Happy Friday!
At some point we decided every human being should remember 47 passwords, which now needs a bunch of special characters, can’t (or shouldn’t) reuse anything, and also somehow not allowed to be written down.
Then we invented password managers because, unsurprisingly, this was insane.
A lot of technology seems to work this way.
First we create complexity. Then we build some ‘innovative’ way to manage that complexity.
Which brings us, oddly enough, to AI in science.
The Cost of Making Science Too Convenient
Reuters reported this week that Anthropic has announced that they have set up a wet lab in the San Francisco Bay Area where it can connect Claude with physical biology experiments and laboratory automation. Which is kind of weird when you remember that Anthropic is the company that makes Claude the chatbot.
Although this isn’t some massive press release, I think that there is something much MUCH more important happening for our future of science.
It’s most likely that maybe Anthropic does not want to become the next Pfizer. I think, maybe it wants to become the Microsoft Office for science.
Anthropic has already been building in that direction. Claude Science is a workbench that connects researchers to scientific software, computing resources and auditable analysis. Claude for Life Sciences goes even further, connecting Claude directly to platforms like Benchling, PubMed, BioRender and 10x Genomics, while Anthropic says its goal is to support work from early discovery through translation and commercialization.
Now add a physical laboratory, and you can imagine where this goes. Claude reads your papers, it knows your previous experiments, analyzes the sequencing run, proposes the next experiment, writes the protocol, sends instructions to a robot and then reads the result… all for you.
From a drug discovery point of view, Anthropic does not need to own the drug. It owns the ‘operating system’ everybody uses to discover one.
And we've seen this before, think about Microsoft. Word, Excel, PowerPoint, Outlook, Teams, SharePoint, etc. None of these things is individually impossible to replace. But once your company runs on Microsoft, leaving Microsoft becomes a pain. Your coworkers use it, your clients use it, your files are there, and everything talks to everything else.
This became real enough that European regulators went after Microsoft over Teams, arguing that bundling it with Office gave Teams a distribution advantage and made it harder for competitors to compete. Microsoft later agreed to changes including selling Office without Teams and opening more interoperability with rivals.
Now take that same business model and apply it to science.
Your ELN connects to Claude, your liquid handler connects to Claude, your CRO uses Claude. Your sequencing pipeline works best with Claude. Seven years of failed experiments are indexed by Claude. Instrument companies make sure their newest hardware plugs into Claude because that is what customers want.
At that point, somebody can build a better scientific AI and it may NOT matter. Everyone is already inside the ecosystem.
And a lot of this could be great. Experiments become easier to reproduce. Metadata get cleaner. Small labs get access to tools that used to require entire computational teams. Failed experiments stop disappearing into someone's old laptop.
But this is where I am truly concerned because science has one problem that Microsoft Office does not.
We actually want people doing things differently.
One researcher attacks a problem differently, someone else builds a weird assay nobody considered, or another researcher thinks the entire premise is wrong.
Now give all of them the same brilliant scientific platform.
The same AI reads the literature, ranks the hypotheses, suggests the controls and gradually gets better at the experiments that plug most cleanly into its own ecosystem.
Nobody has to tell scientists what to study, convenience can do that.
A standardized experiment that costs $800 and runs overnight will generate more data than a messy primary-tissue experiment that costs $75,000 and takes four months. More data make the AI better at the first experiment. That makes scientists use it more. And around we go.
Microsoft lock-in can shape how we work, but scientific lock-in could shape knowledge.
Chart Of The Week
A chatbot company building a wet lab sounds a little strange. But when you look at where the money is going, it starts to look less strange.
The three platform companies in our data that raised the most money, Xaira, Chai and Lila, all pair their models with their own labs. Lila is probably the closest comparison because it is literally building what it calls an autonomous “AI Science Factory.”
And the pattern keeps going. Of the platform companies we track, the nine that operate their own labs raised about $2.3 billion combined. The fourteen without one raised around $270 million.
Now, three companies are doing a lot of the heavy lifting there, so I would not call this some universal rule yet.
But investors do seem to be putting money in that direction, maybe for now it’s all a coincidence…but a model that cannot run its own experiments eventually has to wait for someone else to do them.
What Caught My Eye
Iambic Therapeutics filed to go public, putting another AI-native drug developer in front of public-market investors. The Nvidia- and Qatar Investment Authority-backed company is developing IAM1363, an AI-discovered HER2-targeted cancer drug currently in Phase 1/1b testing, and plans to list on Nasdaq under the ticker IAM. [Link]
The U.S. government announced more than 20 initiatives aimed at replacing some animal testing with human-based research methods. The push includes organoids, tissue models, computational modelling and AI, while NIH is investing more than $88 million in related infrastructure and planning a new lab that combines human organoids, robotics and AI. [Link]
The Guardian took a skeptical look at one of AI’s biggest promises - curing cancer. While AI is already showing value in areas like detection, diagnostics and drug development, researchers interviewed for the piece argue that claims of an approaching “cure” gloss over how many different diseases cancer represents and how slowly biological validation and clinical testing still move. [Link]
Have a Great Weekend!
❤️ Help us create something you'll love—tell us what matters!
💬 We read all of your replies, comments, and questions.
👉 See you all next week! - Bauris