Three threads run through today’s issue.
Agents are being handed real errands. ChatGPT Wallet lets OpenAI’s agents pay with a token instead of a card number, Cognition’s Devin now hosts private apps, and Sierra’s Poppy 1.0 is already switched on for every Shopify merchant.
Models keep promising more for less, mostly on their own tests. Thinking Machines reports a 20 percent faster reinforcement learning stage, Empero trained a full pipeline for about $4,200, and Qwen’s 62 percent cost cut comes from a benchmark it built.
Washington is tightening its grip on compute, with a new license rule for data centers that could serve Chinese firms, while a petition to bring Jensen Huang back to Nvidia passes 115,000 signatures.
Agents Get Errands: Wallets, Apps and Rules of Entry
Software is being trusted to pay, build and walk into a company’s front door, and each vendor is deciding what a person still has to approve.
- ChatGPT Wallet lets agents pay with tokens, not card numbers. OpenAI launched ChatGPT Wallet on November 13 with American Express, Mastercard and Visa, starting with ChatGPT Pro users in the US. The agent shows the items and total, the user approves with Face ID or a fingerprint, and a token replaces the real card number.
- Cognition lets Devin build and host private team apps. Devin Apps turn a plain description into an internal web tool hosted on Devin’s own cloud machine. Everyone who opens one must sign in to Devin and apps are never public, so each viewer needs a login; Cognition gives no hourly rate for the machine’s awake time.
- Sierra ships Poppy 1.0, a rulebook for personal agents visiting companies. Sierra’s Personal Agent Protocol reached 1.0 on November 13, and the customer, not the agent, decides which permissions to grant. Sierra says it is already on for every Shopify merchant and for Meta’s Muse, and its partner list has grown from eight names to more than 100.
- Anthropic lays out five building blocks for agents that work unattended. A guide by William Hu says always-on agents are mostly asleep, woken by events and timers. The design rests on an event inbox, timers, memory and permissions, among other parts, with credentials kept outside the agent’s sandbox.
Cheaper to Train, Cheaper to Run: Small Numbers, Sharp Trade-offs
Four teams show where compute can be saved, from four-bit math to a $4,200 rehearsal, and each says plainly what it gave up.
- Thinking Machines trains a large model with a 4-bit forward pass. The lab says Inkling 2 Small, pretrained in 68 hours, gained an 8 percent pretraining speedup and a 20 percent reinforcement learning speedup over an 8-bit recipe. Pretraining gives up roughly 2 percent of compute efficiency, and the results are the lab’s own.
- Empero trains a 165M-parameter model through a full pipeline for $4,200. Empero’s Quark took about 1,300 H100-hours and ships as four Apache-2.0 checkpoints. It is a rehearsal for a bigger run, and Empero lists what failed: web search never improved and long-input scores dropped during reinforcement learning.
- Sevren’s Embla 1 trades a few benchmark points for phone speed. Sevren says its 4-bit model answers in 1.77 seconds on an iPhone 18 Pro, while five of twelve rivals score higher but take at least twice as long. Part of the lead comes from Sevren’s own Metal runtime, which it discloses.
- Vision pretraining matters less than giving images their own compute. A controlled study finds models trained from scratch match pretrained vision encoders, while models with no dedicated visual parameters trail by 4.1 to 6.4 points on image questions. Putting vision-only layers early in the backbone cut time to first token to 37 milliseconds from 64.
Claims Under Review: Whose Numbers Are These?
A benchmark Qwen wrote itself, a physics proof awaiting experts, a safety score that swings 34 points and an essay on growth limits: each result comes with a source worth checking.
- Qwen4-Max claims a 62% task-cost cut on a benchmark Qwen built itself. Alibaba’s Qwen team says its 2.4-trillion-parameter model cuts the average task cost from $4.43 to $1.67 and scores 52.69 on Work360. Qwen designed and runs that benchmark, there are no independent results, and smaller open models are promised without dates.
- GPT-6 Astra helped build a candidate lattice definition of the Standard Model. Physicist Nathaniel Craig says the model led the detailed proof of a candidate definition of the Standard Model on a grid, and calls it more a beginning than an end. Outside experts have not ruled on it, and the account of the model’s role is his own.
- Anthropic’s playbook for safety evals: a refusal rate alone proves little. Anthropic published how it builds and reports safety tests, scoring whole conversations rather than single replies. For Claude Fable 5.1 the appropriate response rate differed by as much as 34 points between the raw API and claude.ai, where a system prompt applies.
- One writer argues AI cannot outrun a slow world’s 2% growth. Connor Blake, a former physics researcher, says bottlenecks in factories, shipping and regulation will cap AI’s economic effect, as past general-purpose technologies left growth near 2 percent a year. It is an argument built on a stylized model, not data on AI deployments.
Who Holds the Keys: Compute Rules and the Customer Record
One story is about who may rent out chips, the other about who owns the customer relationship once agents take over.
- US adds license rule for data centers that could serve Chinese firms. According to TechCrunch, Commerce’s interim rule follows compute rather than hardware and covers data centers able to supply more than 12 million units of processing performance. Advanced chips bound for Malaysia, Singapore, Indonesia, Thailand and Brazil now need licenses.
- Ex-Freshworks team raises $53M for an AI-first CRM. Dreamteam raised $53 million across a seed round led by Together Fund and a Series A led by Accel, and says it signed more than 100 customers in three months. Those are the company’s own claims, and Andreessen Horowitz has argued that incumbents will keep the customer record.
Today’s Quick Hits
- Musk and tech chiefs back petition to bring Huang back to Nvidia. Elon Musk, Replit’s Amjad Masad, Epic’s Tim Sweeney and Y Combinator’s Garry Tan signed a petition urging Nvidia’s board to reinstate Jensen Huang. TechCrunch reports it passed 115,000 signatures, a count that appears to come from the petition site itself.
- Claude Code’s project coordinator now replies to questions and reports status. The coordinator in Claude Code Projects now answers questions typed into the project chat and reports on each running thread. The update also adds forking, reminders and read-only sharing of threads, and Projects is in public beta on Pro and Max.
- Google Maps now plans your trip and hunts for open appointments. Google’s Ask Maps can now build multi-stop itineraries with live traffic and hours, and search salons, vets and repair shops for open appointments, calling businesses for you. It also expands to 30 European countries, TechCrunch reported.
- Nvidia’s BirdDreams simulates driving and traffic reactions from one photo. Nvidia and KAIST researchers built BirdDreams, a driving simulator that draws a top-down road map beside the video so surrounding cars react to the test car. Nvidia reports 35 percent lower error than an OmniDreams-based baseline in its own evaluations.