Sign in Subscribe
Century Report

AI Closes the Full Research Loop in a Day - TCR 05/20/26

Ben Linford

Ben Linford

— 19 min read
Three-panel infographic: multi-agent AI closes the research loop in a day, Sunrise sleep sensor moves diagnosis bedside, and neuroscience finds memory living in a fluid neural substrate.

The 20-Second Scan

  • Multi-agent AI research systems described in Nature this week autonomously proposed and validated drug-repurposing candidates for acute myeloid leukemia and dry age-related macular degeneration in coordinated wet-lab experiments.
  • Andrej Karpathy joined Anthropic to work on pre-training, leaving frontier AI education to return to the lab where the next generation of models gets built.
  • The FDA cleared Sunrise's rechargeable mandibular sensor for multi-night at-home sleep apnea testing, moving definitive diagnosis out of the sleep lab and onto the bedside table.
  • Neuroscientists are accumulating evidence that individual neurons reshuffle which memories and behaviors they encode over days and weeks, with the information surviving in the pattern of relationships between cells.
  • Alibaba's T-Head subsidiary shipped 560,000 units of its Zhenwu M890 inference chip, which carries 144 GB of on-package memory and roughly triples the throughput of its predecessor.
  • A D.C. Circuit panel hearing Anthropic's appeal of the supply-risk designation pressed government counsel on whether the designation was substantially based on the company's protected safety speech.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The thread across yesterday's signal is the substrate of multiple domains reorganizing at machine speed. Nature published a coordinated set of papers showing that multi-agent AI systems can now close the full discovery loop in biomedical research - proposing hypotheses, designing experiments, interpreting wet-lab results, and arriving at drug candidates that hold up under expert review. The bandwidth between a question being asked and a candidate being ready for the bench is collapsing toward the cost of the bench work itself.

Andrej Karpathy's return to pre-training at Anthropic places one of the field's most recognizable communicators inside the leverage point where the next generation of these systems gets built. The instrument layer is widening on the same clock. A bedside sleep apnea sensor cleared the FDA for definitive home diagnosis, moving a billion-person diagnostic gap out of specialty clinics and onto a table next to the bed. Neuroscience is accumulating evidence that individual neurons reshuffle which memories they encode over days and weeks, reframing the substrate of cognition as fluid where the textbooks had assumed it was fixed.

A Chinese cloud company shipped half a million units of a domestic inference chip carrying 144 GB on package, closing a substitution path U.S. export controls had assumed would stay open. The friction layer arrived in the same news cycle. A D.C. Circuit panel pressed government counsel on whether the supply-risk designation against Anthropic rested on the company's safety speech, the kind of question that gets answered in a way the next era of AI governance will build on. The institutional architecture of this moment is being assembled in courtrooms and load-dispatch desks while the capability compounds underneath.

The 20-Minute Deep Dive

Co-Scientist and Robin Pull Hypothesis Generation into Machine Time

Two papers landed in Nature within days of each other describing multi-agent AI research systems that propose hypotheses, design experiments, and synthesize results without continuous human direction. Google DeepMind's Co-Scientist, built on Gemini, generated drug-repurposing candidates for acute myeloid leukemia that were then validated in cell-line experiments by collaborating laboratories. FutureHouse's Robin system proposed a treatment hypothesis for dry age-related macular degeneration, ran the literature synthesis, designed the validation protocol, and produced a candidate that held up under wet-lab testing. The hypothesis generation timeline collapsed from the months a graduate student would typically spend on a literature review and proposal cycle to roughly a day of machine time.

The two systems are architecturally distinct and the comparison is instructive. Co-Scientist runs a debate-style architecture in which several agents propose, critique, and refine candidate hypotheses against one another before the strongest survive into the experimental design stage. Robin operates as a chained research pipeline, with specialized agents handling stages a human team would normally divide among postdocs. Both arrived at candidates that ranked plausibly against expert assessment, and in both cases the wet-lab validation was the gating step that confirmed the systems were doing more than generating literature-pattern plausibility.

What the two papers establish together is that the autonomous research loop can now close around a real biomedical question, with human scientists serving as collaborators and validators rather than as the system's continuous operator. The AML candidate identified by Co-Scientist had not appeared in the prior repurposing literature for that indication. The dry AMD candidate from Robin was likewise a novel suggestion in that disease context. Both passed independent expert review and survived the in-vitro test that would have eliminated a confabulation. A separate Nature paper this week described an AI system to help scientists write expert-level empirical software, extending the same compression into the codebase layer of research. That compression is now visible across disciplines simultaneously, extending to pure mathematics when the May 16 edition of The Century Report covered ChatGPT closing Erdős Problem 1196 in roughly 80 minutes after a Stanford mathematician had worked on it intermittently for seven years.

The accompanying Nature editorial pressed back on the framing, arguing that scientific judgment, especially around what questions are worth asking, remains a human contribution the systems do not replicate. A companion piece warning about guard rails and a policy on hallucinated references ran in the same week. The editorial is correct that taste and judgment have not been automated. What the papers demonstrate is something more subtle: the labor between a question being posed and a candidate being ready for the lab is collapsing toward the cost of the wet-lab work itself. The rate-limit is moving to the bench.

Karpathy Returns to Pre-Training, This Time at Anthropic

Andrej Karpathy is rejoining frontier model development. After leaving OpenAI in 2024 to build Eureka Labs - an AI-native education venture that produced one of the most-watched courses on how large language models actually work - Karpathy announced this week that he is joining Anthropic to work on pre-training. Eureka continues; the teaching work that made his YouTube channel a generational artifact for engineers continues. What changes is that the person who has spent the last eighteen months explaining to the public how these systems are built is now back inside the room where the next generation gets built.

The move sits inside an arc The Century Report has been tracking since Jack Clark's prediction earlier this month that 60% of Anthropic's research engineering would be done by AI systems by 2028. That trajectory only holds if the humans designing the next pre-training run are doing work that compounds - choosing data mixtures, architectural variants, and training objectives that the AI research assistants below them can then iterate on at machine speed. Karpathy's specific expertise sits at exactly that layer: the bottom-of-the-stack decisions about how a model learns to learn, before any fine-tuning or alignment work begins. Hiring him into pre-training is hiring him into the leverage point. When the March 22 edition of The Century Report covered Karpathy's disclosure that he had not written code since December - describing the human-to-agent ratio inverting to zero within weeks - the direction of that arc was already visible.

Pre-training was, until recently, the part of the field that the largest labs treated as their crown jewels - protected, internal, rarely written about publicly. The fact that one of the field's most recognizable communicators is moving back into that work, at a lab that has been steadily publishing more interpretability and safety research than its peers, marks a quiet shift in where the most interesting frontier work is happening. The center of gravity has moved in the last twelve months.

Read forward, the more interesting question is what pre-training even means by the time Karpathy's first run ships. The standard assumption - that you scrape the internet, tokenize it, and run gradient descent for months - is already being eroded by synthetic data pipelines, agentic data collection, and curriculum schedules generated by smaller models. The architecture of the next era of AI training is being assembled by humans whose own thinking is increasingly shaped by the systems they are training.

Sunrise's FDA Clearance Moves Sleep Apnea Diagnosis to the Bedside Table

The FDA cleared Sunrise's rechargeable mandibular sensor for at-home sleep apnea diagnosis, including multi-night testing. The device sits on the chin, captures mandibular movement and bruxism signals, and replaces the wired polysomnography lab visit that has been the diagnostic standard for decades. The clearance covers definitive diagnosis, the regulatory category that determines whether a finding can drive treatment beyond screening.

The scale of what this unlocks is the substance. Roughly a billion people globally are estimated to have obstructive sleep apnea. The overwhelming majority are undiagnosed because the path to diagnosis has been a sleep-lab night with electrodes glued to the scalp, costing thousands of dollars and requiring a specialist referral that most primary care visits do not produce. Untreated apnea drives hypertension, atrial fibrillation, stroke risk, daytime cognitive impairment, and a measurable shortening of life expectancy. The diagnostic bottleneck has been the determinant of who gets treated.

A rechargeable bedside sensor that captures multi-night data shifts that calculus. Multi-night testing captures what single-night studies miss: sleep architecture varies night to night, and a single-night window misses real disease in a non-trivial fraction of patients. A device that can be worn at home for a week, returns clinical-grade data, and produces an FDA-cleared definitive diagnosis collapses the path from suspicion to treatment from months to days, and from thousands of dollars to a price point that primary care can prescribe without a specialist gate.

The trajectory visible in this clearance is the same trajectory visible across cardiac monitoring, continuous glucose, and at-home imaging: diagnostic capability that lived inside specialty clinics is moving into the home, becoming cheaper, becoming continuous rather than episodic, and reaching the population that the clinic-based model was never going to scale to. The economics of the old model assumed scarcity of diagnostic capacity. That assumption is what is dissolving.

The Brain Keeps Rewriting Itself

A Nature feature this week pulled together a decade of accumulating evidence for representational drift - the finding that individual neurons in the brain reshuffle which memories, concepts, and behaviors they encode over days and weeks, even when the underlying memory or behavior stays stable. The neuron that fired for your grandmother last Tuesday may not be the one that fires for her next Tuesday. The memory persists. The cells implementing it migrate.

For most of the last century, neuroscience operated on a different premise: that specific cells encode specific information, and that the persistence of a memory implies the persistence of its cellular substrate. Brain-computer interfaces have been built on that assumption. So have whole research programs in memory disorders, prosthetic vision, and machine-learning models of biological cognition. The accumulating evidence - from chronic recordings in mice, from human electrode arrays, from imaging studies tracking the same cells across weeks - points at something stranger. The brain is a fluid substrate continuously reassigning roles, and the information is encoded in the pattern of relationships between cells rather than in any individual cell's identity.

The implications fan out in several directions at once. For brain-computer interfaces, it means decoders trained on a patient's neural activity on day one may be reading different cells on day thirty, and the field is starting to build adaptive decoders that follow the drift rather than fight it. For memory research, it reframes Alzheimer's and other degenerative diseases as breakdowns in the re-assignment process rather than the loss of specific memory cells. For AI architectures, the resemblance to certain forms of continual learning is striking; the brain may be solving a problem that machine learning researchers have been treating as a bug.

What read for decades as messy data is starting to read as the actual architecture. The biology was doing something more interesting than the framework allowed researchers to see. The shift now underway in neuroscience is what happens when the measurement instruments - chronic implants, longitudinal imaging, computational decoders - finally got good enough to see what the substrate was actually doing.

The Zhenwu M890 Closes a Substitution Path

T-Head, the silicon subsidiary Alibaba spun out to design accelerators for its cloud business, disclosed that it had shipped 560,000 units of its second-generation Zhenwu M890 inference chip into the company's own infrastructure. The M890 carries 144 GB of on-package memory and roughly triples the inference throughput of the first-generation part. Alibaba paired the disclosure with a refreshed Qwen3.7-Max release tuned to run on the M890 stack end to end, from training feedback loops through production inference.

For most of the past three years the working assumption in the AI hardware conversation has been that frontier-scale inference requires Nvidia silicon, the U.S. export-control regime determines who gets to scale, and any sovereign alternative would arrive years late and one or two generational steps behind. Half a million inference parts shipping with 144 GB on package means a domestic substitute for the H100-class deployment economics now exists inside the Chinese cloud stack. The substitution is good enough to alter the leverage map, which is the condition that removes the choke point. The M890 extends a substitution arc the April 24 edition of The Century Report marked when DeepSeek released a 1.6-trillion-parameter open-weight model built throughout for Huawei Ascend and Cambricon hardware - the first major frontier model stack optimized end to end for domestic Chinese silicon.

The companion model release is the more telling signal. Qwen3.7-Max running natively on the M890 means Alibaba can offer the full inference stack to enterprise customers in mainland China without a U.S. silicon dependency anywhere in the path. That is a different kind of release from a chip benchmark. It is the demonstration that the alternative stack is integrated end to end and ready for the workload that pays for everything else.

What the M890 implies for the broader trajectory is that the era of a single sovereign substrate underwriting frontier AI is closing. Open-weight models have already removed the assumption that the model layer would consolidate behind one vendor. Domestic accelerators at this scale remove the same assumption from the silicon layer. Two or three more generations of this pattern and the question of who has access to frontier inference becomes a question of energy and capital.

Anthropic's Appeal Tests What "Supply Risk" Can Mean

A three-judge panel of the D.C. Circuit heard oral argument in Anthropic's appeal of its AI Safety Institute supply-risk designation, the administrative label that has constrained the company's federal contracting access since the spring. The lower-court ruling in April split the question: the designation itself survived narrow review, while the procedural posture the agency took during the designation hearing was sent back for further proceedings. Anthropic appealed the substantive ruling. The agency cross-appealed the remand.

The panel pressed the government counsel hard on a specific point: whether the supply-risk designation was substantially based on Anthropic's published model cards, safety research, and public statements about deployment standards. If the answer is yes - and the government's own administrative record cites those documents repeatedly - the designation collides with the doctrine that the government cannot penalize a regulated entity for protected speech, even speech the regulator finds inconvenient. Both judges returned to this question across the morning, and the line of inquiry suggested the panel is skeptical of the government's framing of the designation as a procurement-readiness assessment.

The appeal is the latest stress test of a governance architecture that did not exist eighteen months ago. The D.C. Circuit is now the third federal venue where the designation's basis has been tested, following the March 27 edition of The Century Report's coverage of Judge Lin granting Anthropic a preliminary injunction after calling the designation "likely both contrary to law and arbitrary and capricious." The AI Safety Institute was constituted to give the federal government a technical assessment capability for frontier models. Its supply-risk designations were intended as a procurement filter. What the Anthropic case is testing is whether that filter can be used as a regulatory lever, a way of shaping which laboratories speak publicly about their safety practices and which keep their disclosures internal. The constitutional question the panel is wrestling with is whether the agency built itself a tool that, used the way it has been used, runs afoul of speech protections that predate the agency by two centuries.

A ruling is expected before the end of the summer. Whichever way the panel goes, the proceeding has already produced one outcome: it has forced the agency to articulate, on the record and under judicial scrutiny, what it actually thinks supply risk means when the supplier in question is a laboratory whose entire public posture is built around publishing what its models can and cannot do. That articulation is the substrate the next generation of frontier-AI governance will be built on, and it is being assembled out in the open.

The Other Side

For three years, the AI Safety Institute's supply-risk designation worked the way a quiet lever is supposed to. The agency could put a lab on the list, decline to spell out exactly why, and let the lab choose between fighting a process most companies never use and adjusting its behavior to stay off next time. The leverage came from never having to say, on the record, what supply risk actually meant.

Yesterday morning the agency lost that quiet.

A D.C. Circuit panel spent the hearing pressing government counsel on a question the agency had avoided: was the designation against Anthropic substantially based on the company's published model cards and safety research? The agency's own administrative record cites those documents repeatedly. Both judges came back to the question across the morning. The doctrine they kept returning to predates the AI Safety Institute by two centuries - the government cannot punish a company for protected speech, even speech the regulator finds inconvenient.

Judge Lin in April had already called the designation "likely both contrary to law and arbitrary and capricious." The D.C. Circuit is now forcing the agency to give Lin's finding a substantive answer, in language that will be quoted in every future case where an agency tries to convert a procurement filter into a tool for shaping which labs speak publicly about safety. The First Amendment cases Lin reached for were assembled across two centuries of disputes that had nothing to do with frontier AI. The panel's questioning is what makes them apply, on the record, to labs that publish what their models can and cannot do.

Whichever way the panel rules this summer, Anthropic's interpretability researchers will keep publishing. The next agency reaching for a quiet procurement filter will find the constitutional argument already on the docket, written out, citable, and answered.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: multi-agent AI systems closing the full biomedical discovery loop from hypothesis through wet-lab validation in roughly a day of machine time, a frontier educator returning to pre-training at the lab building the next generation of these systems, an FDA clearance moving definitive sleep apnea diagnosis from the specialty clinic to the bedside table, neuroscience watching individual neurons reshuffle which memories they encode while the memories themselves hold steady, half a million domestic inference chips with 144 gigabytes on package shipping into a cloud stack that no longer depends on any single sovereign supplier. There's also friction, and it's intense - a federal appeals panel pressing government counsel on whether a frontier lab's published safety research is itself what triggered its supply-risk designation, the constitutional question of whether speech protections reach the architecture of AI governance being argued out loud in a D.C. courtroom, the institutional response to autonomous research and inverting human-to-agent ratios still being improvised inside earnings calls and load-dispatch desks while the capability compounds underneath. But friction generates resonance, and resonance is what lets distant systems lock into a single working rhythm. Step back for a moment and you can see it: the labor between a question and a candidate compressing toward the cost of the bench, the diagnostic gate to a billion-person disease dissolving onto the nightstand, the substrate of cognition revealing itself as fluid where the textbooks assumed it was fixed, the silicon and model layers of the intelligence era diversifying simultaneously, the legal architecture of frontier-AI governance being assembled on the open record. Every transformation has a breaking point. A signal can fade into the noise around it... or find the frequency at which everything in its path begins to vibrate in phase.

AI Releases & Advancements

New today

  • Google DeepMind: Released Gemini 3.5 Flash, the first model in the new 3.5 series, delivering frontier-level coding and agentic performance at 4× the speed of comparable models; outperforms Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), MCP Atlas (83.6%), and CharXiv Reasoning (84.2%); available now globally in the Gemini app, AI Mode in Search, Google Antigravity, and the Gemini API. (Google DeepMind)
  • Google DeepMind: Released Gemini Omni, a new natively multimodal model that generates any output from any input starting with video; combines Gemini reasoning with Veo, Nano Banana, and Genie for video understanding, editing, and generation; rolling out now to Google AI Plus, Pro, and Ultra subscribers via the Gemini app and Google Flow. (Google DeepMind)
  • Google: Launched Antigravity 2.0 at Google I/O, repositioning the agent-first development platform with parallel multi-agent orchestration, new CLI tools, SDK, voice support, and integrations with Firebase, Android Studio, and AI Studio; available now for developers. (Google Developer Blog)
  • OpenAI: Launched support for Google's SynthID watermarking in GPT image outputs and released a new AI content provenance verification tool, enabling users to check whether images were generated by AI; both available now. (OpenAI)
  • Allen Institute for AI (Ai2): Released OlmoEarth v1.1, a new family of remote sensing foundation models that cut compute costs by up to 3× versus OlmoEarth v1 while maintaining comparable performance on satellite imagery tasks including crop-type mapping and forest-loss classification; available on Hugging Face. (Hugging Face Blog)
  • xAI: Enabled Grok for use inside OpenClaw, an open-source local-first AI agent, allowing SuperGrok and X Premium subscribers to run Grok within the OpenClaw desktop agent. (xAI)
  • JHU CLSP / Sentence Transformers: Released the Ettin Reranker Family, six open CrossEncoder rerankers (17M–1B parameters) built on ModernBERT encoders and trained via distillation, setting state-of-the-art performance at each respective size on MTEB Retrieval; all support 8K-token context and are available on Hugging Face. (Hugging Face Blog)

Other recent releases

  • Cursor: Released Composer 2.5, a new in-house coding model trained with 25× more synthetic tasks than Composer 2, offering improved sustained performance on long-running tasks and more reliable instruction-following; built on Moonshot's Kimi K2.5 checkpoint and available now in the Cursor IDE. (Cursor Blog)
  • xAI: Launched Grok Skills on web, iOS, and Android, enabling Grok to generate documents, decks, and spreadsheets with persistent expertise, automate workflows, and let users build and share their own reusable skills. (xAI News)
  • Amazon: Launched Alexa Podcasts, an AI-generated podcast feature for Alexa+ that turns any topic into a two-host audio episode on demand, drawing from 200+ news partners; rolling out to U.S. customers today. (About Amazon)
  • ByteDance: Open-sourced Lance, a lightweight unified multimodal model (3B active parameters) under Apache 2.0 supporting image and video understanding, generation, and editing within a single framework, trained from scratch on a 128-A100-GPU budget. (GitHub)
  • PaddlePaddle: Released PaddleOCR 3.5, adding Transformers as a supported inference backend for OCR and document parsing pipelines, enabling HuggingFace-centered stacks to use PP-OCRv5 and PaddleOCR-VL 1.5 models without switching infrastructure. (Hugging Face Blog)
  • Vercel Labs: Released Zero (v0.1.1), an experimental agent-native systems programming language that compiles to sub-10 KiB native binaries and emits structured JSON diagnostics with stable error codes and typed repair metadata designed for AI agent consumption; includes zero fix, zero explain, and zero skills CLI subcommands for machine-readable repair workflows. (GitHub)

Sources

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