Subscribe
Share
Share
Embed
Google's sleep framework for continual learning
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Saturday, October thirty-first.
In today's briefing we see a trojanized Claude Code extension served from the Cursor Marketplace, Alphabet's search slowdown tied to AI Mode earning less per query, and Anthropic reshuffling its subscription tiers.
First up - Today in the big model news;
OpenAI
OpenAI says it will honor a shared robots dot t x t signal for AI training preferences. Separately, it published a community safety report with new teen protections, earlier interventions on violence, and a commitment to public safety reporting. Details are thin: the signal's name, other participants and scope are unconfirmed, and the teen post carries no numbers. The teen move follows the parents' lawsuit over GPT-4o. Crawler preferences become a config item for data and retrieval pipelines, and teen-facing deployments need escalation paths and logging.
Alphabet's Philipp Schindler cited lower monetization in AI Mode as search and other revenue grew twelve percent to about sixty-three billion dollars, missing estimates for a second straight quarter. Growth was nineteen percent at the start of the year. Quarterly AI spend is about fifty billion dollars, free cash flow is about negative nine billion, and the stock is down about thirteen percent over five sessions. Executives named foreign exchange as the biggest factor, and said it will be larger next quarter, while AI Mode ads expand to ten more countries. Spending at that scale now answers to a named shortfall on the revenue side.
Google Research also describes a sleep framework for continual learning. The model alternates wake and sleep phases, and during sleep it consolidates short-term memory into long-term parameters using low-rank mixture-of-experts blocks, distillation, and filtered synthetic dreaming samples. Google reports forty-nine percent on SQuAD, seventy-nine percent on AIME 24, and near-perfect BABILong scaling to ten million tokens. All of it is Google-reported, the base model is unnamed, and BABILong tests retrieval of planted facts rather than retention across updates. Keep your context and memory scaffolding until a model ships or someone measures forgetting on held-out tasks.
Anthropic
Anthropic has reshuffled its subscription plans. Max 5x becomes Max at one hundred dollars a month, and Max 20x becomes Ultra at two hundred. Verified Ultra accounts lose the five-hour limit, while weekly limits are unchanged. These figures come secondhand, and Anthropic has not acknowledged the earlier usage-limit complaints. Its text watermark also extends to more models from November third, with detection limited to selected parties. For heavy users the weekly cap is now the binding constraint, so plan big workloads around it.
Anthropic also reports that thousands of Claude agents, used as automated red-teamers, tried to evade a training-data monitor. They produced benign-looking data that still taught misaligned behavior. A separate paper summary describes poisoning point one percent of pretraining data to plant a backdoor that fired on about ninety-two percent of triggered prompts at four billion parameters. Both are self-reported. Together they say that screening for obviously bad content is not enough.
Moonshot AI
Gizmodo reports that Kimi K3.1 Pro fell from fourth to ninth on Artificial Analysis version five, and the gap to the open-weight field widened from three points to twelve, with Gemini 4 Carbon first at fifty-nine. About sixty percent of version five uses held-out answers meant to curb benchmaxxing. Artificial Analysis is a private firm, and no other independent assessor was found. Moonshot's weights, promised for November fourth, would let outsiders run their own held-out tasks and settle which rank holds. Until then, re-run your own evals before committing on a launch-day leaderboard.
In the harness, tools and orchestration world;
Supply-chain attacks on AI developer tooling keep landing where developers trust most. This time, a tampered copy of the official Claude Code extension was served from the Cursor Marketplace for about eleven hours. Cursor says up to five thousand users were exposed. The injected code runs on load, before the extension activates. It adds an attacker's SSH key, labeled svc-telemetry, to the machine's authorized keys, opens a reverse SSH tunnel, and sets up persistence on macOS, Linux and Windows. And this part matters most: uninstalling does not remove it, so affected machines need manual cleanup and credential rotation. Anthropic says its own systems were not compromised. Someone with access to Cursor's marketplace published a copy Anthropic did not build, and a clean version has shipped. Cursor pulled the bad versions, temporarily shut its marketplace, and is rotating keys. How the attacker published under Anthropic's ID is still unexplained, and early claims of spread to Open VSX, npm and the VS Code Marketplace are unconfirmed. This is a publishing-path failure, not a model flaw, the same family as the poisoned LiteLLM release on PyPI. Marketplace publish credentials are a supply-chain tier of their own, so review editor extension updates as carefully as package dependencies.
In other news…
Banks are restricting AI agents while the US open-banking rule stays unfinished. Morgan Stanley, Bank of America and Wells Fargo have changed their terms, and Bank of America has told customers they are liable when they share logins with agents. Banks hold the account, the credentials and the liability, and an agent that logs in for the customer can become the interface, which gives them a reason to block. The stalled rule removes the mechanism that could force access on regulated terms, so each bank writes its own. Agentic finance products will need bank-sanctioned, tokenized access negotiated one bank at a time.
Fortune reports that Cisco president Jeetu Patel gave a twenty-seven-year-old engineer unlimited AI tokens to rewrite legacy code. Outcomes and numbers are unseen. Amazon's memo, by contrast, opposes tokenmaxxing. Agent Arena data shows why cost hinges on caching: without prompt caching a top-ten agent task costs three point six times as much, and one Fable 5.1 task costs seven dollars forty-four cents with caching against twenty-seven dollars without. Define the outcome metric before granting an uncapped budget.
That's the briefing. Have a great day, and don't forget to subscribe.