Anthropic · Claude · AI for science
Claude found something in DNA that scientists had not recognized as a system
Anthropic introduced its life-sciences research group and physical laboratory with an unusually concrete result. Researchers asked Claude to search DNA databases for interesting reverse transcriptases. Roughly 950 agents then spent 21 hours on the computational search, gathered more than 200,000 enzymes, found about 3,500 potential systems and narrowed the field to 20 candidates.
One agent noticed an unusual reverse transcriptase next to an organized array of repeating DNA sequences in bacteriophages. Anthropic's researchers performed the physical experiments and concluded that the components appear to form a previously uncharacterized system, which they named ART. The computational search had high-level human direction; the wet-lab validation remained human work.
The CRISPR comparison needs brakes. The similarity lies in ART's repeat-array architecture, while ART's biological function remains unknown and no gene editing or therapeutic use has been demonstrated. The result is interesting without exaggeration: AI inspected a space no person could manually review, noticed something odd and gave scientists a worthwhile experiment. Claude found something; now everyone gets to learn what it does.
Source: Anthropic ↗United Nations · Frontier AI · Global safety
The people building frontier AI asked the Security Council to help control frontier AI
OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei addressed the United Nations Security Council during a session on AI and international security. Amodei said poorly managed AI could become a risk to humanity as a whole. Altman warned that humanity could lose control of its future to AI and called for shared benchmarks to measure capabilities and safeguards.
The institutional setting matters. The Security Council usually handles wars, nuclear weapons, sanctions, terrorism and threats to international peace. It now has frontier-lab executives asking governments to build common rules around the systems those companies are developing. Hugging Face co-founder Clément Delangue and AI researcher Yoshua Bengio also joined the meeting.
No treaty or binding frontier-development limits emerged, and political disagreement remains enormous. Thursday begins with a smaller but meaningful change: frontier AI risk has entered the same diplomatic room as international security. The debate moved from research papers to company safety teams, national regulation and now the Security Council. Whether governments can move as quickly as the models is the less encouraging question.
Source: Reuters ↗IonQ · Quantum computing · Error correction
Quantum computing solved a problem by giving it to the boring computer
Useful fault-tolerant quantum computers will need continuous error correction. A classical decoder must interpret a stream of error information quickly enough that the quantum computation does not sit waiting. IonQ says its new end-to-end decoder can do that in real time on one conventional commercial CPU.
The company tested a dual-decoder architecture with simulated circuits representing as many as 408 logical qubits across 88 memory blocks and magic factories, executing more than 31.5 million operations. IonQ reports as little as 0.02% stretch time under standard operational noise. The caveat is essential: the test did not demonstrate a functioning 408-logical-qubit quantum computer.
It did demonstrate that one classical scaling problem may be less ugly than expected. IonQ also says its Superion 256 system will be installed at Nvidia's Accelerated Quantum Research Center and connected with GPU infrastructure. Practical quantum computing increasingly looks like a division of labor: CPU handles this, GPU handles that, QPU handles the weird part. The ordinary computer got the job of keeping the quantum computer sane.
Source: IonQ ↗