In today’s newsletter: Anthropic’s IPO filing shows how much revenue came from customers later found to have broken its rules, and why privacy agreements can make misuse harder to investigate. Plus: Block’s Dessa learns directly from human conversation, robots learn from shared experience, and a world model helps them act better even when its predictions look worse.
This Week in Turing Post:
Wednesday / AI Builds AI: The Self-Building Model? Auditing AI’s Role in Recursive’s Flint
Friday / ∇ Guide: DeepSeek in the Open: Architecture, Infrastructure, and the Deployment Gap
Sunday / Library: Best Sources to Explore Harness
Today’s editorial – What Anthropic’s IPO Tells Us About AI Safety
Safety in AI is fake. That’s what Anthropic’s IPO tells us.
When I wrote about Anthropic in 2023, people who left OpenAI had founded it to put safety at the center of AI development. Three years later, Anthropic is preparing for what could be the biggest IPO in history. Again, I find myself looking at the people leaving.
Two people behind Claude Code, Thariq Shihipar and Anthony Morris, announced their departures just before the filing. Shihipar said he was taking a break “before diving into something new”; Morris said he would “share what's next soon”. It’s just so weird they did it right before IPO. Earlier, researcher Jacob Coxon left before his equity vested, warning about the speed of the AI race. And Joe Benton left Anthropic’s safety team to join an independent evaluator. Benton said competition pushes frontier companies to underinvest in safety because falling behind is too costly.
This brings us to Anthropic’s own choices. Its original Responsible Scaling Policy promised to pause training or deployment at certain capability thresholds if the necessary safeguards were missing. In February, the company replaced that commitment and added transparency measures. It argued that stopping alone while less cautious competitors continued could leave the world less safe.
I find this hypocritical: Anthropic built its identity around the dangers of an uncontrolled AI race. When it decided it could not be the only one slowing down, it rewrote its promise. Now it is asking the public to finance the next stage. There is nothing to stop it to rewriting any other promises it makes.
Anthropic’s revenue grew from $41 million in 2023 to $16.3 billion in the first half of 2026; these are insane numbers that can affect almost anyone’s judgement. Its second quarter delivered $974 million in operating profit. The filing also shows its dependence changing shape: its five largest customers brought in about 36% of revenue in the first quarter and 21% in the third. Sales through cloud marketplaces, meanwhile, grew from 11% of revenue in 2023 to 48% in the first half of 2026.
Anthropic leaves those five customers unnamed. Cursor is a plausible candidate: its co-founder called it Anthropic’s largest customer in 2025, and Semafor reported it was still among the biggest in June 2026. Microsoft and Meta are worth watching too. Microsoft had projected spending over $1 billion on its own Claude use in 2026, then reportedly cut that forecast by more than a third; Meta reportedly spent over $105 million on Claude Code in one 28-day period.
These relationships can get complicated pretty quickly. Cursor now belongs to SpaceX, whose AI arm competes with Anthropic and supplies it with computing power. With Elon, going from “Anthropic-misanthropic” to passionate love in one breath, doesn’t add any stability. Amazon, Google and Microsoft invest in Anthropic, sell it computing power, distribute Claude and build competing models. Plus, Anthropic has committed at least $110 billion to Amazon through 2036 and approximately $114.5 billion to Google through 2033, even if it uses less computing than expected. Its customers can reduce spending – but those bills will keep coming.
That pressure extends to the rules for customers. Anthropic estimates that 5–7% of its revenue in the first nine months of 2026 came from customers it later found violating its usage policy or commercial terms. Finding and investigating violations takes time, and the company earns revenue from those customers before restricting or terminating their access. Anthropic says continued enforcement could reduce future revenue and growth. With commitments this large, saying no to customers has a price.
The rules become even more consequential when the customer is the government. Anthropic refused the Pentagon’s demand to make Claude available for all lawful uses, maintaining restrictions on mass surveillance of Americans and lethal autonomous warfare. By then, it had already made exceptions for government uses it denied to private customers, including weapons design and offensive cyber operations.
Image Credit: Reuters
Anthropic is deciding which risks it is willing to accept, which customers get exceptions, and where to draw the line. As the company prepares to sell shares to the public, who gets to define the public good?
When I wrote about Anthropic in 2023, I looked at its ties to effective altruism and its ambition to do the most good. Then, as now, the question is: who decides what “good” means, and how?
The seven founders have pledged 80% of their equity to philanthropy, though that pledge is revocable. A special share gives them 50.1% of the votes on general shareholder matters. An independent trust elects a majority of the board, yet its special share can be removed over the trustees’ objections with the required votes from founders, employees and Class A shareholders. The board will also set its own standards for measuring public benefit.
The AI race is real, and Anthropic is one of the companies driving it. It has acknowledged the pressure from competitors and weakened its promise to pause when safeguards aren’t in place. The people who will live with the consequences need a way to intervene before a dangerous model is released. I don’t trust Anthropic to make that decision for the rest of us. Do you?
Also, watch our latest video on the world DeepSeek is building, and why their breakthrough proves AI really has become the new electricity →
Follow us on
Twitter Library – Exploring System One AI Models Beyond Jev: 15 Outstanding Models
News from the usual suspects ™
AI agents tracked drug development, speech models listened while speaking, and new tools made it easier to run AI on different chips. Such was the week.
OpenAI
OpenAI acquired Orchestra Bio, bringing its drug-development coordination platform together with Rosalind. Orchestra’s agents track studies, dependencies, and timelines, extending OpenAI’s life-sciences work into managing the process between scientific findings and potential treatments.
Google began rolling out Personal Intelligence on Android Auto, letting Gemini draw on connected apps such as Gmail, Calendar, Photos, and Maps for personalized assistance. Access is opt-in for eligible personal accounts; cars with Google built-in follow later.
Meta
Meta published the Muse Glimmer technical report, documenting its 30B model for local agents, four-bit quantization, tool use, and failure recovery.
Separately, its Prometheus engineering account explains how hundreds of thousands of GPUs work across interconnected buildings, with reported training goodput above 85% for pretraining and 90% for RL.
Microsoft
Microsoft explained how Project Perception can supervise enterprise agents, combining permissions, runtime signals, and action-level risk assessment.
Its Saint-Gobain case study reported assembling a knowledge base from 106,000 patents and abstracts in hours, then developing cement-additive hypotheses over six weeks. Candidate molecules were still undergoing testing.
Amazon
Amazon introduced TorchNeuron, an open-source native PyTorch backend for Trainium2 and Trainium3. Currently in beta, it supports eager execution, compilation, and distributed training, reducing the engineering work required to move existing PyTorch workloads onto Amazon’s chips.
NVIDIA
NVIDIA expanded its sovereign-AI framework with the European Broadcasting Union, adding cloud, open-model, and integration partners. The initiative addresses how broadcasters can deploy agents while retaining control over editorial data, infrastructure, and security. NVIDIA also highlighted accelerated Adobe Frame.io video search with producer UFA.
DeepSeek
DeepSeek released DeepInfer, its inference framework built originally around vLLM, under Apache 2.0, with NVIDIA and Huawei hardware support. The release exposes core inference functionality but still needs additional components for a complete serving system.
Separately, a DHS notice ordered DeepSeek products excluded from US federal civilian procurement and removed from agency systems. The order also covers third-party products and services developed using DeepSeek implementations.
MiniMax
MiniMax launched M3.1 Flash for coding, computer use, and agent workflows at $0.15 per million input tokens and $0.60 per million output tokens. MiniMax reports nearly threefold faster inference and 70% lower usage costs than its predecessor.
ElevenLabs
ElevenLabs launched Eleven Realtime Mini and Max, speech-to-speech models that combine recognition, conversational timing, reasoning, and voice generation. Both support more than 90 languages; Max targets demanding tool-use workflows. Developers can access them through ElevenAgents or an API compatible with OpenAI’s Realtime specifications over WebSocket.
fal / Black Forest Labs
fal and Black Forest Labs released FLUX 3 Max, combining the FLUX 3 video model with fal’s post-training and inference optimizations. fal reports generating a ten-second clip in about eleven seconds, versus roughly 126 seconds for the original model, alongside improved human-preference scores.
Unexpected, potentially usual suspects
Inception
Inception launched Mercury 3, a diffusion language model with a one-million-token context window and company-reported generation speed of 1,068 tokens per second. Standard pricing is $0.20 per million input tokens and $0.70 per million output tokens. Its speed makes it worth evaluating for interactive applications and agents with repeated model calls.
Block / Dessa, October 23
Block previewed Dessa, a speech-to-speech model that Block says was trained entirely on human conversation, without text or pretrained models. It can listen while speaking and reportedly runs locally on a MacBook Pro M4 Max. The research question is compelling: how much language, knowledge, and conversational behavior can a model learn directly from sound? Early access is limited to selected safety researchers.
Research highlight
🌟🌟 DRILL: Training world models to improve policies
Image Credit: The original paper
World models are often trained to predict the next frame. DRILL asks a different question: how useful are their imagined rollouts for improving a robot’s policy?
The authors train the world model around the policy’s downstream performance. Across five manipulation simulators and two policy backbones, their stronger method raises mean success by 13.1 percentage points over WMPO, even with higher next-frame prediction error. The result suggests that the best prediction is the one that helps an agent act. →read the paper
Models
🌟 LaLa-World – describes actions and visual changes in language so one latent world model can learn across robots, simulations, images, and videos. In a matched comparison, it uses 6× less training compute and plans 13.9× faster than a video-generation baseline. →read the paper
🌟 Ropedia W0 – learns across humans and robots from 10,000 hours of human video and 3,000 hours of robot data using shared camera and skeleton representations. Distillation enables real-time rollouts. →read the paper
RynnUnify – connects embodied understanding with visual planning in a 16B model, with extensions for testing imagined outcomes and controlling robots. →read the paper
🌟 ALPS – learns a visual representation that combines physical state, language alignment, and pixel reconstruction. Its frozen features reach 85.5% mean policy success on LIBERO. →read the paper
MemT-1.0 – builds recurrent memory into model computation. At 128K history tokens, its Nano model reports 1.86× faster end-to-end inference and 66× less persistent storage than the full-context baseline. →read the paper
WeVisOne-VL – combines detection, counting, OCR, segmentation, and matting in one vision-language model, spanning visual information from objects and boxes to pixel-level masks. →read the paper
ScaleNav – learns navigation dynamics from internet egocentric video and aligns them with robot actions in simulation, helping robots navigate toward a language-specified target visible at the start. →read the paper
Research
Trends we see in this week’s selection:
World models are being judged by the policies they improve as well as the futures they predict.
Human video and fleet experience are becoming training material for robot policies.
World representations are connecting visual appearance with geometry, motion, and executable interactions.
Agents are learning across diverse environments and tool configurations, while memory helps them retain evidence and track state.
Evaluation is moving beyond aggregate scores to test generalization, decision consistency, and whether verifiers accept valid solutions.
World representations, simulation, and planning
CADRE – shows how video compression choices affect motion fidelity and action-response delay in interactive world models, with selected configurations reaching 20 frames per second on one H100. →read the paper
Learning the 4D World through Point Trajectory Forecasting – predicts videos and persistent 3D point trajectories together, improving geometric motion prediction and manipulation in simulated and real-robot experiments. →read the paper
WorldMirror – uses a coding agent and verification feedback to reconstruct robot demonstrations as executable simulations that preserve interactions. →read the paper
Learning from physical experience
🌟 Learning from Fleet Experience – pools successful trajectories from deployed robots to improve a shared policy after expert-supervised initialization. With 96 robots contributing data, phone-packaging success rises from 79.2% to 98.3%. →read the paper
Self-Distilling Future Video Representations – trains future representations for robot actions in a single denoising step. With human-video co-training, it reaches 50% success on skills absent from robot demonstrations. →read the paper
FORCA – gives a coding agent visual observations and wrist force/torque measurements to write and revise control programs for contact-rich robot manipulation. →read the paper
Agent environments, tools, and training
🌟 Agent World Model in the Wild – generates executable environments for graphical interfaces, command lines, and MCP tools. Two models trained in these synthetic worlds improve across eight external benchmarks. →read the paper
Harness Scaling – trains agents across varied tool, context, and orchestration configurations, with gains transferring to unseen combinations and an external harness, Claude Code. →read the paper
OpenVLHarness – combines visual tools, persistent multimodal memory, and code execution, improving average scores by 9.7–12.6 points over matched no-tool baselines across 22 benchmarks. →read the paper
State tracking and autonomous discovery
PAR – teaches recurrent models to advance through intermediate states. A 0.5B model trained on sequences up to 32 moves reaches 95.5% full-state accuracy at 512 moves on a synthetic tracking task. →read the paper
Building AI Research Teams – organizes agents with different research backgrounds, private histories, shared evidence, and adaptive experiment allocation to discover a linear-attention architecture. →read the paper
Generalization, decisions, and verification
AnyGoal – tests whether world models can plan for new goals that combine familiar states and actions, exposing gaps between reproducing expert trajectories and generalizing dynamics. →read the paper
🌟 VerifierBench – tests generated agent verifiers on 222 tasks and 4,993 hidden cases. Many checks reject valid solutions, and self-tests can repeat the same mistaken assumptions. →read the paper
DecisionPivot – evaluates decision accuracy, consistency, probability ordering, and calibration across 48 tasks and 54 models, revealing differences that aggregate scores can hide. →read the paper
StreamEval – evaluates interactive video generation across 300 cases, separating instruction execution, state consistency, visual fidelity, and audiovisual coordination. →read the paper
That’s all for today. Thank you for reading! Please share this newsletter with colleagues who want to understand how AI is developing.
⬅️ FOD#172: Better AI Agents Without a New Model
