DeepSeek V4.1 Flash, Claude's quantum circuits, and review at volume

Five primary pages today. Every number below is the publisher's own, and I have not reproduced any of them.

Claude designs quantum circuits

Anthropic gave Claude plain-language specifications for eight quantum circuits, drawn from chemistry, optimization, physics simulation and number theory, plus an automated evaluator that checks each design and scores it. The score is qubits multiplied by gates, a proxy for run time. Claude Mythos 5 worked in a closed loop: propose a circuit, get the score, improve the next attempt, with no human input between rounds.

The post says it found designs that beat the best published circuits on all eight, in some cases by an order of magnitude or more. Two are described in detail. One is a new way to do modular inversion, the most expensive step in elliptic-curve calculations. The other makes the costliest step of a leading quantum optimization algorithm about 12x cheaper. The authors say the gains came from new structural and mathematical insights, not from tuning existing designs. The work was led by Andi Gu with collaborators at Harvard and QuEra.

What I like is the shape of the task. Correctness can be checked by an ordinary computer, so the model gets a clear number to push on. That pattern keeps showing up in the science posts I cover.

DeepSeek-V4.1-Flash-1017

DeepSeek calls this a minor update to V4.1 Flash, released under the MIT license with a 1M token context window. The card's table compares it with the earlier Flash, K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Against the earlier Flash, Terminal-Bench 4.0 goes from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0 and Apex Bench v1.1 from 64.6 to 72.1. Opus 5 still leads on several rows, such as Terminal-Bench 4.0 at 51.8.

DeepSeek also says the model uses less execution time and shorter outputs. The release has no Jinja chat template. It ships a Python reference encoder and a separate Rust toolkit called deepseek-recipe instead.

How the Claude Code team reviews code

The Claude Code team writes that merged pull requests more than doubled from June to September while the share reverted fell by more than 40 percent. Their answer is layers: Claude verifies its own work before a pull request exists, a Claude review runs on every pull request, a separate security pass follows, and a risk-based classifier decides whether zero, one or two people approve. Rollback sits under all of it.

The point I would copy is the metric. They track human approvals per pull request and time spent waiting on a person, and treat reverts and incidents as the guardrail.

StarSkirmish and the benchmark map

StarSkirmish Bench v0.2 asks each model to write a Protoss bot in C++ within one hour. GPT-6 Astra scores highest, GPT-6.1 Sol is second, and Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1 are tied between 37 and 39. Against the three strongest human-written bots, the best model wins only 7 percent of games.

Mercor's benchmark map is the quieter counterweight. Across 923 occupations, only 208 have a task with at least five benchmark prompts, and the authors estimate 90 percent of US working hours are spent on tasks with little or no coverage.

What I'd watch next

Whether the evaluator-in-a-loop approach shows up in more fields, and whether DeepSeek's gains on agent benchmarks hold up outside its own table.