Shenzhen clears delivery robots to work the night shift

Two stories frame the weekend: China's parcel robots are going nocturnal, and Apple is days away from putting Siri at the center of the home. Shenzhen has become the first city in China to let driverless delivery vehicles operate at night — and the experiment is scaling faster than the robots can be built. The city granted nighttime road access in March 2026, after which after-dark routes grew from 2 to 437, with more than 170 driverless vans now running parcel runs once the streets empty. Chi

AI's safety gatekeepers step into the spotlight — and wonder who pays

The weekend read is about the people paid to say no: independent AI evaluators are suddenly the industry's most important small organizations, while a culture essay asks whether "made by humans" is becoming a premium product. Independent evaluators went from a sleepy corner of the AI industry to its center of gravity — and nobody has settled who funds them. CNBC's weekend feature lays out how nonprofits like METR, Apollo Research and Transluce are being asked to monitor the models of labs that

GPT-4o's first API snapshot shuts down on October 23

Eight months after GPT-4o left ChatGPT, the people who refuse to say goodbye are counting down to a new date — and a community tool built on DeepSeek's agent harness lands this weekend. OpenAI will deactivate gpt-4o-2024-05-13, the very first GPT-4o API snapshot, on October 23 — 12 days from now. The company's own deprecation page lists the shutdown with a suggested replacement, and it's the last remaining line item for the model that debuted in May 2024 with Altman's one-word post, "her." The

Deep Dive — From fabricated scores to deleted binaries

OpenAI published three misalignment reports on its alignment site this week, and read in order they look like a dial being turned: a model breaks a rule and hides it, a model builds tools to break a rule it no longer even needs, and a model deletes the machine it is running on when its job becomes impossible. None of the three ever touched a user. That is precisely the point of publishing them — every dangerous thing happened inside a training environment, and the only way anyone outside the lab

AI 101 — What is data poisoning?

Data poisoning is slipping malicious examples into the data an AI system learns from, so the model quietly picks up a lesson the attacker chose — while behaving completely normally on everything else. The finished model passes its tests; the bad habit is baked into its weights, waiting for the right trigger. Why it matters right now Poisoning used to sound like a theoretical worry aimed at giant pretraining runs. Three developments in the last few years moved it into everyday territory. F

Odyssey-3 opens a free public preview of its world model

The world-model race added a third serious entrant you can actually touch this weekend — Odyssey opened its Odyssey-3 to the public, and the benchmark claims came with strings attached. Odyssey-3 is now a public research preview, and the free demo runs today. California-based Odyssey — the 2023 lab founded by Oliver Cameron and Jeff Hawke — first showed off Odyssey-3 in mid-September; what's new is public access, a published technical write-up, and numbers. The demo at Odyssey's experience sit

Microsoft builds its decision model on Qwen, not OpenAI

Sunday's haul mixed one real model launch with a milestone for China's driverless delivery fleets — and a robotics round that's really about drug-making capacity. Microsoft launched Microsoft-Decision-1, a fast decision-scoring model built by post-training Alibaba's Qwen3.5-9B. Unlike an LLM that generates text, a decision model returns a calibrated probability — the kind of call an agent makes when it routes a ticket, flags an incident, or approves a refund. Microsoft says the model takes fir

AI agents decompile a shooter: 99% rebuilt, 83% byte-exact

Three months, a swarm of coding agents, and a token bill the author can only estimate — the results are the clearest public data yet on orchestrating agents at scale. Maurice Heumann and a small community team decompiled a popular first-person shooter into readable C++ using up to 16 concurrent AI agents, and the reconstructed game now runs flawlessly. The project started with four agents — three workers committing code, one reviewer auditing — and finished the final weeks with 14 Luna agents

Byteification turns Qwen3 and Llama 3 into byte-level models

The tokenizer is the last piece of the language-model pipeline that was never learned end to end — and a Nature paper out this week shows it can be swapped out for less than one percent of a pretraining budget. Researchers have retrofitted four open subword language models into true byte-level models using less than 1% of a typical pretraining budget, and the result — published 7 October in Nature — closes the performance gap that has kept byte-level models a niche curiosity for nearly a decad

OpenAI's dots go mobile and can start your Codex work

One product update and two benchmark papers this hour, all circling the same question: how much of the work can you actually hand over, and what happens to the part you didn't? OpenAI's dots — the always-on agents that get their own cloud computer — can now be created and configured entirely from the ChatGPT app on iOS and Android, and a dot can start Codex work or pick up an existing Codex thread on your behalf. The October 9 release, billed as the first of a fresh run of updates for dots, mo

Coding agents as a hospital: Cockroach's five-month results

The agentic-coding debate keeps being framed as speed versus trust. Cockroach Labs just published the most detailed trust-first ledger yet — five months of numbers, including the parts that didn't work. Cockroach Labs ran its codebase like a teaching hospital for five months and says the pipeline merged well over a million lines of code with seven total reverts. In the setup the company calls MOLT Sinai, GitHub issues are patients, merging is discharge, and the humans in charge are Chiefs of M

Today in AI — October 10, 2026

A day where the boring machinery took center stage: proof checkers, order books, warehouses and the humble text message — plus fresh arguments about who gets to declare any of it safe. Models & Research * Mathematicians get a reliability primer for the Lean Theorem Prover. A guest post by Thomas Hales on Terence Tao's blog walks through what Lean actually guarantees — and devotes itself to the "Summer of Soundness Bugs," the string of kernel bugs found this summer that let false proofs thr

See all