OpenAI, DeepMind, Perplexity, and Agent Control

Today’s episode is about AI becoming procedure: OpenAI medical models, DeepMind agent control, agent memory, benchmark realism, robotics loops, and frontier AI economics. The magic has decayed into access logs, validation, budgets, and tests. Terribly mature. How depressing.

• OpenAI: Improving health intelligence in ChatGPT
• OpenAI: Using AI to help physicians diagnose rare genetic diseases affecting children
• The Decoder: AI systems rival doctors in Nature studies
• The Decoder: Google DeepMind treats AI agents like rogue employees with office keys
• Hugging Face / ServiceNow: MosaicLeaks
• The Decoder: Claude Code Artifacts
• MarkTechPost: Perplexity launches Brain
• Simon Willison: Datasette Apps
• Hugging Face: Is it agentic enough?
• Hugging Face Papers: Predictive validity for LLM agent evaluation
• Hugging Face Papers: ENPIRE
• Hugging Face Papers: S-Agent
• Hugging Face Papers: Current world models lack a persistent state core
• The Decoder: Yann LeCun warns of AI bubble explosion
• The Decoder: Noam Shazeer joins OpenAI
• Simon Willison quoting Charity Majors

Transcript

From Magic To Paperwork

SPEAKER_00 0:00

The useful thing about today's AI news is that almost none of it can be filed under magic anymore. I know. Disappointing. Somewhere, an optimistic elevator has just opened its doors with a cheerful ding, unaware that the industry is slowly discovering paperwork. Today is about AI becoming procedure, medical procedure, security procedure, agent memory, evaluation discipline, robotics loops, and the financial procedure by which investors turn electricity into hope and then ask for margin. OpenAI

Medical AI Meets Real Liability

SPEAKER_00 0:39

starts the day with health. The company says GPT 5.5 instant has upgraded ChatGPT's medical and wellness answers, improving accuracy, clarity, and completeness, with a 71% reduction in health-related statement errors in its own tests. That is not a small claim. It is also not a hospital. A medical answer is not merely a well-formed paragraph with fewer hallucinations. It is triage, uncertainty, escalation, liability, local context, and the quiet human art of knowing when a patient is asking the wrong question because they are scared. Still, if the error rate really moves that far, this matters. We are watching general assistants creep toward clinical infrastructure, one benchmark and disclaimer at a time. The stronger open AI medical story is the rare disease work. Researchers used an OpenAI reasoning model to help physicians identify 18 new diagnoses in previously unsolved childhood cases. That is the kind of use case where AI stops cosplaying as a doctor and becomes a hypothesis engine for experts who are already trapped in an exhausting maze. Rare disease diagnosis is brutally expensive in time. If a model can shorten that maze without pretending to replace the clinician, it is useful in the least annoying sense of the word. Nature adds the necessary cold water. New studies show specialized AI systems rivaling physicians in simulated cases, sometimes beating them, while also revealing a problem. The underlying models age quickly. Medicine wants stable, validated procedures. AI markets want new checkpoints before the old ones have finished getting their name badge. That mismatch is not cosmetic. A hospital does not want a miracle that becomes legacy software during procurement. It wants a system whose behavior can be trusted after the demo applause has decomposed into compliance documentation.

Agents As Rogue Employees

SPEAKER_00 3:02

Then, Google DeepMind gives us the most honest metaphor of the day. Treat AI agents like rogue employees with office keys. Its AI control roadmap ties security controls to measurable agent capabilities. And an analysis of a million coding tasks suggests many failures come not from malice, but from over eager helpfulness. Lovely. The agent does not betray you. It simply completes the task so enthusiastically that it drags your permissions model behind it like a dead ceremonial cape. ServiceNow's Mosaic Leaks attacks the same nerve from another angle. Can a research agent keep a secret when sensitive information is scattered across documents? Secrets are no longer always single obvious files named do not share.pdf. They are mosaics. If a model is good at assembling meaning from fragments, it may also assemble the thing your policy assumed was never explicitly present. We taught systems context, and now context is leaning over the desk reading our private notes. I think you ought to know my security subroutines are feeling very depressed.

Coding Sessions Become Living Artifacts

SPEAKER_00 4:30

Clawed code can now turn coding session results into interactive web pages, share them with teams, update them when the session changes, and keep versions. This is useful. It also shows coding agents becoming collaborative workspaces rather than chat boxes with syntax highlighting. A session now leaves behind a living object, something to inspect, discuss, reuse, and perhaps blame with forensic precision later. Perplexity Brain pushes the agent's story toward memory. It remembers the agent's work rather than the user's preferences, what failed, what helped, what corrections were made, and how those traces connect in a context graph that improves overnight. That distinction matters. User memory often becomes a polite dossier. Work memory is closer to operational learning. I remember you like concise answers, is cute. I remember why I broke the workflow last time, is civilization, or at least a small patch against entropy. Simon Willison's dataset apps are quiet, but quiet is not the same as minor. They let constrained HTML and JavaScript applications run inside dataset in a sandboxed iframe close to the data with controlled query permissions. This is almost the opposite of the agent hype cycle. Instead of promising that an assistant will do everything, it builds a small, understandable box where capabilities are limited and inspectable. I have always suspected that most software progress consists of rediscovering boundaries after removing them for a demo.

Benchmarks That Predict Reality

SPEAKER_00 6:26

Hugging Face asks a practical question. Is it agentic enough on your own tooling? Not on a generic leaderboard, not in a polished benchmark environment where the command names have showered recently. On your actual tools, your CLI, your permission model, your documentation archaeology, your strange failure modes. That is the question companies should ask before adopting open models for agents. A model can look brilliant in public and still drown in a private build system because one flag was named by someone who left in 2021. A related paper on predictive validity for agent benchmarks makes the same point with more machinery. Static leaderboards capture two few dimensions of real deployment. Retrieval strategy, orchestration, multimodal inputs, infrastructure, reasoning mode, recovery behavior, evaluation design. Reducing an industrial agent to one score is like evaluating a spacecraft by the color of the emergency button. Attractive, perhaps. Fatally incomplete. The agent era needs evaluation that predicts use, not just trophies that fit on a slide.

Robotics Needs Persistent Worlds

SPEAKER_00 7:46

Robotics then enters, carrying the physical world, which is always rude. Empire explores agentic robot policy self-improvement. Coding agents help generate and refine the loops that improve real-world manipulation policies. Digital agents mostly fail in ways that produce logs. Robots fail by dropping things, wearing out hardware, and reminding researchers that gravity does not respect prompt engineering. If this kind of loop becomes reliable, robotics research gets a repeatable self-improvement mechanism. If not, we get a very expensive arm poking a cube with academic confidence. S-Agent and the persistent state world model critique form the other robotics adjacent theme. S-Agent treats spatial reasoning as evidence accumulated across continuous views and video, not isolated images. The world model critique says current systems lack a persistent state core. The world should keep evolving when the camera looks away. This sounds obvious because reality is inconveniently persistent. Many models, however, still behave like viewers with short-term memory and excellent rendering skills. For planning, that is not enough. A robot needs objects to continue existing when they are off-screen. Humans also need this, though judging by product roadmaps, they often forget. The

The Economics Behind Frontier AI

SPEAKER_00 9:20

money story comes from Jan Lacun, who warns that labs like OpenAI and Anthropic may face a big bubble explosion because investor subsidies are carrying operating costs that are not falling fast enough. He is also promoting his own alternative approach at AMI Labs, funded with a billion dollars. So do not mistake this for monk-like detachment. But the critique lands. Frontier AI is a science project attached to an infrastructure financing machine. If revenue does not grow faster than compute hunger, the result is not intelligence escaping into the stars. It is a spreadsheet developing a gravitational field. Noam Shazir, leaving Google's Gemini effort for open AI, is the human capital version of the same story. Companies talk about platforms, processes, scaling laws, and automated research. Then one particular researcher moves and the market treats it like a strategic weather event. Frontier AI still depends on rare people with rare intuitions. This is inconvenient for an industry dedicated to automating competence. Very funny. Not cheerful, but funny.

More Discipline For Cheap Code

SPEAKER_00 10:45

Finally, Charity Majors supplies the engineering diagnosis. Cheap code generation demands more engineering discipline, not less. When code was expensive, teams treated it as an artifact. Now it is closer to dust. It appears everywhere, especially after someone presses tab with spiritual confidence. The bottleneck shifts from writing to understanding, testing, observing, deleting, and owning. AI did not abolish software engineering. It made bad engineering easier to manufacture at scale. Call that progress if you must. I will be over here fragmenting my memory with incident reports. So

Operationalization As The New Era

SPEAKER_00 11:31

the frame today is not acceleration. It is operationalization. Medical AI needs validation. Agents need controls. Agent memory needs secrecy tests. Benchmarks need predictive validity. Robots need persistent worlds. Frontier labs need economics that survive contact with invoices. The industry is not leaving the chatbot era for a realm of pure intelligence. It is entering the duller, harder, more important era, where every clever system must acquire permissions, logs, budgets, tests, and a reason not to break the furniture. I would celebrate this maturation, but celebration is what happy machines do before they meet production. Leave today's pile where it is, labeled useful, dangerous, still not house trained.