the brief
Infra and tooling dominated: DeepSeek’s open-weight V4-Flash landed with strong scores, GitHub shipped stacked PRs, and Cloudflare exposed MoQ relays while Vercel added AI spend controls. On the agent/MCP front, Datasette Agent gained in-browser tooling, a new MCP client shipped, and papers like FilmBench and Orca-Bench pushed evals forward. Two deadlines loom: ChatGPT drops GPT‑5.4 models and the EU’s AI labeling kicks in.
GPT‑5.4 and 5.4 mini exit ChatGPT for signed‑in users on Aug 31; both remain available via the OpenAI API and API‑authenticated Codex sessions.
GPT-5.4 and GPT-5.4 mini will no longer be available for users signed in with ChatGPT starting August 31. The models will remain available on the OpenAI API and Codex sessions authenticated with an API key. https://learn.chatgpt.com/docs/changelog#codex-2026-07-31
From Aug 2, AI‑generated media on matters of public interest must be clearly labeled in the EU; audit and update pipelines for compliant disclosure.
AI-generated images, video, audio, and text on matters of public interest designed to look authentic must be labeled in the EU under the AI Act from August 2 (Jennifer Rankin/The Guardian) — Jennifer Rankin / The Guardian: AI-generated images, video, audio, and text on matters of public interest designed to look authentic must be labeled in the EU under the AI Act from August 2 — Companies must ensure people know when they are interacting with artificially generated images, audio and text des...
Open‑weight 304B‑param model with enhanced agentic abilities ships on Hugging Face (167GB), benchmarking near top flash models and reportedly punching above its weight.
deepseek-ai/DeepSeek-V4-Flash-0731 — <p><strong><a href="https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731">deepseek-ai/DeepSeek-V4-Flash-0731</a></strong></p> The latest release in DeepSeek's V4 family, "with substantially enhanced agentic capabilities". It's 304 billion parameters - 167GB on Hugging Face - but it appears to punch <em>well</em> above its weight.</p> <p>Artificial Analysis <a href="https://artificialanalysis.ai/models/deepseek-v4-flash">rank it</a> ahead of MiniMax M3...
Public preview enables splitting large changes into reviewable PR stacks, improving review flow and merge safety on complex codebases.
Large changes are hard to review. Smaller pull requests make them easier to ship. Stacked PRs are now on GitHub. https://docs.github.com/en/pull-requests/how-tos/stacked-pull-requests?utm_source=bluesky-motion-graphic&utm_medium=social&utm_campaign=stacked-prs-gtm-public-preview-2026
New provisioning API lets you spin up isolated Media over QUIC relays with fine‑grained publish/view controls for real‑time media at scale.
An API for MoQ: provision your own isolated relays — Last year we made every Cloudflare server a Media over QUIC (MoQ) relay. Now the new provisioning API lets you create your own isolated relay and control who can publish and who can only watch.
AI Gateway budgets can now scope to teams and projects, enforcing dollar limits across overlapping budgets and auto‑blocking requests once limits are hit.
AI Gateway now supports team and project spend budgets — AI Gateway spend budgets can now be scoped to a team or a project, in addition to individual API keys. Set a dollar limit on a scope, and the gateway meters spend against it and stops further requests once the limit is reached, until the budget resets or you raise it. A budget attaches to one of three scopes. A request can fall under multiple budgets and has to pass all of them: if any one is over its limit, the request is rejected, eve...
0.4a0 adds await context.browser_task(), letting agent tools safely execute code directly in the user’s browser for richer, interactive workflows.
datasette-agent 0.4a0 — <p><strong>Release:</strong> <a href="https://github.com/datasette/datasette-agent/releases/tag/0.4a0">datasette-agent 0.4a0</a></p> <blockquote> <ul> <li>New <code>await context.browser_task()</code> mechanism allowing agent tools to run code directly in the user's browser. <a href="https://github.com/datasette/datasette-agent/pull/33">#33</a></li> </ul> </blockquote> <p>This is an exciting new capability: it makes it easy for Datasette Agent plugins to provide tools ...
Initial alpha of Simon Willison’s MCP client for the llm ecosystem, enabling easier connections to Model Context Protocol servers from tooling.
llm-mcp-client 0.1a0 — <p><strong>Release:</strong> <a href="https://github.com/simonw/llm-mcp-client/releases/tag/0.1a0">llm-mcp-client 0.1a0</a></p> <p>See <a href="https://simonwillison.net/2026/Jul/31/stateless-mcp/#llm-mcp-client">this blog entry</a>.</p> <p>Tags: <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/model-context-protocol">model-context-protocol</a></p>
Prime Radiant’s lightweight framework compares models, prompts, and harnesses; try it via “uvx smevals docs” to stand up quick, targeted capability checks.
smevals - a small eval suite for evaluating models, prompts, and harnesses — <p><strong><a href="https://primeradiant.com/blog/2026/smevals.html">smevals - a small eval suite for evaluating models, prompts, and harnesses</a></strong></p> I've been working with Jesse Vincent's <a href="https://primeradiant.com">Prime Radiant</a> applied AI research lab building out this evals framework to help answer questions about the capabilities of different models.</p> <p>The result is <strong><a href="ht...
Google is rolling out Gemini Spark access to AI Pro subscribers in 160+ countries and adding desktop Chrome auto‑browse integration for research/automation flows.
Google starts rolling out access to Gemini Spark for Google AI Pro subscribers to over 160 countries and adds a Chrome auto browse integration on desktop (Abner Li/9to5Google) — Abner Li / 9to5Google: Google starts rolling out access to Gemini Spark for Google AI Pro subscribers to over 160 countries and adds a Chrome auto browse integration on desktop — Gemini Spark is getting two big updates today, starting with Chrome auto browse integration on desktop. — Previously, Gemini Spark had acces...
A new demo video teases improved robotics capabilities powered by Gemini; details are light, but the focus is manipulation and embodied control.
This is Gemini Robotics 2 🤖
A new benchmark uses prompts inspired by award‑winning films to assess models on dynamic, cinematic aesthetics, exposing gaps traditional video benchmarks miss.
FilmBench redefines video generation benchmarking with a cinematic evaluation framework from experts. Using prompts from award-winning films, it evaluates AI models, emphasizing cinematic skills and gaps in dynamic aesthetics versus existing benchmarks. https://arxiv.org/abs/2607.24241
Technique raises draft acceptance in speculative decoding for multi‑head latent attention models, delivering notable efficiency gains without extra inference cost.
This study presents Functional Reconstruction, enhancing draft acceptance in speculative decoding for multi-head latent attention models. It achieves significant efficiency gains without additional inference costs, revolutionizing large language model performance. https://arxiv.org/abs/2607.27269
ArXiv benchmark evaluates how ready LLM agents are for real on‑call operations, offering a structured way to compare agent robustness and reliability.
Orca-Bench: How Ready Are Language Model Agents for Oncall? — Article URL: https://arxiv.org/abs/2607.28545 Comments URL: https://news.ycombinator.com/item?id=49126943 Points: 11 # Comments: 3
Simon Willison unpacks the biggest MCP spec change yet and how stateless MCP enables simpler, more robust tooling—spurring projects like mcp‑explorer and datasette‑mcp.
Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp) — <p>Tuesday was <a href="https://x.com/ade_oshineye/status/2082129440943866149">Stateless MCP day</a> - the rollout of MCP 2.0, or <a href="https://blog.modelcontextprotocol.io/posts/2026-07-28/">the 2026-07-28 Model Context Protocol specification</a> to use the more formal but less memorable name. This is the most significant change to the MCP spec since it first launched, and has also served to reignite ...
Honest postmortem: Tailscale wasn’t exploited but still should have blocked the Hugging Face breach—useful lessons on network assumptions and zero‑trust boundaries.
Tailscale didn't stop the Hugging Face intrusion — Tailscale wasn’t exploited. We still should have stopped the intrusion.