Isomorphic eyes $50B valuation as Google pushes Gemini into the workplace

The most interesting thing in today's news isn't the $50 billion that investors reportedly want to pour into Isomorphic Labs. It's the fact that a drug-discovery spinoff from Alphabet can command that kind of valuation while the people funding it are, at the same time, nervously asking whether the whole boom is a bubble. Those two instincts coexist in the same portfolios right now, and I find that tension more telling than any single deal. If Isomorphic closes at anything near that figure, it will be a bet that AI's real payoff lies in science rather than chatbots.

There is evidence for that view, and it's modest in a way I like. A new study shows that GPT-style language models can predict how molecules smell, including how a scent shifts with concentration, and reportedly beat traditional molecular-based approaches. Smell is a notoriously messy problem, and chemists have struggled to map structure to perception for decades. A text-trained model doing well here suggests that these systems are picking up something general about how chemistry gets described, not just reciting facts. By the way, this is exactly the sort of result that makes an investor's pulse quicken, and exactly the sort that deserves a skeptical second look before anyone extrapolates from fragrance to cancer drugs.

The same week offers a useful counterweight on trust. Google unveiled a Gemini workplace agent that writes and executes code and plugs into Gmail and other apps to clear routine tasks. Meanwhile, the developer behind the ARTEX agent has gone closed-source after cybersecurity firms linked the tool to an attack tied to a Korean bank. Put those side by side and the lesson is plain: an agent that can act in your inbox and run code is useful precisely because it has power, and that power cuts both ways. Closing the source after the fact looks more like damage control than a security strategy. If you're building with agents, assume anything that can execute code will eventually be pointed at something you didn't intend.

That's why I was drawn to VR-FraudNet, which lets a language model propose fraud explanations but forces a fixed, deterministic verifier to confirm each one before it counts. It's an unglamorous design, and I think it's the right one. Let the model be creative, and let something boring and predictable make the final call. Why aren't more agent deployments built this way?

The mood at the Singapore conferences, where investors and policymakers fretted over both financial and existential risks, and the renewed push for Congress to act on safety and accountability, suggest the appetite for that kind of discipline is growing. The Center for AI Safety is now courting creators to widen the conversation, which tells me the argument is moving from policy circles to the public. I expect the next few months to hinge on whether verification and accountability become engineering defaults or remain afterthoughts, and I suspect the market will punish whoever learns that lesson last.