← All episodes
Daily AI Brief

Uber says AI agents booked just four rides in its busiest week

Wednesday, November 4, 2026 · 99 stories · 6:31 · 775 words
Uber says AI agents booked just four rides in its busiest week

Watch on YouTube · subscribe for the daily brief

On its third-quarter earnings call, Uber said it handled about 320 million trips and orders across Uber and Uber Eats in the week of Halloween, its busiest week of October. AI agents ordered four of the rides, and 500 Uber Eats orders. CEO Dara Khosrowshahi: "It is very, very early and very limited traction that we are seeing so far."

Why agents have nothing to save here

Khosrowshahi's own numbers explain it. Booking a ride in Uber's app takes about 50 seconds on average; re-ordering a favorite meal takes 90. An agent that saves a few seconds, after you have explained what you want, is not a better product. DoorDash reported the same thing a week earlier: only a few hundred orders from AI agents so far.

He is still bullish. He counts about eight major AI assistants, expects more than ten, and says "every single AI system company is talking to us." For builders, the lesson is simple: an agent wins where the task is long, repetitive or spread across tools, not where a good app already makes it one tap.

Where agents do work: code, and two deals on the same day

  • Meta bought Conductor (SiliconANGLE). Its Mac app runs several coding agents at once, each in an isolated copy of the codebase, then lets you review and merge. It says 100,000+ people use it; Meta, Google and Spotify engineers used it before the deal. Terms were not disclosed. If you use it: sign-ups have stopped, Conductor Cloud goes read-only on Nov 13 and shuts on Nov 20, and stored data is deleted after that. Export or move workspaces to your machine first.
  • OpenAI is buying Pierre Computer Company (The Decoder). Pierre's Trees, Diffs and Highlights libraries already power Codex, and its Code Storage tracks history through Git. OpenAI's reason: with agents writing code in parallel, Git tooling has to keep up. The open-source projects stay open; the deal needs regulatory approval.

Both buyers paid for the layer that coordinates many agents on one codebase, not for a better model.

Apple rations tokens

Apple cut the daily AI spending cap for its engineering teams from $300 to $200 in October, Bloomberg reported via TechCrunch, with some allowances heading to $100 a day and a possible $25 cap outside engineering. If your team is still on "use as much as you can", expect a budget conversation. A per-person, per-day cap is the model to plan for.

Voters put AI last

In NBC News and SSRS exit polls, 56% called AI and data centers an important factor in their vote, last of five issues, behind corruption (90%), cost of living (89%), the war in Iran (74%) and immigration enforcement (73%).

GPT-6 Astra edits other models' weights

MIT researchers showed GPT-6 Astra reading raw weights of other networks and changing them directly, which they call "weight-space programming". On Qwen3.5-0.8B, it changed 0.016% of the weights to break thinking loops; MATH-500 accuracy went from 7.9% to 48.1%. Limits: 50 test iterations, a third of answers still hit the length limit, and no gain on Gemma 3 270M, which rarely looped.

Claude's watermark, and its off switch

Anthropic told API customers that Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 get text watermarking from Nov 16. On these four, the mark only applies when temperature, top_p and top_k are at their defaults. If you rely on detecting it, know that one sampling parameter turns it off on those models.

Tencent Hy4

Tencent released Hy4: 770B parameters, 49B active, 1M-token context, and computer use for the first time.

Security

  • APT flaw, CVE-2026-109518: a malicious mirror or proxy can make APT on Debian 12+ and Ubuntu 22.04+ install a swapped package as root. Fixed APT on Debian: 2.6.2, 3.0.4, 3.3.6; Ubuntu is rolling out fixes. Update APT first. Advisory: APT team, Ubuntu.
  • DreamHost: attackers were in from August through a file manager flaw and took session tokens, files and inboxes; "well under 10%" of customers affected (TechCrunch).
  • mSecure: the password manager told current and former customers to change passwords, reset 2FA and remove passkeys, without saying what happened (The Hacker News).

Quick hit

ChatGPT now makes animated GIFs from a prompt or a photo, with Images 2.6, for all users. For developers, GPT-Image-2.6 comes as Flare (efficient) and Sunburst (higher quality).

The through-line

Agents are in every pitch and barely in the numbers: four rides at Uber, rationed tokens at Apple, AI last for voters. The money is moving to where agents already work, which is code. One thing to do today: before you build an agent for a task, time how long the task takes by hand.

Sources covered

1 story · 99 sources (61 news, 34 company, 4 voices)

  1. HubSpot to ClickHouse: Sync Methods, Setup & API Limits Guide
    🏢Estuary↗🏢Estuary↗🎙️Ben Rogojan↗📰TechCrunch↗📰TechCrunch↗🏢Databricks↗📰LWN.net↗🏢Vercel AI↗🏢AWS AI/ML↗🏢Google Research↗📰The New Stack↗🏢Hugging Face↗📰TechCrunch↗📰The Verge↗📰Wired↗📰Engadget↗📰TechCrunch↗📰TechCrunch↗📰SiliconANGLE↗📰TechCrunch↗🏢Vercel AI↗📰SiliconANGLE↗📰SiliconANGLE↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢AWS AI/ML↗📰TechCrunch↗🏢AWS AI/ML↗🏢Fivetran↗📰SiliconANGLE↗📰TechCrunch↗📰Ars Technica↗🏢Confluent↗🏢Confluent↗🏢Confluent↗📰SiliconANGLE↗🏢MongoDB↗🎙️Simon Willison↗📰The Verge↗📰SiliconANGLE↗📰SiliconANGLE↗📰The Decoder↗📰RIA Novosti↗📰The Hackers News↗📰SiliconANGLE↗📰TechCrunch↗🏢OpenAI↗📰BleepingComputer↗📰The Decoder↗📰MIT Technology Review↗📰TechCrunch↗📰The Decoder↗🎙️Ben Lorica↗📰The Verge↗📰The New Stack↗📰The Verge↗📰The Verge↗📰The New Stack↗📰Hacker News↗🏢Neon↗📰The New Stack↗📰TechCrunch↗📰The Decoder↗📰The Hackers News↗📰TechCrunch↗📰SiliconANGLE↗📰The Verge↗📰BleepingComputer↗📰The Verge↗📰The Decoder↗📰Hacker News↗📰The Decoder↗📰The Verge↗📰Hacker News↗📰SiliconANGLE↗🏢Neon↗📰The Hackers News↗📰MIT Technology Review↗🏢OpenAI↗📰The Verge↗🏢OpenAI↗📰Hacker News↗🏢MLflow↗📰The Decoder↗📰Intuition↗📰The Decoder↗📰Hacker News↗🏢LangChain↗🏢ClickHouse↗🏢ClickHouse↗🎙️Simon Willison↗🏢Temporal↗🏢Elastic↗🏢Anthropic Research↗🏢Together AI↗🏢Together AI↗🏢vLLM↗🏢vLLM↗🏢OpenAI↗