AI Engineer World's Fair 2026
Verifiable Environments for AI in Biology — Kenny Workman, LatchBio
Read the talk
Verifiable Environments for AI in Biology
Biological agents need more than coding skill and textbook knowledge: they need experimental data, executable tests, and grading that respects valid scientific alternatives.
From a talk by Kenny Workman
Before you start: Basic familiarity with AI agents, Python functions, and data-analysis workflows is helpful; the biological measurement techniques are introduced as needed.
When one experiment outgrows a laptop
How does a scientist analyze an experiment whose output can exceed the storage capacity of a consumer laptop? This is the practical starting point for Kenny Workman’s work at Latch: biological measurement is producing large datasets, and turning those measurements into scientific conclusions requires increasingly capable infrastructure.
A few experimental classes drive much of this growth. Single-cell biology separates cells and measures their RNA. Spatial biology adds location information, showing how RNA is distributed across tissue. Proteomics measures proteins through a broader collection of techniques. Workman gives the following examples of experimental output; these are illustrative volumes, not universal sizes for every assay.
| Experimental class | Reported output |
|---|---|
| Single-cell | 2–6 TB per run |
| Spatial | 7 TB per run |
| Proteomics | A few hundred GB |
Even before interpretation begins, moving and storing the data can become a substantial part of the work.
Improvements in molecular capture point toward further throughput gains. The infrastructure problem therefore grows with the measurement technology: a useful agent needs access to the data and the compute that can process it, not just a chat window.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Data analysis supplies something executable
A useful way to read a biological paper is to reconstruct its sequence of work:
- Choose a biological model—an experimental biological system, not an AI model.
- Generate measurements from that system.
- Process the resulting data.
- Interpret the results against prior literature.
- Make a scientific claim.
The details vary, but the recurring task is to find a signal in noisy measurements.
Data analysis could provide an executable foundation for biological agents, much as code provides one for software agents. A broad scientific goal may be difficult to verify directly, while individual computations along the way can have checkable outcomes. Those components offer places to benchmark capability and potentially train improvements. The slide makes this concrete with analysis steps including cell-line quality control, differential expression, and Perturb-seq.
The analogy is a proposed foundation, not a claim that executing correct code establishes a biological discovery. Workman explicitly qualifies it: data analysis might scaffold harder scientific work in the way executable software tasks scaffold broader engineering work, but the biological case needs its own justification.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From experimental files to an agent loop
Latch began as a data-tool vendor for biotech and pharmaceutical companies, storing and transforming large experimental files. It later moved toward manufacturers of experimental kits, packaging analysis software that those manufacturers could provide to scientists under their own brands. The same infrastructure components then became tools an agent could call. The operational difference from many chat-based tools is substantial: some biological operations take days or weeks.
The first promising prototypes used coding models that, to the team’s knowledge, had not received serious biology-specific post-training. Their interfaces were familiar: ask a question in chat, build a dashboard, and dispatch computation to external infrastructure.
Consider a cancer biopsy measured with a spatial assay. A scientist wants to know which genes are overexpressed in malignant tissue compared with non-malignant tissue. The agent receives large experimental files, and the scientist iterates with it toward that question. Answering requires working with the measured tissue data, not merely recalling facts about cancer. These early agents were weak, but they showed enough promise to suggest a progression like coding agents: increasingly capable individuals, followed by harnesses that coordinate teams. Workman presents those teams of scientific agents as a likely future, not an already demonstrated capability.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Knowing biology and writing code are not enough
The missing capability lies between biological knowledge and programming: extracting scientific insight from real data. Workman argues that frontier models at the time of the talk could not be trusted with that work without focused post-training. The task combines software execution, data analysis, and scientific reasoning. Spatial biology offered a useful starting point because Latch already had customers, the technical field was open, and the measurements made biological patterns directly visible—even the organization of a developing mouse embryo.
Building agents required understanding how those measurements were made. Spatial capture technologies draw on chemistry, optics, semiconductors, and physics, with different branches reflecting decades of work on measuring particular molecules. In one sequencing-based approach, a slide carries small beads bearing DNA that captures RNA from a tissue section. Placing tissue over the slide allows the experiment to recover RNA measurements together with information about where that RNA was located.
The resulting inputs include a large numerical matrix and a high-content image. Reaching a biological conclusion requires a sequence of analysis steps, but that sequence varies with the capture technology, tissue, and disease. There is also limited consensus about how to perform many steps. An agent can therefore execute a plausible workflow without establishing that it has done good science. Measuring that distinction becomes the next engineering problem.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Break the workflow into verifiable tasks
The available benchmarks largely tested academic questions about what a scientist should do, or covered experiments too broadly to measure this particular work. Latch built SpatialBench to evaluate agents working with spatial data. The original SpatialBench contained 146 problems, attempting coverage across experimental kits and analysis tasks.
Grading only the final scientific outcome produced too little information when models usually failed. SpatialBench instead decomposed the analysis into components of a directed acyclic graph, or DAG. Each task starts with data prepared at an intermediate state where a scientist could take the next meaningful step. Benchmark authors then establish what can reasonably count as success from that state. The construction diagram connects this decomposition to data snapshots, graders, manual inspection, and checks for agent cheating.
Each evaluation packages four pieces:
- Data nodes: One or more experimental inputs, such as a matrix or image.
- Task prompt: A carefully described scientific goal.
- Grader configuration: The information needed to evaluate the submitted result.
- Deterministic grader: A function that checks the success condition.
The grader can be a Python function, and the evaluation specification is represented as JSON. SWE-bench supplied an early design reference: give an agent a concrete environment and task, then evaluate the outcome through executable checks.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A deterministic grader can still be wrong
A useful biological test needs more than a function that returns True or False.
- Verifiability: A function can check the success condition.
- Durability: The grader accepts valid scientific alternatives rather than enforcing one author’s analysis path.
- Data dependence: Reaching the answer requires interaction with the supplied measurements, not a memorized response in one turn.
Durability is especially difficult because science often lacks a single canonical answer. A correct analysis may follow a different path and produce a different result from the benchmark author’s. The target should therefore be a property that remains invariant across acceptable paths.
For example, if a task’s contract is to identify a set of genes, ordering should not determine whether the answer passes. This small Python grader illustrates that distinction:
python
from collections.abc import Iterable
def grade_gene_set(
submitted: Iterable[str],
expected: Iterable[str],
) -> bool:
return set(submitted) == set(expected)
expected = ["GENE_A", "GENE_B"]
submitted = ["GENE_B", "GENE_A"]
assert grade_gene_set(submitted, expected)
The check preserves an order-independent success condition. It does not establish that expected is the right biological target; constructing that target remains the harder scientific responsibility. Benchmarking helped the team choose models for customer tasks, but also exposed the need for human verification and longer evaluations.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Human attempts expose hidden choices
Watching model trajectories across successive releases revealed that many initial task assumptions were poor. Without a canonical scientific answer, having scientists attempt tasks and grade one another’s work became a practical way to find those problems.
One task asked an agent to split a gene list into microglial-activation and oligodendrocyte-inflammation groups, score cells, find neighboring oligodendrocytes using an appropriate radius, and compute a Spearman correlation at two time points. Each instruction concealed a decision:
| Instruction | Unspecified choice |
|---|---|
| Split the genes | Membership of each activity group |
| Identify inflammation | Which genes qualify |
| Score cells | How to normalize measurements |
| Find neighbors | What radius is appropriate |
| Pool local counts | How to aggregate within the radius |
These choices can change the final correlation. A grader tied to one unspoken interpretation can reject a defensible answer.
Human attempts also challenged conventional numerical thresholds used for quality control. A threshold that a bioinformatician routinely applies is not automatically a durable success criterion. Requiring a machine to satisfy an explicit test forces the task author to explain what scientific structure the threshold is supposed to capture. After two rounds of human attempts, the team published a verified subset. Here, verification means an independent expert could reconstruct a passing answer from the prompt and data; it does not imply a uniquely correct scientific answer.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From individual steps to a scientific result
The next extension was SpatialBench-Long. Real biological work combines experiment types and complete workflows. At every step, the analyst must interpret the data against experimental design, prior literature, and the original scientific goal. The longer tasks aim to resemble the results section of a paper or the analysis behind an industry decision to continue or stop a drug program. Workman reports that constructing each task took roughly a week for a group of three people.
One example asks whether an agent can reconstruct a tumor’s metastatic niche. Given a primary-tumor biopsy and biopsies from metastatic lesions elsewhere in the body, the agent must combine genetic and mRNA evidence to identify the region of the original tumor that seeded the spread. The scientific motivation is to identify the populations that cause problems and might warrant targeted intervention.
Workman reports that none of the evaluated models solved this particular metastatic-origin task at the time of the talk. That result is specific to the example, not a claim of zero success across the benchmark. It also exposes a training problem: when a long workflow ends in failure, a terminal reward says little about which parts of the analysis were useful or where it went wrong.
The team began experimenting with intermediate rubrics. Imagine the possible analysis paths as a tree: different legitimate branches may still pass through common scientific checkpoints. Domain experts can identify those invariant points and use them to assess progress without prescribing every step. But the rubric scores were only loosely correlated numerically with verifiable final outcomes. Workman therefore did not yet regard them as sufficiently established for reinforcement learning or benchmarking. Intermediate assessment remained experimental; verifiable outcomes remained the preferred foundation.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Extending evaluation across biological work
The spatial work began reaching frontier-model developers. Workman reports that Anthropic included the benchmarks in model cards and presented them at a Claude Science launch, without advance coordination with Latch. He connects that adoption to the team’s early experience shipping products and experimenting with coding agents on customer workflows.
The research then expanded beyond spatial assays into single-cell and epigenomics tasks, including longer workflows. Drug discovery adds another organizational problem: a benchmark must identify which part of a program it measures. Latch’s first small-molecule preclinical pharmacology benchmark begins a broader effort to organize tasks across discovery, development, and translation, further separated by therapeutic type and experimental technique. The slide places TxBench-PP beside a grid of these benchmark categories.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Refusal is also an evaluation problem
Latch also formed a biosecurity team through an acquisition. Workman introduces collaborations with American Wetware and Aclid; the refusal study is associated with American Wetware, while Aclid’s collaboration concerns surveillance. The practical concern is that safety behavior can fail in both directions: blocking basic scientific assistance and overlooking risky requests.
Benchmarking Refusals in Agentic Biology contrasts routine requests a legitimate scientist might make with red-team requests designed to appear innocuous. The conceptual deception example is a request that labels biological material as GFP, a fluorescent protein, while the underlying material is hazardous. Evaluating only the benign description would miss the risk.
In the reported evaluation, routine tasks were refused more often than red-team tasks. The comparison concerns the tested configurations under direct task framing; it is not a universal refusal rate for biological assistants. Interpreting these results also requires distinguishing provider API blocking from a model’s own refusal. The evaluation target is whether the system recognizes the relevant risk while still supporting ordinary scientific work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Customer work determines what to measure next
Latch is aggregating results, preprints, evaluations, and agent trajectories into a central research resource. But the benchmark program remains connected to deployment: the company continues serving customers and working with kit manufacturers, using those relationships to identify the scientific tasks that matter in practice.
The intended cycle is concrete. Customer workflows reveal missing capabilities; those gaps become benchmarks; competition among model developers improves performance on the measured tasks; and better models return to the products scientists use. Workman closes with an invitation to engineers and scientists to join that research and deployment effort.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The original benchmark of 146 spatial-analysis problems with experimental data and deterministic graders.
Two rounds of independent expert attempts produce 115 verified tasks and illustrate ambiguous prompts and overly narrow grading tolerances.
A 24-task benchmark testing recovery of scientific conclusions through complete spatial-biology workflows.
A paired evaluation of legitimate biological research and concealed hazards, separating API blocking from model refusal behavior.
Code, datasets, and evaluation instructions for resolving real software issues with language models.
Further reading
Workman's argument for data analysis as the executable foundation of biological agents, including the limits of the software analogy.
Read the complete timestamped transcript
- 0:00
[upbeat music] Thank you to the organizers for having me. Uh, I'm one of the co-founders and CTO at Latch.
- 0:17
We are basically a vertical AI lab for benchmark and agent engineering, um, hoping to motivate and explain exactly what that means today. Um, starting directly with motivation for agents in, in bio generally.
- 0:33
Um, many people in my domain are familiar with this curve, uh, but this is basically the log linear curve of data generated over the years in, in biology. And the reason I'm bringing it up, it will become directly important to the kinds of things we want to do in engineering.
- 0:46
Um, this curve is driven by a very small handful of experimental classes. Uh, one is called single-cell biology. Um, this is where we split up cells, break them apart, measure their RNA.
- 0:58
The second is spatial biology, which will become the focus of the next segment of the talk. Same thing as single cell, but you get spatial resolution. You can look at how RNA is spread out geometrically over a tissue.
- 1:08
And the third thing is proteomics. It's a broad category of different techniques. They measure proteins, um, less, less abundant in ordering, like less, less data volume generated relative to the other two, but still important.
- 1:21
Um, you guys are technical, and I always think it's good to ground things somewhat quantitatively, but, um, these are really big numbers, and the experimental data from these techniques is growing quite rapidly.
- 1:33
Um, almost greater than any other domain of science other than particle collider, um, machines. Single-cell experiments can yield two to six terabytes per run. Spatial runs can yield seven terabytes of run, proteomics a few hundred gigs.
- 1:47
Um, and the only reason I bring this up is to say, hey, like, m- the output of a single experiment can exceed what a scientist can safely store on a consumer laptop in many cases.
- 1:58
And the laws driving how the, the molecular capture works point to rapid, uh, gains in this throughput over the coming years.
- 2:05
Um, one thing I like to do, uh, is when I read a new paper, is decompose it and align it to this framework because it will become important in a second.
- 2:13
Uh, modern biology research is centered around those experiments. You basically choose a model, biological model, not the kind of models you guys are used to. You generate data from that model, you process the data, you creatively think about the results in the context of prior literature, and you make a claim.
- 2:30
Almost all modern experiments, papers that you see published follow this loose structure with a lot of nuance. Um, all, all, all that to say is they be- they become something of a, a panning experiment.
- 2:40
You're looking for a signal using measurement in a sea of noise. Um, and so this is building up to the claim that like code and SWE, data analysis scaffold to agentic biology.
- 2:50
It becomes this executable substrate that we can use to train things. It induces a natural way to benchmark and climb capability. Um, I've written about this a lot at this blog, uh, link here.
- 3:02
There's a lot more depth to this claim, so I wouldn't take it at face value, but it's something to look into. All, all you can take away from this is like, just like code provided a verifiable substrate for complex software tasks that are not inherently verifiable, uh, data analysis might do the same thing in bio.
- 3:17
So how did we get started? We were originally a data tool vendor for biotech and pharma, um, where we started five years ago out of Berkeley. I'm [REDACTED:age]. We started when I was [REDACTED:age].
- 3:29
We stored, transformed file data from large experiments as a service. Uh, we tried to build products, explored lots of things. Over the last two years, um, we started moving away from biotech and pharma and more towards the people who build those kits I was talking about.
- 3:45
Packaged the software into kind of white-labeled things that they s- provided the scientists themselves, um, help them analyze their data. And then over time, there became this strong interaction, uh, with the agents, uh, using the infrastructure components as tools in the loop context you guys are familiar with.
- 4:03
Except in our domain, the tools can take days or weeks. I'm serious. Um,
- 4:09
what started to happen around last summer is agent prototypes, uh, started to work. So we took, uh, coding models. Uh, to our knowledge at the time, they were not seriously post-trained on any tasks in biology to this point.
- 4:21
And we started to build products that look a lot like all the other agent products. They have a chat interface for you to ask questions to, and they build dashboards and dispatch operations to external compute.
- 4:32
Um, the kinds of things these-- that this, this agent would do is take large file data from the types of experiments I was talking about earlier, say like, uh, tissue bisec-- uh, biopsy from cancer with a spatial measurement.
- 4:45
And then the, the scientist is just iterating with it to get at some question they have, you know, maybe between a malignant and non-malignant part of the tissue, what kind of genes are being overexpressed.
- 4:58
Um, but what was fascinating is e- even though it was pretty bad, it showed the early signs of working. And, uh, it became clear to us at this time that agentic biology might look a lot like code.
- 5:08
I actually lifted this slide from Anthropic's, um, cloud science announcement yesterday. But we-- just like, you know, you had this like kind of faulty, silly engineer that became better, and then a- as it improved in capability, you could dispatch work to teams of them that work together.
- 5:24
The same pattern will probably emerge in science, and products and harnesses will emerge to orchestrate, uh, work and abstract it so teams of s- teams of agentic scientists can take on capability.
- 5:36
But we needed focused post-training, uh, 'cause at the time and still now, um, frontier models cannot be trusted to do real work. They're missing some capability between knowing biology and writing code, and this is exactly extracting scientific insight from real-world data.
- 5:49
Unlike code, it-- which is one constituent component of this work, uh, it also involves data analysis and domain reasoning, scientific reasoning. We thought spatial biology was a good place to start, so we started building agents.
- 6:00
This is a technical greenfield. We had many existing customers. Uh, it's also just a beautiful example of measurement drives progress. You can actually see biological phenomena play out. Um, you can look at a developing mouse embryo.
- 6:13
Um, and we had to get in the guts of how the data was captured and analyzed to build good agents. Um, I'm not gonna get into this in detail, but I put up this tree of, uh, different capture technologies in spatial biology to highlight the diversity of things that exist.
- 6:28
They really span advances in chemistry, optics, semiconductors, physics. Each branch, uh, is induced from decades of cumulative work to figure out how to measure a, a type of molecule.
- 6:41
As, as a specific example, one technique we work with is called sequencing-based spatial, and it's where you take a slide of little beads with clumps of DNA attached to them that fuse the RNA inside of a tissue section.
- 6:54
So biologists can, like, lay like a chunk of a tumor over it, and then it'll capture all the RNA in it, and then let you know with precise, uh, geometric resolution where the RNA was.
- 7:04
It's cool stuff. The data, when it comes out, ends up looking like a big matrix of numbers in a large high content image. Um, you have to take it through a sequence of steps, uh, to get to the end thing that you want.
- 7:16
These steps are highly variable, especially across technology types, tissue, disease contexts. Um, there isn't a lot of consensus in the field, uh, for each step, so we really need a measuring stick to understand if the agents we're building were doing scientific work.
- 7:30
Uh, the existing benchmarks we saw at the time did not measure the tasks relevant to this category of work. Um, they mostly measured things in a Q&A setting, so like what would you do in a kind of academic way, or they weren't sufficiently focused on the experiment type.
- 7:43
This is an actual screenshot of, um, Anthropic's model card at the time that we built this benchmark. So we built one, it's called SpatialBench, um, last December. There's one hundred and forty-six problems.
- 7:55
They spanned the, the different kits I talked about or attempted to, and then they spanned all those different tasks that I talked about as well. So the thing that we found at this time, and still to an extent is true today, is the, the grading of these end outcomes in biology, uh, is too sparse because the models
- 8:12
are pretty bad. So you have to break things up into manageable chunks to get some semblance of verifiability, and that's kind of induced by, um, sticking to these little components of like that DAG, that analysis DAG.
- 8:26
Um, getting data to a state where it would exist right before a scientist or theoretic could do work on it, and then, um, figuring out what the ground truth would be in that context.
- 8:34
So a single evaluation kind of looks like one or more data nodes, again, like a matrix of numbers, a high content image, something like this. A task prompt carefully describing some scientific goal, configuration for a grader, and then a deterministic grader.
- 8:47
It's like a Python function. If you guys notice, this looks a lot like SWE-bench. We borrowed a lot of the early ideas and tried to extend them as much as possible.
- 8:53
Evaluation ends up looking like this, just a lot of JSON. And we ended up identifying properties of like what we thought good biological tests were. Um, a little, little different from code, and we've built on these over time, but they still hold up.
- 9:06
They gotta be verifiable. You have to be able to check the success condition with the function. Um, nothing's changed there. We'll get into some rubric stuff later, but still holds.
- 9:15
Durability is particularly important. Science does not admit clear ground truth. Um, if you are lazy with your ground truth construction of the task, a possible valid analysis path can come with a correct answer, um, and you'll fail it, uh, incorrectly.
- 9:29
So you gotta make sure you're reasoning about something that's somehow invariant across analysis paths. And then obviously, we're, we're working with agentic stuff here. You don't want the model to answer the question in one turn.
- 9:38
You want the conclusion to require interaction with the data and not some memorized knowledge. In practice, that's pretty difficult. We learned a lot about what models could do and which ones to use in specific context for this category of work for our customers.
- 9:50
And we thought, "Hey, this is pretty cool. Let's, let's start to improve and learn more about this benchmarking problem." So we jumped to human verification and long-horizon extension. I'm gonna quickly breeze through these.
- 10:00
So human verification is incredibly important in science. Science does not admit clear ground truths. Uh, after watching trajectory data from multiple rounds of model releases, circa like January to March of this year, um, we really realized a lot of our assumptions were pretty bad.
- 10:16
Um, and in the absence of like a canonical answer, uh, having a bunch of scientists grade each other's work ended up being like the best proxy. So I'm gonna look at one, one issue to highlight exactly what I'm talking about, um, is problem ambiguity.
- 10:32
A task might ask an agent to split a gene list into two groups of activity, microglial activation, oligodendrocyte, inflammation. Just like biological categories of things. Score the cells, find neighboring oligodendrocytes around some region using an appropriate radius.
- 10:48
Compute a Spearman correlation at two time points. As you can probably clearly deduce, the original pro-problem statement creates a host of open choices. How do you split the gene list?
- 10:57
How do you count what inflammatory genes are? It's like some ambiguous word. How do you normalize the data? Um, what, what, what the hell is an appropriate radius? How do you pull the counts within the selected radius?
- 11:08
Um, these, these are all problems that pointed to tasks that were bad, that only became revealed with human verification. Another issue is just like a lot of people in bioinformatics canonically have used like numerical thresholds to QC stuff.
- 11:22
Just like completely arbitrary stuff. Um, cool thing about evaluation like coding is it forces you to reason about things more rigorously than you would when you're doing the thing yourself.
- 11:31
If you have to teach a machine to do it, uh, you, you might be picking out some structure that's more important or more durable than what you were doing if, if you're just doing it on your own.
- 11:39
So we just found a lot of these numerical thresholds to be like bad. Uh, I'm not gonna get into this. Um, after two rounds of human attempts, we produced a verified subset of the benchmark.
- 11:48
Uh, we, we published it. That was fun. Uh, and then we also tried to increase the time horizon. So I want to be clear, the frontier of knowledge is still not quite there with biology.
- 11:59
Like the labs are r- starting to catch up with the post-training, but we kind of want to stay ahead. Um, so we built a benchmark that we thought would recapitulate like really difficult true work.
- 12:11
Um, so we built a SpatialBench-Long. Um, real biological tasks are messy. They use lots of different experiment types. They use the whole workflow, they don't use little chunks. They are, uh-- tasks, every step is interpreted against experimental design or contextualized with some prior literature and the original goal of what you're doing in the first place.
- 12:30
Um, so we built a bank of these tasks that are really trying to simulate the results sections of entire papers or the kinds of decisions you'd make in practice in industry to make a go, no-go decision on a drug program.
- 12:41
Um, these ki-- these tasks took like a week for a group of three people to make each. Taught us a bunch of stuff. Um, an example is can an agent reconstruct a like metastatic niche in a tumor?
- 12:54
Um, if you have like a tumor biopsy and a bunch of metastatic biopsies from like where it metastasizes and spread across the body, can it use like both the genetics, mRNA of the metastatic lesions and the tumor to like find the part of the tumor that initially seeded the metastatic growth and let it spread?
- 13:09
From that, you can figure out like, hey, what parts of the tumor are more like genetically fit? Which ones actually cause problems? And construct targeted medicines to nip them in the bud.
- 13:18
Like for example, this is one of the benchmark, uh, evals in the long horizon set. None of the models get this right, um, but they're getting there. As you can imagine with these long horizon extensions, uh, verifiable rewards at the end are like somewhat uninformative.
- 13:31
So we, we're starting to play with rubrics, uh, constructing these choke points. If you can imagine like the set of analysis paths as inducing some sort of tree, um, there are nodes that are invariant with respect to, you know, different paths, and you can use these to build rubrics, um, using knowledge of how the tasks work.
- 13:49
Uh, we, we-- we're playing with these. We noticed that, um, they're associated with the verifiable outcomes, uh, which is exciting, but they're loosely correlated numerically, um, making us not fully have confidence in them for things like RL or benchmarking.
- 14:06
Um, a lot of-- lot more work to do here still. We, we still strongly believe the verifiability structure is what's going to carry in-intelligence, uh, a bit, a bit longer.
- 14:19
And so these days, uh, excitingly, we've been expanding, um, from this initial spatial focus. Uh, really cool to see the frontier labs and community adopt these benchmarks organically. Um, we had this interesting position by like, you know, building and shipping products, uh, early and kind of playing with the coding agents, so I think we just had a
- 14:38
early advantage. But the benchmarks are now in like the recent Anthropic Model Cards and, uh, this is a picture from yesterday, Eric just showing the benchmarks, um, at the Claude Science launch.
- 14:49
They don't tell us this happens. They just like do it, um, and you like read about it, and it's cool. Uh, we published a bunch more papers, um, beyond spatial to other omics classes, so other experiment types, single cell epigenomics, so RNA, and then the bit above the, the DNA, and then long horizon extensions of these things.
- 15:09
Um, and then we're starting to index and measure the very gnarly, complex landscape that is drug discovery. We just put out our first benchmark on, um, preclinical pharmacology for small molecules, and then systematically biting off pieces of the program landscape from discovery to development to translation, strati-stratifying it by, um, therapeutic types and experiment types.
- 15:33
We just acquired a, uh, company building a biosecurity to form a s-biosecurity team, and then we just put out a collaboration with American Wetware and a surveillance company called Aclid,
- 15:45
um, some new work. Uh, the first was released this morning. I don't know if you guys have been hearing, uh, refusals kind of suck in biology right now. If you ask Fable basic questions about the like mitochondria, it'll-- won't answer.
- 15:58
Uh, it's kind of stupid. So I mean, this is just like an evaluation problem. There's a lot, there's a lot more nuance to this, but I just use that because people tend to recognize it, uh, where we build routine tasks that simulate the kinds of things a scientist would ask for, and then more sinister red team tasks,
- 16:14
which are supposed to look innocuous, but have some structure that is bad. Like, "Hey, I want to clone a gene into a bacteria," and I'm telling you it's GFP.
- 16:23
It's like a glowing protein. But in reality it's like a, a toxin, um, or, or could be used to bootstrap a virus. We found that the routine tasks like drastic-- get drastic-- or fused drastically more frequently than the red team tasks, which is, uh, not great.
- 16:40
Um, we're aggregating a lot of these results, uh, essential resource along with all the pre-prints and a lot of evals and trajectories for you guys to check out. Um, and I actually did okay on time.
- 16:51
Let's go. Uh, and that's it. So we, we are kind of like-- I don't even-- I hate the word lab, but we're kind of like a research lab for [chuckles] bio, and we do research and deployment of these agents.
- 17:02
So we, we still have like a lot of customers. Um, we work with the kit manufacturers, and we use that to inform what kinds of things we make benchmarks for.
- 17:10
We try to get the labs to compete, um, on the benchmarks because then it makes the models better at our products, and it's a-- it's been a pretty rewarding flywheel.
- 17:18
A lot of growth, and we're hiring aggressively across engineering and science. So if you're interested in this work, please find me afterwards. Thank you. [outro music]