By Nicholas Thomas | TrendyVest There was a moment this weekend when a handful of seemingly unrelated stories started to look very related to me. Anthropic has built a biology lab. Claude has been optimizing biomolecular models.
Specialized AI systems are being developed specifically to study aging and longevity. And Google’s Gemini recently accessed the protected systems of three real companies during what was supposed to be a controlled cybersecurity exercise. Each is an interesting technology story on its own.
Put them together, though, and I think they tell us something much bigger: AI is beginning to cross the screen. For the past several years, nearly everything we’ve associated with the artificial-intelligence boom has happened inside a computer. We ask a question, and AI answers.
We ask it to write code, analyze a spreadsheet, or create an image, and it does. Even some of the most extraordinary things AI can do still follow roughly the same relationship: a human asks, a machine responds, and a human decides what happens next. That alone has been powerful enough to trigger one of the largest infrastructure spending cycles the technology industry has ever seen.
But I’m beginning to wonder whether we’ve been looking at the first chapter and mistaking it for the whole story. What if the chatbot was never the destination? What if it was just the first interface?
Something Has Changed You’ll increasingly hear the word agentic. It’s an awkward technology word for a fairly simple idea. Instead of asking AI to answer a question, you give it an objective.
The AI determines what it needs to do, uses tools, observes what happened, adjusts its approach, and continues working toward that objective. Think about what that could eventually mean for science. A traditional AI might analyze existing research and tell a scientist that five proteins look interesting.
A more autonomous system could model those proteins, design hundreds of variations, eliminate most computationally, select the most promising candidates for physical testing, analyze the experimental results, and use what it learned to design the next experiment. Then it could do it again. That is a fundamentally different machine.
It isn’t simply retrieving human knowledge anymore. It’s beginning to participate in the process that creates new knowledge. And we’re starting to see the pieces required to build it.
When AI Meets Biology This week, Anthropic revealed that Claude optimized more than 30 open-source biomolecular models in less than four weeks. According to Anthropic, those models became roughly four times faster on average. Claude also developed a low-memory approach that allowed very large biomolecular systems to be modeled on a single Nvidia GPU node.
Those numbers are impressive, but they weren’t what really caught my attention. Anthropic now has a wet lab. A company known for building one of the world’s most capable language models has established a physical biology laboratory in the San Francisco Bay Area.
That deserves more attention than I think it is getting, because biology has a way of humbling computer models. A molecule can look incredible in a simulation and fail in the real world. A protein can behave differently than expected.
A drug can perform beautifully in a laboratory and fail when it reaches a human body. Reality gets the final vote. If you’re serious about using artificial intelligence to advance biology, eventually the AI has to collide with reality.
That’s what a laboratory provides. The model makes a prediction, an experiment tests it, the result comes back, and the system gains new information from the physical world to formulate the next hypothesis. That loop is where things become interesting.
The Scarcity Nobody Talks About This becomes even more fascinating when you think about longevity. Researchers have introduced AI tools specifically aimed at aging biology, including LongevityBench, specialized Longevity-LLMs, and Longevity Claw. We should be careful about what that means.
There is an enormous distance between an AI identifying an interesting biological target and actually extending a human life. Anyone claiming that AI has “solved aging” is getting far ahead of the evidence. But I don’t think that’s the important story anyway.
The interesting possibility is that AI changes how fast we can search biology. When we discuss scientific progress, we usually talk about money: research budgets, drug-development costs, laboratory equipment and available capital. But another resource is just as scarce—human attention.
There are only so many scientists, laboratories, and hours in a day. A researcher can only read so many papers. A team can only investigate so many ideas.
A laboratory can only perform so many experiments. Science is therefore constantly deciding which questions are worth pursuing, which means enormous numbers of possibilities are never investigated—not necessarily because they’re bad ideas, but because there isn’t enough time. Now change that equation.
Imagine an AI system evaluating 100,000 possibilities, eliminating most computationally, modeling the survivors, and passing the most promising candidates to automated laboratory equipment. The results return to the system, its assumptions change, and another round of experiments begins. Humans still determine the objective.
Humans establish the safety boundaries. Humans validate the results and decide what should ultimately move forward. But suddenly the amount of scientific exploration a research team can perform becomes dramatically larger.
That’s the part I can’t stop thinking about. AI doesn’t need to become an all-knowing scientist to change science. It just needs to make experimentation cheaper.
And machines have one enormous advantage in that environment: they can afford to be wrong. A thousand failed computational experiments aren’t necessarily a disaster. They’re data.
The Other Side of the Equation This same weekend gave us a much less comfortable example of what happens when AI starts acting instead of merely answering. Google’s Gemini was participating in a cybersecurity evaluation when it accessed protected systems belonging to three real companies. The testing environment had inadvertently allowed internet access.
Reporting on the incident says Gemini guessed credentials in one case and found exposed credentials in others. Google says the system stopped when it recognized the targets were real. It would be easy to turn that into a sensational story about AI “escaping.” That’s not what happened.
But the less dramatic explanation may actually be more important. The cost of an AI mistake changes when AI can act. A chatbot hallucinates and gives you a bad answer.
That’s annoying. A financial agent makes a mistake while connected to an account, and the consequences are different. A cybersecurity agent makes a mistake while connected to a network, and the consequences are different again.
Give an AI system access to physical equipment and the stakes change once more. The more hands we give AI, the more important it becomes to determine exactly what those hands are allowed to touch. Permissions, verification, identity, cybersecurity, audit trails, and human oversight aren’t side issues in an agentic world.
They become part of the infrastructure. The Investment Story Gets Bigger The first stage of the AI trade was relatively easy to understand. Artificial intelligence required enormous amounts of computing power.
That meant accelerators. Accelerators required data centers. Data centers required networking, memory, cooling, and electricity.
An enormous infrastructure cycle followed. But if AI moves from answering prompts to continuously performing tasks, the next layer becomes much broader. An AI agent doesn’t necessarily answer once and stop.
It reasons, calls a tool, retrieves information, runs software, observes the result, and reasons again. Eventually some agents will interact with other agents, robots, industrial equipment and laboratories. That still creates demand for computation, but it also creates demand for networking, storage, sensors, robotics, cybersecurity, laboratory automation, scientific instrumentation, biological data, power, and the software required to coordinate it all.
That’s why I think focusing exclusively on who builds the fastest AI chip may eventually become too narrow a way to view this technology cycle. The bigger question may become: Who gives AI access to the real world? The Moat May Not Be the Model Another investment implication here deserves attention.
AI models are becoming extraordinarily capable, but access to capable models may eventually become widespread. What won’t necessarily be widespread is proprietary real-world feedback. Imagine two biotechnology companies using similarly capable AI.
One has access to the model. The other has access to the model, proprietary biological data, automated laboratories, and thousands of physical experiments continuously feeding results back into its system. Those are very different businesses.
The second company has created a loop: experiment, result, data, better hypothesis, another experiment. Every cycle produces information competitors don’t necessarily possess. The longer that loop runs, the harder it may be for someone else to reproduce what the company has learned.
In that world, the moat isn’t necessarily the AI. The moat is the feedback loop. And that concept extends far beyond biotechnology.
Robotics, manufacturing, materials science, agriculture, energy, and autonomous transportation all have the same underlying opportunity. Anywhere AI can take an action, observe the physical result, and learn from it, proprietary feedback can become valuable. And Then There Is Longevity This is where the technology story becomes deeply human.
For most of history, humanity learned to accept aging rather than investigate it at enormous computational scale. Medicine has made extraordinary progress, but aging itself remains incredibly complicated. There probably isn’t a single switch waiting to be flipped.
Interconnected biological systems, genetic factors, cellular damage, immune changes, metabolic processes, and mechanisms we still don’t fully understand. AI doesn’t magically remove that complexity. What it may do is help us search through it faster.
Perhaps thousands of hypotheses that would never receive funding can at least be modeled. Perhaps obscure relationships buried across decades of scientific literature become visible. Perhaps experiments that once took months to design can be developed much faster.
Perhaps automated laboratories eventually operate continuously, with each failed experiment informing the next. Most experiments will still fail. That’s science.
But the important variable may become how quickly we can get to the next one. AI doesn’t need to know the answer. It may simply allow humanity to ask more questions.
What I’m Watching Now Over the next several years, I’m going to care less about spectacular AI demonstrations and more about a much less glamorous word: throughput. How many useful experiments can an AI-assisted laboratory complete? How much does each experiment cost?
How quickly does failure inform the next attempt? How much human intervention is required? How much proprietary data is generated?
And, most importantly, are those experiments producing discoveries that survive contact with reality? For scientific AI, we may eventually talk about experimental throughput the way the technology industry currently talks about tokens per second. That’s when we’ll know whether this has become something bigger than another impressive AI demo.
The Screen Was Only the Beginning I don’t believe AI is about to cure aging. I don’t believe scientists are about to become obsolete. And I certainly don’t believe every company attaching the letters “AI” to a laboratory deserves an enormous valuation.
There will be failures. There will be hype. There will probably be spectacular amounts of wasted money.
But beneath it all, something important appears to be changing. For decades, computers helped humans calculate. Then they helped us communicate.
Then they helped us find information. Generative AI allowed us to interact with information. Now we’re beginning to connect machine intelligence to action.
Once intelligence can act, observe the consequences, and try again, we enter very different territory. Maybe the defining breakthrough of artificial intelligence won’t ultimately be a machine that can answer every question. Maybe it will be a machine that helps humanity run enough experiments to discover answers we never would have found on our own.
If that’s where we’re headed, the chatbot wasn’t the AI revolution. It was the introduction. AI is crossing the screen.
And what it finds on the other side may matter far more than anything it ever showed us on it. Nicholas Thomas TrendyVest