My AI Agents Can Do More. Why Am I So Anxious?

In today’s newsletter: Let’s talk about high-functioning anxiety and that it might be the real problem of us working with AI. plus to fuel more anxiety more amazing models and outstanding papers you should try/read ;)

Featured

Stop shipping agents that repeat mistakes

With Zep Agent Skills, DeepSeek V4.1 Flash passed more EnterpriseOps-Gym tasks than Claude Opus 5.5 at maximum effort.

With no fine-tuning.

Zep Agent Skills turns your agents' traces into step-by-step skills they read at runtime, over MCP or the Zep SDKs.

Free on every Zep plan until January 4.

· Sponsored by Zep

This Week in Turing Post:

• Wednesday, November 11: Freestyling from AIE Code in SF

• Friday, November 13 / ∇ Guide: AI Routing: Who Answers Your Question, and How Can You Know?

• Sunday, November 15 / Twitter Library: The Race to Real-Time Video – 10+ Forcing Methods for Faster, Longer Video Generation

Today’s editorial

This is an unusual editorial, and not only because I wrote it by hand, in a paper notebook, on a plane (I consulted AI on a few things later). It’s unusual because it’s personal and quite emotional: it’s about the anxiety that has knocked me over these last few weeks.

Every week a new model arrives that can do more, better, faster and cheaper. I no longer believe AI is “normal technology,” and not because I’m falling for “AGI is here” or “the Terminators will kill us.” I think something very different is happening: some of the most efficient, talented and productive people I know are burning out, because there is just too much you can do, right here and all at once.

If you have ever experienced high-functioning anxiety, you know how visceral it is. It messes you up mentally and physically. It doesn’t necessarily stop you from working or wanting things the way depression can. Instead, it turns you into a maniac who needs to achieve more and keep everything under control – or it simply terrifies you while leaving you entirely capable. And then it exhausts you so much that you want to quit.

High-functioning anxiety is uniquely exhausting because you’re running two high-energy programs at once: the visible engine delivering output, and the invisible suppressor burning through glucose and mental bandwidth just to mask the panic and stay steady.

What I’m trying to say is that I’m less concerned about AI taking our jobs and much more concerned about AI taking out some of our most high-functioning minds.

You might ask why in the world I would have anxiety. Everything is great! And it is. My newsletter owns its niche, the YouTube channel just crossed 50,000 subscribers, and I’m invited to most of the major conferences, to speak, or moderate.

I used to work on a newsletter on machine learning. Now it’s a newsletter that touches on ML, AI, Robotics and whatnot with three articles a week and a YouTube channel with two deep dives a week. I’m also writing a book, building a math city for the kids, working on five productivity dashboards and a longevity and fasting system – and that’s just scratching the surface.

Turing Post is still a small human team: two of us full time, a video editor half time, and two consultants.

And there is an army of agents.

My agents know no mercy. They know nothing about work-life balance. They work nonstop, and with them I can do so many, so many things.

And then there is another thing that exhaust each and every one of those who work a lot with AI. At the end of the day, have you ever considered how much you’ve read just communicating with agents? And it’s not only reading or skimming. You have to understand, and you have to make decisions. The agents take the doing off your hands, but the deciding stays with you, and it doesn’t scale the way they do.

The load on the brain is tremendous. And now we carry it almost nonstop, across so many things at once – ten, maybe a hundred times more than in the simpler era of the internet.

Do I sometimes imagine people’s heads just exploding? I do. That would make a great science-fiction movie.

Alvin Toffler wrote about this in Future Shock back in 1970. He described how technological and social change could demand too many complicated decisions in too little time, and he even had a word for part of it: overchoice. Even then, in 1970, it was already too much. I can only imagine what he would write today. It’s overabundant overchoice.

Is that abundance AI labs were talking about?

This editorial is about us, humans. I think we might be reaching the limits of our own capacity without even getting close to realizing the full potential of AI.

And here is a little twist, my own touch of grass if you will.

My beloved friend Michael Murphy helped start the Human Potential movement in the ’60s, cofounding Esalen in 1962. I think it’s time to get back to that. I don’t mean Esalen per se, though it is magical. I mean the question of what human potential really is. In the ’60s, it was about how much more we could become. Now I wonder if it’s also about how much we can hold.

Can we use AI to elevate human potential, or do we have to protect it from AI? Or maybe there is a combination where AI doesn’t exhaust the best of us. I don’t know the answer, but I’m certainly ready to have that discussion.

For the first time in more than five years, I’m experiencing tremendous anxiety. Before, it came from stress and uncertainty. This time, it comes from an overabundance of capabilities.

How do you feel?

Follow us on

Twitter Library – Computer-use AI agents

Which agents can work directly with the screen? We updated our guide to open- and closed-source tools that can navigate a browser, click through a desktop app, and fill out forms. Open the collection →

We are reading / watching

• Not even wrong: why AI won’t conquer theoretical physics by David Louapre

• What is work for? Human wellbeing and the value of work in the AGI era by Stephanie Chan et al.

• The future of software by Thomas Wolf

News from the usual suspects ™

Anthropic and Meta both released models on Monday. If you build with them, one of the first questions is who gets access, and the next is what a million tokens will cost. Such was the week:

Anthropic

• Claude Fable 5.2 and Mythos 5.2 are released today, on Monday. They share a model, a 1M-token context window, and API pricing of $8 per million input tokens and $40 per million output tokens (nice reduction). Fable has stronger cyber and biology safeguards. Organizations doing qualifying work in those fields can apply for Mythos. Meet the models

• Roughly 1,000 Fable agents formalized an existing proof of Fermat’s Last Theorem in Lean in 30 hours, according to Anthropic. A list of intermediate theorem statements let the agents divide the work and track what still needed proving. The Fermat experiment

• In a separate memory study, Claude agents working on contracts overwrote one another’s notes. Giving each contract its own page preserved all 57 facts the researchers checked, up from 41. The components of agent memory

Meta

• Muse Ember, released Monday, handles text, images, audio, and video. Meta priced it for repeated jobs such as tool calls and document processing. Up to 200K input tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens (cheap!!!) Muse Ember · Pricing

• Conductor, the workspace for running coding agents in parallel, is joining Meta. Conductor Cloud becomes read-only on November 13 and shuts down on November 20. Conductor’s announcement

NVIDIA

• GR00T 2, introduced also on Monday, predicts future video and robot actions together. Across 50 tasks not previously shown on the target robot, NVIDIA reports 40% average zero-shot success, rising to 95% after fine-tuning on five demonstrations per task. Code is planned for December. Meet Gr00T 2

• NVIDIA’s board opened an independent investigation into reports that CEO Jensen Huang had been targeted by a foreign intelligence effort. The board said it had no reason to believe national security or company secrets had been compromised. Board statement

OpenAI

• GPT-6 Cyber is available to selected Daybreak Red customers. In a Linux security experiment, OpenAI’s pipeline produced 948 working local privilege-escalation exploits and proposed fixes at an API cost of $1.2 million. Human researchers reviewed the findings and coordinated disclosure. Linux experiment · GPT-6 Cyber

• OpenAI agreed to acquire Pierre, whose code-browsing and review tools already appear in Codex. It also described Pando, the Git backend it built to handle concurrent changes from coding agents. Pierre acquisition · Pando

Google / DeepMind

• Google introduced two TPU systems. The 8tx targets confidential inference, and the 8i Max pairs two chips in one package for workloads with large context windows. Cloud Pathways, the distributed runtime used for Gemini, is now generally available on Google Kubernetes Engine. TPU systems · Cloud Pathways

• Gemini Omni 1.2 Flash can make a 15-second video and take follow-up instructions to edit the same clip. It’s available through the Gemini API. Video release

Microsoft / GitHub

• GitHub Copilot came to JetBrains Air through the Agent Client Protocol, with parallel sessions and code review inside the IDE. It also entered public preview in Xcode’s built-in coding assistant. JetBrains Air · Xcode

Amazon

• Amazon began inviting a small group of US customers to shop by text. Alexa can compare products, fill a cart, and place the order in the conversation. Alexa for Shopping

Cohere

• Model Vault can now host third-party open models, including DeepSeek, Qwen, Gemma, and Mistral. A subset is available in a confidential-computing environment in the US. Model Vault update

Thinking Machines Lab

• Inkling 2 Small accepts images and activates 17 billion of its 594 billion parameters per token. It’s available through Tinker. The lab plans to release the weights under Apache 2.0 after external safety testing. Inkling 2 Small

Unexpected, potentially usual suspects

Ricursive Intelligence

• A lab working on AI-assisted chip design released the weights of its dense Ripple-65B Preview model under Apache 2.0. Ricursive puts total research and development spending at $32 million. Ripple-65B Preview

Around the World

• Mistral

  After releasing Le Chonk’s weights last week, Mistral expanded Studio with tools for deploying agents and apps, building workflows, and evaluating how they perform. Mistral Studio

• Moonshot AI

  The weights for Kimi K3.1 Pro are now downloadable. The multimodal model has 2.8 trillion parameters, activates 104 billion per token, and supports 1M tokens of context. Its license requires a separate agreement for certain larger model-service businesses. Model weights · License

• Tencent

  Tencent released Hy4 through its products and API, adding visual understanding and computer use. It says the new weights and system cards will follow. Hy4 release

• ByteDance

  Seedream 5.5 adds native transparent backgrounds. ByteDance reports that the image model can preserve regions outside the requested edit through as many as 100 rounds of changes. Seedream 5.5

• LG AI Research

  LG opened a free API for K-EXAONE 2.5 on Monday. The model is designed for Korean-language work, coding, and tool use. Announcement · EXAONE platform

Research highlight

🌟🌟 On Agentic Manipulation Policies – A robot picks up a spoon

Fascinating: the job for a robot was to put a mustard bottle in a bowl. The bottle cap was just beyond the reach of the robot’s gripper. A wooden spoon was closer. The robot picked it up, tipped the bottle toward itself, then grasped the cap.

Apple’s team gave Claude Opus 5.5 camera views, a Python environment, and commands to move the arm. Their software, AMP1, had no built-in routine for grasping, pushing, or using a spoon. The model could inspect the scene again after each move and decide what to do next.

The team ran 108 physical trials across 12 tasks on three Franka arms. When a controller error ended the attempt, AMP1 succeeded in 94 trials. When it could clear the error and replan, it succeeded in 104. →read the paper

Models

🌟 marks our must-reads this week.

• 🌟GR00T 2 – NVIDIA’s model learns to predict the next video frames and the robot’s next actions together. On 50 tasks the target robot hadn’t been shown, success went from 40% with no demonstrations to 71% with one per task and 95% with five. NVIDIA plans to release code in December. →read the announcement

• 🌟HALO – A humanoid practiced moving through 1,000 simulated homes before stepping into six unfamiliar real spaces. It could sidestep through half-open doors and crawl under tables. Across 360 single-task trials, the team reports 91% success. →read the paper

• 🌟Open-HumanWAM – The team released 50,000 hours of first-person footage annotated with language and hand motion, along with training code and checkpoints. A model trained on human and robot experience reached 84.6% on RoboTwin’s Clean2Random simulation. →read the paper

• 🌟La-JEPA – A robot can learn to predict what it will see after a move, then skip that prediction when it acts. La-JEPA trains in V-JEPA’s feature space. The optimized policies ran two to five times faster than the Fast-WAM variants the authors compared. →read the paper

• NEO-VLA – Which layer should a robot consult before it grasps an object? In controlled comparisons, the visual clues that helped choose an action often peaked before the last layer. NEO-VLA lets the action model use those intermediate features. →read the paper

• PrimeSim – Given a robot command, PrimeSim generates synchronized camera views of what happens next at 10.1 frames per second. It learned from 30,110 hours of demonstrations across 22 platforms. The team released model weights and code. →read the paper

• Drex – This week’s technical report explains how Drex assigns probabilities to choices you supply, such as whether to approve a request or send it for review. The 27B Max model took a median of 183 milliseconds per request in the authors’ test. →read the paper

Research

Trends we see in this week’s selection:

• Human video is starting to pay off on real hardware. GR00T 2 measures how much a robot can do with one demonstration of its own, then with five.

• A camera turn gives video models two different jobs. They have to remember the glass they saw a moment ago and work out what happened to it while it was out of view.

• The second attempt is becoming a training example in its own right. Harness-S saves runs that another harness rescued, and multi-turn RL measures what agents can recover when they get another turn.

• What counts as progress? RoboColosseum measures how far a robot gets even when it fails. Scientist World gives a research agent a working baseline to beat.

What does a robot need to predict?

• 🌟VFlash – A robot’s plan can be a box around a target or a line showing where to move. VFlash predicts all the coordinates of each shape together. With those plans, a world-action model’s average success rose from 62.6% to 70.2% across four MolmoSpaces task suites. →read the paper

• DreamX – DreamX compresses each video frame’s latent representation from 1,200 tokens to four that describe movement. The authors test whether those four tokens can carry an action from a video of a human performing it to a robot. →read the paper

• Solv3R – A new camera angle can reveal that an earlier estimate was off. Solv3R updates both the camera positions and the 3D map as images arrive, reusing work from views it has already processed. →read the paper

What happens when the camera looks away?

• History Forcing – Two versions of a teacher see the same scene, but only one gets the relevant history. A video generator learns from the difference between their predictions. Its object-consistency score on MemoBench rose from 58.43 to 62.18. →read the paper

• EvoWorld – A camera pans from a street vendor to a building facade as it collapses. What should the vendor be doing when the camera pans back? EvoWorld keeps track of the event and its consequences separately from the model that draws the next frame. →read the paper

How agents learn from a second attempt

• Harness-S – When an agent gets stuck, the researchers let it try again with a different harness. They train the model on the attempts that succeed, then return it to its original harness to see which skills transferred. →read the paper

• 🌟ToolSpace – The team builds practice tasks from real sequences of tool calls, replays each task, and checks the result. Qwen3.6-35B-A3B’s average across seven tool-use benchmarks went from 23.0 to 41.0, then to 43.2 with reinforcement learning. →read the paper

• What Does Multi-Turn Reinforcement Learning Really Improve? – A trained agent’s advantage shows up when it gets another turn to recover from a mistake. Tighten the turn limit and much of the gain shrinks. The authors vary turn budgets during training to help the policy adapt. →read the paper

• 🌟LeanGuard – Before an agent calls a tool, LeanGuard checks that action against a policy written in Lean. It blocked every policy-forbidden call in the Tau-Bench evaluation. On a separate test, DTap, direct prompt-injection success fell from 57.9% to 5.9%. →read the paper

What counts as success on a hard task?

• 🌟RoboColosseum – Across 124 sites, evaluators made 5,699 blinded comparisons of robot policies. On the G1 humanoid, every policy failed all five tasks. Measuring how far each got revealed differences a pass/fail score hid, with average progress ranging from 13.3% to 23.2%. →read the paper

• FrontierPhysics – Forty-two physicists contributed 56 research tasks that ask agents to review the literature, plan and run experiments, and write a manuscript. Across three attempts per task, the strongest of 26 model configurations passed 46 of 168 attempts. →read the paper

• 🌟Scientist World – The authors turned 1,000 empirical AI papers into environments where an agent can run experiments against a reproduced result. On ten lower-compute environments, 74% of attempts to develop a method from scratch beat the baseline. The team also released 6,400 hours of research traces. →read the paper

Memory and the cost of a decision

• Muir – Muir changes how a linear-attention model writes, erases, and retrieves information in a fixed-size memory. Its 1.34B model, trained on sequences of 4,096 tokens, recovered a hidden passkey with near-perfect accuracy at a million tokens. →read the paper

• LowRAM – A model can often reconstruct the values in its attention cache from the keys it already stores. At 262K context, LowRAM’s selective reconstruction compressed the cache fourfold and delivered about five times the uncompressed baseline’s peak decoding throughput. →read the paper

• Route Anything – Across 374,401 scored runs, the authors measured when a system benefits from choosing a different model or robot policy for each input. In manipulation, learned routing gained 1.1 points over the best fixed policy while using 24% less policy time. →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#174: What is Weight-Space Programming and why it suddenly blew up the internet?