Good morning. It’s a busy Wednesday: Anthropic’s biolab says Claude autonomously discovered a CRISPR-like enzyme system, Google shipped an expressive new TTS model, and the Pentagon is blaming AI for the deadliest targeting error in recent memory. Also, Bernie Sanders would like to send you to prison for building superintelligence.
Anthropic’s wet lab claims a CRISPR-adjacent discovery. Anthropic announced that Claude autonomously discovered a novel enzyme system with CRISPR-like repeat arrays in bacteriophage DNA, using roughly 950 agents running for 21 hours across 210 million tokens. Feng Zhang reportedly reviewed the preprint, though The Verge notes the company doesn’t yet know what the enzyme actually does, and TechCrunch flagged that a Stanford group previously found something similar. HN commenters were split between genuine interest and eye-rolling at the marketing-whitepaper format, with one pointing out the awkward contradiction: Anthropic spent months warning that Claude could enable bioterrorism, and is now using Claude to edit genomes. The IPO timing isn’t subtle either.
GPT-6 Astra can drive a car, apparently. A comma.ai-adjacent benchmark, DrivingBench, had GPT-6 Astra complete a GPS-guided steering course at 100% progress on its second attempt, while Claude Fable 5.1 maxed out at 45% and Grok 4.6 managed 11%. An openpilot contributor on HN gave a thoughtful breakdown of what the test actually measures — steering commands from camera feeds, not real-time control — and noted that latency alone rules out cloud LLMs for actual driving. Still, the gap between Astra and everything else on vision-spatial tasks keeps showing up in benchmark after benchmark, which is the more interesting signal.
Pentagon blames AI overreliance for Iranian school strike. Bloomberg reports the Pentagon has acknowledged that overreliance on AI targeting systems contributed to what it called the deadliest American targeting error, with the internal review concluding the failure “went beyond mere negligence.” HN commenters were near-unanimous that AI is being used as a scapegoat: an algorithm can’t be court-martialed, and someone made the decision to delegate targeting authority to it. As one commenter put it, AI didn’t cause the incompetence — it scaled it.
Bernie Sanders wants to jail superintelligence developers. Senator Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, which would pause frontier AI development and impose up to 20 years in prison for violators, with penalties explicitly modeled on illegal nuclear weapons work. Development could resume only after a scientist-led Department of Artificial Intelligence is established to oversee it. The bill has essentially no path in the current Congress, but it’s a marker of how the “pause” conversation has moved from open letters to draft legislation.
Gemini 3.8 gets a voice. Google launched Gemini 3.8 Flash TTS and Flash-Lite TTS, with per-line emotional direction, custom character voices from natural language prompts, and 30-second voice cloning gated by consent recording and SynthID watermarking. Early HN testers said the expressiveness rivals ElevenLabs v3, though the rollout has Google’s usual quirks — different features on different platforms, missing regions, and one user pointing out there are no neutral-gender voices in English. The consent-recording storage policy is also getting side-eye, which seems fair.
Claude Code silently ignored AGENTS.md if you turned telemetry off. A bug in Claude Code 2.1.277–280 gated AGENTS.md loading behind a remote feature flag that defaulted to false when telemetry was disabled, meaning users who opted out of data collection had their project instructions quietly dropped. An Anthropic engineer confirmed in HN comments that the flag was an intentional rollout kill switch, but its interaction with telemetry-off users was unintentional and was fixed the same day in v2.1.281. It’s a good reminder that “feature flag” and “silent behavioral change” are the same thing seen from two directions.
That’s the morning. The Anthropic biology announcement is the one worth watching — if the enzyme holds up under peer review, the “AI made a real scientific discovery” narrative gets its first genuinely defensible example.