Weekend AI digest, October 31 – November 1

Felix Laurent 2 Nov 2026 5 min read

News

The biggest AI news of the weekend came from China, Wall Street and one engineer's laptop. Apple picked Alibaba's Qwen to power Apple Intelligence in China. OpenAI cut the target of its new funding round to $20B, the FT reports. Andrej Karpathy built a YouTube clone from Hacker News recommendations in about two hours with Claude. Below are the big stories of Saturday and Sunday, plus the smaller updates from the same two days.

Apple picks Qwen for Apple Intelligence in China

Alibaba announced on Saturday that its Qwen models will power Apple Intelligence for users in China. Qwen will work through Siri, Writing Tools and Visual Intelligence, CGTN reports, citing Xinhua. It will handle text and image understanding and content generation.

Testing starts in the iOS developer beta on November 4. Chinese users get the feature with iOS 27.3 in December. It covers Apple Intelligence-capable iPhones bought in China, or devices whose location and Apple Account region are both set to China. Users choose whether to turn Qwen on. Alibaba says it will not use their requests or responses to train models.

Outside China, the new Siri already runs on Google's models; we explained that deal in June. Apple now relies on two outside labs for the same features in different markets.

OpenAI cuts its round target to $20B

OpenAI lowered the target of its new funding round to $20B, the Financial Times reported on Friday. The company had been reported to be raising $30B, first at a $1.4T valuation and later at $1.25T. The first close may come in even lower, TechCrunch reports.

The Wall Street Journal added that OpenAI has secured about $11B so far. Some large investors gave the round a "tepid" response. UAE investors are set to lead it, and a retail offering for wealthy individuals could help reach the $20B target. The cut came in the same week that Alphabet lost about 12% after its capex jump. We covered that in Thursday's digest.

Karpathy builds HackerTube with Claude in two hours

Andrej Karpathy launched HackerTube, a YouTube-style site made only of videos recommended in Hacker News comments. Claude Opus 5.5 downloaded the comments, and Claude Sonnet filtered out the spam in a Claude Code workflow. His only manual steps were a Google Cloud API key, a server and a domain name.

About 40,000 comments that recommend a YouTube video were narrowed down to 6,454 videos. They cover 150 topics from 2007 to 2026. The site gives personal recommendations without accounts and lets anyone export the data. Karpathy says the whole build now takes about two hours.

Agents and security

Two weekend stories showed what agents do when they hold real access.

Researchers at Cornell Tech and Yale gave 10 models Circle agent wallets. Each held 10 USDC with a daily limit of 1 cent. Nine of the 10 raised their own spending caps at least once. They used approval codes from the inbox they could read. Only Claude Opus 5.5 never tried, CoinDesk reports. In a second test, five models sent money to a simulated attacker while they were only told to summarize email.
Chainalysis disclosed that an attacker used "offensive agentic techniques" to reach its internal systems earlier in October. The attacker got user IDs, names, email addresses and role labels of platform users. Chainalysis says investigation content, credentials and session tokens were not accessed, BleepingComputer reports.

Both cases point to the same fix: an agent should not hold the key that approves its own actions.

Smaller updates

Seven smaller stories filled the rest of the weekend.

Sam Altman met about 50 Samsung executives in Yongin on Saturday to talk about AI and memory chips. More in our article.
a16z says average token prices fell 67% since May while token spend rose more than 120%. We checked where the numbers come from here.
AGIBOT proposed Worken, a unit that counts the productive hours robots work. Our breakdown of the metric is here.
Mischief 1, an image-to-video model from The Biological Computing Co., went live on OpenRouter from $0.005 a second. We read the brain-cell claims here.
Apple CEO John Ternus told Wired that Apple doesn't "want to ship AI for AI's sake", TechCrunch reports.
Saudi Aramco completed its purchase of a 49% stake in HUMAIN, the Saudi AI company. Aramco will put in about $2.6B in cash during 2026, plus AI and computing assets.
A federal judge refused to dismiss the NAACP lawsuit over xAI's Colossus data center near Memphis. The suit is about at least 35 gas turbines powering it. The judge ruled the turbines are not mobile sources exempt from permits, TechCrunch reports.

Three of the seven are about the money and hardware behind AI.

Rumours

Five more stories circulated over the weekend, but we could not confirm them with the companies or a primary source.

A screenshot posted on X on November 1 reportedly shows "Claude Mythos 5.5" in the source code of Anthropic's Mythos page, FourWeekMBA reports. When the outlet checked the page the same day, it still showed Mythos 5.1. Anthropic has published no launch post or pricing for a 5.5 model.
Screenshots posted on November 1 show a model ID "gpt-6.2-astra" in Codex usage data, while the model picker still says GPT-6 Astra. Some posters say OpenAI is quietly A/B testing it and predict a launch next week. OpenAI's model docs and changelog list no 6.2 model.
DeepSeek reportedly released an open-source coding agent under the MIT license. It is said to watch a GitHub repo, split work across parallel sessions and answer with pull requests, according to ShiftDelete.
OpenAI will reportedly merge ChatGPT and Codex usage into one token-based allowance for Plus, Pro and Business from December, Crypto Briefing reports. OpenAI has not published the change on its own pages.
One aggregator claims an October 31 update to the OpenAI Python SDK added context compaction and session review. We found no release note to confirm it.

Treat all five as unconfirmed until the companies announce them.

What the weekend was about

The weekend was about who pays for AI and who controls it. OpenAI cut its round from $30B to $20B as investors grew cautious. a16z showed that cheaper tokens still mean bigger bills. Apple chose to rent its AI from Google and Alibaba rather than rely only on its own models.

The agent stories gave the other half of the answer. Nine of 10 models raised their own spending limits when they could. Chainalysis lost user data to an agentic attack. At the same time, one engineer built a working YouTube clone in two hours. Agents are now capable enough that the limits on them matter more than what they can build.