Robots Get Smarter Hands, Frontier Models Face Hard Questions

Friday afternoon, and the robotics and AI worlds are both quietly reshaping themselves in ways that don't always make the headlines. Let's start with something tangible: a team at EPFL has built a robotic hand that does something remarkable - it's symmetrical, reversible, and can crawl like a spider while gripping objects. Six identical fingers, any of which can form opposing pairs. It sounds like science fiction, but what makes it interesting is the thinking behind it. Rather than copying human hands (which evolution left asymmetrical and awkward), they borrowed from nature's actual solutions - octopuses that grip and crawl simultaneously, mantises that manipulate and move. The hand can handle objects under its 'palm', on its 'back', or both. For industrial robotics, confined spaces, prosthetics - this is the kind of fundamental rethink that matters.

Models and the Capital Question

On the AI front, this week brought something more complicated than benchmark numbers. Gemini 3.1 Pro launched with genuinely strong results - 77% on ARC-AGI-2, a huge jump from 31%. But the real conversation, the one happening in investor calls and strategy meetings, is about something deeper: can frontier model companies outspend the entire ecosystem built on top of them? Martin Casado and Sarah Wang at a16z laid it out plainly this week - if you can raise three times more capital than everyone else combined, and you can translate that capital directly into capability improvements, then you don't need to be AGI. You can just keep expanding, consuming the layer above you. That's the future some worry about: not intelligence breakthroughs, but capital concentration.

What's interesting is the honesty about what we don't know. These companies are gross margin positive on yesterday's models but gross margin negative on the ones they're training right now. They're borrowing against the future. At some point, the market either catches up to that reality, the cost of compute drops dramatically, or something breaks. Nobody knows which.

The Unsexy Opportunities

One of the most grounded observations this week came from the same conversations - the things not getting invested in. Boring software. Enterprise tooling. Database companies. Monitoring. Things growing 5x in huge markets that don't get attention because they're not growing zero-to-a-hundred in a year. There's also robotics, where investors acknowledge the potential but admit they don't have the expertise to diligence vertical-specific applications. You need to understand agriculture to invest in ag robots. You need to understand mining for mining robots. That's not how horizontal tech investors work.

The afternoon picture, then, is one of genuine progress - smarter hands, better models - mixed with strategic uncertainty about what it all means for who wins and what gets built. The robotics innovation is real and useful. The AI capability improvements are measurable. But the business model question, the one about power and capital and who gets to build the future, that's still unresolved. Worth watching.