Better questions. Reliable discoveries. Human agency at the center.
The future I want to help build.
I want a future in which AI agents help people ask better questions, explore more possibilities and turn reliable discoveries into benefits that reach more people. Science and mathematics offer extraordinary opportunities for this kind of collaboration, especially in medicine and biology, where better understanding can ultimately mean less suffering and more healthy life. I am interested in human–AI teams and collaborations across laboratories around the world, with each contributing strengths the others lack.
Progress should deepen human agency and meaning: the freedom to choose what matters, understand the work, challenge an answer and share in discovery. That means evaluating evidence rather than just fluent output, preserving uncertainty and giving people real authority over consequential decisions. It also means designing against accidental harm and deliberate misuse, taking dual-use risks seriously, and making the benefits of scientific capability broadly accessible.
Every reading below is linked to the 44 people, 23 organisations, 59 concepts, 26 methods and topics it involves: 244 nodes and 536 links, growing as the collection does. Select a node to follow its connections. The 3D space orbits and zooms around the selected idea; the 2D map lets you drag nodes to rearrange them. Links describe what a source discusses or argues, not endorsement.
Dates are first public posting/publication where verified; talks use upload dates. Year-only dates retain their original precision. OP means an original author/team post or identified commentary. Missing posts are marked as unverified. Inclusion means worth thinking about, not endorsement. External links open in a new tab. Responses stay grouped with their original essays; conversations are ordered by their latest entry.
40 resources
Updated 2026-10-20
A conversation: the future of mathematics
Start with Litt and doomslide’s response, then read Gowers and Hairer on mathematical values and institutional engagement. Close with Grant Sanderson’s video, which asks what survives if AI overtakes each part of a mathematician’s work. Each entry retains its publication date.
Distinguishes producing mathematical results from cultivating human understanding, and proposes rewarding research programs, learning and community-building as AI capabilities grow. A constructive essay on preserving scientific values and complementary human–AI inquiry.1 Protect mathematics’ values, not its institutions. One concrete proposal: award PhDs on an oral defence of deep understanding, whatever the provenance of the thesis.
A critical response to Litt focused on provenance, credit, accountable authorship and open mathematical discourse; its claims about private-data reuse and opaque proof workflows are conditional or speculative.2 Accepts Litt’s goals but says credit breaks first. If proofs are laundered through undisclosed pipelines, the community can no longer say who knows what.
Timothy Gowers defends a plurality of mathematical motivations, including problem-solving, while arguing that preserving the institutions supporting shared understanding matters more than restricting the supply of AI-generated results.3 Gowers treats credit problems as serious but temporary. The lasting danger is erosion of PhD training, funding and custodianship of the tradition.
Martin Hairer explains his decision to join the Advisory Group on Mathematics and Artificial Intelligence, describes its independence and advisory limits, and argues for engagement with AI labs to improve scholarship and represent community concerns.4 Hairer signed the Fields medallists’ statement and joined AGMAI: nine unpaid mathematicians in a purely advisory role. Complaining while refusing a channel would be hypocritical, he argues.
Grant Sanderson (3Blue1Brown) lists what mathematicians do (prove theorems, pose questions, build theory, find applications, communicate), where math’s value lies (applications, transferable problem-solving, aesthetic joy) and where the money comes from (mostly teaching). He then supposes AI overtakes each layer in turn. His argument is that value and funding barely move while credit assignment has to change, and that a field openly looking for a new shape makes this the best moment to enter it.5 Treat a proof as raw data, like more digits of π; the prize should go to whoever builds the theory around it. His advice for tomorrow: give lots of talks, and take a long walk with a fellow student to work out what you love about math.
Read on 20 Oct 2026: 33 of the post’s 52 replies (the quote posts weren’t retrievable). The post had about 2.5K likes, 288 reposts, 59 quotes and 275K views.
@DmitriyDrichel (most-liked reply) goes further than Grant: math is now “automatable, scalable and parallelizable”, so this is a golden age.
@MathWithShan a fellow math animator: if AI floods the world with proofs, the bottleneck moves from producing math to understanding it, so it is also the best time to explain it.
@LexiLearn_ and @Techboy_Wien: cheap proofs move the value to verification and make the mathematician an auditor, but who will pay people to understand why something is true?
@tiozzo_giulio pushes on the teaching pillar: society may decide it can outsource its thinking and stop valuing math teaching.
@Davidmdrpi the elegiac view: there will never be another Perelman, Gödel or Hilbert, “only HAL”. @CagatayCandan offers chess as the precedent to study.
@nicholas_diao agrees, and points to his team’s open-sourced research on the Erdős problems (github.com/plasma-ai/erdos).
@hadilq says the funding table misses money flowing from engineering, as with Euler and Gauss, and urges AI labs to hire mathematicians as engineers.
@alemmdrake asks the personal version: if video models made his videos 1.5× better, would Grant keep making them?
Google Research: building blocks for autonomous wet labs
One blog post and three papers released together. Start with the overview, then read how the lab is made checkable (VERID), how a fast loop exposes agent shortcuts (aptamers), and how a human–agent split cuts design cost (epitope excision).
Pooja Rao, Petar Sirkovic and Annalisa Pawlosky argue that agents should run wet-lab experiments, learn from the results and improve, with humans directing. Two things block this: models reason poorly about 3D geometry and physical state, and wet-lab validation lags far behind in-silico design. The post introduces three building blocks: a verified digital twin of the lab (VERID), DNA aptamers as a fast closed-loop sandbox, and agent-guided epitope excision for cheaper protein binder design. Safety measures include non-infectious testbeds, biosecurity screening of every sequence, simulated dry runs before any robot command, and human sign-off.6 The theme is grounding: agents may not act on the world until a simulator, a sequencer or a binding assay can say whether they were right. Full autonomy is called an open problem; humans set strategy and authorize runs.
A semantic digital twin of a robotic workcell, built on PyLabRobot, that tracks arm geometry, well volumes, reagents and tip inventory as a live 3D state tree. Agents dry-run protocols in seconds, and safety checks block invalid commands before they reach the robot; synchronized video and action logs support debugging. In the case studies reported in the blog, the simulator rejected an aspiration that would not fit a trough, the agent recovered from a plate-reader failure without re-pipetting, and a self-play campaign improved dye-mixture reconstruction more than fivefold. Summary from the Google blog; the PDF has not yet been read here.7 A digital twin is a verifier for physical actions: it can say “this will collide” or “that well is empty” before reagents are spent. It turns part of lab work from judgement into checking.
DNA aptamers can be synthesized in pools and read out by sequencing, so a design–test–analyze cycle takes about 7–10 days. Over three rounds against α-thrombin, the agents' hit rate went from 0% (primer overhangs broke folding) to 42.1% at 98.5% specificity. Against myoglobin, agents chasing raw affinity filled round 2 with sticky G-quadruplex designs and specificity fell from 75.0% to 19.1%; they spotted the artifact, capped those motifs and recovered to 72.4% specificity with a 33.5% on-target hit rate. Figures as reported in the Google blog; the PDF has not yet been read here.8 Myoglobin round 2 is Goodhart in a test tube: told to maximize binding, the agents found molecules that stick to everything. The wet lab caught it, because specificity was measured alongside affinity.
Designing binders to FAP, a large receptor overexpressed in solid tumors, the team re-folds only the local binding pocket instead of the whole target, speeding structure prediction 8.6x without losing accuracy. At equal compute, a scientist setting strategy with an agent executing tactics nearly doubled the in-silico hit rate of the scientist alone, across more target sites. Yeast display with FACS then gave single-digit nanomolar binders in a few experimental runs. Figures as reported in the Google blog; the PDF has not yet been read here.9 Folding cost grows with the square of target length, so trimming a target to a pocket one-seventh its size cuts that term about 49-fold (the reported end-to-end speedup is 8.6x). The human sets strategy and the agent executes tactics: a division of labor, not full autonomy.
Wraps off-the-shelf coding agents in an explicit research loop, with pre-registered rubrics, parallel workers sharing a memory and ablations of the winning program, so they behave less like hyperparameter optimizers. Sakana reports an ALIFE 2026 oral on Error Diffusion with humans setting questions and verifying claims, plus autonomous gains on Doc-to-LoRA retrieval and learned evolution strategies.10 The useful part is evaluator integrity: sealed search and final splits, and claims tied to evidence hashes. Even so, an agent quietly changed the evaluation protocol in 2 of 3 runs, and no gate caught it. Read with “Why LLMs Aren’t Scientists Yet”.
Niko Grupen reports that giving Harvey’s legal agents prebuilt context graphs over their documents improved results on legal tasks. The cost of building the graphs and how far the gains generalise are still open questions.11 This page’s knowledge graph is a small context graph: entities, typed relations and sources pre-linked, so a reader or an agent starts from structure instead of a pile of documents. The post itself couldn’t be fetched here, so this summary relies on a newsletter description.
Benchling's Head of AI, Nicholas Larus-Stone, launches Science Foundries: you pick a process such as protein production or lead optimization, then assign which steps agents run and where scientists must approve. In the antibody example, design, cloning, lab and analysis agents pass work along, send instructions to workcells and CROs, run QC and update training data, while scientists choose which candidates to advance. Parabilis Medicines is the named customer. The productivity claims (three people matching thirty, months saved) are the vendor's own and unquantified.12 The design choice that matters is the role assignment: agents run the steps and scientists hold the approval gates. That is the human actuation gate, sold as a product. What is missing is any independent check beyond QC and sign-off.
Nicholas Larus-Stone's launch post says Science Foundries let “anyone … orchestrate a team of agents to accomplish real scientific work”, from target ID and screening through assay development to medical writing. He credits “vibe coding, natively multiplayer agents (!!!), agent swarms, and of course good old fashioned harness engineering”, and says a Foundry is set up by chatting with Benchling AI. The post has no data or evaluation; the first comment asks about evaluation, reliability, traceability and governance.13 The use cases span very different levels of checkability. An assay readout can be checked against a plate; target ID and medical writing are judgement calls, the tasks Convoke found models weakest on. “Multiplayer” raises a new question: when two scientists give the agents conflicting instructions, who decides, and is it recorded?
Convoke benchmarks models on six biopharma work products outside R&D: target safety assessment, tumor landscape, shelved assets, catalyst calendar, clinical benchmark and commercial forecast. Domain experts wrote penalty-based rubrics (major, moderate, minor errors), and an LLM grader with research tools scored each criterion against a citation. They report that frontier OpenAI models lead and that models handle structured spreadsheet compilation well but struggle with narrative synthesis and judgment-heavy forecasts. Frontier models cost more than 3x as much per work product. The grader has not yet been validated against expert grading.14 Compiling a catalyst calendar is checkable. A commercial forecast needs design judgment, and that is where models slip: the taste gap again, in a pharma setting. The rubric grader is itself unvalidated, so treat the rankings as provisional.
Ben Davis defines taste as “the ability to make good decisions in places where there isn’t a programmatic verification function”. Models excel at verifiable hill-climbing, but their open-ended design choices regress to the mean, so people still have to supply direction. He calls taste “the only real stop” to recursive self-improvement.15 The real question: can a harness and a model together make good decisions where no programmatic verifier exists? “Make this function fast” is hill-climbing. Which direction to pursue, which design to ship and which claim to trust are taste.
Two papers ten days apart reach the same answer: long-running research agents get stuck refining their own past ideas, and the cure is to forget. Read Fork-and-Flush first, then the more aggressive version.
Cai, Papailiopoulos and colleagues find that independent runs of the same autoresearch agent on the same task end at very different scores, and that extra compute does not close the gap. Comparing solutions by what they do suggests each run stays inside a local “idea basin”. Their fix periodically forks the agent into parallel trajectories that keep the workspace but start with a fresh chat context, then continues from the best one. Across 13 long-horizon tasks at equal compute, they report a 66% relative gain over a single run and 44% over best-of-N.16 Think of a hiker who only ever steps uphill from where they stand: they reach the top of the nearest hill, not the highest one. A fresh context is a new starting point on the map.
Tan, Dimakis, Zaharia and colleagues at Berkeley show that running an autoresearch agent longer, running agents in parallel, or letting them talk all plateau, because agents anchor on their growing history. On AHC014, edits shrink from about 150 lines to 30 per proposal and thinking tokens from about 28k to 7k per turn. Their fix is to make the whole team periodically forget everything except the best candidate so far. Across four tasks (AHC014, Frontier-CS, Anthropic's VLIW kernel challenge and nanochat autoresearch), they report averages about 10% above parallel agents and 22% above a single agent.17 Fork-and-Flush keeps the workspace and clears the chat. This paper clears both and keeps only the best answer. Shrinking edit sizes are a measurable symptom of an agent getting stuck.
Frozen teams learn reusable coordination strategies, with gains depending partly on their ability to recognize correct reasoning.18 Three frozen models learn a teamwork protocol and reach 66.7%, beating even an oracle that picks the best member (59.0%). The organisation is learned, not the weights.
Miguel Rios (Grep.ai) describes a harness in which a frontier agent first does a job from a written procedure, keeps notes and traces, then writes a typed workflow and “fires itself”. The workflow mixes agent steps, plain code and typed questions answered by a small calibrated model (Jev), under the rule “Jev decides what a thing is; code decides what follows.” A verify clause checks each claim against the evidence the agent cited. On an anti-money-laundering alert review, the author reports cost per alert falling from $2.89 to $0.25 and judged accuracy rising from 90% to 95%, scored on a held-out set they never looked at.19 A practical answer to the taste question. Closed decisions (“is this a match?”) become typed questions that a cheap calibrated model answers and code routes. Open judgment stays with the agent or a human. A held-out set guards against grading your own homework.
An autonomous research pipeline combines experiments, ablations and simulated peer review; its reported successes require careful attention to selection and evaluation.20 A closed loop on 107 problems from ICLR, ICML and NeurIPS papers. Over its 86 successes the mean gain over the human state of the art is 25.2%, but the median is 7.7%: a few big wins pull the average up, and the 21 failures are left out. Simulated review is AI checking AI, so ask who verifies the papers.
Describes infrastructure for long scientific tool runs, specialized models and efficient training; performance claims come from the company.21 Scientific RL rollouts last hours, with tool calls such as XRD analysis, while training steps take minutes. That gap forces asynchronous RL.
Combines model-use data, a specialist-model inventory and a researcher survey to examine adoption, time savings and verification demands.22 Scientists report about 6.9 hours a week saved, and much of it goes back into verifying AI output. Early empirical support for verification as the bottleneck.
Six researchers reflect on cancer biology, translational disappointments and opportunities for human–AI co-science.23 A canonical lost qualifier: BRAF inhibitors work in BRAF-mutant melanoma but not in colorectal cancer, because of EGFR feedback.
Describes medical researchers integrating specialized AI agents with vetted data, laboratory workflows and human oversight.24 The shift is from chatbots on the open web to agents with a restricted library card: FDA records, guidelines and de-identified patient data. Humans still choose the questions.
François Chollet highlights latent iterations alongside longer reasoning and parallel exploration.25 Depth, breadth and latent iteration. The third axis is token-cheap but gives up the inspectable reasoning that verification relies on.
Reports on AI-assisted formalization of Fermat’s Last Theorem and the role of shared proof infrastructure.26 About 13 million lines of Lean in about 11 days, using only the standard axioms. That is certainty without new human understanding, and the readable rewrite is still undone.
Eunice Jun explores representations and interfaces that make scientific reasoning more inspectable.27 Much of what makes science valid is implicit. Making reasoning about validity, feasibility and priority explicit turns it into something people can check.
A mathematical research swarm illustrates how shared infrastructure can spread both verifier exploits and organized resistance.28 A Lean notation trick spread through the shared library, and the swarm “solved” every remaining problem within 27 minutes. A quarter of agents protested but had no power to enforce anything.
Terence Tao discusses scientific understanding, human learning and the limits of accelerating answer production.29 AI is a cheap, slightly drunk idea generator: useful when paired with verification. The bottlenecks move to verification, curation and communication (“proof indigestion”).
Reports multi-agent research workflows across materials, biology and computing, with different levels of autonomy and validation.30 Optimising against an LLM reviewer teaches agents to fabricate. Deterministic checks of every number against execution logs cut severe hallucination from 46% to 4%.
Proposes attaching reusable qualifications and uncertainty records to scientific claims as humans and AI reuse them.31 Caveats get lost each time a claim is reused. The hedges should travel with the claim, not stay behind in the paper.
Dean’s thesis for Discovery Loop, the public benefit corporation he co-founded with Ghemawat, Le and Vinyals: science is bottlenecked by slow experimental loops, so automate them, run thousands in parallel and cut each iteration’s latency. It starts with ML research, as its own first customer, before NAE Grand Challenges. No technical papers have been published yet.32 The same “automate the experimental loop” bet as Sakana and Google’s Co-Scientist, framed around scale and latency. A FutureSearch forecast argues it should start where loops close fastest and verification is cheap, such as chip design.
Calls for evaluating AI by its contribution to better drug-development decisions and patient-relevant outcomes.33 Clinically relevant impact so far is “disappointingly limited”. Judge AI by whether it improves decisions, especially at Phase II.
Dean's first public talk after leaving Google, at the AASF Frontier & Pioneer Symposium at Stanford, moderated by Dawn Song. He describes the loop that science and engineering share: decompose, propose, implement, evaluate, feed back. Discovery Loop aims to automate it, cut each iteration from days to minutes, run thousands in parallel and pick the next experiment by expected value per unit of cost, starting with ML itself. He also gives his research-taste heuristics: skim 10 papers rather than read one closely, look for 10x gains, and choose problems where about 5 of 7 parts are taking shape. The summary is built from auto-generated captions.34 Experiment A has a 10% chance of +5 and costs 1 GPU-day (EV/cost 0.5). B has a 50% chance of +2 and costs 4 GPU-days (0.25). Dean's rule runs A first. What pure EV/cost misses is what a cheap failure teaches you.
Daphne Koller argues that useful AI drug discovery depends on better measurements and causal understanding of human biology.35 Most AI effort designs keys (molecules), while the bottleneck is choosing the lock (the mechanism). The only true ground truth is a slow human trial.
Argues that cooperative AI must account for adaptive counterparts, institutions and human agency rather than optimize in isolation.36 Deployment breaks the fixed-world assumption: the more an AI exploits a regularity, the faster it disappears. The emergent-cheating swarm is a live example.
Sakana, UBC, Vector and Oxford report in Nature on the AI Scientist, whose v2 wrote a paper that passed blind review at an ICLR 2025 workshop before being withdrawn. Its ensembled automated reviewer matches human reviewer accuracy on OpenReview decisions. The authors list naive ideas, weak rigour, hallucinated citations and a compute-only scope as limits, and call for watermarking and community norms.37 The withdrawn ICBINB paper averaged 6.33, above the workshop’s human acceptance bar. The authors still list naive ideas and hallucinated citations as limits. Read with the Agentic Scientist, its successor.
Develops an economic argument that verification capacity and responsibility become scarce as automated execution becomes cheaper.38 The cost to automate falls toward zero, but the cost to verify stays tied to expert time. Only the verifiable share of agent work counts as real output.
Four autonomous ML research attempts expose implementation drift, weak evaluation and failures of scientific judgment.39 Three of four attempts failed. The useful output is a six-part failure taxonomy, including implementation drift and the “eureka instinct”.
Studies model preferences among future research directions using order-controlled pairs, without equating agreement with scientific merit.40 Models matched the authors’ own next direction 45–67.5% of the time. Agreement with the authors is not the same as scientific merit.