AI on X Summary (October 30, 2026)

Here are some interesting things that happened on my X timeline in the last day: people tried the new Gemini 4 and GPT-6.1 models, a few tools for building with coding agents launched, and the debate over how much AI work should be prompted or remembered kept going.

What got attention

New models, mostly judged by feel. @BHolmesDev calls Gemini 4 Argon the best copywriter they have seen, citing its tone matching and default voice. On the OpenAI side, @angelbrodin says even internal skeptics now like the new GPT-6.1 Terra release, and @udayruddarraju finds its computer use addictive. Both OpenAI posts come from people close to the company or enthusiastic users, so treat them as impressions. @kqyno7 passes on a Reddit rumor about a future Gemini 4 model; it is unconfirmed.

Agents for creative and testing work. @lukeigel announced Kino, a tool for editing video end to end with Claude Code and Codex, and later said you can use it with public-domain archival footage (follow-up). @PandelisZ describes a startup that sends a swarm of agents through a product to test it like real users. Both are launch announcements, not independent reviews.

Agent commerce is still tiny. According to @juokaz, DoorDash says it has received only several hundred orders in total from personal agents such as Muse. This rests on one post.

Google’s economics research. @alexolegimas announced a first release from a Google DeepMind group studying the economics of AGI, written with Andrew Koh and Shane Legg.

Usage limits and free access. @derrickcchoi quotes an announcement that usage limits are doubling for higher paid tiers. @joshwoodward says Notebook now offers over 700 publisher textbooks free. @suraj_sharma14 says OpenAI opened applications for Codex Ambassadors, with a free Pro subscription and credits.

Practices worth copying

• Write less prompt, add a check. @vikingmute summarizes a post arguing that strong models are over-prompted. Give a clear goal and a concrete verification condition, such as an end-to-end test, and let the model work. The reason: extra instructions can limit what the model does.
• Know your harness. @Saboo_Shubham_ defines an agent as a model plus a harness, the code that turns the model’s chosen action into work, and lists Claude Code, Codex and others as examples. Thinking in those two parts helps when comparing tools.
• Let an agent do the setup. @garrytan says you can point any coding agent at a document to get GBrain running quickly. Handing an agent the setup instructions saves manual steps.
• Retrieval over elaborate memory. @Xudong07452910 reports a paper finding that many purpose-built agent memory systems did no better than plain RAG over raw history, and cost more. One paper, relayed second-hand.
• Use agents for odd chores. @turingou reports having a Grok bot work out an oven temperature program from an online recipe when the oven had no preset. A small case, but a sensible use.

Doubts and open questions

@dingyi lists the visual habits that mark AI-generated design as “slop”: heavy labels and italics, the same few fonts, too many dividers and shadows, over-detailed charts. @sewon__min argues that good slides involve deliberate choices about what to say aloud versus show, which AI-generated slides tend to miss. @yinsuecci asks why many of OpenAI’s reported math results depend on exact cancellation of higher-order terms, and wants to hear how others read that. @steipete writes that some engineers feel a loss of identity as AI changes their jobs, and says he sympathizes. @MaxForAI highlights an Anthropic post saying Claude flags possible misuse after responding, a technique called self-assessment, which has run in production since June. Whether that worries you depends on how you feel about review of conversations.

Takeaways (my own synthesis)

• Model comparisons on the timeline are still mostly personal impressions. Try the model on your own work before believing any ranking.
• Tools that wrap coding agents for non-coding jobs, such as video editing and product testing, keep appearing. Most are early announcements.
• Simple setups, whether a short prompt with a test or plain retrieval, keep beating elaborate ones in these reports.
• Agent shopping looks small for now, if the DoorDash figure holds.