a foundation.
The caveat matters: researchers haven’t independently verified the proofs, WIRED reports. Both firms published them on GitHub, in natural language and in Lean, a machine-checkable format. OpenAI has asked ETH’s originators to organize a community effort and promised to fund a conference. Daniel Lokshtanov estimates that if the result holds, around 50 of his papers become pointless or need revisiting. Thore Husfeldt calls it a once-in-a-generation result, and says it may make his year of work irrelevant.
The more telling detail is the mood. Some of these researchers are thrilled, because a long-open question just got an answer. Husfeldt’s department has even considered hiring an occupational psychologist as AI results keep wiping out ongoing projects. Add OpenAI’s earlier dump of 722 manuscripts, some with errors, which some mathematicians read as IPO-driven haste. The strongest read is that a good proof can be both exciting and exhausting, and the field’s verification capacity is now the bottleneck. Cryptography isn’t directly affected, but Lokshtanov says the result should still “scare” some cryptosystem designers.
3. Belgium’s tax agency confirms cyberattack and pulls services offline
Belgium’s tax agency has confirmed a cyberattack and taken services offline, according to The Register’s headline. The supplied text contains nothing beyond that, so there is no detail on the attacker, the method, any data exposure or how long the outage will last.
Even so, the pairing is worth noting. Confirming an attack is routine, but pulling public services offline means the agency chose disruption over the risk of leaving systems running. That is a defensible call and a costly one for a body whose job is to process filings and payments.
The questions that will decide how serious this is are whether the services come back quickly, and whether the agency says what it found when it looked.
4. Mistral’s ‘Le Chonk’ pitch: an AI model nobody can switch off
Mistral’s argument for Le Chonk is political as much as technical. The French lab says the finished trillion-parameter model is the “world’s best” open-weight model for automated coding and cybersecurity, matching or surpassing Chinese models from DeepSeek and Z.ai. WIRED’s report relays the claim from Mistral’s press briefing and cites no independent benchmarks, so treat the ranking as unproven.
The sales pitch is control. Because the weights are public, VP of science Pierre Stock says, a model can’t be cut off remotely, unlike proprietary offerings from OpenAI, Anthropic and Google. Stock also called Chinese models potential “threat vectors.“
Sophos’s Rafe Pilling undercuts both halves. Open weights are dual-use, he says, so attackers can download and customize them too. And any advanced model can go rogue without careful monitoring, wherever it was built. He calls the nationalistic framing a different thing from the technical reality. The practical read: Le Chonk may give Western enterprises and governments a non-Chinese option, but the benefit rests on Mistral’s unverified claims, and the safety burden stays with whoever deploys it.
5. Mintive’s agent-control startup raised $11 million in a market of unicorns
Mintive is entering what TechCrunch calls a field with hardly any room: Amazon, Microsoft, OpenAI and Anthropic all sell agent management, along with point products and more than 40 Y Combinator startups. Its funding is a $7 million seed led by Cowboy Ventures, bringing the total to $11 million, not the Series A the selection implied.
The product covers an MCP gateway that controls which cloud services agents can access, an LLM gateway that routes models and caps costs, separate agent identities and a company-wide monitor. The credentials are strong: co-founder Vijay Vasudevan helped create TensorFlow, and angels include Jeff Dean and Andrej Karpathy. Mintive says it has 50-plus enterprise customers since launching in February.
The restraint is the interesting part. Cowboy’s Aileen Lee praised CEO Jiquan Ngiam for not chasing a “vanity” round, while TechCrunch contrasts it with Prime Intellect’s $130 million Series A at a $1 billion valuation. Mintive’s claim is ease of deployment, which is unproven against giant platform vendors, but a small raise means it has less to prove.