AI Daily · 2026-08-01

The headline of the day is DeepSeek V4-Flash-0731, a 304B model that gained massive agentic improvements through pure post-training, rivaling or surpassing larger frontier models on terminal benchmarks while offering extremely low-cost API and open weights—fueling fresh momentum for the open-weight revolution. Simon Willison and others have turned their attention to stateless MCP, simplifying toolchains with projects like mcp-explorer. In other news, OpenAI revealed advances across mathematics and theoretical CS, LangChain introduced ReviewBench for evaluating code review agents, and Willison shipped the lightweight eval framework smevals.

North America · First-hand

OpenAI

⭐⭐ [Research] Ten advances in mathematics and theoretical computer science

OpenAI shares new research results on longstanding open problems in mathematics and theoretical computer science, including advances in geometry, cryptography, and complexity theory.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] deepseek-ai/DeepSeek-V4-Flash-0731

Simon Willison covers DeepSeek's V4-Flash-0731 release, a 304B model with substantially enhanced agentic capabilities. On Hugging Face it is 167GB, yet Artificial Analysis ranks it above the 428B MiniMax M3. Its pricing ($0.14/M input tokens, $0.27/M output tokens) positions it as potentially the best value-for-intelligence model available. Willison tested pelican generation via OpenRouter: the default reasoning level produced a poor result, but increasing reasoning effort to 'high' yielded a much better output.

⭐⭐⭐ [Model Release] [AINews] not much happened today

The piece covers the DeepSeek V4-Flash 0731 update, a post-training-only upgrade that delivers major agentic capability gains without changes to architecture or parameter count. It achieves a large leap on benchmarks like Terminal-Bench, approaching GPT-5.6's performance frontier while offering lower costs via a 98% cache-hit discount. The API and open weights were released on the same day, sparking immediate community quantization and deployment, and the consensus highlights it as a tool-use and long-horizon post-training win rather than a scaling story.

Product Update

⭐⭐ [Product Update] smevals - a small eval suite for evaluating models, prompts, and harnesses

Simon Willison and Jesse Vincent's Prime Radiant lab have released smevals, a small eval suite for running evaluations across different model configurations and grading the results. Users define evals as directories with YAML task files, run them against models via uvx smevals run, grade separately, and view reports through a local server or static HTML. The tool introduces a clear vocabulary: eval, task, config, run, grader, and checks, each with a defined role. This is Willison’s third iteration on eval approaches, and he is pleased with the current design, planning to expand it for his own projects.

⭐⭐ [Product Update] datasette-agent 0.4a0

Datasette tool 'datasette-agent' releases version 0.4a0, introducing an `await context.browser_task()` mechanism. This allows agent tools to execute code directly in the user's browser, making it easier for Datasette Agent plugins to provide custom JavaScript execution capabilities. This is a minor alpha release focused on this new feature.

Research

⭐⭐ [Research] Evaluating code review agents with ReviewBench

LangChain built ReviewBench, a code review agent benchmark grounded in real defects from trusted reviewers on the LangSmith monorepo. Raw comments were filtered and curated into verifiable findings, then packaged as 59 reproducible Harbor tasks covering 64 baseline issues. The benchmark tests whether an agent can catch substantive defects — such as missing tenant filters or API parity regressions — that require understanding broader system contracts, not just changed lines. Scoring measures coverage and precision via LLM-as-judge comparison against the curated set.

Opinion

⭐⭐ [Opinion] Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)

Simon Willison has regained interest in MCP 2.0 (stateless MCP), noting that the new spec enables tool calls in a single HTTP request, dramatically simplifying client and server implementations without session state, making it more scalable for web applications. He built mcp-explorer, a CLI tool for interactive MCP server probing, and also mentions datasette-mcp. The post illustrates the request format differences between legacy and stateless MCP.

⭐⭐ [Opinion] Oxide and Friends: The Open Weight Revolution with Simon Willison

Simon Willison appeared on the Oxide and Friends podcast to discuss the open weight model revolution. Key topics included Kimi K3 showing open weight models can compete with proprietary frontier ones, recent accidental cybersecurity attacks, and a public letter on open weights and American AI leadership signed by most big names in AI. Willison noted the conversation was already outdated, as it would have covered DeepSeek V4 Flash 0731 and Anthropic’s security incident had they recorded a few days later. The wide-ranging chat also touched on Golden Gate Claude, the Zizians, other digressions, and revisited predictions from January, adding a new prediction that the Pope will say something about open models by year’s end.