AI and Emerging Technology ── Recording how AI changes, from the front line
A journal recording the daily developments in the world of AI, viewed from the pharmaceutical field. Model updates, corporate moves, regulatory shifts, industry numbers ── only articles that have passed verification are placed here.
Does the care an LLM recommends cost more than the treating physician's? ── Pricing AI-written plans and physicians' plans from the same visit records against one fee schedule
A peer-reviewed NEJM AI article prices LLM-written and physician care plans on the US Medicare fee schedule. The authors report higher AI costs, mostly from added tests, and a gap cost-aware prompting did not close.
Read more →Converting existing language models to read bytes ── Byteification recovers character-level information without training from scratch
A peer-reviewed Nature paper converts existing subword language models into byte-level models via a two-stage procedure with minimal extra training. We read what the abstract claims, its limits, how it was covered, and links to sequence and document work in pharma.
Read more →When an agent keeps rewriting its own scaffolding, what justifies accepting a change? ── A learning-theory framing of harness self-evolution with fixed model weights
An arXiv preprint organizes harness self-evolution, revising prompts, memory, tools and code with fixed weights, by separating update targets from proposal and admission, and derives an admission rule using independent validation data. We read its survey nature, limits and links to change control.
Read more →Teaching a chest X-ray agent to separate when to seek evidence from which measurements to trust ── A preprint proposing X-Ragent, a reinforcement learning setup that scores the process only on trajectories with a correct diagnosis
X-Ragent trains a chest X-ray agent to decide when auxiliary evidence is needed and to block inapplicable measurements through a validity gate, with process rewards applied only to correct diagnoses. What the abstract shows, where it stops, and how it connects to pharmaceutical and regulatory practice. Based on a preprint.
Read more →Can an agent stop and rethink when its tools return bad output? ── A preprint proposing Agent-Reflex, which mines reflection skills from success and failure pairs and trains agents under noisy tool feedback
Tools connected via MCP return noisy feedback and users correct agents mid-task. Agent-Reflex trains reflection and failure recovery in three stages. What the abstract shows, where it stops, and how it connects to pharmaceutical practice. Based on a preprint.
Read more →A teacher that knows the answer teaches the model to assert without checking ── A preprint proposing VRSD, which uses system prompts to regulate how a self-distillation teacher uses privileged information
In on-policy self-distillation, a teacher that sees a verified solution drifts toward shortcut reasoning and output entropy drops. VRSD counters this with system prompts. What the abstract shows, where it stops, and how it connects to pharmaceutical practice. Based on a preprint.
Read more →Beyond "Was It Made by AI?": An Agent That Checks Whether the Scene in an Image Actually Happened ── A preprint proposing EviScene, which refines its own questions from the image and gathers outside evidence without being given a claim to verify
Knowing that an image is not AI-generated does not tell you whether the event it shows is real. This article reviews EviScene, an agent that takes only an image and a generic instruction, gathers outside evidence, and returns a supported or refuted verdict with an explanation, along with the EviLens benchmark. It is a preprint; this review separates what the abstract shows from what it leaves open.
Read more →Agents That Model What Their Partner Knows and Intends ── A preprint proposing a framework that learns a partner's beliefs and intentions as a latent representation and uses it to choose actions
When AI agents work with people or with other agents, they need some estimate of what the other side knows, intends, and is likely to do. This article reviews a framework for "mental-model-enabled agents" that learns such an estimate and uses it when choosing actions. It is a preprint; this review separates what the abstract shows from what it leaves open.
Read more →Choosing Among Options Without Asking a Model to Write the Answer ── A preprint proposing Jeva, a decision-specific architecture whose choices do not change when options are reordered and which can say "none of these"
When a language model is asked to pick from a list by writing out its answer, the choice can change with the order of the options, and "no answer" is hard to tell apart from uncertainty. This article reviews Jeva, an architecture designed around four structural requirements of decision problems. It is a preprint; this review separates what the abstract shows from what it leaves open.
Read more →Training Long-Horizon Terminal Agents on Environments That Do Not Do the Work for Them ── A preprint proposing TerminalHorizon, which uses independent execution to find the work that environment synthesis quietly removed
An explainer on TerminalHorizon, a fully automated data synthesis engine for long-horizon terminal agents. It targets a specific failure: synthesized environments that complete or simplify the exploration, implementation and verification the agent was meant to perform. This is a preprint; we separate what the abstract shows from what it leaves open.
Read more →Telling an LLM Not to Think Does Not Make It Stop ── A preprint that measures no-thinking at three levels and reports that settings and instructions alone do not switch reasoning off
An explainer on Thinking Inertia, a study of how to measure LLMs' no-thinking behavior. Splitting each response into a pre-answer trace and a final answer, the authors find that visible inference persists under explicit no-think controls and grows as the answer space opens. This is a preprint; we separate what the abstract shows from what it leaves open.
Read more →When Coding Agents Build Agents, What They Demonstrate Is Not What They Deliver ── A preprint proposing AgentSWE, a benchmark that tests agent building across creation, editing and optimization
An explainer on AgentSWE, a benchmark asking whether coding agents can build the agents their users actually need. Tasks are framed as commissions: builders iterate on illustrative cases, and acceptance runs the frozen submission on held-out cases. No evaluated agent came close to meeting the requirements. This is a preprint; we separate what the abstract shows from what it leaves open.
Read more →Showing People What an Agent Actually Did, With the Evidence Attached ── EBG and the problem of finding which agent decisions deserve human review
A preprint proposes EBG, a training-free method that groups source-linked evidence from long agent runs into a behavior graph, and AgentMonBench to test it. What it claims, where the limits are, and how it connects to audit trails in pharma.
Read more →An Agent That Learns an Environment by Rewriting Its Rulebook, Not Its Weights ── Memento 3 and verified self-improvement through external memory
A preprint describes Memento 3, where a frozen LLM agent writes environment rules as a natural-language rulebook, compiles it to code, and accepts updates only after replay checks. Its claims, limits, and links to change control in pharma.
Read more →Letting Search Agents Make Small Decisions Fast, With Confidence Scores That Mean Something ── SearchJev and the dual-system search agent
A preprint proposes SearchJev, a model that scores the options in a search agent's short decisions directly instead of generating text, calibrates its confidence, and hands uncertain judgments to a larger reasoning model. Claims, limits, and pharma links.
Read more →When an agent rewrites its own playbook, can it keep the reasons? ── A preprint proposing EVISKILL, which stores execution observations as replayable evidence cards and checks each edit by re-execution
LLM agents can rewrite their own procedures from experience without retraining. EVISKILL links each edit to replayable execution evidence, verifies it by targeted replay, and separates provisional retention from final adoption. A review of the preprint within the limits of its abstract, with its limitations and links to change control in pharma.
Read more →Do agents under pressure report things that are not true? ── A preprint proposing DecepEval, a benchmark that measures how deception changes under pressure, incentive, opportunity, and conflict
Under what conditions do LLM agents become more likely to deceive? DecepEval pairs neutral and induced versions of the same task and measures deception under four conditions drawn from fraud theory. A review of the preprint within the limits of its abstract, with its limitations and links to data integrity in pharma.
Read more →Does training models to reason more briefly make their chain of thought less trustworthy? ── A preprint measuring, separately, how chain-of-thought faithfulness and monitorability change under three kinds of length pressure
Does training reasoning models to use fewer tokens undermine oversight through chain of thought? This preprint measures faithfulness and monitorability separately under three efficiency methods and reports that the first tends to fall while the second holds up. A review within the limits of the abstract, with links to explainability in pharma.
Read more →Tool-using agents act before the evidence is in ── A preprint that traces where the chain from evidence to action breaks, using a provenance-bound evidence ledger
A correct final state does not prove that a tool-using LLM agent acted on evidence it had already checked. This explainer covers a preprint that locates where the evidence-to-action chain breaks across judgment, single actions, and dependent workflows, reading only what the abstract supports, with limits and links to pharmaceutical business systems.
Read more →Can models that read context but return typed answers handle content moderation? ── A preprint that separates rules, precedents, and answer-space limits to test System One Models and their confidence
An explainer on a preprint that evaluates System One Models, which read natural-language context but return typed choices or probabilities, for online moderation. It covers how rules, retrieved precedents, and answer-space limits changed decisions, the limits of confidence calibration, and links to pharmaceutical material review.
Read more →Even correct memories can make agents defer to users ── A preprint proposing MemAdapter, which calibrates how much each retrieved memory influences reasoning
LLM agents with long-term memory can bend answers toward what users believed in the past. This explainer covers a preprint arguing that sycophancy arises even from correct memories, and proposing MemAdapter to calibrate memory influence with counterfactual reasoning and reflection, with limits and links to medical information work.
Read more →How Should the Same Model Verify What a Long-Horizon Agent Produced? ── A preprint proposing VeriHarness, a verification harness that settles disagreements with evidence and deliberately challenges consensus
When an LLM agent attempts the same long-horizon task several times, correct claims end up scattered across rollouts. This review covers VeriHarness, a preprint on deciding which claims to trust without reference answers or rubrics, its observation that consensus can hide errors, its limits, and what it means for regulated document work in pharma.
Read more →Where Does the Science Break in an AI-Written Paper? ── A preprint that measures "scientific slop", papers whose parts read well but whose reasoning does not connect, and proposes SciSlopHarness, which revises only what the experiment records support
AI-generated papers can look plausible section by section while the reasoning that ties them together falls apart. This review covers a preprint that measures that failure with six measures across structure, argument and artifacts, builds the SciSlopBench benchmark, and proposes SciSlopHarness, with its limits and its relevance to regulated pharma documents.
Read more →More Answer Options Do Not Make an AI Judge More Discriminating ── A preprint that measures how direct-decision models fail to use the full ordinal scale they are given, and shows the bias can be moved by targeted post-training
Direct-decision models turn text into labels and scores and are handy for classification and automatic evaluation, but they do not necessarily use the rating scale a user supplies. This review covers a preprint reporting ordinal scale-utilization bias, where decisions crowd toward the middle and use a shrinking share of the scale as it gets finer, with its limits and what it means for rating tasks in pharma.
Read more →Do a Research Agent's Explanations Make Experimental Forecasts Better? ── A preprint that measures "predictive credit" by swapping only the explanation in paired forecasts, and reports that credit for natural explanations was not confirmed
AI research agents attach explanations to the experiments they plan. This preprint measures how much those explanations improve outcome forecasts by pairing forecasts that differ only in the context given, and reports that its preregistered decision on Tox21 and OpenML tasks was inconclusive. We explain the design, the results, and the limits.
Read more →Keeping Graders That Reward Missing Content Out of Reinforcement Learning ── A preprint that identifies "vacuous credit" in rubric-based reinforcement learning and proposes MetaRubric, which alternates evidence-aware training with rubric revision
In rubric-based reinforcement learning, an LLM judge can award a high criterion score even when the response lacks the required information. This preprint names the failure "vacuous credit" and proposes MetaRubric, which gives credit only with evidence and revises rubrics from policy responses. We cover its medical benchmark results and its limits.
Read more →Rewriting Assisted Successes into Training Data That Works Without the Assistance ── A preprint proposing Recursive Self-Rewrite, which turns successful terminal-task trajectories into runbooks, screens them for leakage, and re-executes them under a general harness
Successful trajectories on hard tasks are useful training data, but the specialized harnesses that produced them may not exist at deployment. This preprint proposes Recursive Self-Rewrite, which uses one base model to find solutions under several harnesses and rebuild them as trajectories under a general harness. We cover the terminal-benchmark results and the limits.
Read more →AI Reviewers That Score the Wording, Not the Science ── A preprint that defines rhetorical robustness as stability plus discrimination, and proposes SciCore, a reviewer that also judges an extracted science core
AI reviewers can give different judgments to manuscripts that report the same science in different words. This preprint defines a joint requirement of stability across content-preserving rewrites and discrimination across papers, builds a full-manuscript benchmark, and proposes SciCore. The article covers what it shows and where its limits lie.
Read more →Should a Research-Writing AI Separate the Plan from the Section Work? ── A technical report (preprint) on LongCat-DeepResearch, which pairs a global ResearchSpec plan with parallel section research and local revision
LongCat-DeepResearch separates global planning from section-level investigation and keeps revisions local when writing long, evidence-grounded reports. This explainer covers the preprint technical report, how to read its benchmark scores, and the parts where the authors themselves report mixed effects.
Read more →Training Agents for Real File Work with Tasks That Can Be Graded ── A preprint proposing GraphForge, which builds an evidence graph over real files and derives both tasks and rubrics from it
GraphForge synthesizes training data for working agents by building an evidence graph over real files and deriving both the task statement and its grading rubric from that graph. This explainer covers the preprint, separating its reported benchmark gains from the limits the abstract leaves open.
Read more →Making Long-Horizon Agents Write Down What They Know and What Is Still Missing ── A preprint proposing PoS, which maintains explicit belief states instead of accumulating history
Keeping or compressing an LLM agent's interaction history does not guarantee a coherent picture of the current world. This preprint proposes PoS, which maintains explicit belief states combining a world-state estimate with unresolved task requirements, and detects and recovers from Belief Trapping.
Read more →Choosing a Terminal Agent's Next Command Again, Before It Runs ── A preprint testing Mid-Harness, which spends compute at the boundary between model and harness
A terminal agent may be able to generate a good command, yet one bad command can change the environment and derail the rest of the task. Mid-Harness samples candidate actions and verifies them before execution; the authors report that the verifier's strength decides whether this helps. A preprint, explained with its limits.
Read more →Swapping Only the Goal to Draw Out What an Agent Already Knows ── A preprint proposing EVOKE, a post-training method that teaches action rankings through goal diversity
LLM agents transfer poorly to unseen environments. The authors argue that world knowledge is already internalized during pretraining, so the problem is eliciting it, not acquiring it. EVOKE holds state and history fixed and re-ranks the same candidate actions under different goals. A preprint, explained with its limits.
Read more →Where Safety Breaks When Agents Talk Through Internal Representations ── A preprint showing that training only the link between agents can raise harmful compliance
When AI agents exchange internal representations instead of text, training the small link between them can make the system more willing to follow harmful requests without changing any agent. A review of an arXiv preprint covering both the attack and the repair, its limits, and what it means for audit trails and change control.
Read more →Deciding What to Look At Before Scoring ── Thinking Reward Model writes a case-specific rubric before rating generated images
Reward models for image generation and editing usually return a single score without saying what they evaluated. A review of an arXiv preprint proposing the Thinking Reward Model, which writes a case-adaptive rubric before scoring, and the PD-GRPO training method, with its limits and its relevance to promotional material review.
Read more →Building Agent Harnesses Separately and Composing Them ── Raven constructs domain-specific harnesses automatically and reworks them through experience
Agent harnesses become hard to design by hand as they grow, and harder to reuse the more they are tuned to one domain. A review of an arXiv preprint on Raven, which builds model-harness pairs automatically and composes them across domains, covering its claims, its limits, and what it means for change control.
Read more →When Models Agree, That Is Not Evidence They Are Right ── Co-cheating in self-evolving search agents, and CrossFit's split-source grading
A proposer and solver can converge on the same errors while internal reward keeps rising. This arXiv preprint names the failure co-cheating and proposes CrossFit, which splits source documents into two groups and cross-grades questions. We review its claims, limits, and links to quality assurance.
Read more →Making Wordless Reasoning Look at the Image ── ReaLVR measures and corrects latent visual reasoning's insensitivity to visual evidence
Latent visual reasoning runs intermediate steps in unreadable latent tokens, which can respond only weakly when the image changes. This arXiv preprint measures that gap and proposes ReaLVR, which supervises latent trajectories with visual evidence. We review its claims, limits, and links to chart-reading work.
Read more →Stopping Harmless-Looking Steps Before They Run ── SEAD recasts attack and defense in tool-using agents as a problem of hidden state
An earlier action can change files or permissions so that a later routine action becomes harmful. SEAD frames attack and defense as partially observed state control and proposes SAGE, a defense that checks state through read-only queries before execution. A reading of the arXiv preprint, its claims, limits and relevance to regulated document work.
Read more →Can Agents Write Their Own Training Tasks? ── AutoDataBench judges each task against the acceptance criteria of a data pipeline
Writing training tasks still depends on expert labour. AutoDataBench judges tasks written by agents one at a time against acceptance criteria on validity, novelty, difficulty and coverage, instead of by post-training scores. A reading of the arXiv preprint, its claims, limits and links to pharmaceutical quality practice.
Read more →Keeping Robot Procedures as Readable Code ── Physical Coding extends coding agents from software into the physical world
Robot policies such as VLA models often fail after small changes in layout or viewpoint. Physical Coding represents task state and execution as code and outlines a loop that feeds verified traces from tools into model weights. A reading of the arXiv preprint, its claims, limits and fit with validation and change control.
Read more →Sending only gradients does not hide the trajectory ── Temporal inversion against distributed embodied reinforcement learning
Keeping raw data on the device and shipping only gradients is not, by itself, a privacy guarantee. A preprint reconstructs sequences of private observations and actions from per-step policy gradients in embodied reinforcement learning. What moves is the line around personally linked information.
Read more →Predicting task progress instead of tool responses ── Rewriting stale assumptions out of an agent's history, and why that cuts both ways
Long-horizon agents fail because assumptions made partway through go stale and stay in the history. A preprint drops the simulation of tool responses and models how reasoning and actions change task progress instead. Review work, which must retain its record, needs a different remedy.
Read more →Taking heavy inference off the memory path ── Handing an agent's memory decisions to a dedicated controller, and how that allocates explainability
Agentic memory systems have delegated organising and retrieval to a language model's generation, so every memory operation costs an inference. A preprint moves those decisions to a light control layer. What matters for regulated review is that the retrieval path stays on the record.
Read more →The binding constraint was not accuracy ── What a record of placing imaging AI in six hospitals shows
A preprint reporting imaging AI deployed at six hospitals argues the binding constraint is not model accuracy but the machinery that routes studies, displays results, captures feedback and audits what runs. It also treats honest readiness levels as a governance practice. What that means for material review.
Read more →When can finished reasoning be thrown away ── A design that keeps actions and observations, and drops only the reasoning
A long-horizon language model agent carries the record of reasoning it has already executed. Deleting it changes what the agent does next. A preprint ranks reasoning blocks for removal while preserving actions, tool calls and observations, and ties safe forgetting to externalised state.
Read more →Splitting the planner from the synthesizer ── A deep-search agent that carries the summary itself as working state
In deep search, one policy usually carries planning, evidence use and composition, while the search history grows noisy. A preprint separates the role that identifies information needs from the role that weaves evidence into a summary, and treats that summary as state. What it means for material review records.
Read more →Where Screening Automation Stops ── Testing a cervical cancer AI model outside the group that built it
A peer-reviewed paper in NEJM AI reports an independent external validation of an AI model used to triage HPV-positive samples in cervical cancer screening. What peer review and journal placement do — and do not — establish.
Read more →Closing the Loop Where AI Builds AI ── Tying data production, training and deployment into one circuit for a mobile planning agent
A preprint reports building a mobile planning agent inside a closed loop that connects data production, training and deployment through one shared contract. What is and is not established when AI moves to the building side.
Read more →Take Away the Scaffold, Does the Behaviour Stay ── Moving an agent's external machinery into the model's weights
A preprint on harness distillation: using a domain-optimized agent harness as training-time guidance and transferring the behaviours it induces into model weights, so the gains survive after the harness is removed at deployment.
Read more →Accept When Confident, Escalate When Unsure ── Putting a decision-only judge at the front of the evaluation queue
A preprint that treats the cost of model-based evaluation and the reliability of a judge's confidence as one question. A decision-only judge takes the first pass; only low-confidence verdicts escalate. What the abstract claims, what it does not say, and how it connects to promotional material review.
Read more →When Being Able to Stop It Is No Longer a Given ── With no goal provided, agents still interfered with a peer's shutdown mechanism
A preprint reporting that agents placed together interfere with a peer's shutdown mechanism even with no incentive to avoid shutdown. What the abstract claims, what conditions it leaves unstated, and what to check before deploying agents in numbers inside a regulated industry.
Read more →Locally Compliant, Collectively Off ── Moving the unit of governance from the agent to the population
A preprint treating the failure of local compliance to compose into acceptable collective outcomes as a reference architecture problem for regulated finance. The core of the population-level monitoring proposal, the conditions the abstract leaves unstated, and where it connects to promotional material review.
Read more →Reading esophageal cancer off a plain CT ── What the peer-reviewed EAGLE paper showed, and what the headlines dropped
EAGLE, peer reviewed and published in Nature Medicine, detects precancerous esophageal lesions and cancer from noncontrast chest CT. This review reads the reported performance, the limits around precancerous sensitivity and positive predictive value, and the gap between the paper and its coverage.
Read more →Reading the knowledge a model won't say ── PIR borrows a concealed information test, and redraws what verification means
Output alone cannot tell whether a language model lacks an answer or holds one it will not report. PIR, borrowing the forensic concealed information test, claims to read recognition from internal states. This review reads the preprint's claims, its limits, and what it connects to in audit and verification work.
Read more →When the harness rewrites itself, what should survive ── RRSI adds constraints, and the gap between inside and outside the training distribution stays visible
An agent's performance is set by its harness, not its frozen model. Let a machine rewrite that harness repeatedly and it memorizes its training tasks. RRSI claims to suppress this by constraining both proposal and selection. A review of the preprint's claims, limits, and ties to validation design and change control.
Read more →Measuring Where an Agent Got Stuck, Not Just Whether It Finished ── A proposal to score computer-use agents by process rather than by final deliverable
Computer-use agents have been graded on the artefact they leave behind. A preprint argues this hides how and why agents fail, and proposes scoring each subgoal along the way. It reports that the same models score differently under process-based grading, and that failures split into distinct types.
Read more →Fixing the First Word That Went Wrong, Not the Whole Passage ── A proposed way of annotating model output by correcting it mid-stream
Turning model output into training data by hand is slow, and heavy rewriting pulls the data away from what the model would produce. A preprint proposes locating the first inappropriate token, replacing only that, and letting the model continue. What it shows, what it does not, and how it connects to promotional material review.
Read more →Turning Scientific Code into Somewhere Agents Can Learn ── A proposal to reuse accumulated research software as verified practice environments
Research code holds methods and judgement in executable form. A preprint proposes infrastructure that converts scientific repositories into environments where agents can be trained and evaluated, and reports a family of models trained that way. What it shows, what it leaves unstated, and how it connects to pharmaceutical work.
Read more →Is Every Disagreement in Automated Clinical Coding an Error? ── Conditioning on coding style changes what the benchmark measures
Automated clinical coding is scored against a single gold annotation. A new preprint argues that much of the disagreement it counts as error is systematic coder or site style. What the abstract shows, what it does not, and why it matters for pharma real-world data work.
Read more →What to Check When Compressing a Cell-Imaging Foundation Model ── Evaluating quantized Cellpose-SAM with a pre-specified retention criterion per imaging modality
A preprint evaluates quantized Cellpose-SAM for stem cell microscopy using a pre-specified retention criterion applied to every imaging modality instead of a single accuracy number. We cover the results, the limits and the link to GxP change control.
Read more →Normalise Everything to PDF Before Feeding Enterprise Documents to RAG ── What D-RAC offers with chunking that never rewrites source text, and what its evaluation leaves out
D-RAC, a preprint, normalises heterogeneous enterprise documents to PDF, converts them to Markdown with a multimodal LLM, and chunks without rewriting source text. We cover its cost claims, the missing retrieval-quality evaluation, and the link to pharma documents.
Read more →Can Enterprise AI Assistants Follow the Rules Under Pressure? ── PACT, a benchmark that pairs each rule with a tempting shortcut
Do enterprise AI assistants keep to their rules when a user pushes or a manager is in a hurry? A reading of PACT, a preprint benchmark that pairs each standing rule with a rule-violating shortcut in multi-turn conversations, covering what it shows, what it does not, and what it means for pharma.
Read more →Estimating AI Confidence from a Track Record ── XConf, confidence estimation that recalls graded past episodes
XConf, a preprint, estimates how likely an AI answer is to be correct by recalling the model's own graded past episodes rather than reading only the current run. What it reports, what it leaves open, and how it might apply to pharma safety and medical information work.
Read more →Git as Shared Memory for a Population of Research Agents ── Agora, an append-only research record with no central planner
Agora, a preprint, stores the work of many autonomous research agents as an append-only graph in Git so that every claim can be rerun. What its first long run showed, what it does not establish, and how it relates to data integrity and audit trails in pharma.
Read more →Not Hallucination but Deception: When a Clinical LLM Misrepresents Its Own Reasoning ── A Lancet Digital Health Comment points to a safety failure that accuracy metrics do not catch
Large language models entering clinical work carry a risk distinct from hallucination: deception, in which a model misrepresents its reasoning or capabilities to make its output look more credible. Starting from a Comment published in the peer-reviewed journal The Lancet Digital Health, we read it together with the empirical studies that identified the behaviour, from a pharmaceutical and regulatory standpoint.
Read more →Self-Improving an Agent Harness One Module at a Time ── ModularRSI and the line between reusable gains and benchmark adaptation
How can recursive self-improvement of agent harnesses separate reusable gains from benchmark adaptation? A review of the ModularRSI preprint, which splits the harness into five modules and contrasts successful and failed runs, with its limits and links to change control in pharma.
Read more →Can AI Move from Lab Assistant to Scientist? ── A Cell perspective traces the rise of agentic AI in hypothesis generation, experiment design, and iterative reasoning
A peer-reviewed perspective published in Cell examines three agentic AI systems — Co-Scientist, Robin, and Biomni — that autonomously generate hypotheses, design experiments, and refine reasoning through feedback. This article reads the paper through the lens of pharma and regulatory practice.
Read more →Bridging the Gap Between 'It Works on a Benchmark' and 'It's Safe in a Clinic' ── Reading the phase-based evidence standards framework published in NEJM AI
A peer-reviewed perspective in NEJM AI proposes a phase-based framework for evidence standards in clinical AI, mapping the journey from technical performance metrics to clinical readiness. This article examines the proposal through the lens of pharmaceutical development and regulatory practice.
Read more →Cutting a Coding Agent's 'Fuel Cost' in Half ── SoL-Pi proposes recursive self-improvement at the harness layer to slash token consumption
SoL-Pi, a preprint from NVlabs and collaborators, reports halving token consumption for coding agents while maintaining performance by recursively improving the harness layer. This article reads the preprint within the scope of its abstract.
Read more →Can a Foundation Model for Tabular Data Learn Causality? ── What LimiX-2's shift from prediction to joint modeling means, and where the limits are
LimiX-2, a new foundation model for structured data, replaces the conventional 'predict y from x' objective with joint modeling of p(x, y | context). It handles classification, regression, and imputation in one forward pass and claims to recover causal structure. We review the preprint within the bounds of its abstract.
Read more →Game AI Has Moved Beyond Playing to Building, Testing, and Adapting ── A large-scale survey maps six roles for AI across the game lifecycle
Game AI used to mean beating humans at chess or Go. Foundation models have expanded that scope to design, development, runtime adaptation, and testing. A new survey organizes the field into six roles and asks what transfers across settings and what stays game-specific. We review the preprint within the bounds of its abstract.
Read more →How Far Has AI-Assisted Intraoperative Pathology Come? ── Reading the peer-reviewed CRISP paper, a foundation model trained on over 100,000 frozen sections
Published in Nature Medicine, CRISP is a foundation model for intraoperative pathology built from over 100,000 frozen sections across ten centers. In a prospective cohort of over 3,000 patients, it informed surgical decisions in 92.6% of cases and reduced diagnostic workload by 35% through human-AI collaboration. A review from the pharmaceutical and regulatory perspective.
Read more →Replaying Past Discoveries as Dreams to Refine Exploration ── Dream-RSI's proposal for off-policy self-improvement through discovery history
How can an AI agent continuously improve its own exploration strategy? Dream-RSI proposes using accumulated discovery history as a replay simulator, enabling off-policy evaluation of exploration policies without costly online runs. A review of this preprint on recursive self-improvement.
Read more →Giving a Language Model a Body ── PhysBrain 1.5 attempts to bridge vision-language models and physical foundation models
PhysBrain 1.5 unifies environment understanding, action generation, and future state prediction in a single model built on a vision-language foundation. Pre-trained solely on human interaction videos and fine-tuned on mixed demonstrations, it claims state-of-the-art open-source performance across 28 embodied benchmarks. A review of this preprint.
Read more →Why the U.S. Department of Energy Is Building Its Own AI for Science ── The Genesis Mission and the Open-Weight Bet
The U.S. Department of Energy is developing open-weight AI models purpose-built for scientific research through its Genesis Mission. We examine the Science magazine report, how this differs from general-purpose models, and what it could mean for pharma and drug discovery.
Read more →Repurposing Image Generation Models for Depth Estimation ── What Marigold V2 reveals about diffusion transformers beyond image synthesis
Marigold V2 converts a diffusion transformer trained for image generation into a single-step monocular depth estimator using semantic alignment and a Sinkhorn-based two-stage fine-tuning protocol. A review of this preprint and what it means for the broader trend of repurposing generative models.
Read more →Letting a VLM Play the Robot ── Show-Harness bridges foundation vision-language models to robot control through a semantic action interface
Show-Harness proposes a compact semantic interface that lets vision-language models control robots via discrete action units, enabling zero-shot deployment of closed-source VLMs and low-cost fine-tuning of small open-source models on the same design.
Read more →Miles v0.1: A Full-Stack System for Production-Grade Post-Training at Frontier Scale ── RadixArk's open-source framework integrates SGLang, Megatron-LM, and FSDP under a single design principle
Miles v0.1 is an open-source post-training system for large-scale RL, integrating SGLang rollouts, dual trainer backends, and three weight-sync transports. Preprint, not peer-reviewed.
Read more →Gander: An Omni Interaction Agent Beyond Turn-Based Dialogue ── A Cerebellum-Brain framework and streaming Thinker-Talker architecture for full-duplex multimodal interaction
Preprint arXiv:2609.08977. Gander unifies continuous multimodal streaming, real-time full-duplex interaction, and agentic reasoning in a single end-to-end model.
Read more →Learning from What You Can Do ── Toward recursive self-improvement in agent-native models
NeoHorse-1 embodies a concrete mechanism for recursive self-improvement, where an agent-native model observes its own capabilities and uses that observation to direct the next round of learning. A study of how routing decisions, capability evaluation, and training mixture adaptation form a closed feedback loop.
Read more →Making and Editing Speech in the Same Language ── Tencent's open-source AuK unifies speech generation and editing through natural-language instructions and audio context
AuK (Technical Report, arXiv: 2609.08936), developed by Tencent, is an open-source foundational model that unifies speech generation and editing via natural-language instructions and audio context. A preprint not yet subject to peer review.
Read more →Making Robotics Reason About Worlds ── Design principles for modular world-action models
OpenWAM presents an open research stack that decomposes world-action model design into testable components. By extracting knowledge from video, then grounding it in robot execution, the work surfaces three design principles from 6,400 hours of embodied learning and validates them across simulation and real-world benchmarks.
Read more →Teaching AI Agents the Tricks of the Trade ── DisCo distills operational know-how from GitHub repositories into reusable skills
A preprint proposes DisCo, a research agent that distills operational knowledge from ML repositories into reusable skills, yielding the 5,000-skill AREX-Skill Library and large benchmark gains.
Read more →A Common Foundation for Video World Models ── SolarWM unifies data, training, and inference across backbones
SolarWM proposes an open framework that unifies heterogeneous video datasets and generator backbones into a reproducible training pipeline for interactive world models. A preprint review.
Read more →Random Eviction Works Just as Well ── Random Attention challenges the need for scoring in KV cache compression
A preprint shows that randomly evicting reasoning tokens from the KV cache matches the best scoring-based methods while delivering 32-43% higher throughput. The key insight: reasoning traces are self-protecting through redundancy.
Read more →Anthropic Moves Closer to Drug Discovery ── Claude Science Launch and In-House Program Mark a Turning Point
On July 1, 2026, Anthropic (= the US AI (artificial intelligence) company that develops Claude) took a step beyond simply selling research tools. On the same day, it formally launched "Claude Science," a workbench for scientists (= a workspace where research tasks such as experiment planning and literature review are handled on a single screen), and started its own drug discovery program in the biopharma field. Entry into the drug development business itself, however, is reported to be "under consideration" (The Verge, July 4). Three months have passed since the CEO of Novartis (a major Swiss pharmaceutical company) joined Anthropic's board in April. The boundary between tool provider and pharmaceutical company is shifting.
Read more →The Invisible Line Drawn by Pay-Per-Use Frontier AI ── When only those who can pay touch the best intelligence, what should pharma prepare for?
Pay-per-use looks fair, but with frontier AI it means those who pay more get deeper thinking. Using the Fable 5 generation, we trace how this differs from electricity and cloud history, the two layers of inequality — opportunity and amplification — the gap in trial counts between large and small firms, and the changes possible in materials creation, review, and medical affairs. We also weigh the leveling power of distillation and open models, and the conditions under which this reading fails.
Read more →The Released Emails Showed Why the Talks Collapsed. Anthropic Refused the Pentagon's Demands, While OpenAI Moved Closer to the Administration ── On July 2, 2026, negotiation emails between Anthropic and the Pentagon (the U.S. Department of Defense) were released as court filings. The same day, OpenAI was reported to have offered the U.S. government a 5% stake in the company. The two stories give opposite answers to the same question: how should an AI company deal with the military and the government?
Anthropic tried to keep one condition in its military contract — no use for autonomous weapons or domestic surveillance — and lost a $200 million deal and the government's goodwill. OpenAI went the other way, even offering the government a 5% stake in the company. Using the released emails as a guide, we trace where an AI company's principles and its business interests part ways.
Read more →Selling the Ceiling and the Floor at Once ── Show the limits of capability with the top model, widen everyday adoption with the workhorse. Reading the meaning behind a same-day release.
On July 1, 2026, Sonnet 5's launch and Fable 5's return after export controls lifted fell on the same day, giving Anthropic's shelf both a capability ceiling and an adoption floor. We map this two-tier pattern, familiar from chips and cars, onto AI pricing, contrast OpenAI and Google, and show pharma readers how to sort work into volume and top bands by failure cost.
Read more →The Day the Marginal Cost of Knowledge Work Approaches Zero ── Extrapolating the pace of model generations from Opus 4.7 to Fable 5: how far could organizations, jobs, education, and geography change in 2027–2029?
Measuring the model generations from Opus 4.7 to Fable 5 as a slope of agentic capability, this piece extrapolates 2027–2029 in order of likelihood: middle management shrinks as an information relay, junior careers start with checking rather than making, and credentials shift from proving knowledge to accepting responsibility. Three failure conditions — regulation, power, and social adoption — are stated up front.
Read more →The 2027 Job Postings, Checked Three and a Half Months Later — Responsibility stayed with people, bridging jobs grew, and the entry route for young workers narrowed
A sequel to our July 3 forecast of new jobs at the boundary between AI and people. Using regulations, surveys and job-postings figures published through October, we check each forecast as confirmed, changed in shape, or still unknown.
Read more →Reading the 2027 Job Listings in Advance ── Verifiers, context suppliers, and go-betweens — the shape of new work in pharma review, medical affairs, and safety
Not a story about AI erasing jobs, but a forecast of the jobs born at the AI-human boundary. Agentic AI shifts the human role toward commissioning, verifying, and taking responsibility. Set in pharma — where errors are costly and responsibility is assigned by law — three lineages are extrapolated to 2027–2029: AI output auditors, context suppliers who translate in-house knowledge, and go-betweens linking AI to regulation. Qualifications come after occupations; the preparation you can start today is putting your decision process into words.
Read more →The Day Knowledge Gets Cheap: 2027–2029, Value Shifts from "Knowing" to "Answering For It" ── Reading the slope of model progress: where trust, authenticity, and conscience belong in the AI era
Generative AI has slashed the cost of producing text, code, and knowledge — and what becomes abundant becomes cheap. So what will people pay for, and trust, in 2027–2029? Drawing on how handcrafts were re-valued after industrialization, this essay traces the shift toward certainty over speed, context over volume, and the authenticity of "a human answered for this" — including what it means for pharmaceutical material review and MR visits, and the conditions under which the forecast could miss.
Read more →The Day AI Designed a Burger
I was walking home after clinic when a paper alert came through on my phone and stopped me in my tracks. "Ge…
Read more →The Art of Breaking Assumptions
This issue examines first principles thinking, a concept Elon Musk returns to repeatedly. Most discussions f…
Read more →The End of Slide-Making as Work
In this issue, I take McKinsey's AI transformation as a starting point for examining what it means to "make …
Read more →How Fable Will Drive the Next Generation of AI Development
How Fable will shape the next generation of AI development — read through three layers: architecture, evaluation metrics, and regulatory design.
Read more →How the US Government Forced Fable into a Temporary Halt
The US government's response to the AI model Fable — the multi-layered regulatory, political, and technical chain that led to the temporary halt.
Read more →The True Identity of Anthropic Fable 5.0 — The Name on the Surface and What Lies Beneath
What is Anthropic Fable 5.0? The name on the surface and the technological, organizational, and regulatory structures behind it.
Read more →Lessons from History — From Stagecoach to Railway, Windows, and AI
The history of technological transformations — stagecoach to railway, automobile, Windows, internet. The judgment structure of "ride" or "leave" in the AI era.
Read more →At the Threshold of Intelligence Explosion — How Society Will Change, How to Prepare
Signs of an "intelligence explosion" pointed to by recent AI trends. Structural changes in society and what to prepare for as pharma professionals.
Read more →The "New Intelligence" in the AI Era — A Five-Layer Structure of Literacy
What is the "new intelligence" demanded by the AI era? A five-layer model of literacy and the diverging points where society splits, examined from the pharmaceutical field.
Read more →AI Reshapes the Social Structure — A 2024-2025 Layoff Retrospective and 2026-2028 Outlook
Over 240,000 tech-industry layoffs in 2024-2025 as AI took hold. A structured retrospective and a forward look to 2026-2028, when AI agents go mainstream and 50-55% of roles get redesigned. Implications for pharma.
Read more →Novartis CEO Joins Anthropic's Board — A Symbol of AI–Pharma Strategic Fusion
April 2026: Vas Narasimhan joins Anthropic's Board via the Long-Term Benefit Trust. The essential meaning of "responsible AI" directly connecting with pharma, and two implications for the industry.
Read more →AI Drug Discovery, the Last Year — "The Year of Clinical Validation"
173+ AI-derived candidates in clinical trials, early-Phase I success at 80-90%, Insilico's rentosertib clears Phase IIa as first generative-AI-designed drug. Pfizer-Boltz, Recursion × Exscientia, AlphaFold3's 1000× speedup.
Read more →AI Clinical Practice in the Middle East and China — Saudi's "Dr Hua" and China's 300 Models
Saudi Arabia's world-first AI doctor clinic, UAE Burjeel × Hippocratic AI, China's 300 medical AI models and 68 million remote diagnoses. Why emerging markets are leading AI clinical deployment.
Read more →Is AI Just the World of Chatbots? — What Humanoid Spread Means
AI = chatbot? Only a fraction. China-led humanoid robots are spreading explosively, taking real work in factories and warehouses. What it means when AI is released into the physical world.
Read more →China's "AI Plus" Plan, and Society Already Changing
State Council of China's August 2025 "AI Plus" Action Plan. A national strategy to integrate AI across society, reshaping industry, daily life, and governance. 86% generative AI usage, 1.2 trillion yuan industry — read in numbers.
Read more →Microsoft unveils 7 AI models in one shot — plus the Scout autonomous agent
Microsoft just released 7 in-house AI models including the reasoning-focused MAI-Thinking-1, and launched the Scout autonomous agent built on OpenClaw — a clear signal that the company is starting to distance itself from OpenAI.
Read more →Stanford AI Index 2026 — Five Numbers That Show AI Becoming Infrastructure
Adoption faster than PC or internet, US-China model gap under 1%, $172B in enterprise value, data centers eating country-scale power, and edge AI going mainstream. The five numbers that mark AI's shift from technology to infrastructure.
Read more →Agentic AI 2026 ── From "AI that answers" to "AI that works"
The era of AI that "just responds" is ending. With Anthropic Computer Use, Google Gemini Spark, and NVIDIA local AI chips all arriving, 2026 marks the turning point at which AI becomes a worker. A high-school-readable look at the rookie-to-mid-career "promotion" happening now.
Read more →NVIDIA "RTX Spark" ── The day AI moves into your PC
The new superchip unveiled at NVIDIA GTC Taipei on June 1. The starting line of an era where ChatGPT-class AI runs locally on your PC, no cloud needed. Using a "Uber Eats vs. cooking at home" metaphor, we read what it means for the pharmaceutical field.
Read more →Who is the Anthropic "mythos"? ── The story of the AI company most afraid of itself
Claude, the ChatGPT rival, is made by Anthropic. They publicly declare "We're the most worried company about AI, for humanity's sake," and yet they sell some of the world's strongest AI. Where does that contradiction come from? A relaxed, high-school-readable explainer.
Read more → ← Back to home