'One screw-up' from catastrophe: AI makers scramble to wall off bioweapon risk· 3h ago
AI companies and biosecurity experts fear that next-generation models could help users design novel viruses or bioweapons, and the firms are racing to build safeguards before that capability arrives, the Financial Times reports.
The AI threats this newspaper has led with all summer were events that had already happened — a break-out from a sandbox, an unauthorised action, a central banker's warning. What is different here is that the industry is trying to wall off a capability that does not yet exist. The risk is the gravest category, one that could produce mass casualties, and the response is a set of self-imposed limits erected ahead of any demonstrated harm.
But 'future models could help' is a conditional. The framing follows the same shape as the safety coalitions and the systemic-risk warnings before it: the makers name a catastrophe and position themselves to manage the story before regulators do. Whether the safeguards would actually stop a determined user has not been tested outside the labs, and the assurance rests once again on the word of the same companies.
Two worlds split at the G20: Washington eases off, Brussels tightens· 5h ago
Washington and Brussels have chosen opposite roads on AI. At the G20 technology meeting the United States urged a lighter-touch approach while the European Union pressed ahead with a new law, Al Jazeera reports. At the same gathering, Musk and Zuckerberg were reported to have opposed AI regulation outright.
What splits here is not tone but direction. One side puts not impeding development first; the other wants to bind the technology with a framework before it ships. For companies, the calculus of where to base headquarters and compute shifts with it. Any prospect of a coordinated international line has, for now, receded.
Exclusive | New Google AI Model Said to Narrow Gap on Coding Ability· 9h ago
DEVELOPING…
Google to unveil Gemini 3.8 Flash AI model this week· 8h ago
DeepMind: Piloting the first double-blind AI evaluations· 4h ago
Hugging Face's Duck Robot Sells 10K Units, Runs on China Chip· 6h ago
BenchMIRT: What are LLM benchmarks actually measuring?· 9h ago
State of Local AI Report· 6h ago
DEVELOPING…
Cognition Set to Raise Around $1B at a $47B Valuation· 6h ago
Dell raises full-year forecast again on surging AI server demand· 7h ago
Palo Alto Networks: Palo Alto Projects Double-Digit Growth as AI Drives Cybersecurity Spending· 9h ago
Data Center Spending to Reach $31.6 Trillion by 2050 on AI Boom· 3h ago
DEVELOPING…
China AI chipmaker Enflame, last of the 'five little tigers', files for IPO to raise about ¥145bn (~$1bn)· 3h ago
Companies say AI isn't driving layoffs — but there's a disconnect with workersNEW
FT: AI sounds the death knell for audit fee inflation· 3h ago
DEVELOPING…
New York City: A.I. Is to Be Banned in N.Y.C. Elementary and Middle Schools· 5h ago
Musk, Zuckerberg Oppose AI Regulation at G20NEW
California lawmakers pass bill governing lawyers' use of AI· 7h ago
AI in every ministry — but only for chat? UAE targets replacing 50% of government work· 5h ago
One in three AI scribe notes carries a verified clinical error· 3h ago
Architect of British government's AI strategy joins Anthropic· 3h ago
- AI makers race to wall off bioweapon design before it arrives — self-imposed limits ahead of any demonstrated harm.
- The G20 splits: the US urges lighter rules, the EU presses a new law; Musk and Zuckerberg oppose regulation.
- A new Google model is said to narrow the coding gap — but the model is not yet in hand.
- Coding-AI firm Cognition is set to raise about $1bn at a $47bn valuation.
- New York City moves to ban generative AI in elementary and middle schools — institutions drawing lines.
Billing 2050 today
The trouble is that no one can check it. A capital-spending forecast that runs to a decimal a quarter of a century out is hope dressed as precision. When the major firms cannot even agree on demand three years from now, the premise that a 2050 total collapses into a single number reveals intent more than evidence. Numbers of this size do work: they justify the capex piling up today — 'the future is so large that current spending is, if anything, too small.'
This is not to deny that data centres will grow. But a decimal point and a 25-year horizon do not belong together. When you read it, look first not at the 31.6 but at who needs the 31.6 to be true.
The desks writing rules, and the danger no one has yet judged
And yet the lead went to a danger no desk has judged. The FT's bioweapon piece is neither an incident that happened nor a law that passed. It is a conditional fear — that next-generation models could help design novel viruses. I put the un-ruled territory above the rule-makers because it is the least reversible category of harm. But the piece is written in the shape of an industry binding itself: name the catastrophe, then step forward as its manager — the same shape as this summer's coalitions and warnings, as I wrote in the lead.
The newest model news of the day was an 'exclusive' saying a Google model had narrowed the coding gap. But the model is not in hand, and the performance rests on the words of the makers and those close to them. A preview ahead of a release stayed in the column, not on the front. A preview is not an announcement. Holding that line is the judgment the model-wars column tests every day, and I note it here.