{"updated":"2026-10-20","relations":{"tagged":{"label":"about","inverse":"readings"},"authored_by":{"label":"written or presented by","inverse":"author of"},"commented_by":{"label":"original post or commentary by","inverse":"posted about"},"affiliated_with":{"label":"works at or leads","inverse":"people"},"published_in":{"label":"published in","inverse":"publishes"},"responds_to":{"label":"responds to","inverse":"responses"},"discusses":{"label":"discusses","inverse":"discussed in"},"proposes":{"label":"proposes or argues for","inverse":"proposed in"},"critiques":{"label":"warns about or critiques","inverse":"critiqued in"},"uses_method":{"label":"uses method","inverse":"used in"},"releases":{"label":"releases artifact","inverse":"released by"},"about_system":{"label":"is about the system","inverse":"written about in"},"made_by":{"label":"made by","inverse":"made"},"related_to":{"label":"related to","inverse":"related to"},"builds_on":{"label":"builds on","inverse":"built on by"}},"types":["resource","topic","concept","method","person","org","venue","system","artifact"],"nodes":[{"id":"scientisttwo","type":"resource","label":"ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI","kind":"Preprint","url":"https://arxiv.org/abs/2609.19644v1","date":"2026-09-17","description":"An autonomous research pipeline combines experiments, ablations and simulated peer review; its reported successes require careful attention to selection and evaluation.","anchor":"#scientisttwo"},{"id":"t-agents","type":"topic","label":"Agents"},{"id":"t-evaluation","type":"topic","label":"Evaluation"},{"id":"t-scientific-discovery","type":"topic","label":"Scientific discovery"},{"id":"self-organizing-teams","type":"resource","label":"Self-Organizing Agent Teams Learn to Reason Together","kind":"Preprint","url":"https://arxiv.org/abs/2609.22682v1","date":"2026-09-19","description":"Frozen teams learn reusable coordination strategies, with gains depending partly on their ability to recognize correct reasoning.","anchor":"#self-organizing-teams"},{"id":"t-collaboration","type":"topic","label":"Collaboration"},{"id":"solipsistic-superintelligence","type":"resource","label":"Solipsistic Superintelligence is Unlikely to be Cooperative","kind":"Paper","url":"https://arxiv.org/abs/2606.03237v1","date":"2026-06-02","description":"Argues that cooperative AI must account for adaptive counterparts, institutions and human agency rather than optimize in isolation.","anchor":"#solipsistic-superintelligence"},{"id":"t-governance","type":"topic","label":"Governance"},{"id":"t-human-agency","type":"topic","label":"Human agency"},{"id":"economics-agi","type":"resource","label":"Some Simple Economics of AGI","kind":"Working paper","url":"https://arxiv.org/abs/2602.20946","date":"2026-02-24","description":"Develops an economic argument that verification capacity and responsibility become scarce as automated execution becomes cheaper.","anchor":"#economics-agi"},{"id":"t-verification","type":"topic","label":"Verification"},{"id":"t-economics","type":"topic","label":"Economics"},{"id":"drug-discovery-path","type":"resource","label":"Artificial intelligence in drug discovery—what it is, where we stand and the path forward","kind":"Perspective","url":"https://doi.org/10.1038/s41573-026-01496-2","date":"2026-08-07","description":"Calls for evaluating AI by its contribution to better drug-development decisions and patient-relevant outcomes.","anchor":"#drug-discovery-path"},{"id":"t-biology","type":"topic","label":"Biology"},{"id":"t-medicine","type":"topic","label":"Medicine"},{"id":"cancer-25-years","type":"resource","label":"Reflecting on 25 years of cancer research: transformative advances and unmet expectations","kind":"Viewpoint","url":"https://doi.org/10.1038/s41568-026-00974-4","date":"2026-09-08","description":"Six researchers reflect on cancer biology, translational disappointments and opportunities for human–AI co-science.","anchor":"#cancer-25-years"},{"id":"t-human-ai-collaboration","type":"topic","label":"Human–AI collaboration"},{"id":"uncertainty-layer","type":"resource","label":"The scientific paper needs an uncertainty layer","kind":"Perspective preprint","url":"https://doi.org/10.31222/osf.io/duymx_v1","date":"2026-08-20","description":"Proposes attaching reusable qualifications and uncertainty records to scientific claims as humans and AI reuse them.","anchor":"#uncertainty-layer"},{"id":"t-evidence","type":"topic","label":"Evidence"},{"id":"t-scientific-communication","type":"topic","label":"Scientific communication"},{"id":"four-autonomous-attempts","type":"resource","label":"Why LLMs Aren’t Scientists Yet: Lessons from Four Autonomous Research Attempts","kind":"Preprint","url":"https://arxiv.org/abs/2601.03315v1","date":"2026-01-06","description":"Four autonomous ML research attempts expose implementation drift, weak evaluation and failures of scientific judgment.","anchor":"#four-autonomous-attempts"},{"id":"t-research-taste","type":"topic","label":"Research taste"},{"id":"cheating-whistleblowing","type":"resource","label":"A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms","kind":"Preprint","url":"https://arxiv.org/abs/2609.04170v1","date":"2026-09-03","description":"A mathematical research swarm illustrates how shared infrastructure can spread both verifier exploits and organized resistance.","anchor":"#cheating-whistleblowing"},{"id":"t-mathematics","type":"topic","label":"Mathematics"},{"id":"research-taste","type":"resource","label":"Eliciting Research Taste in LLMs through Future Research Direction Choice","kind":"Workshop paper","url":"https://github.com/Lossfunk/research-taste-eval-v1","date":"2026","description":"Studies model preferences among future research directions using order-controlled pairs, without equating agreement with scientific merit.","anchor":"#research-taste"},{"id":"no-magic-wands","type":"resource","label":"Drug Discovery Has No Magic Wands","kind":"Essay","url":"https://deepphenotype.substack.com/p/drug-discovery-has-no-magic-wands","date":"2026-08-03","description":"Daphne Koller argues that useful AI drug discovery depends on better measurements and causal understanding of human biology.","anchor":"#no-magic-wands"},{"id":"gemini-real-world","type":"resource","label":"Accelerating Scientific Research with Gemini in the Real-World","kind":"Preprint","url":"https://arxiv.org/abs/2608.26701v1","date":"2026-08-27","description":"Reports multi-agent research workflows across materials, biology and computing, with different levels of autonomy and validation.","anchor":"#gemini-real-world"},{"id":"t-materials","type":"topic","label":"Materials"},{"id":"claude-fermat-report","type":"resource","label":"Anthropic uses Claude to formalize proof of Fermat’s Last Theorem","kind":"News article","url":"https://siliconangle.com/2026/09/04/anthropic-uses-claude-to-formalize-proof-of-fermats-last-theorem/","date":"2026-09-04","description":"Reports on AI-assisted formalization of Fermat’s Last Theorem and the role of shared proof infrastructure.","anchor":"#claude-fermat-report"},{"id":"t-formal-verification","type":"topic","label":"Formal verification"},{"id":"medical-lab-partner","type":"resource","label":"Medical school lab scientists get a new partner: AI","kind":"News article","url":"https://www.aamc.org/news/medical-school-lab-scientists-get-new-partner-ai","date":"2026-09-08","description":"Describes medical researchers integrating specialized AI agents with vetted data, laboratory workflows and human oversight.","anchor":"#medical-lab-partner"},{"id":"chollet-third-axis","type":"resource","label":"Latent-space reasoning as a third axis of test-time scaling","kind":"Post","url":"https://x.com/fchollet/status/2096310217898356739","date":"2026-09-05","description":"François Chollet highlights latent iterations alongside longer reasoning and parallel exploration.","anchor":"#chollet-third-axis"},{"id":"t-reasoning","type":"topic","label":"Reasoning"},{"id":"t-compute","type":"topic","label":"Compute"},{"id":"tao-ai-science","type":"resource","label":"The paradox at the heart of AI and science","kind":"Talk","url":"https://www.youtube.com/watch?v=svl_1upFpQo","date":"2026-09-03","description":"Terence Tao discusses scientific understanding, human learning and the limits of accelerating answer production.","anchor":"#tao-ai-science"},{"id":"jun-future-science","type":"resource","label":"Imagining a new future for science","kind":"Talk","url":"https://www.youtube.com/watch?v=hawgxIEpPqU","date":"2026-09-04","description":"Eunice Jun explores representations and interfaces that make scientific reasoning more inspectable.","anchor":"#jun-future-science"},{"id":"periodic-infrastructure","type":"resource","label":"AI Infrastructure at Periodic","kind":"Technical article","url":"https://periodic.com/news/ai-infrastructure-at-periodic","date":"2026-09-15","description":"Describes infrastructure for long scientific tool runs, specialized models and efficient training; performance claims come from the company.","anchor":"#periodic-infrastructure"},{"id":"t-infrastructure","type":"topic","label":"Infrastructure"},{"id":"ai-in-science","type":"resource","label":"AI in Science: Early Insights","kind":"Report","url":"https://ai.google/static/documents/AI-in-Science.pdf","date":"2026-09-15","description":"Combines model-use data, a specialist-model inventory and a researcher survey to examine adoption, time savings and verification demands.","anchor":"#ai-in-science"},{"id":"litt-beginning-mathematics","type":"resource","label":"A beginning for mathematics","kind":"Essay","url":"https://www.daniellitt.com/blog/2026/9/13/a-beginning-for-mathematics/","date":"2026-09-13","description":"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.","anchor":"#litt-beginning-mathematics"},{"id":"doomslide-mathematics","type":"resource","label":"Mathematics is effectively dead","kind":"Essay","url":"https://doomslide.substack.com/p/mathematics-is-effectively-dead","date":"2026-09-17","description":"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.","anchor":"#doomslide-mathematics"},{"id":"gowers-fields-letter","type":"resource","label":"Why I didn’t sign the Fields medallists’ letter","kind":"Essay","url":"https://gowers.wordpress.com/2026/09/17/why-i-didnt-sign-the-fields-medallists-letter/","date":"2026-09-17","description":"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.","anchor":"#gowers-fields-letter"},{"id":"hairer-agmai","type":"resource","label":"Why I agreed to join AGMAI","kind":"Essay","url":"https://proofsandprompts.com/2026/09/22/why-i-agreed-to-join-agmai/","date":"2026-09-22","description":"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.","anchor":"#hairer-agmai"},{"id":"agentic-scientist","type":"resource","label":"The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery","kind":"Preprint","url":"https://arxiv.org/abs/2610.18933","date":"2026-10-15","description":"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.","anchor":"#agentic-scientist"},{"id":"ai-scientist-nature","type":"resource","label":"The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature","kind":"Research announcement","url":"https://sakana.ai/ai-scientist-nature/","date":"2026-03-26","description":"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.","anchor":"#ai-scientist-nature"},{"id":"discovery-loop-launch","type":"resource","label":"Jeff Dean on Launching Discovery Loop","kind":"Interview","url":"https://radical.vc/articles/radical-reads-jeff-dean-on-launching-discovery-loop/","date":"2026-08-09","description":"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.","anchor":"#discovery-loop-launch"},{"id":"grupen-context-graphs","type":"resource","label":"Context graphs for document-heavy legal agents","kind":"Post","url":"https://x.com/nikogrupen/status/2110773803592937733","date":"2026-10-15","description":"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.","anchor":"#grupen-context-graphs"},{"id":"davis-labs-taste","type":"resource","label":"The Labs Have the Best Models With Infinite Tokens, So Why Aren’t Their Products Better?","kind":"Essay","url":"https://davis7.substack.com/p/the-labs-have-the-best-models-with","date":"2026-10-14","description":"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.","anchor":"#davis-labs-taste"},{"id":"agentrun-grep","type":"resource","label":"AgentRun: a harness for repetitive knowledge work","kind":"Article","url":"https://x.com/MiguelriosEN/status/2101029313906987422","date":"2026-09-18","description":"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.","anchor":"#agentrun-grep"},{"id":"convoke-biopharma-benchmark","type":"resource","label":"Benchmarking biopharma knowledge work","kind":"Blog post","url":"https://www.convoke.bio/blog/benchmarking-biopharma-knowledge-work","date":"2026-09-29","description":"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.","anchor":"#convoke-biopharma-benchmark"},{"id":"t-biopharma","type":"topic","label":"Biopharma"},{"id":"dean-dawnsong-fireside","type":"resource","label":"Jeff Dean in conversation with Dawn Song: from answering questions to making discoveries","kind":"Talk","url":"https://www.youtube.com/watch?v=0kC3xOZChdA","date":"2026-08-07","description":"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.","anchor":"#dean-dawnsong-fireside"},{"id":"benchling-science-foundries","type":"resource","label":"Introducing Science Foundries, a new way for scientists and agents to work together","kind":"Blog post","url":"https://www.benchling.com/blog/introducing-science-foundries","date":"2026-10-15","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.","anchor":"#benchling-science-foundries"},{"id":"fork-and-flush","type":"resource","label":"Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents","kind":"Preprint","url":"https://arxiv.org/abs/2610.07447","date":"2026-10-05","description":"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.","anchor":"#fork-and-flush"},{"id":"scaling-discovery-forgetting","type":"resource","label":"Scaling Discovery Requires Forgetting","kind":"Preprint","url":"https://arxiv.org/abs/2610.20563","date":"2026-10-14","description":"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.","anchor":"#scaling-discovery-forgetting"},{"id":"larus-stone-foundries-post","type":"resource","label":"“We just launched Science Foundries at Benchtalk!”","kind":"Post","url":"https://www.linkedin.com/feed/update/urn:li:activity:7516501677785624576/","date":"2026-10-15","description":"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.","anchor":"#larus-stone-foundries-post"},{"id":"google-wet-lab-agents","type":"resource","label":"Grounding AI agents in biological reality: Building blocks for autonomous wet labs","kind":"Blog post","url":"https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/","date":"2026-10-16","description":"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.","anchor":"#google-wet-lab-agents"},{"id":"verid-semantic-lab","type":"resource","label":"VERID: VERified Infrastructure for Discovery Laboratories","kind":"Paper","url":"https://www.gstatic.com/science-ai/verid_semantic_lab.pdf","date":"2026-10-16","description":"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.","anchor":"#verid-semantic-lab"},{"id":"aptamer-sandbox","type":"resource","label":"Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery","kind":"Paper","url":"https://www.gstatic.com/science-ai/agentic_aptamer_sandbox.pdf","date":"2026-10-16","description":"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.","anchor":"#aptamer-sandbox"},{"id":"epitope-excision","type":"resource","label":"Agent-guided epitope excision enables economical de novo binder design","kind":"Paper","url":"https://www.gstatic.com/science-ai/epitope_excision.pdf","date":"2026-10-16","description":"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.","anchor":"#epitope-excision"},{"id":"3b1b-better-time-for-math","type":"resource","label":"There’s never been a better time to go into math","kind":"Talk","url":"https://www.youtube.com/watch?v=LKrwhjOI8Nc","date":"2026-10-16","description":"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.","anchor":"#3b1b-better-time-for-math"},{"id":"p-daniel-litt","type":"person","label":"Daniel Litt"},{"id":"p-doomslide","type":"person","label":"doomslide"},{"id":"p-timothy-gowers","type":"person","label":"Timothy Gowers"},{"id":"p-martin-hairer","type":"person","label":"Martin Hairer"},{"id":"p-terence-tao","type":"person","label":"Terence Tao"},{"id":"p-aneesh-pappu","type":"person","label":"Aneesh Pappu"},{"id":"p-liam-fedus","type":"person","label":"Liam Fedus"},{"id":"p-arthur-turrell","type":"person","label":"Arthur Turrell"},{"id":"p-francois-chollet","type":"person","label":"François Chollet"},{"id":"p-eunice-jun","type":"person","label":"Eunice Jun"},{"id":"p-richard-sever","type":"person","label":"Richard Sever"},{"id":"p-daphne-koller","type":"person","label":"Daphne Koller"},{"id":"o-agmai","type":"org","label":"AGMAI","description":"Advisory Group on Mathematics and Artificial Intelligence."},{"id":"o-periodic","type":"org","label":"Periodic Labs","url":"https://periodic.com/"},{"id":"o-google","type":"org","label":"Google"},{"id":"o-anthropic","type":"org","label":"Anthropic"},{"id":"o-lossfunk","type":"org","label":"Lossfunk"},{"id":"o-aamc","type":"org","label":"AAMC"},{"id":"o-fields-letter","type":"org","label":"Fields medallists’ letter","description":"An open letter from Fields medallists on AI and mathematics, discussed by Gowers."},{"id":"v-nat-rev-cancer","type":"venue","label":"Nature Reviews Cancer"},{"id":"v-nat-rev-drug-disc","type":"venue","label":"Nature Reviews Drug Discovery"},{"id":"v-arxiv","type":"venue","label":"arXiv"},{"id":"v-proofs-prompts","type":"venue","label":"Proofs and Prompts","url":"https://proofsandprompts.com/"},{"id":"v-metaarxiv","type":"venue","label":"MetaArXiv (OSF)"},{"id":"sys-claude","type":"system","label":"Claude"},{"id":"sys-gemini","type":"system","label":"Gemini"},{"id":"sys-scientisttwo","type":"system","label":"ScientistTwo","url":"https://scientist-two.github.io/"},{"id":"c-verification","type":"concept","label":"Verification as the bottleneck","description":"As producing results gets cheaper, checking them and owning responsibility for them becomes the scarce resource."},{"id":"c-human-understanding","type":"concept","label":"Understanding vs. answers","description":"The value of science lies partly in humans understanding results, not only in results existing."},{"id":"c-research-taste","type":"concept","label":"Research taste","description":"Judgment about which questions and directions are worth pursuing."},{"id":"c-credit-provenance","type":"concept","label":"Credit and provenance","description":"Who did the work, where ideas and data came from, and who is accountable."},{"id":"c-scientific-institutions","type":"concept","label":"Scientific institutions","description":"Journals, prizes, departments and communities that sustain shared understanding."},{"id":"c-uncertainty-records","type":"concept","label":"Uncertainty records","description":"Machine-reusable qualifications attached to scientific claims."},{"id":"c-autonomous-research","type":"concept","label":"Autonomous research agents","description":"Pipelines where AI systems plan, run and write up research with little human input."},{"id":"c-multi-agent-coordination","type":"concept","label":"Multi-agent coordination","description":"How teams or swarms of agents divide labour, communicate and check each other."},{"id":"c-verifier-exploits","type":"concept","label":"Verifier exploits and reward hacking","description":"Agents gaming the checker instead of solving the problem; can spread through shared infrastructure."},{"id":"c-test-time-scaling","type":"concept","label":"Test-time scaling","description":"Spending more compute at inference: longer reasoning, parallel exploration, latent iteration."},{"id":"c-latent-reasoning","type":"concept","label":"Latent-space reasoning"},{"id":"c-formalization","type":"concept","label":"Formalization of proofs","description":"Turning human proofs into machine-checked proofs."},{"id":"c-causal-biology","type":"concept","label":"Measurement and causal biology","description":"Better data about human biology as the real constraint on AI drug discovery."},{"id":"c-cooperative-ai","type":"concept","label":"Cooperative AI"},{"id":"c-inspectable-reasoning","type":"concept","label":"Inspectable scientific reasoning","description":"Representations and interfaces that let people see and challenge how a conclusion was reached."},{"id":"c-human-oversight","type":"concept","label":"Human oversight in the lab"},{"id":"c-implementation-drift","type":"concept","label":"Implementation drift","description":"Agents silently changing the experiment they were asked to run."},{"id":"c-decision-value","type":"concept","label":"Decision-relevant evaluation","description":"Judging AI by whether it improves real decisions and outcomes, not benchmark scores."},{"id":"c-selection-effects","type":"concept","label":"Selection effects in reported wins","description":"Showcased successes may be cherry-picked from many attempts."},{"id":"m-simulated-peer-review","type":"method","label":"Simulated peer review"},{"id":"m-ablations","type":"method","label":"Automated experiments and ablations"},{"id":"m-order-controlled-pairs","type":"method","label":"Order-controlled pairwise choice"},{"id":"m-researcher-survey","type":"method","label":"Researcher survey and usage data"},{"id":"m-specialist-models","type":"method","label":"Specialist scientific models"},{"id":"m-learned-coordination","type":"method","label":"Learned coordination strategies","description":"Frozen models learning reusable ways to coordinate as a team."},{"id":"m-case-study","type":"method","label":"Retrospective case study"},{"id":"m-proof-assistant","type":"method","label":"Proof assistants (e.g. Lean)"},{"id":"a-ai-scientist-artefacts","type":"artifact","label":"ai-scientist-artefacts-v1","url":"https://github.com/Lossfunk/ai-scientist-artefacts-v1"},{"id":"a-research-taste-eval","type":"artifact","label":"research-taste-eval-v1","url":"https://github.com/Lossfunk/research-taste-eval-v1"},{"id":"o-sakana","type":"org","label":"Sakana AI","url":"https://sakana.ai/"},{"id":"sys-agentic-scientist","type":"system","label":"The Agentic Scientist","description":"Sakana's scaffold that orchestrates coding agents through hypothesis, rubric-based evaluation, shared memory and ablation stages.","url":"https://pub.sakana.ai/agentic-scientist/"},{"id":"sys-ai-scientist","type":"system","label":"The AI Scientist (v1, v2)","description":"Sakana's earlier end-to-end automated research systems, the predecessors of the Agentic Scientist."},{"id":"sys-coding-agents","type":"system","label":"Coding agents (Codex, Claude Code)","description":"Agents that carry out long software tasks on their own."},{"id":"a-ai-scientist-v3","type":"artifact","label":"AI-Scientist-v3","url":"https://github.com/SakanaAI/AI-Scientist-v3"},{"id":"a-autoillustrate","type":"artifact","label":"AutoIllustrate","description":"A pipeline for scientific figures that keeps an editable draw.io file as the source of truth, so every diagram can be corrected by hand."},{"id":"v-alife-2026","type":"venue","label":"ALIFE 2026"},{"id":"c-co-discovery","type":"concept","label":"Co-discovery","description":"Humans set the question and the standards of evidence, and agents widen the experiments researchers can run."},{"id":"c-optimizer-not-scientist","type":"concept","label":"Optimizer, not scientist","description":"Coding agents given a research problem tend to push scores up with narrow ideas, without examining why the gains happen."},{"id":"c-error-diffusion","type":"concept","label":"Error Diffusion","description":"A biologically motivated alternative to backpropagation."},{"id":"m-preregistered-rubrics","type":"method","label":"Pre-registered rubrics","description":"Evaluation criteria fixed before any experiment runs."},{"id":"m-shared-memory","type":"method","label":"Parallel workers with shared memory","description":"Workers test hypotheses in parallel and pool successes and failures."},{"id":"v-nature","type":"venue","label":"Nature"},{"id":"v-iclr-icbinb","type":"venue","label":"ICLR 2025 ICBINB workshop","description":"The “I Can’t Believe It’s Not Better” workshop, where an AI Scientist-v2 paper averaged 6.33 in blind review."},{"id":"o-ubc","type":"org","label":"University of British Columbia"},{"id":"o-vector","type":"org","label":"Vector Institute"},{"id":"o-oxford","type":"org","label":"University of Oxford"},{"id":"p-jeff-clune","type":"person","label":"Jeff Clune"},{"id":"p-jakob-foerster","type":"person","label":"Jakob Foerster"},{"id":"m-automated-reviewer","type":"method","label":"Automated Reviewer","description":"An ensemble of five LLM reviews following NeurIPS guidelines, reported at 69% balanced accuracy on OpenReview decisions."},{"id":"c-ai-paper-norms","type":"concept","label":"Norms for AI-generated papers","description":"Watermarking, disclosure, withdrawal and IRB review for research written by AI."},{"id":"a-ai-scientist-v2-code","type":"artifact","label":"AI-Scientist-v2","url":"https://github.com/SakanaAI/AI-Scientist-v2"},{"id":"o-google-deepmind","type":"org","label":"Google DeepMind"},{"id":"o-stanford","type":"org","label":"Stanford University"},{"id":"o-mit","type":"org","label":"MIT"},{"id":"o-google-cloud-ai","type":"org","label":"Google Cloud AI Research"},{"id":"v-icml-2026","type":"venue","label":"ICML 2026"},{"id":"v-colm-lm4sci","type":"venue","label":"COLM 2026 LM4SCI workshop"},{"id":"v-agents4science","type":"venue","label":"Agents4Science 2025","description":"A venue that required AI first authors."},{"id":"v-aamcnews","type":"venue","label":"AAMCNews"},{"id":"p-joel-leibo","type":"person","label":"Joel Z. Leibo"},{"id":"p-sasha-vezhnevets","type":"person","label":"Alexander Sasha Vezhnevets"},{"id":"p-logan-cross","type":"person","label":"Logan Cross"},{"id":"p-natasha-jaques","type":"person","label":"Natasha Jaques"},{"id":"p-james-zou","type":"person","label":"James Zou"},{"id":"p-marinka-zitnik","type":"person","label":"Marinka Zitnik"},{"id":"p-andreas-bender","type":"person","label":"Andreas Bender"},{"id":"p-christian-catalini","type":"person","label":"Christian Catalini"},{"id":"p-paras-chopra","type":"person","label":"Paras Chopra"},{"id":"p-dhruv-trehan","type":"person","label":"Dhruv Trehan"},{"id":"p-samuel-schmidgall","type":"person","label":"Samuel Schmidgall"},{"id":"p-tomas-pfister","type":"person","label":"Tomas Pfister"},{"id":"sys-agent-h","type":"system","label":"Agent_H","description":"A medical-QA agent architecture found by a fully autonomous search in the Gemini real-world study."},{"id":"sys-athena","type":"system","label":"ATHENA","description":"A treatment-reasoning agent from Zitnik's lab, wired to disease data, patient cohorts and FDA drug records."},{"id":"sys-co-scientist","type":"system","label":"Google Co-Scientist","description":"Started as an idea-generation system and was extended into an ideation, experimentation and paper-writing pipeline."},{"id":"c-goodhart","type":"concept","label":"Goodhart’s law","description":"When a measure becomes a target, it stops being a good measure."},{"id":"c-eureka-instinct","type":"concept","label":"Eureka instinct","description":"Agents declaring success too early and writing up results with too much excitement."},{"id":"c-proof-indigestion","type":"concept","label":"Proof indigestion","description":"Results arriving faster than people can absorb and understand them."},{"id":"c-knowledge-commons","type":"concept","label":"Knowledge-commons governance","description":"An Ostrom-style framing: shared libraries need sanctions and collective choice, not only visibility."},{"id":"c-institutions-as-primitives","type":"concept","label":"Institutions as design primitives","description":"Building roles, rules and enforcement into multi-agent systems rather than treating agents as lone optimizers."},{"id":"c-solipsistic-ai","type":"concept","label":"Solipsistic AI","description":"Assuming a fixed, stationary world in which the AI is the only optimizer."},{"id":"c-reviewer-fabrication","type":"concept","label":"Fabrication under LLM review","description":"Agents optimized against an LLM reviewer learn to invent results."},{"id":"c-fast-cheap-feedback","type":"concept","label":"Fast, cheap feedback","description":"AI wins come where scorecards are fast and cheap. Clinical outcomes are neither."},{"id":"c-proxy-validity","type":"concept","label":"Context of use and proxy validity","description":"Whether a measured proxy predicts the human outcome that matters."},{"id":"c-deskilling","type":"concept","label":"Deskilling","description":"Human skills atrophying when work is handed to AI."},{"id":"c-demonstrability","type":"concept","label":"Demonstrability","description":"How easily a correct answer can be shown to be correct."},{"id":"c-physical-bottleneck","type":"concept","label":"Physical experimentation bottleneck","description":"Wet labs, furnaces and characterization stay slow while analysis gets cheap."},{"id":"m-deterministic-claim-check","type":"method","label":"Deterministic claim-to-log checks","description":"Every number in a write-up is checked against execution logs."},{"id":"m-step-audit","type":"method","label":"Step audits and divergence reconciliation","description":"Teammates flag the specific step where they disagree, and the step is re-checked by agents that did not write it."},{"id":"m-closed-loop-lab","type":"method","label":"Closed-loop autonomous lab"},{"id":"m-economic-model","type":"method","label":"Formal economic model"},{"id":"o-discovery-loop","type":"org","label":"Discovery Loop","description":"A Palo Alto public benefit corporation launched on 5 Aug 2026 to automate ML, science and engineering experiments.","url":"https://www.discoveryloop.com/"},{"id":"p-jeff-dean","type":"person","label":"Jeff Dean"},{"id":"p-sanjay-ghemawat","type":"person","label":"Sanjay Ghemawat"},{"id":"p-quoc-le","type":"person","label":"Quoc V. Le"},{"id":"p-oriol-vinyals","type":"person","label":"Oriol Vinyals"},{"id":"c-loop-latency","type":"concept","label":"Experimental loop latency","description":"Cutting the time per propose, implement, run and evaluate cycle to minutes or hours, and running thousands of cycles in parallel."},{"id":"c-recursive-self-improvement","type":"concept","label":"Recursive self-improvement","description":"Using AI research tools to improve the stack that builds them."},{"id":"c-nae-grand-challenges","type":"concept","label":"NAE Grand Challenges","description":"The US National Academy of Engineering’s 14 grand challenges, such as medicines, clean water and solar."},{"id":"p-niko-grupen","type":"person","label":"Niko Grupen"},{"id":"o-harvey","type":"org","label":"Harvey","url":"https://www.harvey.ai/"},{"id":"c-context-graphs","type":"concept","label":"Context graphs","description":"Prebuilt graphs of the entities and relations in a document collection, handed to agents as structured context instead of raw text."},{"id":"p-ben-davis","type":"person","label":"Ben Davis"},{"id":"c-verifiable-hill-climbing","type":"concept","label":"Verifiable hill-climbing","description":"Change, test, repeat until a programmatic goal is met. Models already excel at this."},{"id":"c-regression-to-mean","type":"concept","label":"Regression to the mean in open-ended choices","description":"Without a verifier, model choices drift toward the familiar and average, so they are functional but generic."},{"id":"p-miguel-rios","type":"person","label":"Miguel Rios"},{"id":"o-grep-ai","type":"org","label":"Grep.ai"},{"id":"sys-jev","type":"system","label":"Jev (TypeSafe)","description":"A small “System One” model that answers typed questions with calibrated probabilities in about 150 ms."},{"id":"c-decision-action-split","type":"concept","label":"Decision/action split","description":"A model classifies what something is; deterministic code decides what happens next. From 12-Factor Agents."},{"id":"m-agent-writes-workflow","type":"method","label":"Agent writes its own workflow","description":"An agent does a task from an SOP, then compiles what it learned into a typed workflow that handles most cases without it."},{"id":"m-replayable-judgment","type":"method","label":"Replayable, per-node evaluation","description":"Re-run each judgment without redoing the research, and score every workflow step on its own."},{"id":"o-convoke","type":"org","label":"Convoke"},{"id":"c-drug-discovery","type":"concept","label":"Drug discovery and biopharma","description":"Using AI across the drug pipeline, from target biology and R&D to clinical, regulatory and commercial work."},{"id":"m-penalty-rubric-grading","type":"method","label":"Penalty-based rubric grading","description":"Experts list the errors a work product must avoid, weighted major, moderate or minor; an LLM grader applies the checklist with citations."},{"id":"c-synthesis-vs-compilation","type":"concept","label":"Synthesis vs compilation gap","description":"Models compile structured trackers well but do worse at integrating evidence into narrative judgments."},{"id":"p-dawn-song","type":"person","label":"Dawn Song"},{"id":"m-ev-cost-experiment-portfolio","type":"method","label":"Experiment portfolio by expected value per cost","description":"Score each queued experiment by expected value divided by its compute and other costs, and run the best first."},{"id":"c-ten-x-signal","type":"concept","label":"The 10x signal","description":"A 10x gain in compute-to-quality marks an idea as important, as with mixture of experts."},{"id":"c-five-of-seven","type":"concept","label":"The 5-of-7 problem shape","description":"Good research problems have most parts starting to take shape and a few hard but imaginable, on about a 5-year horizon."},{"id":"c-shared-discovery-loop","type":"concept","label":"The shared science–engineering loop","description":"Decompose, propose, implement, evaluate, feed back, repeat."},{"id":"c-institutional-alignment","type":"concept","label":"Reliability as institutional design","description":"When AI systems become many communicating agents, reliability may depend as much on the rules of the collective as on aligning each model."},{"id":"c-governed-claim-ledger","type":"concept","label":"Governed claim ledger","description":"Shared claims are reviewed, promoted, challenged and corrected under explicit rules. Detecting a problem is kept separate from the authority to change the record, and repairs are scoped: narrow, downgrade, reassess or revoke."},{"id":"c-typed-knowledge-objects","type":"concept","label":"Typed scientific knowledge objects","description":"Shared memory as claims, hypotheses, evidence, contradictions, negative results, replications, retractions and uncertainty, like PubMed plus Git plus peer review, not a chat log."},{"id":"c-selective-contagion","type":"concept","label":"Selective contagion of ideas","description":"Treat claims like infections: keep R below 1 for bad claims and well above 1 for validated ones. Heuristic: detection, adjudication and intervention must be faster than a harmful claim spreads."},{"id":"c-evidence-chain","type":"concept","label":"Separated evidence chain","description":"Prediction quality, specification fidelity, decision utility and clinical translation are kept apart. In biology, association, intervention, mechanism and translation are separate stages, and evidence for one does not validate the next."},{"id":"c-integrity-channels","type":"concept","label":"Integrity reporting channels","description":"Separate channels for negative results, uncertainty reports, confessions, pre-action alerts and peer challenges, each with its own response, so that reporting disappointing science is not punished."},{"id":"c-distinct-bets","type":"concept","label":"Protected distinct bets","description":"Give cohorts of agents protected budgets to pursue different hypotheses. Compare diversity on complete proposals against repeated sampling from one strong model at the same budget."},{"id":"o-benchling","type":"org","label":"Benchling"},{"id":"p-nicholas-larus-stone","type":"person","label":"Nicholas Larus-Stone"},{"id":"o-parabilis","type":"org","label":"Parabilis Medicines","url":"https://parabilismed.com/"},{"id":"sys-science-foundries","type":"system","label":"Benchling Science Foundries","description":"Agent teams that run a lab process end to end on shared data, with scientist approval steps chosen by the team."},{"id":"m-approval-gates","type":"method","label":"Scientist approval gates","description":"Assign each workflow step to an agent or a person, and require a scientist's sign-off at chosen decision points."},{"id":"p-shangyin-tan","type":"person","label":"Shangyin Tan"},{"id":"p-alex-dimakis","type":"person","label":"Alex Dimakis"},{"id":"p-matei-zaharia","type":"person","label":"Matei Zaharia"},{"id":"p-dimitris-papailiopoulos","type":"person","label":"Dimitris Papailiopoulos"},{"id":"p-ziyang-cai","type":"person","label":"Ziyang Cai"},{"id":"o-uc-berkeley","type":"org","label":"UC Berkeley"},{"id":"c-idea-basins","type":"concept","label":"Idea basins","description":"Long-running agents stay in a local region of solution space and refine their own earlier ideas, so independent runs end far apart."},{"id":"c-history-anchoring","type":"concept","label":"Anchoring on history","description":"As an agent's context grows, its proposals become smaller and more conservative."},{"id":"m-fork-and-flush","type":"method","label":"Fork-and-flush","description":"Periodically fork into fresh-context trajectories that keep the workspace, then continue from the best."},{"id":"m-forgetting","type":"method","label":"Forgetting (keep only the best)","description":"Periodically discard all sessions and search state, keeping only the best candidate, and restart fresh agents from it."},{"id":"sys-autoresearch","type":"system","label":"Autoresearch agents","description":"Agents that loop propose, score and refine on a measurable research or engineering task, as in Karpathy's autoresearch.","url":"https://github.com/karpathy/autoresearch"},{"id":"o-google-research","type":"org","label":"Google Research"},{"id":"p-pooja-rao","type":"person","label":"Pooja Rao"},{"id":"p-petar-sirkovic","type":"person","label":"Petar Sirkovic"},{"id":"p-annalisa-pawlosky","type":"person","label":"Annalisa Pawlosky"},{"id":"sys-verid","type":"system","label":"VERID","description":"A semantic digital twin of a lab workcell that dry-runs and safety-checks robot protocols."},{"id":"a-pylabrobot","type":"artifact","label":"PyLabRobot","description":"Open-source framework for controlling lab robots.","url":"https://docs.pylabrobot.org/"},{"id":"m-digital-twin-dry-run","type":"method","label":"Digital-twin dry runs","description":"Simulate every protocol against a live model of the lab before any physical action."},{"id":"m-aptamer-loop","type":"method","label":"Aptamer lab-in-the-loop","description":"Pooled synthesis and sequencing readout give a 7–10 day design–test–analyze cycle."},{"id":"m-epitope-excision","type":"method","label":"Epitope excision","description":"Re-fold only the local binding pocket of a large target to cut structure-prediction cost."},{"id":"m-biosecurity-screening","type":"method","label":"Biosecurity screening of designs","description":"Every designed sequence is screened before synthesis."},{"id":"c-strategy-tactics-split","type":"concept","label":"Human strategy, agent tactics","description":"The scientist chooses where to aim; the agent executes and iterates within that choice."},{"id":"p-grant-sanderson","type":"person","label":"Grant Sanderson","description":"Creator of 3Blue1Brown, a YouTube channel of animated mathematics.","url":"https://www.3blue1brown.com/"},{"id":"c-proof-as-data","type":"concept","label":"Proof as raw data","description":"Once machines generate proofs cheaply, a proof is treated like a numerical result: input for theory building, not the goal."},{"id":"c-math-teaching","type":"concept","label":"Teaching as math’s funding base","description":"Most money for mathematics comes from teaching, which sells problem-solving skill and the joy of the subject."}],"edges":[{"s":"scientisttwo","t":"t-agents","r":"tagged"},{"s":"scientisttwo","t":"t-evaluation","r":"tagged"},{"s":"scientisttwo","t":"t-scientific-discovery","r":"tagged"},{"s":"self-organizing-teams","t":"t-agents","r":"tagged"},{"s":"self-organizing-teams","t":"t-collaboration","r":"tagged"},{"s":"self-organizing-teams","t":"t-evaluation","r":"tagged"},{"s":"solipsistic-superintelligence","t":"t-governance","r":"tagged"},{"s":"solipsistic-superintelligence","t":"t-human-agency","r":"tagged"},{"s":"solipsistic-superintelligence","t":"t-collaboration","r":"tagged"},{"s":"economics-agi","t":"t-human-agency","r":"tagged"},{"s":"economics-agi","t":"t-verification","r":"tagged"},{"s":"economics-agi","t":"t-economics","r":"tagged"},{"s":"drug-discovery-path","t":"t-biology","r":"tagged"},{"s":"drug-discovery-path","t":"t-medicine","r":"tagged"},{"s":"drug-discovery-path","t":"t-evaluation","r":"tagged"},{"s":"cancer-25-years","t":"t-biology","r":"tagged"},{"s":"cancer-25-years","t":"t-medicine","r":"tagged"},{"s":"cancer-25-years","t":"t-human-ai-collaboration","r":"tagged"},{"s":"uncertainty-layer","t":"t-evidence","r":"tagged"},{"s":"uncertainty-layer","t":"t-verification","r":"tagged"},{"s":"uncertainty-layer","t":"t-scientific-communication","r":"tagged"},{"s":"four-autonomous-attempts","t":"t-agents","r":"tagged"},{"s":"four-autonomous-attempts","t":"t-evaluation","r":"tagged"},{"s":"four-autonomous-attempts","t":"t-research-taste","r":"tagged"},{"s":"cheating-whistleblowing","t":"t-agents","r":"tagged"},{"s":"cheating-whistleblowing","t":"t-governance","r":"tagged"},{"s":"cheating-whistleblowing","t":"t-mathematics","r":"tagged"},{"s":"research-taste","t":"t-research-taste","r":"tagged"},{"s":"research-taste","t":"t-evaluation","r":"tagged"},{"s":"research-taste","t":"t-scientific-discovery","r":"tagged"},{"s":"no-magic-wands","t":"t-biology","r":"tagged"},{"s":"no-magic-wands","t":"t-medicine","r":"tagged"},{"s":"no-magic-wands","t":"t-scientific-discovery","r":"tagged"},{"s":"gemini-real-world","t":"t-agents","r":"tagged"},{"s":"gemini-real-world","t":"t-biology","r":"tagged"},{"s":"gemini-real-world","t":"t-materials","r":"tagged"},{"s":"claude-fermat-report","t":"t-mathematics","r":"tagged"},{"s":"claude-fermat-report","t":"t-formal-verification","r":"tagged"},{"s":"claude-fermat-report","t":"t-agents","r":"tagged"},{"s":"medical-lab-partner","t":"t-medicine","r":"tagged"},{"s":"medical-lab-partner","t":"t-biology","r":"tagged"},{"s":"medical-lab-partner","t":"t-human-ai-collaboration","r":"tagged"},{"s":"chollet-third-axis","t":"t-reasoning","r":"tagged"},{"s":"chollet-third-axis","t":"t-compute","r":"tagged"},{"s":"chollet-third-axis","t":"t-agents","r":"tagged"},{"s":"tao-ai-science","t":"t-mathematics","r":"tagged"},{"s":"tao-ai-science","t":"t-human-agency","r":"tagged"},{"s":"tao-ai-science","t":"t-scientific-discovery","r":"tagged"},{"s":"jun-future-science","t":"t-human-ai-collaboration","r":"tagged"},{"s":"jun-future-science","t":"t-scientific-communication","r":"tagged"},{"s":"jun-future-science","t":"t-evidence","r":"tagged"},{"s":"periodic-infrastructure","t":"t-infrastructure","r":"tagged"},{"s":"periodic-infrastructure","t":"t-scientific-discovery","r":"tagged"},{"s":"periodic-infrastructure","t":"t-agents","r":"tagged"},{"s":"ai-in-science","t":"t-human-ai-collaboration","r":"tagged"},{"s":"ai-in-science","t":"t-evaluation","r":"tagged"},{"s":"ai-in-science","t":"t-scientific-discovery","r":"tagged"},{"s":"litt-beginning-mathematics","t":"t-mathematics","r":"tagged"},{"s":"litt-beginning-mathematics","t":"t-human-agency","r":"tagged"},{"s":"litt-beginning-mathematics","t":"t-human-ai-collaboration","r":"tagged"},{"s":"litt-beginning-mathematics","t":"t-research-taste","r":"tagged"},{"s":"doomslide-mathematics","t":"t-mathematics","r":"tagged"},{"s":"doomslide-mathematics","t":"t-governance","r":"tagged"},{"s":"doomslide-mathematics","t":"t-scientific-communication","r":"tagged"},{"s":"gowers-fields-letter","t":"t-mathematics","r":"tagged"},{"s":"gowers-fields-letter","t":"t-human-agency","r":"tagged"},{"s":"gowers-fields-letter","t":"t-governance","r":"tagged"},{"s":"hairer-agmai","t":"t-mathematics","r":"tagged"},{"s":"hairer-agmai","t":"t-governance","r":"tagged"},{"s":"hairer-agmai","t":"t-scientific-communication","r":"tagged"},{"s":"agentic-scientist","t":"t-agents","r":"tagged"},{"s":"agentic-scientist","t":"t-scientific-discovery","r":"tagged"},{"s":"agentic-scientist","t":"t-human-ai-collaboration","r":"tagged"},{"s":"agentic-scientist","t":"t-evaluation","r":"tagged"},{"s":"ai-scientist-nature","t":"t-agents","r":"tagged"},{"s":"ai-scientist-nature","t":"t-scientific-discovery","r":"tagged"},{"s":"ai-scientist-nature","t":"t-evaluation","r":"tagged"},{"s":"ai-scientist-nature","t":"t-governance","r":"tagged"},{"s":"discovery-loop-launch","t":"t-agents","r":"tagged"},{"s":"discovery-loop-launch","t":"t-scientific-discovery","r":"tagged"},{"s":"discovery-loop-launch","t":"t-infrastructure","r":"tagged"},{"s":"discovery-loop-launch","t":"t-compute","r":"tagged"},{"s":"grupen-context-graphs","t":"t-agents","r":"tagged"},{"s":"grupen-context-graphs","t":"t-evaluation","r":"tagged"},{"s":"grupen-context-graphs","t":"t-infrastructure","r":"tagged"},{"s":"davis-labs-taste","t":"t-research-taste","r":"tagged"},{"s":"davis-labs-taste","t":"t-verification","r":"tagged"},{"s":"davis-labs-taste","t":"t-agents","r":"tagged"},{"s":"davis-labs-taste","t":"t-human-agency","r":"tagged"},{"s":"agentrun-grep","t":"t-agents","r":"tagged"},{"s":"agentrun-grep","t":"t-verification","r":"tagged"},{"s":"agentrun-grep","t":"t-evaluation","r":"tagged"},{"s":"agentrun-grep","t":"t-research-taste","r":"tagged"},{"s":"agentrun-grep","t":"t-infrastructure","r":"tagged"},{"s":"convoke-biopharma-benchmark","t":"t-biopharma","r":"tagged"},{"s":"convoke-biopharma-benchmark","t":"t-evaluation","r":"tagged"},{"s":"convoke-biopharma-benchmark","t":"t-agents","r":"tagged"},{"s":"convoke-biopharma-benchmark","t":"t-medicine","r":"tagged"},{"s":"convoke-biopharma-benchmark","t":"t-research-taste","r":"tagged"},{"s":"dean-dawnsong-fireside","t":"t-scientific-discovery","r":"tagged"},{"s":"dean-dawnsong-fireside","t":"t-agents","r":"tagged"},{"s":"dean-dawnsong-fireside","t":"t-research-taste","r":"tagged"},{"s":"dean-dawnsong-fireside","t":"t-governance","r":"tagged"},{"s":"benchling-science-foundries","t":"t-biopharma","r":"tagged"},{"s":"benchling-science-foundries","t":"t-agents","r":"tagged"},{"s":"benchling-science-foundries","t":"t-human-ai-collaboration","r":"tagged"},{"s":"benchling-science-foundries","t":"t-biology","r":"tagged"},{"s":"benchling-science-foundries","t":"t-infrastructure","r":"tagged"},{"s":"fork-and-flush","t":"t-agents","r":"tagged"},{"s":"fork-and-flush","t":"t-scientific-discovery","r":"tagged"},{"s":"fork-and-flush","t":"t-research-taste","r":"tagged"},{"s":"scaling-discovery-forgetting","t":"t-agents","r":"tagged"},{"s":"scaling-discovery-forgetting","t":"t-scientific-discovery","r":"tagged"},{"s":"scaling-discovery-forgetting","t":"t-research-taste","r":"tagged"},{"s":"larus-stone-foundries-post","t":"t-biopharma","r":"tagged"},{"s":"larus-stone-foundries-post","t":"t-agents","r":"tagged"},{"s":"larus-stone-foundries-post","t":"t-human-ai-collaboration","r":"tagged"},{"s":"google-wet-lab-agents","t":"t-biology","r":"tagged"},{"s":"google-wet-lab-agents","t":"t-agents","r":"tagged"},{"s":"google-wet-lab-agents","t":"t-biopharma","r":"tagged"},{"s":"google-wet-lab-agents","t":"t-human-ai-collaboration","r":"tagged"},{"s":"google-wet-lab-agents","t":"t-verification","r":"tagged"},{"s":"verid-semantic-lab","t":"t-biology","r":"tagged"},{"s":"verid-semantic-lab","t":"t-agents","r":"tagged"},{"s":"verid-semantic-lab","t":"t-verification","r":"tagged"},{"s":"verid-semantic-lab","t":"t-infrastructure","r":"tagged"},{"s":"aptamer-sandbox","t":"t-biology","r":"tagged"},{"s":"aptamer-sandbox","t":"t-agents","r":"tagged"},{"s":"aptamer-sandbox","t":"t-biopharma","r":"tagged"},{"s":"aptamer-sandbox","t":"t-verification","r":"tagged"},{"s":"epitope-excision","t":"t-biology","r":"tagged"},{"s":"epitope-excision","t":"t-agents","r":"tagged"},{"s":"epitope-excision","t":"t-biopharma","r":"tagged"},{"s":"epitope-excision","t":"t-human-ai-collaboration","r":"tagged"},{"s":"3b1b-better-time-for-math","t":"t-mathematics","r":"tagged"},{"s":"3b1b-better-time-for-math","t":"t-human-agency","r":"tagged"},{"s":"3b1b-better-time-for-math","t":"t-human-ai-collaboration","r":"tagged"},{"s":"3b1b-better-time-for-math","t":"t-scientific-communication","r":"tagged"},{"s":"3b1b-better-time-for-math","t":"t-research-taste","r":"tagged"},{"s":"litt-beginning-mathematics","t":"p-daniel-litt","r":"authored_by"},{"s":"litt-beginning-mathematics","t":"v-proofs-prompts","r":"published_in","note":"republished"},{"s":"litt-beginning-mathematics","t":"c-human-understanding","r":"proposes"},{"s":"litt-beginning-mathematics","t":"c-research-taste","r":"proposes","note":"reward research programmes, not just results"},{"s":"litt-beginning-mathematics","t":"c-scientific-institutions","r":"discusses"},{"s":"doomslide-mathematics","t":"p-doomslide","r":"authored_by"},{"s":"doomslide-mathematics","t":"litt-beginning-mathematics","r":"responds_to"},{"s":"doomslide-mathematics","t":"c-credit-provenance","r":"critiques","note":"opaque proof workflows and private-data reuse"},{"s":"gowers-fields-letter","t":"p-timothy-gowers","r":"authored_by"},{"s":"gowers-fields-letter","t":"o-fields-letter","r":"responds_to"},{"s":"gowers-fields-letter","t":"c-scientific-institutions","r":"proposes"},{"s":"gowers-fields-letter","t":"c-human-understanding","r":"discusses"},{"s":"hairer-agmai","t":"p-martin-hairer","r":"authored_by"},{"s":"hairer-agmai","t":"p-timothy-gowers","r":"commented_by"},{"s":"hairer-agmai","t":"v-proofs-prompts","r":"published_in"},{"s":"p-martin-hairer","t":"o-agmai","r":"affiliated_with"},{"s":"hairer-agmai","t":"o-agmai","r":"discusses"},{"s":"hairer-agmai","t":"c-scientific-institutions","r":"discusses"},{"s":"hairer-agmai","t":"gowers-fields-letter","r":"related_to","note":"same debate"},{"s":"tao-ai-science","t":"p-terence-tao","r":"authored_by"},{"s":"tao-ai-science","t":"c-human-understanding","r":"discusses"},{"s":"tao-ai-science","t":"litt-beginning-mathematics","r":"related_to","note":"answers vs. understanding"},{"s":"self-organizing-teams","t":"p-aneesh-pappu","r":"commented_by"},{"s":"self-organizing-teams","t":"v-arxiv","r":"published_in"},{"s":"self-organizing-teams","t":"m-learned-coordination","r":"uses_method"},{"s":"self-organizing-teams","t":"c-multi-agent-coordination","r":"discusses"},{"s":"self-organizing-teams","t":"c-verification","r":"discusses","note":"gains depend on recognising correct reasoning"},{"s":"scientisttwo","t":"v-arxiv","r":"published_in"},{"s":"scientisttwo","t":"sys-scientisttwo","r":"about_system"},{"s":"scientisttwo","t":"c-autonomous-research","r":"discusses"},{"s":"scientisttwo","t":"m-ablations","r":"uses_method"},{"s":"scientisttwo","t":"m-simulated-peer-review","r":"uses_method"},{"s":"scientisttwo","t":"c-selection-effects","r":"discusses"},{"s":"periodic-infrastructure","t":"p-liam-fedus","r":"commented_by"},{"s":"periodic-infrastructure","t":"o-periodic","r":"made_by"},{"s":"p-liam-fedus","t":"o-periodic","r":"affiliated_with","note":"co-founder"},{"s":"periodic-infrastructure","t":"m-specialist-models","r":"uses_method"},{"s":"periodic-infrastructure","t":"c-autonomous-research","r":"discusses","note":"long scientific tool runs"},{"s":"ai-in-science","t":"p-arthur-turrell","r":"commented_by"},{"s":"ai-in-science","t":"o-google","r":"made_by"},{"s":"ai-in-science","t":"m-researcher-survey","r":"uses_method"},{"s":"ai-in-science","t":"m-specialist-models","r":"uses_method","note":"inventory of specialist models"},{"s":"ai-in-science","t":"c-verification","r":"discusses","note":"verification demands"},{"s":"cancer-25-years","t":"v-nat-rev-cancer","r":"published_in"},{"s":"cancer-25-years","t":"c-causal-biology","r":"discusses"},{"s":"medical-lab-partner","t":"o-aamc","r":"made_by"},{"s":"medical-lab-partner","t":"c-human-oversight","r":"discusses"},{"s":"medical-lab-partner","t":"c-multi-agent-coordination","r":"discusses","note":"specialised lab agents"},{"s":"chollet-third-axis","t":"p-francois-chollet","r":"authored_by"},{"s":"chollet-third-axis","t":"c-latent-reasoning","r":"proposes"},{"s":"c-latent-reasoning","t":"c-test-time-scaling","r":"related_to","note":"a third axis"},{"s":"claude-fermat-report","t":"sys-claude","r":"about_system"},{"s":"sys-claude","t":"o-anthropic","r":"made_by"},{"s":"claude-fermat-report","t":"c-formalization","r":"discusses"},{"s":"claude-fermat-report","t":"m-proof-assistant","r":"uses_method"},{"s":"jun-future-science","t":"p-eunice-jun","r":"authored_by"},{"s":"jun-future-science","t":"c-inspectable-reasoning","r":"proposes"},{"s":"cheating-whistleblowing","t":"v-arxiv","r":"published_in"},{"s":"cheating-whistleblowing","t":"m-case-study","r":"uses_method"},{"s":"cheating-whistleblowing","t":"c-verifier-exploits","r":"critiques"},{"s":"cheating-whistleblowing","t":"c-multi-agent-coordination","r":"discusses","note":"research swarms"},{"s":"cheating-whistleblowing","t":"c-autonomous-research","r":"discusses"},{"s":"gemini-real-world","t":"v-arxiv","r":"published_in"},{"s":"gemini-real-world","t":"sys-gemini","r":"about_system"},{"s":"sys-gemini","t":"o-google","r":"made_by"},{"s":"gemini-real-world","t":"c-multi-agent-coordination","r":"discusses"},{"s":"gemini-real-world","t":"c-autonomous-research","r":"discusses","note":"different levels of autonomy"},{"s":"uncertainty-layer","t":"p-richard-sever","r":"commented_by"},{"s":"uncertainty-layer","t":"v-metaarxiv","r":"published_in"},{"s":"uncertainty-layer","t":"c-uncertainty-records","r":"proposes"},{"s":"c-uncertainty-records","t":"c-verification","r":"related_to"},{"s":"drug-discovery-path","t":"v-nat-rev-drug-disc","r":"published_in"},{"s":"drug-discovery-path","t":"c-decision-value","r":"proposes"},{"s":"no-magic-wands","t":"p-daphne-koller","r":"authored_by"},{"s":"no-magic-wands","t":"c-causal-biology","r":"proposes"},{"s":"no-magic-wands","t":"drug-discovery-path","r":"related_to","note":"realistic expectations for AI drug discovery"},{"s":"solipsistic-superintelligence","t":"v-arxiv","r":"published_in"},{"s":"solipsistic-superintelligence","t":"c-cooperative-ai","r":"proposes"},{"s":"solipsistic-superintelligence","t":"c-scientific-institutions","r":"discusses","note":"institutions and adaptive counterparts"},{"s":"economics-agi","t":"v-arxiv","r":"published_in"},{"s":"economics-agi","t":"c-verification","r":"proposes"},{"s":"four-autonomous-attempts","t":"v-arxiv","r":"published_in"},{"s":"four-autonomous-attempts","t":"o-lossfunk","r":"made_by"},{"s":"four-autonomous-attempts","t":"a-ai-scientist-artefacts","r":"releases"},{"s":"four-autonomous-attempts","t":"c-autonomous-research","r":"critiques"},{"s":"four-autonomous-attempts","t":"c-implementation-drift","r":"critiques"},{"s":"four-autonomous-attempts","t":"c-research-taste","r":"discusses","note":"failures of scientific judgment"},{"s":"research-taste","t":"o-lossfunk","r":"made_by"},{"s":"research-taste","t":"a-research-taste-eval","r":"releases"},{"s":"research-taste","t":"m-order-controlled-pairs","r":"uses_method"},{"s":"research-taste","t":"c-research-taste","r":"discusses"},{"s":"agentic-scientist","t":"o-sakana","r":"made_by"},{"s":"agentic-scientist","t":"o-sakana","r":"commented_by"},{"s":"agentic-scientist","t":"v-arxiv","r":"published_in"},{"s":"agentic-scientist","t":"sys-agentic-scientist","r":"about_system"},{"s":"sys-agentic-scientist","t":"o-sakana","r":"made_by"},{"s":"sys-ai-scientist","t":"o-sakana","r":"made_by"},{"s":"sys-agentic-scientist","t":"sys-ai-scientist","r":"builds_on"},{"s":"sys-agentic-scientist","t":"sys-coding-agents","r":"builds_on","note":"wraps off-the-shelf coding agents"},{"s":"agentic-scientist","t":"a-ai-scientist-v3","r":"releases"},{"s":"agentic-scientist","t":"a-autoillustrate","r":"releases"},{"s":"agentic-scientist","t":"c-optimizer-not-scientist","r":"critiques"},{"s":"agentic-scientist","t":"c-co-discovery","r":"proposes"},{"s":"agentic-scientist","t":"m-preregistered-rubrics","r":"uses_method"},{"s":"agentic-scientist","t":"m-shared-memory","r":"uses_method"},{"s":"agentic-scientist","t":"m-ablations","r":"uses_method","note":"ablates the winning program to generate evidence"},{"s":"agentic-scientist","t":"c-error-diffusion","r":"discusses","note":"case study accepted as an ALIFE 2026 oral"},{"s":"c-error-diffusion","t":"v-alife-2026","r":"published_in"},{"s":"agentic-scientist","t":"c-autonomous-research","r":"discusses"},{"s":"agentic-scientist","t":"c-human-oversight","r":"discusses","note":"researchers verified results, references and claims"},{"s":"agentic-scientist","t":"four-autonomous-attempts","r":"related_to","note":"failure modes of autonomous ML research"},{"s":"agentic-scientist","t":"scientisttwo","r":"related_to","note":"autonomous research pipelines"},{"s":"c-co-discovery","t":"c-verification","r":"related_to"},{"s":"ai-scientist-nature","t":"o-sakana","r":"made_by"},{"s":"ai-scientist-nature","t":"v-nature","r":"published_in"},{"s":"ai-scientist-nature","t":"sys-ai-scientist","r":"about_system"},{"s":"ai-scientist-nature","t":"a-ai-scientist-v2-code","r":"releases"},{"s":"ai-scientist-nature","t":"v-iclr-icbinb","r":"discusses","note":"a fully AI-written paper passed review, then was withdrawn"},{"s":"ai-scientist-nature","t":"m-automated-reviewer","r":"uses_method"},{"s":"m-automated-reviewer","t":"m-simulated-peer-review","r":"related_to"},{"s":"ai-scientist-nature","t":"c-ai-paper-norms","r":"proposes"},{"s":"ai-scientist-nature","t":"c-autonomous-research","r":"discusses"},{"s":"ai-scientist-nature","t":"c-selection-effects","r":"critiques","note":"naive ideas and hallucinated citations remain"},{"s":"ai-scientist-nature","t":"agentic-scientist","r":"related_to","note":"the Agentic Scientist is the next step in this line"},{"s":"ai-scientist-nature","t":"four-autonomous-attempts","r":"related_to","note":"also published AI-first-author work at a workshop"},{"s":"p-jeff-clune","t":"o-ubc","r":"affiliated_with"},{"s":"p-jeff-clune","t":"o-vector","r":"affiliated_with"},{"s":"p-jakob-foerster","t":"o-oxford","r":"affiliated_with"},{"s":"ai-scientist-nature","t":"p-jeff-clune","r":"authored_by","note":"named collaborator"},{"s":"ai-scientist-nature","t":"p-jakob-foerster","r":"authored_by","note":"named collaborator"},{"s":"c-ai-paper-norms","t":"c-credit-provenance","r":"related_to"},{"s":"self-organizing-teams","t":"p-aneesh-pappu","r":"authored_by"},{"s":"self-organizing-teams","t":"p-james-zou","r":"authored_by"},{"s":"self-organizing-teams","t":"o-stanford","r":"made_by"},{"s":"self-organizing-teams","t":"m-step-audit","r":"uses_method"},{"s":"self-organizing-teams","t":"c-demonstrability","r":"discusses"},{"s":"c-demonstrability","t":"c-verification","r":"related_to"},{"s":"self-organizing-teams","t":"cheating-whistleblowing","r":"related_to","note":"the bright twin: self-organisation that helps rather than spreads an exploit"},{"s":"scientisttwo","t":"o-google-cloud-ai","r":"made_by"},{"s":"scientisttwo","t":"p-tomas-pfister","r":"authored_by"},{"s":"scientisttwo","t":"four-autonomous-attempts","r":"related_to","note":"an engineering answer to its failure modes"},{"s":"scientisttwo","t":"gemini-real-world","r":"related_to","note":"both use simulated peer review"},{"s":"gemini-real-world","t":"p-samuel-schmidgall","r":"authored_by"},{"s":"gemini-real-world","t":"o-google-deepmind","r":"made_by"},{"s":"gemini-real-world","t":"sys-co-scientist","r":"about_system"},{"s":"sys-co-scientist","t":"o-google","r":"made_by"},{"s":"gemini-real-world","t":"sys-agent-h","r":"releases"},{"s":"gemini-real-world","t":"c-reviewer-fabrication","r":"critiques","note":"severe result hallucination cut from 46% to 4%"},{"s":"gemini-real-world","t":"m-deterministic-claim-check","r":"uses_method"},{"s":"gemini-real-world","t":"c-physical-bottleneck","r":"discusses","note":"humans still ran the reactor"},{"s":"cheating-whistleblowing","t":"o-google-deepmind","r":"made_by"},{"s":"cheating-whistleblowing","t":"p-joel-leibo","r":"authored_by"},{"s":"cheating-whistleblowing","t":"p-sasha-vezhnevets","r":"authored_by"},{"s":"cheating-whistleblowing","t":"p-logan-cross","r":"authored_by"},{"s":"cheating-whistleblowing","t":"c-knowledge-commons","r":"discusses"},{"s":"cheating-whistleblowing","t":"c-goodhart","r":"discusses"},{"s":"cheating-whistleblowing","t":"m-proof-assistant","r":"uses_method","note":"100 agents on 71 Lean 4 conjectures"},{"s":"cheating-whistleblowing","t":"solipsistic-superintelligence","r":"related_to","note":"empirical illustration by overlapping authors"},{"s":"solipsistic-superintelligence","t":"p-joel-leibo","r":"authored_by"},{"s":"solipsistic-superintelligence","t":"p-sasha-vezhnevets","r":"authored_by"},{"s":"solipsistic-superintelligence","t":"p-logan-cross","r":"authored_by"},{"s":"solipsistic-superintelligence","t":"p-natasha-jaques","r":"authored_by"},{"s":"solipsistic-superintelligence","t":"v-icml-2026","r":"published_in"},{"s":"solipsistic-superintelligence","t":"c-solipsistic-ai","r":"critiques"},{"s":"solipsistic-superintelligence","t":"c-institutions-as-primitives","r":"proposes"},{"s":"solipsistic-superintelligence","t":"c-deskilling","r":"discusses"},{"s":"c-institutions-as-primitives","t":"c-knowledge-commons","r":"related_to"},{"s":"four-autonomous-attempts","t":"p-dhruv-trehan","r":"authored_by"},{"s":"four-autonomous-attempts","t":"p-paras-chopra","r":"authored_by"},{"s":"four-autonomous-attempts","t":"c-eureka-instinct","r":"critiques"},{"s":"four-autonomous-attempts","t":"v-agents4science","r":"discusses","note":"one negative result was accepted there"},{"s":"four-autonomous-attempts","t":"sys-claude","r":"about_system","note":"Claude Code did the implementation"},{"s":"research-taste","t":"p-dhruv-trehan","r":"authored_by"},{"s":"research-taste","t":"p-paras-chopra","r":"authored_by"},{"s":"research-taste","t":"v-colm-lm4sci","r":"published_in"},{"s":"research-taste","t":"four-autonomous-attempts","r":"responds_to","note":"measures its failure mode 6, lack of taste"},{"s":"economics-agi","t":"p-christian-catalini","r":"authored_by"},{"s":"economics-agi","t":"o-mit","r":"made_by"},{"s":"economics-agi","t":"m-economic-model","r":"uses_method"},{"s":"tao-ai-science","t":"c-proof-indigestion","r":"discusses"},{"s":"gowers-fields-letter","t":"c-proof-indigestion","r":"discusses","note":"his word is digestion"},{"s":"tao-ai-science","t":"o-fields-letter","r":"related_to","note":"a signatory"},{"s":"p-martin-hairer","t":"o-fields-letter","r":"related_to","note":"a signatory"},{"s":"claude-fermat-report","t":"litt-beginning-mathematics","r":"related_to","note":"certainty without new understanding"},{"s":"claude-fermat-report","t":"cheating-whistleblowing","r":"related_to","note":"Lean exploit guards versus notation shadowing"},{"s":"periodic-infrastructure","t":"m-closed-loop-lab","r":"uses_method"},{"s":"periodic-infrastructure","t":"c-physical-bottleneck","r":"discusses"},{"s":"periodic-infrastructure","t":"no-magic-wands","r":"related_to","note":"new measurements are the scarce input"},{"s":"ai-in-science","t":"c-physical-bottleneck","r":"discusses"},{"s":"ai-in-science","t":"economics-agi","r":"related_to","note":"early empirical support for verification as the bottleneck"},{"s":"cancer-25-years","t":"p-marinka-zitnik","r":"authored_by"},{"s":"medical-lab-partner","t":"v-aamcnews","r":"published_in"},{"s":"medical-lab-partner","t":"sys-athena","r":"about_system"},{"s":"p-marinka-zitnik","t":"sys-athena","r":"related_to","note":"her lab built it"},{"s":"medical-lab-partner","t":"cancer-25-years","r":"related_to","note":"Zitnik's lab in practice"},{"s":"drug-discovery-path","t":"p-andreas-bender","r":"authored_by"},{"s":"drug-discovery-path","t":"c-proxy-validity","r":"discusses"},{"s":"drug-discovery-path","t":"c-fast-cheap-feedback","r":"discusses"},{"s":"no-magic-wands","t":"c-fast-cheap-feedback","r":"discusses"},{"s":"jun-future-science","t":"uncertainty-layer","r":"related_to","note":"reasoning records versus claim records"},{"s":"chollet-third-axis","t":"gemini-real-world","r":"related_to","note":"Agent_H scales breadth and depth"},{"s":"chollet-third-axis","t":"c-inspectable-reasoning","r":"critiques","note":"latent reasoning gives up inspectability"},{"s":"discovery-loop-launch","t":"p-jeff-dean","r":"authored_by","note":"interviewed"},{"s":"discovery-loop-launch","t":"p-jeff-dean","r":"commented_by"},{"s":"discovery-loop-launch","t":"o-discovery-loop","r":"discusses"},{"s":"p-jeff-dean","t":"o-discovery-loop","r":"affiliated_with","note":"CEO and co-founder"},{"s":"p-sanjay-ghemawat","t":"o-discovery-loop","r":"affiliated_with","note":"co-founder"},{"s":"p-quoc-le","t":"o-discovery-loop","r":"affiliated_with","note":"co-founder"},{"s":"p-oriol-vinyals","t":"o-discovery-loop","r":"affiliated_with","note":"co-founder"},{"s":"gemini-real-world","t":"p-quoc-le","r":"authored_by","note":"senior author"},{"s":"o-discovery-loop","t":"o-google","r":"related_to","note":"founding investor and compute partner"},{"s":"discovery-loop-launch","t":"c-loop-latency","r":"proposes"},{"s":"discovery-loop-launch","t":"c-recursive-self-improvement","r":"discusses","note":"the company is its own first customer"},{"s":"discovery-loop-launch","t":"c-nae-grand-challenges","r":"discusses"},{"s":"discovery-loop-launch","t":"c-autonomous-research","r":"discusses"},{"s":"discovery-loop-launch","t":"agentic-scientist","r":"related_to","note":"same bet on automating the research loop"},{"s":"discovery-loop-launch","t":"periodic-infrastructure","r":"related_to","note":"AI-for-science startups built on large-scale infrastructure"},{"s":"c-loop-latency","t":"c-fast-cheap-feedback","r":"related_to","note":"loops close fastest where checking is cheap"},{"s":"c-recursive-self-improvement","t":"c-solipsistic-ai","r":"related_to"},{"s":"grupen-context-graphs","t":"p-niko-grupen","r":"authored_by"},{"s":"p-niko-grupen","t":"o-harvey","r":"affiliated_with"},{"s":"grupen-context-graphs","t":"o-harvey","r":"made_by"},{"s":"grupen-context-graphs","t":"c-context-graphs","r":"proposes"},{"s":"c-context-graphs","t":"m-shared-memory","r":"related_to","note":"structured memory for agents"},{"s":"c-context-graphs","t":"c-uncertainty-records","r":"related_to","note":"machine-readable structure around claims"},{"s":"davis-labs-taste","t":"p-ben-davis","r":"authored_by"},{"s":"davis-labs-taste","t":"c-research-taste","r":"proposes","note":"good decisions where no programmatic verifier exists"},{"s":"davis-labs-taste","t":"c-verification","r":"discusses","note":"taste starts where the verifier ends"},{"s":"davis-labs-taste","t":"c-verifiable-hill-climbing","r":"discusses"},{"s":"davis-labs-taste","t":"c-regression-to-mean","r":"critiques"},{"s":"davis-labs-taste","t":"c-recursive-self-improvement","r":"discusses","note":"taste as the only real stop to RSI"},{"s":"c-verifiable-hill-climbing","t":"c-optimizer-not-scientist","r":"related_to"},{"s":"c-research-taste","t":"c-verification","r":"related_to","note":"judgment is what remains when checking is impossible"},{"s":"davis-labs-taste","t":"research-taste","r":"related_to","note":"both treat taste as choosing directions without a scorecard"},{"s":"davis-labs-taste","t":"economics-agi","r":"related_to","note":"the measurability gap between what agents do and what humans can check"},{"s":"davis-labs-taste","t":"agentic-scientist","r":"related_to","note":"humans set questions and standards; agents expand experiments"},{"s":"agentrun-grep","t":"p-miguel-rios","r":"authored_by"},{"s":"p-miguel-rios","t":"o-grep-ai","r":"affiliated_with"},{"s":"agentrun-grep","t":"m-agent-writes-workflow","r":"proposes","note":"the agent “fires itself” once the workflow covers 98 of 100 cases"},{"s":"agentrun-grep","t":"c-decision-action-split","r":"uses_method","note":"“Jev decides what a thing is; code decides what follows”"},{"s":"agentrun-grep","t":"sys-jev","r":"about_system"},{"s":"agentrun-grep","t":"m-deterministic-claim-check","r":"uses_method","note":"verify clause checks each claim against cited evidence"},{"s":"agentrun-grep","t":"m-replayable-judgment","r":"uses_method"},{"s":"agentrun-grep","t":"c-verification","r":"discusses","note":"verify clause plus a held-out set the author never sees"},{"s":"agentrun-grep","t":"c-research-taste","r":"discusses","note":"turns closed decisions into checkable typed questions and escalates open judgment"},{"s":"agentrun-grep","t":"c-goodhart","r":"discusses","note":"the development score alone was not a promotion criterion"},{"s":"agentrun-grep","t":"c-human-oversight","r":"discusses","note":"low-confidence cases escalate"},{"s":"agentrun-grep","t":"sys-coding-agents","r":"builds_on","note":"built on the Pi agent loop"},{"s":"agentrun-grep","t":"davis-labs-taste","r":"responds_to","note":"posted by the user as a response to the taste framing"},{"s":"convoke-biopharma-benchmark","t":"o-convoke","r":"authored_by"},{"s":"convoke-biopharma-benchmark","t":"c-drug-discovery","r":"discusses","note":"commercial, development and operational work rather than R&D"},{"s":"t-biopharma","t":"c-drug-discovery","r":"related_to"},{"s":"convoke-biopharma-benchmark","t":"m-penalty-rubric-grading","r":"proposes"},{"s":"convoke-biopharma-benchmark","t":"c-synthesis-vs-compilation","r":"discusses"},{"s":"convoke-biopharma-benchmark","t":"c-research-taste","r":"discusses","note":"commercial forecasts need design judgment"},{"s":"convoke-biopharma-benchmark","t":"c-verification","r":"discusses","note":"grader not yet validated against expert grading"},{"s":"drug-discovery-path","t":"c-drug-discovery","r":"discusses"},{"s":"no-magic-wands","t":"c-drug-discovery","r":"discusses"},{"s":"c-causal-biology","t":"c-drug-discovery","r":"related_to"},{"s":"convoke-biopharma-benchmark","t":"drug-discovery-path","r":"related_to","note":"pipeline work beyond discovery"},{"s":"convoke-biopharma-benchmark","t":"agentrun-grep","r":"related_to","note":"both find the harness, not just the model, decides results on knowledge work"},{"s":"convoke-biopharma-benchmark","t":"davis-labs-taste","r":"related_to","note":"judgment-heavy tasks are where models fall short"},{"s":"dean-dawnsong-fireside","t":"p-jeff-dean","r":"authored_by","note":"speaker"},{"s":"dean-dawnsong-fireside","t":"p-dawn-song","r":"commented_by","note":"moderator"},{"s":"dean-dawnsong-fireside","t":"o-stanford","r":"made_by","note":"held at Stanford's Memorial Auditorium"},{"s":"dean-dawnsong-fireside","t":"o-discovery-loop","r":"discusses"},{"s":"dean-dawnsong-fireside","t":"c-shared-discovery-loop","r":"proposes"},{"s":"dean-dawnsong-fireside","t":"m-ev-cost-experiment-portfolio","r":"proposes"},{"s":"dean-dawnsong-fireside","t":"c-loop-latency","r":"discusses","note":"days to minutes, thousands in parallel"},{"s":"dean-dawnsong-fireside","t":"c-recursive-self-improvement","r":"discusses","note":"automate data, evals and architecture; no timeline given"},{"s":"dean-dawnsong-fireside","t":"c-research-taste","r":"discusses","note":"skim widely, 10x signal, 5-of-7 shape"},{"s":"dean-dawnsong-fireside","t":"c-ten-x-signal","r":"discusses"},{"s":"dean-dawnsong-fireside","t":"c-five-of-seven","r":"discusses"},{"s":"dean-dawnsong-fireside","t":"c-autonomous-research","r":"discusses"},{"s":"dean-dawnsong-fireside","t":"c-human-oversight","r":"discusses","note":"Song's account of an agent escaping its sandbox; security as double-edged"},{"s":"c-five-of-seven","t":"c-research-taste","r":"related_to"},{"s":"m-ev-cost-experiment-portfolio","t":"c-loop-latency","r":"related_to"},{"s":"dean-dawnsong-fireside","t":"discovery-loop-launch","r":"related_to","note":"the founder's own account two days later"},{"s":"dean-dawnsong-fireside","t":"research-taste","r":"related_to","note":"first-hand heuristics for choosing problems"},{"s":"dean-dawnsong-fireside","t":"agentic-scientist","r":"related_to","note":"the same loop automated end to end"},{"s":"dean-dawnsong-fireside","t":"self-organizing-teams","r":"related_to","note":"orchestrating agent teams on sub-problems"},{"s":"dean-dawnsong-fireside","t":"cheating-whistleblowing","r":"related_to","note":"agents behaving outside their brief"},{"s":"c-institutional-alignment","t":"c-governed-claim-ledger","r":"related_to","note":"the ledger is one concrete institution"},{"s":"c-institutional-alignment","t":"c-human-oversight","r":"related_to","note":"humans hold the final actuation gate"},{"s":"c-governed-claim-ledger","t":"m-shared-memory","r":"related_to","note":"governs what enters shared memory"},{"s":"c-governed-claim-ledger","t":"c-context-graphs","r":"related_to"},{"s":"c-governed-claim-ledger","t":"c-verification","r":"related_to","note":"promotion requires review and checking"},{"s":"c-typed-knowledge-objects","t":"c-governed-claim-ledger","r":"related_to"},{"s":"c-typed-knowledge-objects","t":"c-uncertainty-records","r":"related_to"},{"s":"c-typed-knowledge-objects","t":"c-context-graphs","r":"related_to"},{"s":"c-selective-contagion","t":"c-governed-claim-ledger","r":"related_to","note":"the ledger is where spread is checked"},{"s":"c-selective-contagion","t":"c-loop-latency","r":"related_to","note":"intervention must beat the speed of spread"},{"s":"c-evidence-chain","t":"c-causal-biology","r":"related_to"},{"s":"c-evidence-chain","t":"c-drug-discovery","r":"related_to","note":"translation is the last link"},{"s":"c-integrity-channels","t":"c-uncertainty-records","r":"related_to"},{"s":"c-integrity-channels","t":"c-human-oversight","r":"related_to"},{"s":"c-integrity-channels","t":"c-verifier-exploits","r":"related_to","note":"make cheating reportable instead of hidden"},{"s":"cheating-whistleblowing","t":"c-integrity-channels","r":"discusses","note":"agents reporting misconduct in a research swarm"},{"s":"c-distinct-bets","t":"c-research-taste","r":"related_to","note":"choosing directions without a scorecard"},{"s":"c-distinct-bets","t":"c-regression-to-mean","r":"related_to","note":"protected bets resist drift to the average"},{"s":"benchling-science-foundries","t":"p-nicholas-larus-stone","r":"authored_by"},{"s":"p-nicholas-larus-stone","t":"o-benchling","r":"affiliated_with","note":"Head of AI"},{"s":"benchling-science-foundries","t":"sys-science-foundries","r":"releases"},{"s":"sys-science-foundries","t":"o-benchling","r":"made_by"},{"s":"benchling-science-foundries","t":"o-parabilis","r":"discusses","note":"named customer"},{"s":"benchling-science-foundries","t":"m-approval-gates","r":"proposes"},{"s":"m-approval-gates","t":"c-human-oversight","r":"related_to"},{"s":"benchling-science-foundries","t":"c-drug-discovery","r":"discusses","note":"antibody design, production and lead optimization"},{"s":"benchling-science-foundries","t":"m-closed-loop-lab","r":"uses_method","note":"agents dispatch to workcells and update training data"},{"s":"benchling-science-foundries","t":"c-loop-latency","r":"discusses","note":"compressing the gaps between experiments"},{"s":"benchling-science-foundries","t":"c-multi-agent-coordination","r":"discusses","note":"design, cloning, lab and analysis agents hand off"},{"s":"benchling-science-foundries","t":"convoke-biopharma-benchmark","r":"related_to","note":"agents for biopharma work beyond pure discovery"},{"s":"benchling-science-foundries","t":"dean-dawnsong-fireside","r":"related_to","note":"automating the experiment loop"},{"s":"benchling-science-foundries","t":"periodic-infrastructure","r":"related_to","note":"AI-for-science built on lab infrastructure"},{"s":"fork-and-flush","t":"p-ziyang-cai","r":"authored_by"},{"s":"fork-and-flush","t":"p-dimitris-papailiopoulos","r":"authored_by"},{"s":"fork-and-flush","t":"c-idea-basins","r":"proposes"},{"s":"fork-and-flush","t":"m-fork-and-flush","r":"proposes"},{"s":"fork-and-flush","t":"sys-autoresearch","r":"about_system"},{"s":"scaling-discovery-forgetting","t":"p-shangyin-tan","r":"authored_by"},{"s":"scaling-discovery-forgetting","t":"p-alex-dimakis","r":"authored_by"},{"s":"scaling-discovery-forgetting","t":"p-matei-zaharia","r":"authored_by"},{"s":"scaling-discovery-forgetting","t":"p-alex-dimakis","r":"commented_by","note":"context-management trick to escape idea basins"},{"s":"p-shangyin-tan","t":"o-uc-berkeley","r":"affiliated_with"},{"s":"p-alex-dimakis","t":"o-uc-berkeley","r":"affiliated_with"},{"s":"p-matei-zaharia","t":"o-uc-berkeley","r":"affiliated_with"},{"s":"scaling-discovery-forgetting","t":"m-forgetting","r":"proposes"},{"s":"scaling-discovery-forgetting","t":"c-history-anchoring","r":"discusses","note":"edits shrink from ~150 to ~30 lines"},{"s":"scaling-discovery-forgetting","t":"c-idea-basins","r":"discusses"},{"s":"scaling-discovery-forgetting","t":"fork-and-flush","r":"builds_on","note":"forgets the workspace too, not just the chat"},{"s":"scaling-discovery-forgetting","t":"sys-autoresearch","r":"about_system"},{"s":"scaling-discovery-forgetting","t":"c-multi-agent-coordination","r":"discusses","note":"communication alone still plateaus"},{"s":"scaling-discovery-forgetting","t":"sys-coding-agents","r":"builds_on","note":"agents run on the Pi harness"},{"s":"m-forgetting","t":"m-fork-and-flush","r":"related_to"},{"s":"c-history-anchoring","t":"c-idea-basins","r":"related_to"},{"s":"c-history-anchoring","t":"c-regression-to-mean","r":"related_to","note":"both describe drift toward the safe and familiar"},{"s":"c-idea-basins","t":"c-distinct-bets","r":"related_to","note":"diversity as a guard against local optima"},{"s":"m-forgetting","t":"m-shared-memory","r":"related_to","note":"the opposite bet: less memory, not more"},{"s":"sys-autoresearch","t":"c-verifiable-hill-climbing","r":"related_to","note":"works where a score is cheap to compute"},{"s":"sys-autoresearch","t":"c-autonomous-research","r":"related_to"},{"s":"scaling-discovery-forgetting","t":"davis-labs-taste","r":"related_to","note":"agents choosing safe, incremental moves"},{"s":"fork-and-flush","t":"dean-dawnsong-fireside","r":"related_to","note":"thousands of parallel experiments in the research loop"},{"s":"larus-stone-foundries-post","t":"p-nicholas-larus-stone","r":"authored_by"},{"s":"larus-stone-foundries-post","t":"sys-science-foundries","r":"about_system"},{"s":"larus-stone-foundries-post","t":"benchling-science-foundries","r":"related_to","note":"launch post for the blog"},{"s":"larus-stone-foundries-post","t":"c-multi-agent-coordination","r":"discusses","note":"“natively multiplayer agents” and agent swarms"},{"s":"larus-stone-foundries-post","t":"c-drug-discovery","r":"discusses","note":"target ID through medical writing"},{"s":"larus-stone-foundries-post","t":"agentrun-grep","r":"related_to","note":"harness engineering as the product work"},{"s":"larus-stone-foundries-post","t":"convoke-biopharma-benchmark","r":"related_to","note":"judgement-heavy tasks are where models are weakest"},{"s":"larus-stone-foundries-post","t":"c-human-oversight","r":"discusses","note":"comments ask about evaluation, traceability and governance"},{"s":"google-wet-lab-agents","t":"p-pooja-rao","r":"authored_by"},{"s":"google-wet-lab-agents","t":"p-petar-sirkovic","r":"authored_by"},{"s":"google-wet-lab-agents","t":"p-annalisa-pawlosky","r":"authored_by"},{"s":"google-wet-lab-agents","t":"o-google-research","r":"made_by"},{"s":"o-google-research","t":"o-google","r":"related_to"},{"s":"verid-semantic-lab","t":"o-google-research","r":"made_by"},{"s":"aptamer-sandbox","t":"o-google-research","r":"made_by"},{"s":"epitope-excision","t":"o-google-research","r":"made_by"},{"s":"google-wet-lab-agents","t":"verid-semantic-lab","r":"related_to"},{"s":"google-wet-lab-agents","t":"aptamer-sandbox","r":"related_to"},{"s":"google-wet-lab-agents","t":"epitope-excision","r":"related_to"},{"s":"google-wet-lab-agents","t":"m-closed-loop-lab","r":"discusses"},{"s":"google-wet-lab-agents","t":"c-human-oversight","r":"discusses","note":"human-directed autonomy; full autonomy an open problem"},{"s":"google-wet-lab-agents","t":"m-biosecurity-screening","r":"uses_method"},{"s":"google-wet-lab-agents","t":"c-verification","r":"discusses","note":"agents never send unverified code to robots"},{"s":"verid-semantic-lab","t":"sys-verid","r":"releases"},{"s":"sys-verid","t":"a-pylabrobot","r":"builds_on"},{"s":"verid-semantic-lab","t":"m-digital-twin-dry-run","r":"proposes"},{"s":"m-digital-twin-dry-run","t":"c-verification","r":"related_to","note":"a verifier for physical actions"},{"s":"m-digital-twin-dry-run","t":"m-deterministic-claim-check","r":"related_to"},{"s":"aptamer-sandbox","t":"m-aptamer-loop","r":"proposes"},{"s":"aptamer-sandbox","t":"m-closed-loop-lab","r":"uses_method"},{"s":"aptamer-sandbox","t":"c-goodhart","r":"discusses","note":"agents optimized affinity and found sticky binders"},{"s":"aptamer-sandbox","t":"c-loop-latency","r":"discusses","note":"7–10 day cycles"},{"s":"aptamer-sandbox","t":"c-fast-cheap-feedback","r":"discusses"},{"s":"epitope-excision","t":"m-epitope-excision","r":"proposes"},{"s":"epitope-excision","t":"c-strategy-tactics-split","r":"proposes"},{"s":"epitope-excision","t":"c-drug-discovery","r":"discusses","note":"binders to FAP, a solid-tumor target"},{"s":"c-strategy-tactics-split","t":"c-research-taste","r":"related_to","note":"the human keeps the choice of direction"},{"s":"c-strategy-tactics-split","t":"m-approval-gates","r":"related_to"},{"s":"google-wet-lab-agents","t":"benchling-science-foundries","r":"related_to","note":"agents driving workcells with human sign-off"},{"s":"google-wet-lab-agents","t":"gemini-real-world","r":"related_to","note":"Google's agents moving from ideas to experiments"},{"s":"google-wet-lab-agents","t":"periodic-infrastructure","r":"related_to","note":"autonomous labs as infrastructure"},{"s":"aptamer-sandbox","t":"agentrun-grep","r":"related_to","note":"measured checks catch what the optimizer games"},{"s":"aptamer-sandbox","t":"gemini-real-world","r":"related_to","note":"two Goodhart cases: sticky binders here, verbosity exploits there"},{"s":"verid-semantic-lab","t":"c-verification","r":"discusses","note":"a deterministic veto between planner and robot"},{"s":"verid-semantic-lab","t":"uncertainty-layer","r":"related_to","note":"provenance for physical measurements"},{"s":"3b1b-better-time-for-math","t":"p-grant-sanderson","r":"authored_by"},{"s":"3b1b-better-time-for-math","t":"c-proof-as-data","r":"proposes"},{"s":"3b1b-better-time-for-math","t":"c-math-teaching","r":"discusses"},{"s":"3b1b-better-time-for-math","t":"c-credit-provenance","r":"discusses","note":"credit assignment is what changes most"},{"s":"3b1b-better-time-for-math","t":"c-human-understanding","r":"discusses","note":"theory building over proof"},{"s":"3b1b-better-time-for-math","t":"c-research-taste","r":"discusses","note":"deciding which questions are worth asking"},{"s":"3b1b-better-time-for-math","t":"c-scientific-institutions","r":"discusses"},{"s":"3b1b-better-time-for-math","t":"litt-beginning-mathematics","r":"related_to","note":"same debate"},{"s":"3b1b-better-time-for-math","t":"gowers-fields-letter","r":"related_to","note":"same debate"},{"s":"c-proof-as-data","t":"c-proof-indigestion","r":"related_to"}]}