AI news, October 31, 2026: one robots.txt switch for AI training, Gemini edits whole decks, Claude Code write guard
Six stories from Thursday and Friday, October 29โ30, 2026 (US time) for solo makers who build with AI. One changes how you can talk to AI crawlers from your own site, two are open models you can download today, one is a safety setting for coding agents, and two show where AI features are and aren't landing with real users. Figures are as the companies and outlets state them.
Google and OpenAI plan to honor one robots.txt signal for AI training
Google and OpenAI both said they will respect a shared "policy token" in robots.txt that lets a site state its AI-training preference to every participating provider at once, instead of writing separate rules for each crawler. The mechanism comes from an Internet-Draft in the IETF AI Preferences working group: you add a normal robots.txt group with User-Agent: __usage_ai-training and use Disallow or Allow paths, and the draft says a User-Agent: * group must not count as a preference. Google says it plans to support it in 2027 and will treat an opt-out as disallowing Google-Extended unless that token is set separately; OpenAI says it aims to implement it in the coming months. Search Engine Land reports that Bing and Anthropic announced support too, and existing per-company controls keep working.
Why it matters for makers: once this ships, one short block in robots.txt decides whether your blog or docs can be used for training, without tracking every new crawler name.
Sources: Google Search Central, OpenAI, IETF draft, Search Engine Land
Gemini in Google Slides can now edit a whole presentation from one prompt
Gemini in Google Slides can now change many slides or an entire deck at once from the side panel. You can ask it to restyle fonts, backgrounds and colors to match another file, rewrite content based on files in Drive, merge or add slides, or make titles consistent, and it edits the native slides directly before showing a summary of what changed. Edits can be undone or rolled back with version history. It is English-only at launch, is rolling out to Workspace Business and Enterprise Standard and Plus plans and Google AI Pro and Ultra, and comes with higher promotional limits through at least December 1, 2026.
Why it matters for makers: pitch decks and client reports that need a quick restyle or update are becoming a one-prompt job instead of slide-by-slide work.
Source: Google Workspace Updates
Clover-35B-A3B-Preview: an open diffusion language model with 3B active parameters
DiffuSpace released Clover-35B-A3B-Preview, a mixture-of-experts language model that generates text by diffusion rather than one token at a time. It has 35B total and 3B active parameters, supports a thinking mode, and is published under an Apache license with weights on Hugging Face and code on GitHub. The team says it beats the similarly sized Qwen3.6-35B-A3B on several benchmarks while generating noticeably faster, and inference is supported in SGLang and llama.cpp.
Why it matters for makers: diffusion models were mostly closed APIs until now, so an open one that runs in llama.cpp lets you test the speed claims on your own hardware.
Sources: Hugging Face blog, GitHub
SenseTime open-sources SenseNova-Loop-9B-Preview, which reuses its layers
SenseTime published SenseNova-Loop-9B-Preview under Apache 2.0, a model that takes text and images and is aimed at long-context, coding and tool-using agent tasks. According to the model card, a block of 24 of its 32 decoder layers runs twice with the same weights, so each pass does 56 layer computations while storing only about 9B parameters. The technical report has not been published yet, and running it needs the custom model code that ships with the weights.
Why it matters for makers: looping layers is one way to get more reasoning out of a model that still fits on modest GPUs, which is the size range most solo projects can afford to host.
Sources: Tech City Authority, Hugging Face model
Claude Code v2.1.314 can lock file writes to your working directories
Claude Code v2.1.314 adds a permissions.blockWritesOutsideWorkingDirectories setting that makes its Write, Edit and NotebookEdit tools refuse any path outside the working directories, in every permission mode. The release also makes claude -p print an error and exit with code 1 when startup fails, instead of exiting 0 with no output, and lowers the effort level and retries when an admin has capped it. Long idle conversations in VS Code and Claude Desktop now get compacted automatically.
Why it matters for makers: if you run Claude Code unattended in scripts or CI, the new setting is a cheap guard against edits outside the repo, and the exit-code fix means failed runs no longer look like silent successes.
Source: GitHub release notes
DoorDash says AI agents have placed only "several hundred" orders so far
On its third-quarter earnings call, DoorDash said third-party AI agents have sent it only "several hundred" orders in total, while TechCrunch notes the company completes about 3 million deliveries a day. CEO Tony Xu said customers still order mainly through the DoorDash app, but the company wants to learn from agent use cases and may one day advertise inside agents. DoorDash already runs its own text-to-order agent, a CLI agent and an MCP for corporate ordering from other chatbots.
Why it matters for makers: agent-driven shopping gets a lot of attention, but real usage is still tiny, so building for agents makes sense as an extra channel rather than a replacement for your app or site.
Source: TechCrunch