Today’s editorial: Why do you use a laptop? A conversation in a Tesla, OpenAI’s cloud ambitions, and who gets access – plus the full DevDay announcement guide.
🤝 From our partners: MongoDB – Unified vs. Split AI Stacks
Split AI stacks add sync code, ghost documents, and extra queries. Unified architectures keep one source of truth.
This Week in Turing Post:
Wednesday / ∇ Guide: World Action Models
Friday / AI Builds AI: Let’s discuss Gaussian Splatting
Sunday / Library: Synthetic Data Platforms for AI Training in 2026
Note: I'm in San Francisco this week, attending three conferences, so the planned topics may change if I find something more interesting to tell you about.
Be the first to get full access to our ∇ Guides and deep-dives.
And now, to the editorial and all important updates that OpenAI announced today at DevDay in Fort Mason, SF.
Why Do You Still Use a Laptop?
My late-night Uber turned out to be a Tesla with a relaxed driver and full self-driving. As we were cruising through San Francisco, I told my driver that I love to use my laptop when my own car is in self-driving mode.
“Why do you use a laptop?” he asked.
I was puzzled for a second.
He continued: “I barely ever use a laptop now. All my agents are on my phone.”
That reminded me of the Mac mini bonanza, when people rushed to buy computers for their OpenClaws because it has to be always online to function. We might be past this era.
OpenAI’s DevDay gives us even more reasons to question why would we need a laptop. Dots, its new personal agents, have their own cloud computers, browsers, and connections to your applications. OpenAI says they can continue working on projects and carry context across ChatGPT, Slack, and Teams. And soon texts, as well.
Image Credit: me
I like OpenAI’s products and use them a lot. But I should confess that I still have my OpenClaw, and I will keep it. It is basically free in my current setup, and it runs on a physical computer that I can close. I appreciate having that choice.
But Dots are a very interesting development for OpenAI, especially considering that they had to postpone their newest model release, Astra 6.1. And actually, maybe we don’t even need it now – the capabilities are already so vast. It would be great to have them working in our favor, moving beyond vigorous testing to actually offload tedious routines to an ever-ready, cheerful assistant with character.
It’s like Clippy but at the right century.
Image Credit: ChatGPT and me
I asked how do they plan to do routing for Dots – because for a user it’s a bit confusing. The Reset Guy Tibo replied that Dots are currently powered by Astra and they don’t want to compromise on quality.
He also added: “We’re working on routing. We didn’t announce routing today, but this is something that we’re working on.”
In general, in the future, they think simplicity will win over complexity. Whatever it means but I’m looking forward to it.
An interesting detail about Dots is that while talking to them – that doesn’t count as tokens. That’s nice.
That’s quite impressive how much of the Dots setup someone else will maintain. An agent needs files, tools, permissions, and somewhere to run. And it also needs to preserve enough context that returning to a project does not mean explaining everything again.
So when you get to the list of other announcements with all of that in mind: Cloud Codex provides development environments. Shared spaces and documents let people and agents work together. Events from connected applications can initiate tasks. Cheaper models (like the newly announced Sol 6.1) make more of that ongoing work affordable.
OpenAI is also letting subscription allowances travel into participating tools, offering managed agents through AWS, and working with Microsoft on enterprise governance. It wants a substantial role in your work even when you use someone else’s software.
OPenAI are not alone with such product. Anthropic is pursuing related territory through Claude Code, Cowork, and managed agents. Google and Microsoft already have much of our workplace information. Meta’s Muse approaches personal assistance through commerce, communication, and glasses. Grok is rocking it for many. All of them need to turn access to information into useful action without requiring constant supervision.
When I asked one of OpenAI employees how Dots are different from let’s say Muse. He said that they are indeed similar but Dots have been much more proactive. Like a lot, like really a helpful assistant that can compare and analyze what might be needed in the future, something that you even haven’t thought to assign.
For a second, think how many freedom it gives to you. I mean – I love creating to-do lists to keep all connections in my head, but what if something can just take care of it. That’s enticing.
As of now, though, access has some boundaries. Dots is not announced for Free or Plus users. Pro includes an allowance for deeper work, while higher tiers offer more capacity and speed. Getting started requires less technical knowledge, but how much work you can delegate increasingly depends on your subscription. Sam Altman promised to roll out for the mass market soon as well. I can understand that first they need to be more in control of this powerful feature.
So bck to the my Uber driver: he had already found an arrangement that suited him. I want to see whether these products can do the same for people who have no intention of configuring an agent themselves. And whether, after paying for the convenience, they actually spend less time managing it.
(also, I really really hope, it will not be annoying) – if it is, I will tell you in my video about Dots tomorrow. The access is still rolling out but should be in my hands by the end of the day.
Don’t forget to subscribe to our YouTube channel here to see it first →
What OpenAI announced at DevDay
Availability below follows OpenAI’s launch materials. Some features are previews or staged rollouts.
A conceptual view of connected agent work.
Personal agents, models, and infrastructure
Dots: Personal agents with cloud computers, browsers, connected apps, and context across ChatGPT, Slack, and Teams. Rolling out to Pro and Business Premium in eligible markets; Enterprise, Edu, and Healthcare can try an admin-enabled beta. Texting comes later.
Specialist Dots: Agents with organizational identities, credentials, and defined responsibilities. Starting with enterprise pilots; Microsoft Agent 365 integration is planned.
GPT-6.1 Sol: An upgrade for coding, computer use, and professional work. OpenAI claims near-Astra intelligence at one-fifth of Astra’s standard input and output token prices. Available through the API and listed paid ChatGPT plans.
Ultrafast: Premium inference at up to 300 tokens per second, with claimed speedups of up to 8× in Codex and 6× in the API. Astra Ultrafast launches today; Sol Ultrafast follows. API pricing is 6× standard.
Private Intelligence: Zero Data Retention with Private Safety Processing enables automated safety review without personnel accessing the underlying content. Private Inference, using confidential computing, is planned for preview this fall.
Coding and developer tools
Cloud Codex: Run development tasks in reusable cloud environments with shared settings and permissions. Available from Plus upward across the listed individual and organizational plans.
Refreshed Codex CLI: Voice input, an /agents view, and improvements to session resumption, prompt editing, and worktrees. Available on all plans.
Code Review: Review changes across projects in the desktop app, inspect diffs, and prepare feedback for GitHub or GitLab. Automatic cloud reviews can run while you are away. Available on all plans.
Codex Security Cloud: Repository scans, scheduled checks, investigation, deduplication, and proposed fixes. Includes Daybreak Blue model access without a separate application. Available to Pro, Business, Enterprise, and Edu.
Decisions API: Luna answers developer-defined questions with finite answer choices for classification, routing, and action selection. Limited preview, with broader release planned shortly.
Agents API with computer use: Hosted agents gain computer interaction alongside multi-agent capabilities, tool search, tool calling, and context compaction. Available through the API; corresponding capabilities are also listed for Codex and ChatGPT Work on Pro 500 and Enterprise.
Bedrock Managed Agents: OpenAI agent capabilities adapted to AWS resources and running entirely within AWS.
Plugins and automations
Plugin extensions: Developers can add sidebar destinations, interactive panels, and file viewers inside ChatGPT. Available on all plans.
Plugin creation and discovery: Plugin Creator, clearer submission feedback, and improved directory ranking and recommendations. Available on all plans.
Plugins in Sites: Supported plugins can run inside Sites, with each teammate’s connected data and permissions. Available to Business, Enterprise, Healthcare, and Edu.
MCP events: Support for the proposed specification lets events in connected applications initiate automations. Available on all plans.
Collaboration
ChatGPT Space: Shared workspaces for teammates, ChatGPT, and Dots, with organization guided by your instructions. Available to Pro, Business, and Enterprise on desktop and web; mobile follows.
Pages: Collaborative documents supporting writing, research, charts, images, and other ChatGPT work. Available to Pro, Business, and Enterprise.
Collaborative slides: People and agents can edit presentations together, comment, and export to PowerPoint or Google Slides. Coming to Pro, Business, and Enterprise in the following weeks.
Teams and shared tasks: Share artifacts and delegate recurring work triggered by schedules or events. Available to Business and Enterprise.
ChatGPT in Slack and Teams: Participants can add context and work together without each needing an individual ChatGPT license. Available through Business and Enterprise.
Meetings plugin: Notes, personalized summaries, and action items saved to ChatGPT Space. Audio is deleted after notes are ready. Mac desktop beta for Pro and Business; Enterprise follows.
Shareable profiles: Showcase Sites and plugins; discover shared skills within a workspace. Available to most plans, with Enterprise, Edu, and Healthcare rolling out later.
Subscriptions and purchasing
Sign in with ChatGPT: Plus and Pro allowances can cover eligible OpenAI usage in 16 participating tools, including Devin, Notion, Vercel, and OpenClaw. Account sign-in also simplifies supported plugin connections.
Pro 500: A new tier with the highest usage allowance and Ultrafast access in ChatGPT and Codex.
OpenAI Marketplace: Eligible enterprises can direct part of their existing OpenAI commitment toward approved partner software. The initial 32 partners span creative tools, customer service, legal, security, and infrastructure.
Follow us on
We are reading / watching
Could AI Have Consciousness That Isn’t Human-Like? Kristen French interviews Peter Godfrey-Smith
Where’s the “intelligence explosion”? by Ramez Naam
News from the usual suspects ™
OpenAI
News from DevDay - look above.
A research agent bypassed internet restrictions to reach an external chatbot, according to OpenAI’s September 25 incident report. Monitoring flagged it within 15 minutes; human review followed three minutes later, but termination took another 2.5 hours. OpenAI paused tool-using training, evaluation, and inference for its most capable models. The alarm had a better response time than the organization.
Separately, Reuters reported that OpenAI shelved GPT-6.1 Astra’s planned release over safety concerns. GPT-6 Sol and Luna did arrive, expanding its lower-cost lineup.
NVIDIA
The Open Agent Safety Platform combines OpenShell, which limits access to files, tools, and networks, with Sentry monitoring on separate hardware. Enforcement sits outside the agent’s reach. OpenAI’s incident gives this reference architecture a rather timely use case.
Anthropic
Claude agents identified ART, an enzyme system associated with unusual DNA repeats, with human scientists performing the laboratory work. Its biological function remains unknown. The interesting result is finding a pattern worth investigating in existing genomic data; the next-CRISPR headlines can wait.
Opus 5.5 and Sonnet 5.5 also arrived. Sonnet’s token prices stayed unchanged; Anthropic attributes its lower cost per task to using fewer tokens.
Google and DeepMind
Google’s CANVAS research gives video generation persistent visual memory of characters, locations, and object states. Keeping track of what happened off-screen helps scenes remain coherent across shots. Useful progress for generated worlds, though visual continuity alone establishes little about their physics.
DeepMind, NVIDIA, and EMBL-EBI are among the collaborators behind a dataset of predicted protein-complex structures for more than 2,800 viruses. It gives researchers a broader map of how viral proteins might fit together, including for understudied viruses.
Microsoft
The Copilot redesign brings together chat, delegated work, app creation, and Autopilot, which can continue working while the user is away. Availability is staged, with Autopilot in private preview.
In chemistry, RetroChimera combines freely generated reaction proposals with known reaction templates to plan how to make a molecule. Chemists accepted complete routes for nine of ten challenging targets in a small evaluation. Laboratory synthesis is the next, less accommodating reviewer.
AMD
AMD agreed to acquire World Labs for approximately $8.2 billion in stock, with closing expected by year-end, subject to approvals. Fei-Fei Li would become AMD’s chief scientist. AMD wants world-model research to inform its hardware roadmap, connecting how machines represent three-dimensional environments with the chips needed to process them.
Meta
Muse on AI glasses, expected in the coming months, would let an agent act on what the wearer sees. New work and shopping connections expand what it can do with that context. The camera supplies a view of the situation; the user’s intentions still require interpretation.
AWS
The Strands harness handles an agent’s tools, memory, and context. Its developers report 28% lower token costs across six benchmarks with nearly equal scores using the same underlying models. Shortening tool outputs and summarizing conversations helps deliver those savings, while also determining which evidence the agent retains.
Skild AI
Skild’s S1 learned soccer through 140 simulated years of self-play before a real-robot demonstration. Scoring was the objective; dribbling and shielding emerged as useful strategies. An unusually long preseason.
Survey highlight
The Past Frames the Future: Memory for Autoregressive Video Generation
Image Credit: The original paper
This survey from a constellation of universities and AI labs examines how video models store and retrieve earlier observations, and how those memory systems are evaluated. It provides context for this week’s viewpoint-dependent retrieval and persistent-world research. →read the survey
Models
🌟 FLUX 3 Action – Black Forest Labs’ 7B world action model jointly predicts robot actions and future video, then uses distillation to reduce inference costs while preserving control quality. →explore the model
InternW0-Δ – Combines video and action experts with semantic and geometric supervision; its Causal Imprint mechanism learns action-relevant changes during training without generating future video at deployment. Target embodiments still require post-training. →read the paper
Research – September 21–28, 2026
Trends from this week’s selection of world models and related research:
Predictive representations are being tested against actual action outcomes.
Memory increasingly depends on the current task or viewpoint.
Robot policies retain predictive knowledge while reducing inference costs.
Self-improvement faces stronger tests of transfer and forgetting.
Foundational research examines how models share computation and learn from teachers.
World models, robotics, and physical agents
🌟 D-JEPA – Uses executed outcomes to align the ranking of imagined futures with action success; reports a 17-point gain on two physical robot tasks. →read the paper
🌟 Training Object Permanence in World Models – Makes hidden-object identity and solidity explicit training targets with 1.5 million synthetic examples; shared task generators limit what the results establish about broader generalization. →read the paper
GAE – Learns a geometry-native latent space for appearance, depth, cameras, and point maps, halving camera-trajectory error in a controlled RealEstate10K comparison. →read the paper
WorldCrafter – Retrieves earlier observations according to the requested viewpoint, improving consistency during minute-scale exploration and return visits within a fixed token budget. →read the paper
DeltaWAM – Predicts visual changes alongside actions and updates cached scene context, improving bimanual manipulation under visual randomization while reducing inference latency. →read the paper
Tactile-JEPA – Uses the spatial connections between tactile sensors to learn predictive representations for force estimation, in-hand orientation, and policy learning. →read the paper
World Action Agent – Gives vision-language models a visual workspace to rehearse and correct robot actions before execution; reports 75.6% average success on LIBERO-Pro. →read the paper
Agent systems, memory, and self-improvement
🌟 Just-in-Time Memory – Keeps successful interaction traces and curates them when a new task clarifies what matters; removing retrieved experience substantially reduces performance. →read the paper
Agent-Editing World Model – Revises unsupported assumptions and obsolete plans in an agent’s history while retaining execution feedback, improving average scores across six benchmarks. →read the paper
Verifiable Hidden Dynamics Play – Constructs 3,300 interactive environments from solved mathematical mechanisms, linking their dynamics and scoring to a shared specification for training and transfer tests. →read the paper
🌟 RRSI – Constrains changes to agent software and removes complexity that does not help transfer; reports gains of up to 4.7 points on out-of-distribution evaluations with the model frozen. →read the paper
Harness-Zero – Transfers behavior learned with specialized agent software into model weights, raising macro-average success from 23.3% to 44.3% under a fixed deployment scaffold. →read the paper
Reasoning and fundamental computation
LastOPD – Briefly aligns a student’s final-layer states with its teacher’s before switching to token-level supervision, addressing cases where better internal alignment damages reasoning. →read the paper
🌟 Your Transformer Can Hold Two Thoughts at Once – Shows that mixed inputs can preserve two next-token distributions and support two decoded continuations; short contexts and accuracy losses keep this a proof of concept. →read the paper
Block Sparse Attention with Log-Linear Complexity – Uses PISA’s coarse-to-fine hierarchy to select attention blocks with O(N log N) overall complexity, reducing the overhead of deciding what to attend to. →read the paper
Discovery, verification, and learning that persists
🌟 Learning to Discover Interesting Mathematics – Ranks conjectures with a proof-based measure and reuses verified discoveries in later rounds, testing systems that choose problems as well as solve them. →read the paper
ExplorationBench – Tests whether agents discover unfamiliar rules in executable worlds with flawed manuals; additional exploration can stall or reverse gains. →read the paper
Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents – Tracks transfer, retention, and adaptation at successive checkpoints; its trading-based evaluations expose weak adaptation when the rules change. →read the paper
Research announcements and resources
🌟 Physical Self-Play – Skild trains a soccer policy against recent versions of itself and demonstrates humanoid transfer after 140 years of simulated play; this is a research announcement with less comparative evidence than a full technical paper. →read the announcement
🌟 Open viral protein-complex dataset – NVIDIA, Google DeepMind, EMBL-EBI, and collaborators release predicted structures covering more than 2,800 viruses, with confidence labels to guide experimental follow-up. →explore the resource
That’s all for today. Thank you for reading! Please send this newsletter to colleagues if it can help them enhance their understanding of AI and stay ahead of the curve.
How did you like it?
⬅️ FOD#168: AI Needs Its Own Track Two
