Agents Miss the Ride
Agents Miss the Ride
Four. That is how many Uber rides were booked by outside AI agents in the whole third quarter of 2026, according to figures Uber executives shared on their earnings call, the first time the company has disclosed this number at all. Uber Eats did a little better, with around 500 agent-placed orders in the week of Halloween. The combined total for rides and deliveries was 320 million. The fear in Silicon Valley has a long name, disintermediation, which just means customers asking an AI assistant like Meta's Muse or OpenAI's Dots to book things so they never open the Uber app. CEO Dara Khosrowshahi said "We hope that there will be more traction," while making clear Uber thinks its own app will stay the nicer experience. Uber also has every reason to keep that door narrow, and it is building its own bots to keep you inside, like a Cart Assistant that turns a recipe photo into a grocery list. The Verge: Uber only fulfilled 4 rides from AI agents last quarter. (You can count them on one hand and still hold your phone.)
+ Same week, same lesson from a bigger company: people are using AI more than anyone has figured out how to make money from it. Google's version of the problem is in section 3. The Verge: Google hasn’t figured out how to make AI pay like its search engine. (The agents are coming. Slowly. By taxi, apparently not.)
If you work online: If a client thinks AI agents will soon handle bookings and errands, Uber logged four agent rides in a quarter. That VA work stays human for now.
GIF Them What They Want
OpenAI has given ChatGPT the one skill every Facebook page admin quietly wanted: animated GIFs, made from a typed prompt or a photo you upload. It arrives with Images 2.6, OpenAI's new image model, and lets you change the playback speed, edit individual frames and share the result, which the chatbot could not do in one step before. OpenAI's demo was a dancing cat. Of course it was. OpenAI also says regular image generation is now faster, and the update is rolling out to every ChatGPT and Codex user on web, desktop and mobile, so this is not locked behind a premium tier. For a social media VA or an online seller, that means turning a product photo into a looping post without a separate editor or a separate subscription. It also means the client who used to pay extra for an animated version will expect it within the hour, at the same rate. The human edge is diskarte: knowing which GIF actually sells. The Verge: ChatGPT will now make you animated GIFs. (The cat will be fine. Your rate card needs a look.)
+ The scale behind it: OpenAI says people now generate over 4 billion images a week across ChatGPT and its API, up from 3 billion in September. It is also launching two API models for developers, GPT-Image-2.6 Flare and Sunburst, pitched as faster and higher quality. The Verge: ChatGPT will now make you animated GIFs. (That is a lot of dancing cats.)
If you work online: Test ChatGPT GIFs on one product photo today. Clients will soon expect animated posts at no extra charge, so price your judgment, not the file.
Search Me
Here is the strange part. People are searching Google more than at any point in five years, and Google is making money from it more slowly. Alphabet says Search revenue grew 12 percent year over year in the third quarter, down from roughly 17 percent, while growth in paid clicks, the ad clicks that actually pay Google, fell from 13 percent to 8. Business chief Philipp Schindler blamed currency swings first, but said AI Mode, the chat-style version of search, earns less per query than classic results. All of this while Alphabet plans to spend about $200 billion this year on infrastructure, largely for AI, and its shares have fallen about 14 percent since the report. Google's answer is more ads inside AI Overviews and AI Mode, including outside the US, with ads in the Gemini app floated but not scheduled. An eMarketer analyst summed it up: "No one has it all figured out." The Verge: Google hasn’t figured out how to make AI pay like its search engine. (Two hundred billion dollars buys a very expensive shrug.)
+ The rival's numbers are not better. An October working paper found US visits from ChatGPT ads brought retailers 64 percent less revenue per session than Google Search ads, $0.89 versus $2.45. Tinuiti's Andy Taylor cautions that this may reflect intent, since someone asking a chatbot about running shoes is often still researching, not buying. The Verge: Google hasn’t figured out how to make AI pay like its search engine. (Chatting about shoes is not the same as buying them.)
+ For anyone running ads: Tinuiti says its advertisers raised Google search spend 14 percent and kept hitting their return targets, but Google does not let them target AI ad placements separately or see how those perform on their own. The Verge: Google hasn’t figured out how to make AI pay like its search engine. (Trust the total, Google says. Advertisers have heard that before.)
If you work online: If you run Google Ads for a seller, AI placements can't be tracked on their own yet, so judge by total return and tell your client that plainly.
Hands Off the Mouse
Honest verdict first: this is a developer tool, not something for your morning. Vercel Labs' agent-desktop lets an AI agent see and click through desktop apps on Mac, Windows and Linux by reading the accessibility tree, the hidden map of buttons and menus that screen readers use, instead of squinting at screenshots. It has no AI inside. You bring the agent, it supplies the hands, and setup happens on the command line. The README lists no price. It drew 2,463 GitHub stars in a day, and its own docs warn that snapshots can capture private desktop content. One to watch. GitHub Trending: vercel-labs/agent-desktop. (Your cursor just got a coworker.)
If you work online: If a client wants AI that operates their desktop apps, this is what developers are testing. It needs setup, so quote it as dev work, not VA work.
Extra, Extra
+ Screen Time Champions: Rappler's Decoded newsletter notes that Filipinos lead the world in social media and online video time, averaging 36 hours and 12 minutes a week according to the Digital 2027 report from Meltwater and We Are Social, nearly double the global figure. It also flags growing worry about child safety online. Rappler: [[DECODED] AI is changing our lives more than we realize](https://www.rappler.com/voices/newsletters/decoded-artificial-intelligence-change-filipinos/). (A full-time job, unpaid, with ads.)
+ Airline Booking Data Exposed: Singapore Airlines says about 46,000 booking records, mainly names, booking references and itineraries, may have been affected by a security incident at its reservation provider Amadeus, joining Cathay Pacific, Japan Airlines and Asiana in reporting exposed customer data. Amadeus says it found a vulnerability and that no card or bank details leaked, and the airline says its own systems were not breached; if anyone contacts you about changing a booking, verify directly with the airline. Rappler: Singapore Airlines says 46,000 booking records affected by Amadeus breach.
+ Spot the Fake: Pangram has released Pangram Image 1.0, which it claims is state-of-the-art at detecting AI-made images and video, plus a smaller Image Flash 1.0 for video. That is the company's own claim, and the coverage we have shows no independent test. Techmeme: Pangram releases Pangram Image 1.0. (Detectors and generators, still chasing each other's tails.)
+ Parts and Parcel: A Show HN project grades a prebuilt PC's asking price by estimating what each part would cost: paste a store link or a photo of the spec sheet. It is free with no account, but it only covers US listings and is still in beta. pcdealcheck.com: Show HN: I built a tool that prices a prebuilt PC part by part. (Shopee version, please.)
+ Bug Off: Simon Willison shipped datasette-llm 0.1a10, fixing a bug where his agent.datasette.io demo offered AI models beyond the ones it was configured to allow. Developer housekeeping, but a fair reminder to check that only these models settings actually work. Simon Willison: datasette-llm 0.1a10. (Settings are suggestions until someone tests them.)