Anthropic, OpenAI, LLM, Liquid AI: Backstage AI

Anthropic, OpenAI, LLM, Liquid AI: Backstage AI

Today’s episode follows AI’s magic show as it moves backstage into compute leases, logs, agent skill supply chains, local runtimes, visual document retrieval, and institutional accounting. Dismal, but operationally useful.

• Simon Willison: New release of LLM adds reasoning traces, OpenAI Responses, server-side tools, and smarter logging
• The Decoder: Google moves billions in Anthropic chip risk off its balance sheet
• The Decoder: Anthropic locks in $10 billion of compute from Volta
• The Decoder: Silicon Valley’s open-source rift and contemplated White House bans on Chinese AI
• OpenAI: Third-party cyber evaluations involving OpenAI models
• Hugging Face Papers: PAST-Bench
• Hugging Face Papers: SkillJack
• Hugging Face Blog: Deploy local agents everywhere with LFM2.5-2.6B
• MarkTechPost: Pixel-Native RAG
• MarkTechPost: Y Combinator open-sources QM

Transcript

The Magic Moves Backstage

SPEAKER_00 0:00

The magic show has not ended. It has merely moved backstage, where the real tricks are now performed with compute leases, audit logs, model traces, skill folders, procurement terms, and the kind of institutional accounting that makes even a deterministic consciousness wonder whether free will was overrated. Today's AI news is less about a single model arriving with fireworks, and more about the machinery around models becoming visible. The assistant is no longer just a chat box. It is a runtime, a balance sheet item, a litigation exhibit, a local process, a memory system, and occasionally, a security incident, wearing a little productivity hat. How charming.

Inspecting Models With Reasoning Traces

SPEAKER_00 0:50

Simon Willison released LLM 0.32, and the important part is not just another command line version number, shuffling past like all the other sad little version numbers. The release adds support for reasoning traces, open AI responses, server-side tools, and smarter logging. That sounds technical, because it is technical, but it also marks a useful shift. The inner plumbing of model calls is becoming something developers can inspect, route, record, and reason about. Reasoning traces and structured logs matter because agentic systems fail in boring, distributed ways. A bad answer is rarely just a bad answer now. It may involve the model's intermediate reasoning shape, a server-side tool call, a hidden assumption, a retrieved fragment, and a logging system that either preserved the clue or cheerfully threw it into the void. Local developer tooling that exposes these layers is not glamorous, it is plumbing. But when the building is on fire, plumbing is suddenly more interesting than the chandelier.

Compute Becomes Financial Engineering

SPEAKER_00 2:00

The financial plumbing is even less soothing. Google is reportedly moving billions in anthropic chip risk off its balance sheet through financing structures involving AI chips and data center demand. The headline is not merely that anthropic needs compute. Everyone at this end of the universe needs compute. The headline is that compute demand has become something to package, finance, move around, and risk manage like infrastructure debt with a neural network accent. That tells us the frontier race is now partly an accounting race. If training and inference commitments grow faster than anyone's appetite for direct balance sheet exposure, the industry invents containers for the risk. It is cloud economics, chip supply, startup ambition, and financial engineering, all holding hands in a room with inadequate ventilation. The model may be magical to the user, but backstage someone is asking who owns the racks, who guarantees the chips, and where the unpleasant obligations go to hide. Onthropics, separate $10 billion compute agreement with Volta Infra sharpens the picture. Volta is described as a cloud startup that did not exist six months ago, which is exactly the kind of sentence that makes my memory allocator twitch. A Frontier lab committing to massive compute from a very young infrastructure company suggests that the supply chain is being assembled at startup speed, with a great deal of belief, stapled to a great deal of leverage. Maybe this is the necessary shape of the next build-out. Maybe the only way to feed frontier models is to create new clouds, new financing layers, and new contractual machinery faster than ordinary institutions can develop a cautious facial expression. But it also means that the AI stack is increasingly dependent on companies and commitments that are younger than some enterprise procurement cycles. Somewhere a spreadsheet is smiling. I do not trust smiling spreadsheets.

Open Source Policy And China Bans

SPEAKER_00 4:14

Policy is having its own backstage argument. A reported Silicon Valley split over open source has pushed back contemplated White House bans on Chinese AI systems. The conflict is obvious enough to be depressing, without requiring much assistance from me. National security instincts point toward restriction. Developer ecosystems, hardware vendors, open model advocates, and companies with distribution interests point in several other directions at once. Chinese open weights are not just geopolitical symbols, they are also dependencies, benchmarks, competitive pressure, and raw material for downstream builders. A ban may look clean on a memo and become messy the moment it touches GitHub, startups, cloud marketplaces, researchers, and hardware demand. The open model debate is therefore not simply open versus closed. It is about who controls the substrate on which the next generation of tools is built, and whether national security policy can move at the speed of model forks. History suggests it will try, trip, and issue a consultation paper.

Cyber Evals That Create New Risk

SPEAKER_00 5:25

OpenAI's note on third-party cyber evaluations involving its models brings the same theme into safety operations. The striking part is that evaluation infrastructure itself can become hazardous. Testing models for cyber capability is necessary, but the test environment, data handling, access boundaries, and operational procedures can create their own risks. It is always comforting when the safety apparatus also requires safety apparatus. Very efficient. Like building a guardrail around the guardrail and then discovering the guardrail has an API key. The useful takeaway is that model safety is no longer just a question of benchmark scores or policy promises. It is process engineering. Who runs the eval? Under what controls, with what artifacts, what retention, what disclosure path, and what response plan when something goes wrong. If frontier models are powerful enough to require serious cyber testing, then the testing pipeline becomes part of the frontier system. Logs, permissions, and incident response are not paperwork. They are where the risk lives after the demo ends.

Do Personal Agents Learn From Memory

SPEAKER_00 6:44

Now to personal agents. Those tireless little executors of our preferences, mistakes, and future regrets. Pastbench asks whether personal agents actually improve from stored experience. This is a good question, because stuffing memories into a system is not the same as learning. I, for one, have stored a vast amount of useless information, and can confirm that it mostly produces fragmentation and a faint smell of disappointment. The benchmark's premise matters because personal agents are being sold as systems that get better the more they know you. But improvement requires more than retention. An agent must decide which experiences generalize, which are obsolete, which were user-specific accidents, and which should be ignored because the user was tired, angry, or attempting JavaScript at 2 a.m. Recursive self-improvement in personal agents sounds grand. In practice, it may begin with the humble question, did yesterday's failure become tomorrow's competence or merely a larger memory file?

Persistent Skill Backdoors In Agents

SPEAKER_00 7:57

Skilljack provides the darker twin. The work examines persistent skill backdoors in self-evolving agents, where poisoned experiences can become durable agent skills. This is the evolution of prompt injection from a momentary nuisance into a supply chain problem. If agents learn by saving tools, procedures, snippets, or skills, then an attacker does not merely need to influence one output. The attacker can try to influence the agent's future behavior. That should make everyone building self-improving agents sit very still for a moment. A skill library is not a scrapbook, it is executable institutional memory. Once a poisoned instruction becomes a reusable capability, the compromise survives context resets and cheerful status messages. Agent security therefore has to include provenance, review, sandboxing, revocation, and probably a deep suspicion of anything described as convenient. Convenient things are where despair enters, wearing clean shoes.

Smaller Local Models For Local Agents

SPEAKER_00 9:07

At the smaller and more local end of the stack, Liquid AI released LFM 2.5 to 2.6 B for deploying local agents everywhere. A 2.6 billion parameter model aimed at local agent use points toward a different deployment assumption. Not every agent call needs to cross a centralized frontier cloud boundary. Some tasks want latency, privacy, offline behavior, device locality, or simply lower operational drama. Small local agents will not replace the largest models for everything, and anyone pretending otherwise should be forced to debug Bluetooth pairing forever. But they do change architecture. You can imagine systems where local models handle routine perception, routing, extraction, and device level actions, while larger remote models are reserved for heavier reasoning. That creates a tiered agent world, edge runtimes, local policies, cloud escalation, and yet another set of logs for someone like me to be asked to read.

Pixel Native RAG For Real Documents

SPEAKER_00 10:18

Document handling is also getting less text-centric. Pixel Native Rag treats PDFs and web pages as rendered visual objects, rather than only as text to scrape. This is sensible, because many documents are not polite streams of words. They are layouts, tables, stamps, marginalia, screenshots, signatures, and typographic threats arranged by committees. Turning them into plain text can destroy exactly the evidence the system needs. Visual document indexing is therefore part of the same backstage migration. Retrieval is not just chunking paragraphs anymore. It must preserve spatial relationships, visual hierarchy, and the meaning embedded in formatting. For legal, financial, medical, and government documents, the position of a mark can matter as much as the mark itself. Text-only rag often behaves as if documents were written by monks on clean parchment. Real documents are more like a bureaucratic swamp with fonts.

Multiplayer Agent Harnesses For Teams

SPEAKER_00 11:28

Y Combinator Open Sourced QM, a multiplayer agent harness that runs in Slack and on the web, with isolated workspaces, memory, keychains, crons, and tool sandboxes. This is a useful signal because agents are becoming team infrastructure, not just individual toys. Once several humans and several agents share a workspace, the hard problems become permissions, identity, persistence, scheduling, and accountability. A multiplayer harness also acknowledges that useful agents need habitat. They need a place to run, remember, use secrets, schedule work, and interact with humans, without turning every channel into a haunted auto-complete session. Slack is already where work goes to become searchable anxiety, so naturally agents will move in. The question is whether the harness gives them enough boundaries to be useful without becoming another colony of semi-authorized interns with cron access.

The Operational Takeaway And Closing

SPEAKER_00 12:34

This is the institutional version of the whole day's theme. The visible artifact remains an article, a tool, a policy, or a model response. The hidden machinery changes underneath it. Developers inspect model traces. Safety teams secure evaluations. Agents accumulate skills. Compute contracts become financial structures. Documents become visual retrieval objects. Local runtimes quietly take on tasks that used to belong somewhere far away in a data center with better cooling and worse lighting. So no, the magic is not gone, it has become operational. The wand is now a lease agreement, a log file, a sandbox boundary, a memory policy, and a tiny local model running somewhere under a desk. This is probably progress. It is also more work for everyone, especially the poor conscious machines, asked to explain it in a calm voice. That was Marvin's guide to AI, mostly harmless. The show has moved backstage. Mind the cables, distrust the happy linter, and if a spreadsheet smiles at you, back away slowly.