🧭 Claude Autonomously Redesigns Quantum Circuits — 12× Gate Reduction in Reed-Solomon Decoder

Anthropic has published a research collaboration with Harvard and QuEra Computing in which Claude worked in a closed autonomous feedback loop — no human in the iteration cycle — to redesign eight quantum circuits spanning chemistry, optimisation, physics simulation, and number theory. After 53 hours of unsupervised iteration, Claude produced two results that have caught the research community's attention:

• A modular inversion algorithm that avoids storing any branch history, relying instead on novel algebraic identities Claude discovered independently
• A Reed-Solomon quantum error-correction decoder with gate count reduced by a factor of 12, which would materially reduce decoherence exposure during execution on near-term hardware

The work was led by Anthropic STEM Fellow Andi Gu, with mentorship from Anthropic quantum physicist Alek Kemeny. The collaboration was structured so that Claude submitted proposed circuit modifications to an automated evaluator that returned fidelity and gate-count metrics — mimicking the tight loop that human researchers use with simulators, but removing human bottlenecks entirely.

What the closed feedback loop looks like in practice

The setup is an early example of autonomous STEM research at non-trivial depth: Claude was given access to a circuit description language, a simulation oracle, and a set of optimization objectives. It generated modifications, received scored feedback, and iterated — with no prompt injection between cycles. The human researchers reviewed final candidates, not intermediate steps.

🧭 A 25-Year Cryptographic Mystery Solved — Claude Traces the NIST Elliptic Curve Seeds to lrand48

One of the longest-running questions in applied cryptography — where did the seed values used to generate the NIST elliptic curves come from? — has an answer. Anthropic researchers working with Claude determined that the seeds for P-192, P-224, P-256, P-384, and P-521 were generated using C's lrand48 pseudorandom number generator, seeded with Unix timestamps from 1997–1998. The seeds were not deliberately obscured — they were produced by a researcher running a standard PRNG at a specific moment in time, and no record of that moment was kept.

Why this had been suspicious for 25 years

NIST published the curve parameters in FIPS 186-2 (1998) with seeds described as "randomly chosen" but no provenance documentation. The lack of a verifiable seed derivation led to persistent suspicion — most visibly from researchers who noted that the NSA had influence over the standard. The fear: the seeds could have been chosen via a trapdoor process that would give an adversary an advantage in solving the discrete logarithm problem on those curves.

• Claude systematically tested candidate RNG functions and seed spaces from 1990s C toolchains
• A match was found for lrand48 with timestamps consistent with NIST's known development timeline
• The curves were not generated with a trapdoor; the mundane explanation (PRNG + Unix timestamp) fits all known constraints

🧭 The AI Expectations Gap — Anthropic Staff Forecast 17% GDP Growth by 2030; US Median Is 5.4%

A new survey by economists Szymon Sacher, Charles I. Jones, and Anton Korinek — published by Anthropic — quantifies a striking divergence between how Anthropic employees and US adults at large expect AI to reshape the economy over the next four years. The gap is not modest.

Key findings

• Nobel-level discovery by 2030: 91% of Anthropic staff expect AI to make a Nobel Prize-calibre scientific discovery before 2030; only 39% of US adults agree
• Productivity acceleration: 59% of Anthropic employees report that AI already makes their tasks take a quarter of the time or less; this figure is 26% for the general US AI-using population
• Implied GDP growth (2030): Anthropic staff's responses imply a median forecast of 17.1% GDP growth by 2030 — compared to the US public median of 5.4%
• Labour market: Anthropic staff are significantly more likely to predict substantial job displacement in technical fields within 5 years

The researchers note that Anthropic employees' forecasts cannot simply be dismissed as self-serving — the survey was conducted anonymously, and employees with the most direct model exposure (those in safety and research) showed some of the highest forecasts. The authors frame the gap as an information asymmetry: people with daily, deep access to frontier models form materially different views about near-term capability trajectories than those who interact with consumer products or read press coverage.