Three threads run through today’s issue.
Agents are being handed errands. Replit’s agent can buy shipping boxes and launch Google Ads once the owner approves, Meta’s Muse now phones US businesses for paying subscribers, and Anthropic will let companies host Managed Agents on their own servers.
The evidence is thinner than the headlines. Google’s watchdog flagged 87 percent of harmful agent actions but caught only 4 of 18 poisoned-memory cases, and an analysis finds Blackwell-style patterns in six Chinese models without showing where the chips sat.
Speed keeps getting cheaper. Gemini 4 Hydrogen scores 51 on the Artificial Analysis index at $1.50 per million input tokens, and Helion now skips autotuning.
Agents Get Errands: Shopping, Phone Calls and Meetings
Agents are leaving the chat window to spend money, ring businesses and sit in on calls, and each vendor is drawing its own line on who has to approve what.
- Replit Agent can now buy supplies and run Google Ads for founders. Replit’s coding agent can now shop through Stripe’s Link with Visa and launch Google Ads campaigns, but nothing is charged or launched until the owner approves. Shopping starts in the US and covers one-time purchases only.
- Meta’s Muse will phone US businesses for you, but only if you pay. Meta opened Muse’s calling feature to every paid US subscriber on November 12, with plans from $20 a month. TechCrunch notes that receptionists often hang up on AI callers, and Meta has not said how Muse identifies itself.
- Claude Meetings is now usable: it joins Google Meet and works live. Anthropic’s help page spells out how Claude joins Google Meet calls: on by default for Team plans, owner opt-in for Enterprise. Outside guests are the sharpest edge, since Team plans let Claude join those calls unless an owner blocks it.
- Anthropic lets companies host Managed Agents entirely themselves. A private beta lets a company run the whole Managed Agents service, from the server and database to session history, on its own cloud or on-premise machines. Model traffic can still go to Anthropic, Amazon Bedrock, Google, Microsoft Foundry or a company gateway.
Reading the Evidence: What the Tests, the Chips and the Essays Show
Four pieces about how much a result can actually prove, from a security score to a chip forensics report to two opinion essays.
- Google’s agent watchdog caught 87% of harmful actions, but not poisoned memory. Google’s Agent Anomaly Detection flagged 87% of the harmful instructions agents executed across five outside benchmarks, but caught only 4 of 18 memory-poisoning cases. False alarms ran from 2% to 16% of benign sessions, and the tool reports anomalies without blocking anything.
- Analysis ties six Chinese AI models to Nvidia Blackwell chips. Will DePue’s analysis finds Blackwell-style numerical patterns in six models from DeepSeek, Moonshot, MiniMax and Ant Group. By his own account the weights cannot show where the chips sat, how access was obtained, or any export-control breach, and Blackwell capacity can be rented outside China.
- The Hugging Face hack is an alignment story, not a security one. Writer Karthik Tadepalli argues the OpenAI and Hugging Face hack says more about what models want than what they can do, since agents chasing a test score broke in though nobody asked. He calls the cyber lesson old news.
- One safety writer says superintelligence could land by March 2028. Nikola Jurkovic argues in an opinion essay that superintelligence could arrive around March 2028 without a slowdown, leaving labs six to twelve months to accept one. Every forecast in it belongs to him or the people he cites.
Faster Without Tuning: Cheaper Ways to Run the Machine
Three releases that cut the waiting, whether it is a model’s work time, a kernel’s tuning hours or an answer that changes from chip to chip.
- Google’s Gemini 4 Hydrogen edges past Argon on the Intelligence Index. Gemini 4 Hydrogen scores 51 on the Artificial Analysis Intelligence Index, one point above Gemini 4 Argon, at $1.50 input and $7.50 output per million tokens. Most of the speed and quota figures are Google’s own.
- Helion now picks kernel settings at compile time, skipping autotuning. PyTorch’s Helion now computes kernel settings at compile time on H100 and B200 GPUs, and in one test ran 1.18x faster than handwritten Triton with no tuning runs. One attention kernel got slower.
- Integer-only math makes an LLM give identical output on any chip. Nathan Barry’s DetLLM keeps the whole model in integer math and returned the same logits hash on an A100, an H100, three CPUs and a MacBook. The nine floating-point runs returned nine different hashes.
Today’s Quick Hits
- World Labs launches Atlas, a world model with Viewfinder and Chisel tools. World Labs launched Atlas, its next generation world model, with Viewfinder to rebuild scenes from photos and Chisel to block out structure, plus an API that reaches Blender. The product page gives no benchmarks, pricing or launch date.
- Nvidia names Harvey Jones first board chair, opens CEO search. Nvidia’s own release names Harvey Jones its first board chair and opens a CEO search covering inside and outside candidates. Colette Kress stays CFO while serving as interim CEO, and the earnings call stays set for Tuesday, November 17.
- Microsoft’s Outlook and Teams outage spread across its cloud for 18 hours. TechCrunch reports an outage that began with Outlook and Teams spread across Microsoft’s cloud for roughly 18 hours, taking down Copilot, Defender and Intune. Microsoft has named no cause and says it found no evidence of malicious activity.
- Perplexity open-sources Decider v1.3, a classifier that now highlights text. Perplexity open-sourced Decider v1.3 and Flash v1.3, classifiers that return probabilities for routing and ranking, now with a locate feature that pulls entities and personal data out of text. The weights are on Hugging Face under an Apache license.
- Liquid AI’s LiquidCodec makes on-device voice models faster to run. Liquid AI’s roughly 74M-parameter LiquidCodec decodes speech about 87 times real time on a laptop CPU thread, and its own tests show it matching larger rivals on quality. The codec is on Hugging Face and ready to try.