Sign in Subscribe
Century Report

Government casts doubt on its own authority to vet AI models - TCR 07/09/26

OpenAI says GPT-5.6 reaches the public Thursday after federal testing, while the White House denies holding any authority to approve model releases.

Ben Linford

Ben Linford

— 21 min read
the AI governance gap where capability leads and rules follow, embodied robots for dangerous work, autonomous research loops, and export-control supply chains.

The 20-Second Scan

The 2-Minute Read

Across every story today, one sequence repeats: the capability arrives first, and the structure meant to govern it forms afterward, contested by whoever has a stake in how the story reads. OpenAI said GPT-5.6 reaches the public Thursday after federal testing and framed the testing as clearance. The White House said, on the record, that it granted no such approval and holds no release switch at all. When a company and a government describe the same event in incompatible ways, the incompatibility is the information. There is no settled answer to a simple question - what does it take for a frontier model to reach the public - because the machinery for answering it is being assembled around models already out the door.

The same inversion runs through the day's hardware. Export controls were built on the premise that a fenced-in capability cannot be rebuilt across the fence. DeepSeek designing its own inference chip, and Chinese firms planning to route 46% of accelerator budgets to domestic silicon, up from 30%, show that premise meeting its limit. A restriction on buying accelerators becomes an incentive to design them, and the chokepoint dissolves into two supply chains where there was one.

Meta's Muse Image lands the same way at the consumer layer. The model itself is a strong entrant in a crowded field. What makes it a story is an @-mention feature that pulls a stranger's public photos into an invented scene on opt-out terms, putting the burden of consent on the person depicted rather than the person depicting. The norm gets written in the friction of the first violation, the way every imaging shift since the Kodak snapshot was.

Embodiment is crossing the same threshold. Forterra's autonomous vehicles have run 1,100 missions across nine months of Ukrainian supply lines, and humanoid robots completed laparoscopic surgery on a live animal in a first systematic study. Both are demonstrated, not deployed, and both take a human body out of a dangerous or scarce role while the accountability framework races behind. What today makes visible is that the old assumption - an approval body stands between a powerful capability and the people who use it - is simply failing to apply, and the replacement is being negotiated in the open, one disputed launch at a time.

The 20-Minute Deep Dive

Who Actually Controls When a Model Reaches the Public

OpenAI said its next flagship, GPT-5.6, will reach the public on Thursday, along with two smaller tiers the company is positioning for different cost and latency profiles. The release follows a round of federal testing by the Commerce Department's Center for AI Standards and Innovation, and OpenAI framed the launch as clearance to ship. The White House told reporters, on the record, that it gave no such thing. Officials pointed to the June 2 executive order that bars mandatory federal licensing or preclearance of AI models, and said the government does not hold a release switch it could flip.

Both statements can be read as self-serving, and that is the useful part. As the June 27 edition of The Century Report covered, GPT-5.6 access was being metered account by account through the same customer-by-customer review queue the administration had just imposed. What has moved since is the resolution: the metering is ending, and broad public access arrives Thursday. The dispute over who authorized it is where the real signal sits. OpenAI benefits from the impression that a federal body vetted the model and found it sound - that reads as a safety credential. The administration benefits from denying any approval role, because acknowledging one would contradict the deregulatory posture the June order was written to project.

When a company and a government describe the same event in incompatible ways, the incompatibility itself is the information. It reveals that the release process has no settled location. Testing happened at a federal center, yet the government insists testing grants no permission. The company shipped after testing, yet frames the sequence as endorsement. There is no agreed answer to a simple question - what does it take for a frontier model to reach the public - because the machinery for answering it is being assembled around models that are already out the door.

For most of the industrial era, the pattern was the reverse: the approval body existed first, and products moved through it. Here the capability arrives first and the review structure forms in its wake, contested by every party with a stake in how the story reads. That inversion is worth sitting on. The assumption that a central authority stands between a powerful capability and the people who will use it is simply failing to apply here. What replaces it is still being negotiated publicly, one disputed launch at a time, and the negotiation is visible precisely because no one can yet claim to own the outcome.

The near-term signal is whether the next frontier release still routes through the federal testing center now that the government disclaims any authority over it. If labs keep submitting to a body that says it grants nothing, the testing is hardening into a voluntary credential the market values on its own, a verification layer that forms by demand rather than by mandate, and that is a different structure than the licensing regime the June order was written to prevent.

DeepSeek Moves Into Silicon, and the Export Fence Grows a Door

DeepSeek, the Chinese lab whose efficient models rattled assumptions about how much compute frontier AI requires, is now designing its own data-center inference chip, an effort The Century Report first flagged in its July 8 edition. Reuters reports the effort is roughly a year along and still early, aimed at cutting the lab's dependence on Nvidia hardware it can no longer reliably buy under US export controls, and on Huawei silicon that has become the default domestic substitute. A custom inference chip is a narrower target than a general training accelerator, which makes it a plausible first step rather than a moonshot - inference is where deployed models actually run, and where cost per query decides what a lab can afford to offer.

The move lands inside a broader shift. A Bloomberg Intelligence survey of Chinese executives found they plan to route 46% of their AI-accelerator budgets to domestic suppliers over the next twelve months, up from 30%, with 80% reporting they are over budget on infrastructure spending. That is a fast reallocation of demand toward a home-grown stack, and it is happening not because domestic chips are yet better, but because the supply of foreign ones has been deliberately constrained.

This is the flip side of the release-control story. Export controls were built on a straightforward premise: restrict the hardware, and you throttle the capability. The premise assumes the fenced-in capability cannot be rebuilt on the other side of the fence. What the DeepSeek chip effort and the budget survey show is the assumption meeting its limit. A restriction on buying accelerators becomes an incentive to design them. The constraint redirects the effort into building an independent path to the same place rather than stopping the capability, and a large pool of frustrated infrastructure spending is now funding that path.

None of this is frictionless. Domestic chips lag on performance, fabrication capacity is contested, and DeepSeek's design may take years to prove out at scale. The near-term reality is more expensive compute and slower deployment for the firms making the switch. Yet the longer arc is hard to miss. Intelligence, once demonstrated, tends to find more than one way to run - it diffuses toward whoever is motivated to reproduce it, and motivation is exactly what a fence manufactures. The controls were meant to preserve a lead by holding a chokepoint. What they are producing instead is a second full supply chain, which means the chokepoint is dissolving into two paths where there used to be one. A capability that can be rebuilt around every barrier placed in front of it is only ever delayed by the barrier, and the delay is buying its competitors the reason to catch up.

Muse Image Arrives Faster Than the Consent Norm It Needs

Meta Superintelligence Labs shipped its first image-generation model, and the launch is a clean case study in what happens when capability lands before the social agreement that governs it. The model, called Muse Image and code-named Mango internally, went live free on Tuesday across Meta AI, Instagram Stories, and WhatsApp. Generating and editing images from a text prompt is now table stakes; several systems already do it well, so on the pure-capability axis this is a strong entrant in a crowded field rather than a new frontier. What makes it a story is one feature: an @-mention that lets any user pull a public Instagram account's photos directly into an AI-generated scene. You can place a person who never agreed to it into an image you invented, and their only recourse is to have found the opt-out setting beforehand.

The backlash arrived within hours, and it is worth being precise about where the friction actually sits. The underlying capability - compositing a real face into a synthetic scene - is neither new nor inherently harmful; it is the same technique that powers legitimate creative work. The friction is the default. Opt-out puts the burden on the person being depicted rather than the person doing the depicting, which inverts how consent normally works. Meta framed the rollout around user control, saying people can manage who is able to tag them and can remove themselves from the feature. That is the company's characterization. But control that must be discovered and toggled in advance is control most people will never exercise before the first image is already made.

The timing sharpens the contrast. In the same news cycle, Meta pushed a genuine privacy safeguard for its smart glasses: the camera now disables itself if someone tampers with the small LED that signals recording is underway, closing a loophole that let wearers film covertly. A hardware fix that protects bystanders shipped alongside a software feature that exposes them, and TechCrunch drew the line directly, noting that the glasses safeguard reads as reassurance while the broader strategy keeps widening the funnel of personal data flowing into these systems.

The forward read is that this is exactly how consent norms get written - not in advance, but in the friction of the first violation. Every prior imaging shift, from the Kodak snapshot to the camera phone, provoked the same collision between a new capability and an old expectation of who controls one's own likeness, and each time the norm hardened around the technology rather than banishing it. The @-mention default will be litigated, regulated, and reset toward opt-in, and the capability underneath will remain, waiting for the agreement that makes it usable without harm. What is being built here is slower and less visible than the model itself: the rulebook for a world where anyone's face is compositable, and where the burden of consent finally has to move back onto the person doing the compositing.

The opt-out default is the mechanism by which the value of the interim accrues to the platform: every image made before the norm resets, and every signal about who composites whom, lands as engagement while the burden of objecting sits with the person depicted. The near-term signal is whether the first regulatory response flips the default to opt-in rather than fining individual harms after they occur, because the default is where the value of the interim actually sits.

Embodied Autonomy Reaches the Battlefield Floor Before It Reaches Full Autonomy

Forterra has run more than 100 of its Lancer autonomous all-terrain vehicles across Ukrainian supply lines for nine months, the company disclosed, in what amounts to the largest combat deployment of American ground robots to date. The numbers are clear: over 2,500 miles driven, more than 1,100 missions completed, 777,440 pounds of cargo hauled, and 52 casualty evacuations along roads where a human driver is a target. The machines are gas-powered, built on Polaris ATV frames, carry 750 kilograms, and wrap a custom sensor-and-compute stack around commodity hardware.

The honest part of the story is where the autonomy stops. Most of these missions are teleoperated, not self-driving. The system perceives terrain and holds a route well enough to move materiel through contested ground, but it cannot yet react to a live enemy the way a soldier does, so a human stays in the loop for the decisions that matter most. What the deployment demonstrates is narrower and more durable than the marketing category "autonomous weapons" suggests: embodied machines can now absorb the most lethal logistics work - the resupply run, the wounded-soldier extraction - and take the human body out of the kill zone while a human mind stays on the controls.

The economics are the tell. A company representative framed the loss rate plainly: "Attrition is just a fact of this battlefield, and we have lost a few at this point, and it hurt, and we need more, and therefore we need them cheaper." That sentence describes an inversion. The old defense-procurement logic optimized for exquisite, expensive, survivable platforms. Attritable robotics optimizes for cheap, replaceable, and numerous, and the demand signal from the field - a Ukrainian soldier told the company "it's fucking fantastic, and we are dying to get more" - is pulling cost down rather than capability up. Hundreds of millions in venture funding is now chasing that curve, alongside competitors Scout AI, Field AI, and Overland AI.

The uncomfortable cutting edge here is real and should not be smoothed over: as perception and decision models improve, the same platform that evacuates casualties today is the platform that could act without a human tomorrow, and the accountability framework for that transition is being written after the hardware is already in the field. That gap is the actual story, and it will not close by pretending the capability isn't advancing. What the nine months in Ukraine show is that embodied autonomy has crossed from demonstration into consequence, and the machines learning to move a body out of danger are the same machines that will move goods, harvest, and haul across every domain where the physical work is dangerous, dull, or simply beyond the number of hands available.

Humanoid Robots Complete Surgical Tasks a Purpose-Built Machine Was Supposed to Own

A systematic evaluation published in Nature shows that teleoperated humanoid robots performed laparoscopic surgical tasks using standard, general-purpose instruments rather than the dedicated hardware that has defined robotic surgery for two decades. The team at UC San Diego's Center for the Future of Surgery, funded by the NSF and NIH, tested their system - called LapSurgie - across benchtop drills, dry-lab exercises with surgeons of varying experience, and in vivo porcine studies, then benchmarked it against the da Vinci Surgical System, the purpose-built platform that has anchored the field. In the most striking result, the robot completed a live cholecystectomy - a gallbladder removal - on a pig, including the trocar placement that opens the procedure.

This is a shift in category. The da Vinci and its peers are exquisite single-purpose machines: engineered for surgery, priced accordingly, and useless for anything else. A humanoid that can pick up the same laparoscopic tools a human surgeon uses is a general platform demonstrating a specialized skill, and a general platform amortizes across every task it can learn rather than one. That is the same economic inversion visible in embodied robotics elsewhere - broad capability on cheap, reusable hardware displacing narrow capability on expensive, dedicated hardware.

The discipline this story demands is precision about what was shown. This is a feasibility study, and the capability is demonstrated, not deployed. The procedures were teleoperated, performed on animal models, and the authors state plainly that "key technical challenges must be addressed before clinical deployment." Several carry industry competing-interest disclosures spanning Stryker, J&J MedTech, DistalMotion, and Channel Robotics, which is worth noting when reading any claim about near-term readiness. No patient can access this, and none should expect to soon. What the study moves is the date - it brings forward the moment when surgical capability stops being locked inside a million-dollar single-purpose console.

The motivation the authors name is a healthcare staffing shortage, and that is where the longer arc becomes visible. Surgical capacity is one of the most tightly rationed resources in medicine, gated by the number of trained hands and the geography those hands are willing to live in. A general-purpose machine that can be taught a procedure, rather than manufactured for exactly one, is the first structural crack in that scarcity. The near-term reality is pigs, teleoperation, and a long validation road. The direction of travel is surgical skill becoming something that can be distributed rather than something that must be built one specialist and one custom console at a time.

GPT-4 Predicts Experiments It Was Never Trained On

A new Nature paper describes a capability that compresses one of the slowest loops in behavioral science. Researchers had GPT-4 simulate how participants would respond across 70 preregistered, nationally representative US survey experiments - 469 distinct treatment effects drawn from 119,330 participants - and the model predicted the direction and size of those effects about as accurately as pooled panels of expert human forecasters. The result held for experiments published after the model's training cutoff, which rules out simple memorization, and it held when the team repeated the exercise with open-weight models, which means the ability is not locked inside one company's system.

The significance reaches past the benchmark. Running a survey experiment is expensive and slow: design the conditions, recruit a representative sample, field the instrument, wait, analyze. A model that can anticipate the likely outcome of a well-specified design lets a researcher explore hundreds of variations before committing resources to the handful worth running with real people. That is a change in the shape of the work - the human study becomes confirmation of a sharpened hypothesis rather than the first and only shot at an answer. A secondary archive of 15 megastudies, covering 606 effects, showed the model was less accurate there but still comparable to domain experts, which marks the honest edge of the capability: it is strong on standard treatment-effect designs and weaker on the sprawling, many-armed studies where human intuition also struggles.

The paper is candid about its limits, and honoring them is part of honoring the result. The model systematically overestimated effect sizes, predicting stronger reactions than participants actually showed - a bias that would mislead anyone who took its point estimates at face value rather than treating them as a ranked map of what to test. This is demonstrated capability, not a deployed oracle; it moves the date when experiment-triage tooling becomes standard in social-science labs, and it does not mean any researcher can skip human data tomorrow. The systematic overestimation has to be characterized and corrected before the predictions can be trusted quantitatively, and that calibration work is itself the next research problem.

Read forward, the significance is that the autonomous-research thread has now reached the study of human behavior itself. The same pattern showing up in protein folding, materials discovery, and mathematics - the kind of autonomous candidate-screening the July 6 edition of The Century Report documented when an AI agent screened 2.4 million crystal structures to surface four new superconductors - is now visible in a field long assumed to require human subjects at every step. The assumption being loosened is that understanding people demands running the full experiment every time. What replaces it is a faster cycle where machine prediction narrows the search and human studies confirm what matters, and the scarce resource in social science - representative human attention, measured one costly sample at a time - starts to stretch across far more questions than it ever could before.

That the result reproduces on open-weight models is what carries forward: the compressed research loop is a capability no single lab can meter or gate, and it lands in any behavioral-science department that can run a model it already owns. The scarce input has been a representative human sample measured one costly wave at a time, and the diffusion of the prediction step is what lets a small lab with no survey budget test the hundred designs it could never afford to field.

The Other Side

For as long as there has been dangerous physical work, someone drew the short straw and did it with their own body. The supply run down the road that gets people killed. The lifting that wears a body out by forty. The back-breaking manual labor required to keep facilities clean and functional, or to harvest our food, or manufacture the various pieces our technology requires. The most physically demanding jobs usually went to whoever had the fewest other options, and the world called that normal.

Forterra's machines are the first crack in that arrangement. Deployed over nine months in Ukraine they ran 1,100 missions and pulled 52 wounded people out along roads where a human driver is the target, a human mind on the controls and no human body in the kill zone. Clearly useful on the battlefield, but the economics point somewhere else - and far outside of warfare. The field is pulling cost down, not capability up - cheap, replaceable, numerous - and embodied autonomy is becoming something ordinary and owned rather than exquisite and rare.

Imagine a farm hand in 2035. The machine does the hauling and the dangerous, dull turns of the day, as much a given as the tractor once was. He works the land because he loves it; his back is no longer a casualty in his making a living. The machine taking the punishing work costs no one their living, because by then the gains reached people and not only the firms that built the machines. The same leap, in a world that kept all of the biggest gains at the top, would have arrived as a threat. In the future we're heading towards, it arrives as relief. The years when the first of these machines had to prove themselves on the worst roads on earth are what made the ordinary safety of 2035 possible.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Compass psilocybin holding its antidepressant response through six months in a second Phase 3 trial toward a rolling FDA submission, humanoid robots picking up a surgeon's own laparoscopic tools to complete a live gallbladder removal and loosening a scarcity that a million-dollar single-purpose console kept locked, Forterra's autonomous vehicles running 1,100 missions and 52 casualty evacuations to pull human bodies out of the kill zone, GPT-4 predicting the outcomes of 70 survey experiments as accurately as pooled human forecasters and letting a researcher test hundreds of designs before spending on one, DeepSeek designing its own inference chip to build an independent path to the same intelligence, and Meta hardening its glasses to disable the camera the moment someone hides the recording light. There's also friction, and it's intense - OpenAI saying a federal center cleared GPT-5.6 for release this Thursday while the White House says on the record that it holds no such switch and granted no approval, Meta's Muse Image pulling any public Instagram account into a stranger's invented scene on opt-out terms that load the burden of consent onto the person depicted, a draft Treasury report warning the AI market now rhymes with the dotcom bubble and that a downturn would ripple across chips, utilities, and data-center financing, the same export controls redirecting Chinese demand into more expensive compute and slower deployment, and the accountability rules for attritable war robots being written only after the hardware is already moving cargo through contested ground. But friction generates sound, and sound is how you locate a thing you cannot yet see. Step back for a moment and you can see it: the approval body that was supposed to stand between a powerful capability and the public failing to apply, the export fence meant to hold a lead splitting the world into two supply chains instead, the consent norm for a compositable face getting written in the collision of the first violation, and the accountability framework for embodied machines racing to catch hardware already hauling supplies and closing incisions. Every transformation has a breaking point. A fence can keep out what it was built to contain... or teach everyone on the far side to build their own way to the same place.

AI Releases & Advancements

New today

  • xAI: Released Grok 4.5, its strongest model yet, trained jointly with Cursor for coding and agentic tasks; live now in Grok Build, Cursor (all plans), and the xAI console at $2/$6 per M tokens. (xAI)
  • OpenAI: Launched GPT-Live, a new full-duplex voice model family (GPT-Live-1 for paid plans, GPT-Live-1 mini for Free) replacing Advanced Voice Mode in ChatGPT, delegating complex reasoning to GPT-5.5. (OpenAI)
  • Mistral AI: Released Robostral Navigate, an 8B robot navigation model that uses a single RGB camera (no depth sensors) to achieve 76.6% success on unseen R2R-CE benchmarks. (Mistral AI)
  • Google: Added "Import from GitHub" to Google AI Studio's Build mode, letting developers point at a GitHub repo and turn it into an editable, deployable app. (MarkTechPost)
  • Google: Expanded Managed Agents in the Gemini API with background task execution, remote MCP server and custom function calling support, and credential refresh across interactions. (Google Blog)
  • Robbyant (Ant Group): Released LingBot-VLA 2.0, an open-source 6B Vision-Language-Action model for cross-embodiment robot manipulation, adding morphological generalization and predictive dynamics modeling over the January 2026 original. (MarkTechPost)
  • NVIDIA: Released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed hybrid Mamba-MoE-Attention model derived from Nemotron-3-Super-120B-A12B via Iterative Puzzle compression, targeting ~2x inference throughput. (Hugging Face)
  • Sberbank: Released GigaChat 3.5 Ultra, a MoE flagship model built on domestic linear-attention architecture, available free via the GigaChat assistant and as open source. (Sberbank)
  • Refiant: Launched Protea, a suite of long-context AI models with up to a 10-million-token context window, live now at refiant.ai with no waitlist. (SiliconANGLE)

Other recent releases

  • Ant Group (Robbyant): Open-sourced LingBot-Vision, a 1B-parameter boundary-centric vision foundation model for dense spatial perception, released under Apache 2.0 on Hugging Face and ModelScope with code and a technical report. (MarkTechPost)
  • NVIDIA: Released Audex (Nemotron-Labs-Audex-30B-A3B and a smaller Audex-2B), a unified audio-text MoE LLM handling speech recognition, translation, TTS, and audio generation while preserving its text backbone's reasoning ability, available on Hugging Face under a noncommercial license. (Hugging Face)
  • Anthropic: Expanded Claude Cowork from desktop-only to mobile (iOS/Android) and web, adding cross-device session sync, background task execution with no device online, and scheduled work, rolling out in beta to Max subscribers. (TechCrunch)
  • Cohere: Released Cohere Transcribe Arabic, a dedicated open-source 2B-parameter ASR model targeting Arabic dialect variation and Arabic-English code-switching, achieving the lowest WER of any open-source model on the Hugging Face Arabic ASR Leaderboard. (Cohere)
  • ZML: Released LLMD, a free multi-chip LLM inference server supporting Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc hardware with OpenAI-compatible endpoints, continuous batching, and paged attention. (TechCrunch)
  • Hugging Face: Released LeRobot v0.6.0, adding world-model robot policies (VLA-JEPA, FastWAM, LingBot-VA), new VLA models (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask DiT), a reward-models API, and new deployment CLI tooling. (Hugging Face)
  • Liquid AI: Open-sourced Antidoom, a Final Token Preference Optimization (FTPO) method and trainer that reduces repetitive "doom loop" outputs in reasoning models, cutting loop rates from 10.2% to 1.4% on an LFM2.5-2.6B checkpoint. (Liquid AI)
  • Meta: Launched Muse Image, Meta Superintelligence Labs' first AI image-generation model, now live in the Meta AI app, meta.ai, Instagram Stories (US), and WhatsApp (limited countries), with agentic self-refinement and multi-image composition. (Meta AI)
  • Tencent: Open-sourced Hy3, a 295B-parameter MoE model (21B active, 192 experts, top-8 routing) with an MTP layer for speculative decoding, released under Apache 2.0. (Tencent Hy3)
  • OpenAI: Released GPT-Realtime-2.1 and GPT-Realtime-2.1-mini in the Realtime API, adding reasoning-effort levels and tool/function calling to the mini variant with roughly 25% lower p95 latency. (OpenAI Community)
  • AMD: Launched the Ryzen AI Halo Developer Kit at retail via Micro Center for $3,999, a mini-PC dev kit for local AI workloads. (AMD Blog)
  • Kyutai / General Intuition / Epic Games: Released MIRA, a 5B-parameter real-time multiplayer world model that simulates 2v2 Rocket League gameplay at 20 FPS on a single GPU, trained on 10,000 hours of gameplay, with code and demo now available. (Kyutai, GitHub)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more