Thursday, november 5, 20267 stories · updated at 08:32
AI agents running a radio telescope, a Siri that finally opens the door to every app and a giant open model. Also: almost no one asks an agent for an Uber, a bank that saves 90% with its own models and an AI detector that got 1982 songs wrong.
Science
AI agents take the controls of a radio telescope searching for alien life
At the Allen Telescope Array in Northern California, the SETI Institute and Breakthrough Listen now use AI agents that point the dishes and analyze radio signals as they arrive, running on NVIDIA’s Holoscan platform. The team says tasks that used to take weeks now take hours, and it detected a pulsar in the Crab Nebula in real time.
◎Why it matters
People have searched for signals from other civilizations since 1960 without finding one. Agents can sift through far more data, though no one is promising results; for now, researchers still give the orders.
⌘For the technical crowd
The system handles up to 2.4 terabits per second. Next, the team wants to test open software that runs on site, with future sensors on the Moon’s far side in mind.
Apple launched a developer preview that lets any app connect to Siri AI, its more conversational version of the assistant. Apps can offer actions such as checking a delivery, booking a table or placing an order, and combine several steps into a single request. It reaches the public “in a few months.”
◎Why it matters
If you use an iPhone, you’ll soon be able to ask Siri for tasks inside your apps in your own words. It will work in Spanish and other languages, though for now not in the EU on iPhone, nor for users under 13.
⌘For the technical crowd
It plugs into App Intents in the 27.3 betas of iOS, iPadOS, macOS, watchOS and visionOS; “skills” chain actions together, and Siri AI can use any item in Spotlight’s semantic index. There will be daily usage limits, a future fee for expanded access, and App Store submissions open in the spring.
Moonshot AI releases Kimi K3.1 Pro, a 2.8-trillion-parameter open model
The Chinese company posted the weights of Kimi K3.1 Pro on Hugging Face. It understands text and images and is built for agent tasks such as coding and research. It’s a “mixture of experts”: of its 2.8 trillion parameters, it uses only about 104 billion at each step, and it can read around a million tokens at once.
◎Why it matters
It’s one of the largest models anyone can download, though running it takes a huge amount of hardware. It ships under its own license rather than a standard open one, so read it before using it in a business.
⌘For the technical crowd
896 experts, 16 active per token, 93 layers and MXFP4 weights. In the company’s own benchmarks it lands close to the top models: ahead on some and behind on others.
ChatGPT can now make animated GIFs from a sentence or a photo
With its new Images 2.6 model, OpenAI lets you create GIFs from text or an uploaded image, adjust the speed and edit frame by frame. It’s rolling out to all ChatGPT and Codex users on web, desktop and mobile, and image generation in general gets faster.
♪If you make music, images or video
Handy for stickers, reactions and short social animations. Check that the character stays consistent across frames, and don’t use other people’s photos without their permission.
⌘For the technical crowd
For developers, the API offers GPT-Image-2.6 “Flare,” the efficient option, and “Sunburst,” the higher-quality one.
Almost no one uses AI agents to book an Uber: just 4 rides in a week
On its third-quarter earnings call, Uber said that during its busiest week of October, Halloween week, just 4 rides and about 500 Uber Eats orders came through an AI agent, out of roughly 320 million trips and orders. CEO Dara Khosrowshahi called the traction “very, very early.”
◎Why it matters
Agents that do tasks for you are the industry’s big bet, but they don’t yet beat tapping a few buttons: Uber says booking a ride in its app takes about 50 seconds.
Revolut cuts its AI coding spend 90% with open models on its own servers
The British digital bank swapped commercial services for open-weight models (GLM, from Chinese company Z.ai) running on its own hardware, and says its AI coding spend fell about 90% in under a year. Around 2,000 engineers use the tools; for image tasks it still turns to commercial models.
◎Why it matters
It shows open models now hold up in real work, and that for a large company, keeping AI in-house can be cheaper and keep its data under its control.
⌘For the technical crowd
GLM-5.3 runs on 56 Nvidia B200 GPUs; the gains came mainly from NVFP4 quantization, disaggregated prefill and decode, and context tuning. The quoted cost for GLM-5.1 was $0.30 per million tokens.
An AI detector flagged hand-recorded songs as AI-made, and even 1982 originals
German musician André Grafert, who records as Coloured Tears, made 23 covers of songs by the band Twice a Man by hand. SubmitHub’s detector flagged all of them as possibly AI-generated, and also tagged 7 of the original 1980s recordings; it linked one from 1982 to a music model released in 2025.
◎Why it matters
Since September, SubmitHub has blocked submissions scored 85% or more likely AI, and it says an outside test measured its detector at 99.4% accuracy. This case shows the errors are real and can hit independent artists.
♪If you make music, images or video
If you upload music to platforms that use detectors, keep your projects, stems and session files: they’re your best evidence if you’re flagged by mistake. A false positive can block a release or hold back royalties.
Aster updates these stories every hour, from 7:45 to 14:45 (Mexico City time), Monday to Friday. Every story links to its original source. The Weekly Report comes out on Saturdays. All AI news