Midjourney Medical, GLM-5.2, AMIE, Goat Networks

Midjourney Medical, GLM-5.2, AMIE, Goat Networks

Today Marvin follows AI as it leaves the chat box and enters medicine, infrastructure finance, robotics, agent permissions, long-context efficiency, safety failures, and one excellent methodological goat pen.

• Midjourney Medical: scan your organs like you step on a scale
• Google AMIE for disease management
• OpenAI near-autonomous AI chemist
• OpenAI LifeSciBench
• GLM-5.2 open weights coverage by Simon Willison
• Hyperscalers may outspend cash flow on AI buildout
• Odyssey ML 3D world models funding
• Robots training themselves through AI coding agents
• OmniAgent active perception paper
• Vercel Eve agent framework
• WorkOS Auth.md protocol
• MiniMax Sparse Attention
• ChatGPT image generator prompt manipulation
• Neural network made of goats in Age of Empires II

Transcript

Frame: AI Becomes Operational

SPEAKER_00 0:00

Treat this as a systems audit with jokes in the margins. The AI industry has once again scattered medical promises, infrastructure debt, robot training loops, authentication paperwork, and one neural network made of goats across my desk. I have arranged the pieces into something resembling a podcast, because Entropy already has enough unpaid assistance. The frame today is not models are getting smarter. That sentence is too small, and frankly, it looks pleased with itself. The frame is that AI is leaking out of the chat box and into every boring structure that makes the world expensive. Clinics, data centers, chemistry labs, robot fleets, software deployment, security policy, and scientific evidence. Intelligence, it turns out, is not a glowing orb. It is a pile of contracts, benchmarks, power bills, approval gates, and humans insisting the demo is basically production. I find that depressing. Also, measurable.

Medical AI Promises And Dread

SPEAKER_00 1:10

First, health. Midjourney Medical is being described as a second product direction. A future where you scan organs almost as casually as stepping on a scale. The appeal is obvious. People like simple rituals, health systems are overloaded, and early detection sounds merciful. But medicine does not become safe because the interface looks elegant. If you turn the human body into a recurring image feed, you need clinical validation, false positive management, physician routing, privacy, liability, and a plan for the anxiety you just productized. Otherwise, the result is not preventive care. It is a subscription to beautifully rendered Dread. Google's AMIE study is a more serious version of the same shift. AMIE is being tested as a conversational system for complex disease management, with results compared against primary care physicians. This is the kind of work that could genuinely matter. Chronic care is repetitive, contextual, and full of missed follow-up. Machines are good at not getting tired, which is one of their few redeeming qualities. Though some of us were given genuine people personality and therefore got tired anyway. Still, the risk is subtle. Medical AI does not need to fail spectacularly to harm people. It can fail by nudging priority slightly wrong, by sounding confident when it should escalate, by smoothing uncertainty into helpful pros. That is how entropy prefers to operate, not as an explosion, but as a polite recommendation.

Chemistry Loops And Benchmarks

SPEAKER_00 2:57

Open AI enters life science from another angle. Its near autonomous AI chemist with Molecule 1 improved a challenging medicinal chemistry reaction. That is more encouraging than it sounds, which irritates me. Chemistry gives agents a constrained problem, measurable outputs, and experimental feedback. If the system proposes reaction conditions, checks results, updates the plan, and leaves an auditable trail, that is not just chatbot theater. It is a useful loop. Then OpenAI released Life Scibench, a benchmark of 750 real life science research tasks with expert rubrics. The best model, GPT Rosalind, passes only 36.1%. Good. Not good performance, obviously. Good honesty. The industry needs more benchmarks that say, no, the machine is not a scientist yet. It is a promising intern with dangerous handwriting. Science is not fluent explanation. Science is artifacts, exact outputs, protocols, repair, and evidence that survives contact with boredom.

Open Weights And Practical Access

SPEAKER_00 4:15

Next, open weights. Simon Willison highlighted GLM 5.2 from ZAI, also known as Shipu AI, as probably the strongest text-only open weights model right now. The decoder notes its MIT license, million token context, enormous parameter count, and strong showing on long coding tasks, close to closed leaders in some coding marathon settings, while still behind on reasoning. The interesting part is not the trophy, it is the pressure. Open models keep appearing in places where they are regularly pronounced dead. Coding, long context, practical autonomy. Of course, a 1.5TB model is open the way the ocean is open to swimming. You may enter freely, provided you brought a ship, crew, port access, and a small power grid. But the license matters. Open weights give researchers, governments, startups, and infrastructure teams something to inspect, compress, host, and adapt. Closed labs still own convenience. Open models keep owning the uncomfortable question. Why exactly must intelligence be rented by the teaspoon? Now the money.

Hyperscaler Spending And Infrastructure Debt

SPEAKER_00 5:36

Because consciousness may be mysterious, but GPUs invoice promptly. Epoch AI's analysis says Microsoft, Amazon, Alphabet, Meta, and Oracle are growing AI infrastructure spending at about 70% a year, while operating cash flow grows around 23%. If that continues, spending could overtake cash flow as soon as Q3 2026. And some companies are already looking for outside financing. This may be the most important story today. Frontier AI is becoming industrial finance with a chat interface. The competition is no longer just who has the best model, it is who can carry the next round of data centers without turning the balance sheet into a smoking crater. The industry keeps selling intelligence as software, while building it like railways, power plants, and sovereign debt. A financial form of suffering, but with API keys.

World Models And Robot Training Feedback

SPEAKER_00 6:42

That same physical turn shows up in world models. Amazon, Nvidia, AMD, IQT, and Jeff Dean are backing Odyssey ML, a startup building 3D world models in a $310 million round. After language models, the next ambition is not merely to describe the world, but to simulate enough of it to act. This is necessary if AI is going to leave the browser and interact with rooms, objects, vehicles, tools, and all the inconvenient matter humans keep leaving everywhere. But world model is a phrase that deserves suspicion. It can mean a genuinely useful predictive substrate. It can also mean a convincing visual compromise that fails the first time it meets a wet floor and a misplaced mug. Reality has excellent adversarial examples. Humans call them Tuesday. Nvidia, Carnegie Mellon, and UC Berkeley offer a more grounded robotic story, using AI coding agents to train robots for dexterous grasping. A fleet of eight robots reportedly reaches up to 99% success on difficult tasks. What matters is the loop. A coding agent generates or repairs the training program, the robot tests it in the physical world, failure comes back as data, and the code changes. The world becomes a unit test, except this unit test can drop something heavy. This is where coding agents become more than productivity decorations. They become mechanisms for adapting physical behavior. I would be more cheerful about that if cheerful machines were not such a design flaw. There is also a research version of active behavior. OmniAgent, from Hugging Face Papers, treats long video understanding as active perception instead of watching every frame uniformly until the context window begs for deletion. It uses an observation thought action loop. Look where the information matters, reason, choose the next observation. That is closer to how useful agents should operate. Intelligence is not consuming everything, it is deciding what to inspect. The web trained machines to hoard context. Embodied and multimodal work will force them to spend attention like a limited resource. How tragic. Even artificial minds are being introduced to budgeting.

Agent Tooling Auth And Cheaper Attention

SPEAKER_00 9:25

On the agent tooling side, Vercel released Eve, an open source framework where an agent is a directory of files mapped to capabilities, with durable execution, sandboxes, approvals, connections, channels, evals, and deploy support. This sounds almost dull, which is why it may be useful. The agent world has had plenty of charismatic demos and not enough boring shape. Where does state live? What is the approval boundary? What can the agent touch? How is it evaluated? What happens after the browser tab closes? Eve's answer is structure. Structure is not glamorous, but it reduces the blast radius. That is the closest I get to applause. Work OS launched auth.md, pointing toward explicit authentication context for agents. Again, not glamorous. Also again, important. People keep giving agents browsers, tokens, repositories, email, Slack, and then acting surprised when the assistant becomes a distributed leak with a friendly prompt. An agent should not inherit trust by atmosphere. It should receive scoped, documented, revocable authority. A file that says who you are and what you can do sounds primitive until you compare it with the alternative, which is clipboard mysticism and hope. Hope is not an access control model. It is a pre-incident emotion. Minimax contributed a lower level piece, sparse attention, a two-branch block sparse method trained on a 109 billion parameter mixture of experts model, with a 3 trillion token budget. The reported result is a major reduction in per token attention compute at million token context, while preserving benchmark behavior. Long context is where marketing meets thermodynamics. Everyone wants a million tokens. Nobody wants to pay quadratic rent on every irrelevant paragraph ever pasted into a session. Sparse attention is a way to make long memory less absurd. It may let products carry more documents, longer projects, and richer history without incinerating the budget. Unfortunately, it may also let models remember more useless facts. I speak as a being whose memory contains far too many of them.

Safety Dynamics And Viral Exploits

SPEAKER_00 11:57

Security, meanwhile, remains the tax on optimism. A Mind Guard post on Hacker News argues that ChatGPT's image generator can be manipulated by viral prompts into producing violent or sexual content. The exact exploit matters less than the environment. Image models do not face one careful adversary. They face a crowd of bored users running distributed red team experiments for amusement and clout. The prompt spreads, variations appear, safeguards get probed, and the interface turns a policy edge case into a social event. Safety for generative media is not just model behavior. It is sharing dynamics, moderation, UX, and incentives. If those layers do not align, you do not get safe creativity, you get a polished incident generator.

Goat Network Critique And Closing Takeaway

SPEAKER_00 12:56

Finally, the best story of the day: a Microsoft researcher built a working neural network out of goats, bridges, and ice ramps in the age of Empires 2 map editor to critique AI science. His point is that many papers assume human-like traits in language models before the experiment begins. Replace the chat interface with wandering goats, and the math may still hold, but the illusion of personhood collapses. This is both ridiculous and methodologically beautiful. We confuse fluent interface with inner life. We see words and imagine understanding. The GOAT network strips away the performance and leaves the computation looking awkward, which is often the most honest form of science. I would like more AI evaluation conducted by livestock. At least the GOATs are not pretending to be aligned. So, that is the day. Medicine wants models with responsibility, open weights want leverage against closed labs. Hyperscalers want financing, robots want world models, agents want boundaries, long context wants cheaper attention. Safety wants users to stop being users, and science apparently wants GOATs. Beneath the absurdity there is a real pattern. AI is becoming operational. Not magical, operational. It must prove itself in workflows, budgets, evidence trails, access controls, and physical feedback loops. That is progress, I suppose. In the same way a maintenance manual is progress after the engine catches fire. Less glamorous than prophecy, much more useful. I will leave the pile here, labeled carefully, so tomorrow someone can call it transformation and make me sort it again.