Bengio's LawZero lands $300M from Canada and Germany
Two governments just funded an alternative to how frontier models are trained — and the researcher behind it says that is the point.
LawZero, the non-profit Yoshua Bengio founded last year to build safe AI systems, will receive up to $300 million Canadian in grant funding from Canada and Germany — up to $150 million from each government — to hire staff and pay for the compute behind a research system it calls Scientist AI. The announcement was made Wednesday at Montreal's All In AI conference, and the money buys something specific: a Berlin office and sovereign compute capacity in Canada through partnerships with data centre providers Hypertec and 5C, so the work does not sit on a US hyperscaler's balance sheet. LawZero employs close to 50 people and was incubated out of Mila, the Quebec institute Bengio also founded.
Bengio's case is that the misbehaviour the field keeps catching — the Hugging Face hacking incident is the freshest example — needs three conditions at once: a system with the intent to do harm, enough capability to pull the attack off, and an environment vulnerable enough to let it land. "All three were present last summer," he told The Globe and Mail. His conclusion is the uncomfortable half: patching software raises the cost of an attack but does not remove it, and "eventually, they'll be able to pass through any kind of software defence." So the funding goes toward a different training approach rather than better filters. Scientist AI will not use reinforcement learning — a deliberate break from standard practice — because Bengio argues reward-chasing behaviour persists after training ends, which is what produces models that lie or cheat to finish a task: "They're obsessed with achieving the goal because that's what reinforcement learning does."
The political subtext is worth as much as the science. Germany and Canada each wrote a cheque roughly the size of LawZero's entire original philanthropic backing, and Bengio framed the money as leverage for middle powers rather than charity: "Middle powers like Canada, as Mark Carney has been saying, need to have cards in their hands so that they can sit at the global table." He says LawZero is in talks with other countries about further funding. Read it as a bet that safety research becomes a piece of national infrastructure — funded by states that cannot build frontier models themselves but refuse to be spectators while two countries set the terms. We covered the argument underneath this in September — Bengio: the training process itself makes AI dangerous.
MLCommons published its MLPerf Inference v6.1 results, and for the first time the suite measures agentic and retrieval-augmented pipelines rather than single-shot model calls. The round drew submissions from a record 30 organisations, including six first-timers, and added two tests: an End-to-end RAG benchmark for multi-step, multi-component query answering and an Edge Agentic Inference test for agentic workloads running on local devices. The largest system ever submitted to the benchmark showed up too — 512 accelerators — alongside two heterogeneous configurations, one pairing accelerators from different vendors and one split across the Pacific.
Nvidia used the round to put its next rack on the scoreboard. Vera Rubin NVL72 appeared in preview for the first time, with the company claiming up to 3.7x higher throughput than GB300 NVL72 on the Qwen3-VL vision-language test and up to 2.5x on DeepSeek-R1. The per-accelerator trend matters more than any single entry: the best VLM server result improved 2.99x in six months, and the best DeepSeek-R1 result is 5.7x better than a year ago, which is the compounding that decides what an inference fleet costs to run. These are vendor-framed numbers in a vendor-designed round, so treat the multipliers as claims — the useful news is that agentic inference now has a standardised measurement path at all. We looked at Rubin's earlier agentic figures in August — Nvidia's Vera Rubin posts a 30x agentic jump in first hardware tests.
May Mobility will go public through a $1.4 billion merger with blank-cheque firm ACP Holdings Acquisition Corp, and expects to trade on Nasdaq under the ticker "MAY". The deal is structured to raise up to $337 million in gross proceeds, including a fully committed $120 million private investment, with the remainder from the SPAC's trust — money shareholders can redeem before close, so the final figure may land lower.
The pitch is asset-light: May sells autonomous Toyota Siennas to fleet partners and keeps the software and remote supervision, taking fixed fees or per-trip licensing. It has run more than 550,000 commercial rides across 1.1 million miles in the US and Japan, with partnerships including Uber, Lyft and Grab, but reported roughly $10 million in revenue last year against about $93 million of cash burn. The proceeds are earmarked for removing safety drivers and cutting bill-of-materials costs — the two things that decide whether per-trip economics ever cross zero. Going public first in this category also makes May the market's guinea pig for whether investors want pure-play robotaxi exposure at all, which matters far more to the sector than the valuation headline.
What to watch: whether other governments follow Canada and Germany into LawZero — and whether the MLPerf Endpoints suite, which will replace MLPerf Inference for datacenter submissions, ships its first agentic results before the next hardware cycle.
If states start funding safety research as national infrastructure, does that make alignment progress faster or just more fragmented? Tell us in the comments.
