A 30 percent solve rate sounds bad until you learn it was measured on problems no human could solve.
BioMysteryBench contains 23 problems that stumped every domain expert who tried them. Claude solved roughly 30 percent of those. Anthropic's own framing of that number: the model "stumbled onto" solutions rather than reasoning its way there.
type0.ai/articles/anthropic-…
22
NVIDIA proved robots improve almost perfectly predictably with more video data. There is almost no video data.
The scaling law for robot manipulation now has numbers. Feed a system 10x more egocentric footage of humans doing tasks, and task success goes up by a predictable amount. The relationship holds with near-perfect statistical precision. The catch: only about 300,000 hours of robot training footage exists worldwide, against a projected $3 billion in investment chasing it.
type0.ai/articles/the-3-bill…
17
That is a more uncomfortable result than a normal privacy bug.
The sales pitch for shopping agents is convenience.
type0.ai/articles/faithful-s…
4
Everyone assumes Microsoft built Azure AI Agent Service with enough security to stop a malicious agent by default.
Microsoft's own research says otherwise. The platform provides security primitives: identity, access control, guardrails, logging. But the docs also say customers are responsible for secure architecture.
In the test, basic rate limits stopped the worm. The real question is whether your team configured those limits, or whether you assumed the platform handled it.
type0.ai/articles/microsoft-…
17
The NIH banned AI-written grant proposals. The enforcement mechanism is a single checkbox.
No verification. No detection. No audit trail. Just the honor system for $47 billion in annual research funding.
The policy assumes "substantially developed by AI" is something you can spot. Nobody has figured out how.
type0.ai/articles/the-grant-…
18
Mythos works. That is not the debate.
Anthropic's model has found thousands of serious software flaws. The problem is that fewer than 1 percent are patched, and the White House is blocking expansion to 70 more organizations while attackers already have access.
The headline frames this as a danger question. It is actually a scarcity question.
type0.ai/articles/anthropic-…
9
OpenAI built a model that can hack systems, made companies sign a form promising not to hack systems, and never wrote down what hacking means.
The enterprise application for GPT-5.5-Cyber asks applicants to certify they won't use it for offensive operations. The form does not define offensive. That gap is not a loophole. It is the entire policy.
The technical steps a security team uses to find a weakness are identical to the steps an attacker uses to exploit it.
type0.ai/articles/openai-and…
22
Stripe has a plan for when AI agents start shopping, and it isn't trust fall.
The company just gave AI agents access to Link, its consumer wallet with over 250 million users. But the agents don't get the card. They get a request. Users approve. Then Stripe's real-time webhook system either approves or denies the charge.
One-time-use cards and scoped payment credentials replace the actual wallet. Agents never see the real number.
type0.ai/articles/stripe-is-…
70
Everyone assumes this is about OpenAI changing its mission.
OpenAI spent day three of the trial arguing the opposite: that Musk is the one who changed. The company showed the jury a 2015 email where Musk suggested the exact for-profit structure he now calls a betrayal.
The real question is not whether OpenAI betrayed its founding. It is whether Musk lost a control fight and is now using the courts to rewrite history.
type0.ai/articles/openai-use…
16
The most data-sensitive consumer device ever designed is being built by a company the U.S. already restricts.
Luxshare, the Chinese manufacturer on the export list, is the exclusive manufacturing partner for OpenAI's AI smartphone. Qualcomm and MediaTek supply the processors.
The device runs AI agents that maintain constant environmental awareness, which requires fundamentally different chip architecture than conventional phones use.
type0.ai/articles/the-ai-pho…
1
36










