Jensen Huang spy plot 🕵️, Gemini writes, Claude judges 🧑‍⚖️, What really slows agents 🏎️

US officials concluded that Alice Huang, a top executive at Singapore data centre firm Megaspeed, was a Chinese intelligence asset with a close relationship to Nvidia's chief, The New York Times reported on 29 October. President Trump was briefed roughly a year ago and chose not to act, people briefed on the surveillance said. Nvidia says Megaspeed, which rents computing power to Alibaba, gets ordinary business information, not trade secrets.

Hugging Face's 29 October post puts the weights in FP8 and NVFP4 checkpoints under the Mistral Ecosystem Open-Weight License, an MIT derivative that limits offering the model as a service. The model runs about 49 billion of its parameters per token and holds a 1M-token context, with day-0 support in transformers, vLLM and llama.cpp. Its 2-bit quants run on a single 512 GB Apple Silicon Ultra machine; other quants need distributed inference.

Reddit said on 29 October it will become primarily logged-in, testing outside the US first, and warned daily users could fall by a low-double-digit percentage in Q4 if search referral trends persist. CEO Steve Huffman said in an interview the change aims to improve the product, lift ads and stop bots scraping its data, while content reached through search or shared links stays visible. It says it will keep licensing content to third parties.

OpenAI's developer docs say owners of eligible ChatGPT plans can link them to an API organisation they own: each Pro tier ($100, $200, $500) earns its own price in credits, and Business earns $20 or $100 per paid seat, capped at $500 per workspace. Unused credits expire at the end of each billing cycle. They don't cover Codex through API keys, and with no credits left, requests fail with HTTP 429.

Prime Minister Mark Carney announced the federal investment at Montréal's Mila on 29 October; Cohere says the round, to be led by Germany's Schwarz Group and mostly secured, would value it at nearly $36 billion post-money. AI minister Evan Solomon said the money is staged against milestones, with jobs and revenue targets key. Cohere plans to spend on computing and on models for enterprises and governments, including $1.7 billion in Canadian computing infrastructure by 2030.

Spawn, which builds a multiplayer game creation engine, runs permanent named agents that each own part of the company, and its founder says they made 77,000 substantial changes in 30 days. Agents push to production hourly without pull requests, and 0.9% of changes were reverted over a year. He says only Anthropic's Fable makes this work, and AI judges built from his past decisions now settle most product questions.

In Epoch AI's tests, GLM 5.3, an open-weight model from Z.ai, turned Firefox security fixes that were public before users received them into webpages that stole visitors' files, on all six releases from 17 September to 26 October. Each new version cost about $1,000 to break, no provider refused, and Firefox 159.0 now patches one flaw the attacks relied on. Epoch expects more software to be exposed as open models improve.

ClickHouse said on 29 October that GroeneAI picks its own work from CI failures, fuzzer findings and user reports, then fixes, tests and opens pull requests with no human in the loop until review. It uses Claude for its main sessions and Codex for separate reviews, and stores reviewer lessons in persistent memory. Among decided pull requests, 86.5% merged, and its agent runtime, Nerve, is open source.

Built in one long Claude Code session, the pipeline has Gemini 4 Argon write every word from his voice notes. Claude judges each draft against a voice guide drawn from about 44,000 words of his published writing and passes back only the top three fixes. The judge also predicts his edits in advance, and that hit rate, 38% in the latest batch, tells him when he can step out.

In a write-up on agent performance, Alderson estimates that halving inference time takes only about 20% off a task when tool calls, provisioning and suspend-and-resume stay the same. He traces slowness to process-heavy shell commands, bloated shell start-up files, networked disks, sleep-and-poll loops and broken IPv6 in sandboxes. He also reads open-weight models' reasoning traces to learn which tools an agent wishes it had.

Cramer says most of the predictability in Junior, Sentry's open-source Slack agent and now how he ships software, came from a man-in-the-middle egress proxy. When the agent opens a pull request through the GitHub command line, the proxy rejects it and points to Junior's own tool, which adds the required footer. He argues skills files and system prompts almost never enforce behaviour an agent must show every time.

Armin Ronacher, building a game narrator with his kids, explains that Qwen3-TTS stores a voice as 2,048 numbers that can be averaged into new voices. Qwen documents no way to direct a cloned voice, but passing its numbers to the CustomVoice model makes instructions like "speak sad" work, less reliably than on built-in voices. He also got a halfway convincing clone from 28 seconds of a conference talk.

Sam Altman told Cheeky Pint he now hopes for many AIs keeping each other in check, as people do in a democracy. He said OpenAI knew nothing about how teams of agents scale when it ran what he called its Navier-Stokes proof, and deliberately let that maths spending get "out of control" for a week. He also wants the five-year cost of a gigawatt of compute to fall from roughly $70 billion to $20 billion.

Bodnar told Search Engine Journal much of that traffic was never going to turn into revenue, which grew nearly 20% in 2025 because HubSpot had spent years building channels that don't depend on search. Qualified leads from AI answers grew 1,850% from Q1 2025 to Q1 2026 and convert about three times better than search, he said, calling the signal correlative. He now measures revenue per visit and AI share of voice instead of sessions.

A Google DeepMind essay argues that AI skill gains shift wages and prices gradually, but agents coordinating past a "supercooperative" threshold would trigger abrupt jumps in market concentration, employment and wages. In its quiz of 100 people, complementary pairs beat random pairs by 32%, while copies of one model add capacity without that breadth. The authors want benchmarks for organisations of agents, saying solo leaderboards overstate how models compete with human teams.

Laycock says developers who resist handing coding agents all the execution know that built isn't the same as working in production. In a week-long Thoughtworks experiment where agents wrote all the code for an airline app, the team spent about two-thirds of its time at the whiteboard, and the build was often failing. She asks leaders to give cautious developers room to experiment and to stop judging success by agent code output.

Google's Gemini API CLI is a zero-dependency command-line client for AI agents with shell access, also usable from scripts and CI, working with Claude Code, Codex, Cursor or any shell. It generates text and media, analyses PDFs, recordings and YouTube URLs, and hands research to background Gemini managed agents, writing media to disk and printing only the file path. It suits teams calling Gemini from agents without adding SDK code.