--- schema: 1 type: video title: 'Verifiable Environments for AI in Biology — Kenny Workman, LatchBio' description: >- A single spatial biology run can yield two to six terabytes of data, far more than a scientist can eyeball, and Kenny Workman argues that this is the raw material for teaching AI to actually do science. LatchBio, five years deep in pharma, treats experimental biology as a verifiable substrate: lay a chunk of a tumor over a sequencing surface and you get a giant matrix of numbers whose analysis has a right answer, which is exactly what you need to benchmark and improve a model. They adapted coding models into biology tools and built benchmarks like sequencing based spatial analysis, and found what everyone in post training now knows, that frontier models cannot yet be trusted with this and that measurement is what drives progress. The reason biology is hard is that it is messy and the field rarely agrees on the answer, so much of the work is designing tasks where reasoning, not memorized knowledge, is what gets rewarded. Workman walks through trajectory data from real scientists, tasks like finding the part of a tumor that matters, and why an ambiguous prompt quietly makes a benchmark uninformative. He also gets candid about biosecurity, where model refusals are their own evaluation problem and red team tasks have to be handled carefully, and frames the whole effort as a flywheel: better benchmarks, better tools, more of the program landscape indexed, repeat. Speaker info: - https://x.com/kenbwork - https://www.linkedin.com/in/kennyworkman - https://kenbw.com/ Timestamps: 0:00 - Terabytes of experimental data 1:41 - Decomposing a new paper into tasks 2:44 - A verifiable substrate for science 3:23 - Five years in pharma 4:13 - Coding models as biology tools 5:40 - Why frontier models can't be trusted yet 6:44 - Sequencing based spatial analysis 9:16 - Reasoning, not memorized knowledge 10:07 - Trajectory data from real scientists 12:14 - Why biology tasks are messy 15:36 - Biosecurity and refusals canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/verifiable-environments-for-ai-in-biology-kenny-wor-d28f54a94e22 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/verifiable-environments-for-ai-in-biology-kenny-wor-d28f54a94e22.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T06:29:24.584Z' language: null children: [] videoId: 3ZMUiFaQ3qg durationSeconds: 1062 publishedAt: '2026-07-31T20:00:35.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 101 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 24 endSeconds: 101 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=24' heldReason: rights - ordinal: 2 text: null startSeconds: 111 endSeconds: 165 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=111' heldReason: rights - ordinal: 3 text: null startSeconds: 165 endSeconds: 176 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=165' heldReason: rights - ordinal: 4 text: null startSeconds: 218 endSeconds: 253 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=218' heldReason: rights - ordinal: 5 text: null startSeconds: 263 endSeconds: 349 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=263' heldReason: rights - ordinal: 6 text: null startSeconds: 349 endSeconds: 361 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=349' heldReason: rights - ordinal: 7 text: null startSeconds: 405 endSeconds: 417 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=405' heldReason: rights - ordinal: 8 text: null startSeconds: 417 endSeconds: 561 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=417' heldReason: rights - ordinal: 9 text: null startSeconds: 417 endSeconds: 561 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=417' heldReason: rights - ordinal: 10 text: null startSeconds: 561 endSeconds: 571 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=561' heldReason: rights - ordinal: 11 text: null startSeconds: 571 endSeconds: 615 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=571' heldReason: rights - ordinal: 12 text: null startSeconds: 615 endSeconds: 625 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=615' heldReason: rights - ordinal: 13 text: null startSeconds: 625 endSeconds: 735 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=625' heldReason: rights - ordinal: 14 text: null startSeconds: 735 endSeconds: 746 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=735' heldReason: rights - ordinal: 15 text: null startSeconds: 746 endSeconds: 944 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=746' heldReason: rights - ordinal: 16 text: null startSeconds: 746 endSeconds: 944 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=746' heldReason: rights - ordinal: 17 text: null startSeconds: 746 endSeconds: 944 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=746' heldReason: rights - ordinal: 18 text: null startSeconds: 956 endSeconds: 1062 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=956' heldReason: rights - ordinal: 19 text: null startSeconds: 956 endSeconds: 1062 watch: 'https://liquidgold.cc/watch/3ZMUiFaQ3qg?t=956' heldReason: rights --- # Verifiable Environments for AI in Biology — Kenny Workman, LatchBio A single spatial biology run can yield two to six terabytes of data, far more than a scientist can eyeball, and Kenny Workman argues that this is the raw material for teaching AI to actually do science. LatchBio, five years deep in pharma, treats experimental biology as a verifiable substrate: lay a chunk of a tumor over a sequencing surface and you get a giant matrix of numbers whose analysis has a right answer, which is exactly what you need to benchmark and improve a model. They adapted coding models into biology tools and built benchmarks like sequencing based spatial analysis, and found what everyone in post training now knows, that frontier models cannot yet be trusted with this and that measurement is what drives progress. The reason biology is hard is that it is messy and the field rarely agrees on the answer, so much of the work is designing tasks where reasoning, not memorized knowledge, is what gets rewarded. Workman walks through trajectory data from real scientists, tasks like finding the part of a tumor that matters, and why an ambiguous prompt quietly makes a benchmark uninformative. He also gets candid about biosecurity, where model refusals are their own evaluation problem and red team tasks have to be handled carefully, and frames the whole effort as a flywheel: better benchmarks, better tools, more of the program landscape indexed, repeat. Speaker info: - https://x.com/kenbwork - https://www.linkedin.com/in/kennyworkman - https://kenbw.com/ Timestamps: 0:00 - Terabytes of experimental data 1:41 - Decomposing a new paper into tasks 2:44 - A verifiable substrate for science 3:23 - Five years in pharma 4:13 - Coding models as biology tools 5:40 - Why frontier models can't be trusted yet 6:44 - Sequencing based spatial analysis 9:16 - Reasoning, not memorized knowledge 10:07 - Trajectory data from real scientists 12:14 - Why biology tasks are messy 15:36 - Biosecurity and refusals Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/verifiable-environments-for-ai-in-biology-kenny-wor-d28f54a94e22 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.