Summary

Sajith Wickramasekara sees biotech emerging from its “dot-com bust” into a cycle governed by speed and cost. The 2021 platform-company playbook left startups stranded when investors pivoted faster than companies could, while China began producing clinical-stage molecules “really fast and really cheap”; major pharma companies increasingly buy those assets instead of sourcing only from US biotechs.
The industry’s core economic problem is not merely expensive clinical trials but weak molecules reaching them. Elad Gil calls trials partly a “red herring” because many molecules are simply not good; Wickramasekara agrees that biotech needs better molecules and faster feedback. A medicine generally takes about 10 years and more than $2 billion to commercialize, with many failures arriving after hundreds of millions have been spent.
AI’s near-term payoff is cumulative workflow compression, not typing in a disease and receiving a drug. Benchling is pairing predictive models with agents that recover institutional knowledge, recommend experiments and automate scientific work. Sajith describes a long-term design-make-test-analyze loop, while Sarah Guo is more bullish on augmentation over the next 1–2 years, with a goal of moving timelines from “7 to 10 years down to 2 to 3 years.”
Benchling’s structured context may be central to making scientific AI usable and trustworthy. Wickramasekara says biology has “GPT but there’s no chat”: impressive capabilities exist, yet scientists still face accuracy, IP, security and workflow barriers. In a regulated vertical, “90% of the work is actually translation,” and “the AI that wins is going to be the one that people actually use.”
Pure-play biological model vendors may struggle as model building commoditizes. Wickramasekara considers it unlikely that many will survive on five pharma customers and $100 million upfront deals; plausible endpoints include becoming therapeutics companies, distributing models like SaaS, or enabling transactions in normalized scientific data—including, speculatively, pools or sales of negative results.
Large pharma has not yet transformed its R&D workflows, but its proprietary data-generation capacity is a strategic advantage. Most companies are still piloting copilots and agents, while their laboratories can generate training data at a scale startups cannot match. The more useful adoption metric may therefore be the share of experiments touched by prediction, simulation or AI—not the number of drugs marketed as “AI discovered.”

Deep dive

1. Benchling turned fragmented laboratory work into an R&D system of record

Wickramasekara co-founded Benchling 13 years ago after moving from software engineering into a biology lab and finding scientists working from paper notebooks and desktop spreadsheets. Software developers had rich tools for design and collaboration; scientists, despite working in an advanced field, largely did not.
Benchling now helps scientists design molecules, plan and run experiments, organize and analyze results, and connect each molecule to the work that produced it. It serves about 1,300 biotech and pharma companies plus scientists at more than 7,000 academic institutions.
The data is far broader than sequences or yes/no assays: molecular designs, cell and animal studies, experimental relationships and manufacturing-scale fermenter outputs all emerge across “9,999 steps” after a molecule is conceived, alongside regulatory navigation. Benchling’s premise is that these heterogeneous records must become searchable and reusable so scientists can make decisions from them and AI can increasingly work over them.

2. Biotech’s bust exposed the fragility of the platform-company playbook

Wickramasekara describes the past few years as biotech’s equivalent of the dot-com bust. COVID-era excitement around mRNA drew generalist capital, then changing interest rates, tariffs, regulatory uncertainty, China and slower-than-hoped commercialization of new modalities pushed the sector into a trough.
Gene editing, cell therapies, gene therapies and RNA medicines are real categories with approved products—not scientific mirages. His distinction is commercial timing: investors funded them as platforms capable of treating many diseases, but adoption and returns arrived more slowly and expensively than expected.
In 2021, companies were told to build platforms, address numerous diseases and accept hundreds of millions because “capital is free.” When investors stopped valuing platforms, companies could not pivot nearly as quickly; programs selected as small-population proofs of concept suddenly became the assets defining the whole company.
China changes the competitive axis from modality invention to speed and cost. Major companies including Merck, Pfizer and Lilly are buying molecules from Chinese biotechs that can reach early clinical testing quickly and cheaply; Wickramasekara says China as a major biotech player is “very much here to stay.”

3. Better molecules matter more than shaving the visible trial bill

Wickramasekara’s starting point is that “medicines are magic.” Guo notes that prescription drug sales are about 9% of US healthcare spending and that successful drugs eventually become generic, get cheaper and remain effective, unlike much of labor-intensive healthcare.
The counterweight is brutal development economics: more than $2 billion and generally about 10 years to bring a medicine to market, with many programs failing after seven to 10 years and hundreds of millions of dollars. “It is probably easier at this point to send things to space or to put people on the moon than it is to get a new medicine approved.”
Elad Gil argues that clinical trials are an obvious cost center but can also be a “red herring”: recruitment and study design matter, but “a lot of molecules are just not good.” The higher-leverage task is improving them and reaching human feedback sooner; Wickramasekara agrees that the industry must compress the time and cost of those decisions.
GLP-1s illustrate the return to insight and conviction: obesity was nearly unfundable five years earlier because large, lengthy trials followed repeated failures, although core science had existed since the 1990s. Keytruda similarly passed through acquisitions and nearly reached an out-licensing list before a competitive threat revealed its promise and Merck went all-in.

4. Benchling’s agents are designed to recover scientific memory

Benchling AI begins with simulation: open-source, proprietary and customer-internal models appear at the relevant point in a scientist’s workflow, already connected to Benchling’s data model. The aim is to let wet-lab scientists use computational capabilities without specialized skills and eventually receive recommendations for the next experiment.
Its deep-research agent applies foundation-model-style research to a company’s Benchling records. Scientists can ask questions that previously demanded weeks or months of manual reconstruction and receive an answer in hours, with the experimental context linked to the organization’s data.
One customer was preparing to test 20 mouse models in a study that would have taken eight months. The agent found that a number of models had already been studied years earlier inside a lab notebook from an acquired company; the relevant employees were gone, but the experiment no longer needed repeating.
Wickramasekara’s framing is organizational memory: much of science survives as “folklore and institutional knowledge” and disappears with personnel. Longer term, agents could generate reports, answer questions and compose experiments through voice and vision while a scientist is working in the lab.

5. Augmentation should arrive before the autonomous AI scientist

Sajith says that “AI scientists” evokes a fully automated “design, make, test, analyze loop” in which people sit back and let robots give them drugs. Sarah Guo says she hopes that arrives on a longer time scale, but is more bullish on augmentation over the next 1–2 years.
Guo’s analogy is Waymo versus Tesla: full autonomy may be achievable with enough capital and patience, while broader science can progress through large but staged improvements. The ambition is still substantial—compressing seven-to-10-year programs into two or three years, with fewer specialized roles and lower cost.
Radiology offers a cautionary parallel. Predictions that machine learning would eliminate radiologists have persisted for decades, but the durable pattern has been a copilot; Guo argues that clinical systems also need a human accountable for decisions—someone who can “get sued if something goes wrong.”
Guo is more optimistic about automating experimental decisions than clinical ones. The hosts also ask how specialists can evaluate outputs outside their expertise, comparing the problem to generated code receiving a casual “looks good to me” despite unresolved architectural questions; biological decisions require legibility, provenance and calibrated trust.

6. Biology has powerful models but still lacks its breakout interface

Wickramasekara’s diagnosis is that biology has “GPT but there’s no chat.” Capability has advanced through foundation labs, bio-AI companies and open-source work, yet the interface that makes those capabilities routine for scientists has not emerged.
After returning from a month visiting Boston, London and other scientific centers, Sarah Guo says she found that most people still were not using much AI in R&D. Elad Gil notes that concerns about accuracy, intellectual property, security and legal exposure grow more pronounced farther from San Francisco; large pharma has copilots and pilots, but Sajith has not yet seen transformed R&D organizations.
“In a vertical, I think 90% of the work is actually translation”: earning trust, placing a tool at the right workflow moment, making it simple and validating accuracy. His product rule follows directly: “The AI that wins is going to be the one that people actually use.”
Pharma’s countervailing advantage is experimental scale. Large companies can generate proprietary training data beyond most startups’ reach, suggesting distinctive internal predictive models may mature before agentic workflows reshape the whole organization.

7. AI progress should be measured across the pipeline, not by a branded drug

Open source has become a meaningful force in biology over roughly the past two years, with progress in structural prediction and antibody developability. Wickramasekara also points to Eli Lilly’s announced TuneLab project as a “give-to-get” approach: outside scientists use internal models while Lilly trains them without exposing the raw scientific data. He describes the arrangement as federated “or something like that.”
Important gaps remain, especially models predictive of what happens in patients. Even dramatically better discovery creates downstream pressure: every molecule still needs a scalable manufacturing process optimized for yield, speed, cost, safety and quality—and faster discovery leaves less time to build it.
Guo voices the naysayer’s strongest evidence: years after AlphaFold-era enthusiasm, no marketed drugs have emerged from the pipeline clearly labeled AI-discovered. Wickramasekara rejects the implied vending-machine standard—“type a disease in” and receive a molecule—because drug development is a chain of many independently cumbersome steps.
His preferred metric is the share of experiments touched by prediction, simulation or AI, which he bets is getting higher every day. The thesis is not a single breakthrough model but accumulated compression across targeting, molecule design, experimentation, analysis, manufacturing and development.

8. Model commoditization shifts value toward assets, distribution and data

Wickramasekara doubts many biological model companies can remain pure vendors serving five pharma customers through giant upfront partnerships: “model building is probably commoditizing too fast.” One logical path is using the expertise to become better research organizations and ultimately biopharma companies with proprietary pipelines.
Another path resembles software distribution: Benchling already exposes mostly open models, including Chai and AlphaFold, inside customer workflows; paid models could follow a SaaS or fee-for-service structure and reach the broader industry rather than five negotiated partners.
Better normalization could also create a scientific-data market. Companies rarely buy preclinical data because formats, provenance and experimental quality are difficult to trust; Wickramasekara wonders whether tooling could eventually support transactions—even pools of negative data that let others learn from failed work.

9. The AI pivot drew on founder authority and customer intimacy

Benchling’s co-founder Ashu gave up his direct reports to work full-time on AI, prompting some customers to think he had literally quit. Wickramasekara calls a co-founder’s special power “moral authority”: the willingness to say, “I don’t need any of my Legos anymore,” and risk looking foolish around a controversial bet.
The timing was uncomfortable because biotech customers were laying off staff or shutting down, and employees reasonably asked whether AI distracted from the basics. Ashu’s hands-on model experimentation convinced the founders that Benchling had to translate Bay Area capabilities into a complex, regulated scientific domain.
Wickramasekara probably spends 30% to 50% of his time with customers because product-market fit moves with the market. Five to 10 customers often reveal the industry’s shared needs beneath claims that every laboratory is a snowflake; successful products came from going extremely deep until that cohort was delighted.
The method also applies to AI. One customer used Benchling’s deep-research tool around an FDA meeting, turning a report that took a day into a five-minute task. For Wickramasekara, vertical software makes “do things that don’t scale” work “10x better.”

10. Science and software need each other’s storytelling and discipline

Building an interdisciplinary company left Wickramasekara with “serious battle scars.” Academia treats graduate labor as extremely cheap and papers as currency; a software company must sell products. Benchling repeatedly reinforces that commercial success funds its mission—and he says Benchling is the only independent, scaled player doing this at this point.
Biotech can learn tech’s ability to “go direct” and make people care about its protagonists. Wickramasekara argues that the public knows tech CEOs by first name but rarely knows pharma leaders or scientists, obscuring how difficult medicines are to create and making faceless companies easier to resent.
Guo says tech can learn biopharma’s rigor, validity and accuracy. “Move fast and break things” fits some domains, but credibility with regulators and safety in patients demand something else; she calls the US the gold standard for delivering medicines safely.
Sajith’s personal closing example captures AI’s widening interface: after eight or nine years away from coding, agentic tools restored the “whimsy” of building quickly. His mother uses ChatGPT as a kind of Google Search++, and natural language finally gives less technical users capable software without turning family members into holiday IT support.

GE Aerospace: Full Throttle [Business Breakdowns Episode 235]Sequoia's Transition, Burry Shorts NVIDIA, Gamma Raises $100M at $2BN