Bishamon

I can share an interesting experience from last week. We have a person who is incharge of buying hardware, software and data sets. This might sound stupid but when you are buying 100s of servers, workstations and laptops a month, it's complicated. This dude used Claude to create an entire tracking and maintanence portal that inventoried everything. He even managed to integrate the portal with our monitoring software to display the status of every server vm. He then modified it to store invoices and so on. He's been at it for a couple of weeks and we've been able to identify wastage and needs. Without Claude, this would have been a maze of spreadsheets and a lot of manual labor. But we wouldn't have hired a developer for this. To me, this kind of software is the killer use case for AI. Enough to simplify your life, but not enough to justify hiring someone or buying a product. Is the code great? Is it scalable? Is it good software engineering? No, no and no. But that's besides the point.

the model alone is no longer the product

Jane Street's head of technology just explained the full spectrum of how fast their trading decisions are made. the fastest systems turn around a packet in under 100 nanoseconds. at that speed, if you attached an oscilloscope to the wire going in and the wire going out, you'd see the response start to leave before the incoming packet has finished arriving. at that speed, you can't use a CPU. you can't use any programming language. you're on an FPGA direct wired to the network. and the decisions you're making are incredibly simple. because you literally can't compute anything complex in that time. but here's the part most people miss: that's just one end of the spectrum. Jane Street runs an ensemble of systems operating at every timescale simultaneously. some decisions happen in nanoseconds. some in microseconds. some in milliseconds. some take hours or a full day. "the right way to build an optimal trading strategy is an ensemble approach. for some decisions you're making very simple decisions very quickly. for others, you're operating at the scale of microseconds, milliseconds. and in some cases, if you can get that decision turned around in an hour, that's totally fine." the faster you need to respond, the simpler the decision has to be. the slower you can afford to go, the smarter the model can be. this is why "Jane Street is just a speed game" is wrong. speed is one dimension. intelligence is the other.

Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵

Doing biotech is so metal

Neural networks might speak English, but they think in shapes. Understanding their rich *neural geometry* is key to understanding how they work – and to debugging and controlling them with precision. Starting today, we’re releasing a series of posts on this research agenda. 🧵