@edzitron@danielnewmanUV The constraint is cost, not science. In terms of the thermodynamics: space has plenty of cold regions to radiate into. And, if you’re in the right spot, there’s also plenty of sunlight.
@TheRealDanSaedi “We conjecture that in five years 10,000x more AI inference will be done worldwide, with these gains hypothetically coming from multiplicative progress of 50x in AI algorithms, 50x in system/hardware optimization/specialization, and 4x from further data center growth.”
Do not design systems assuming privilege escalation is hard. It never was. Anything local can become root. Every OS has had trivial privesc bugs, and any serious attacker keeps a few. Treat user separation as hygiene; not security. Disposable instances, minimal persistence.
Well, we have speculative execution and speculative decoding. Why not add speculative network responses to the mix to hide network latency?
Main downside seems to be that there’s a lot of stack to unwind if you guess wrong.
Inference Chips for Agent Workflows
@sdianahu
Most AI chips are designed for "prompt in, response out." Agents don't work that way. They loop, branch, and hold context across dozens of steps, and current GPUs hit 30–40% utilization as a result.
That gap is where purpose-built silicon wins.
@BucknSF There’s plenty of room to consume more tokens by exploring more branches. Snapshot, try a bunch of things, pick the best one. Budget is the only true constraint.
@shakoistsLog@mean_field_zane I think it's pointing at something real, the distinction between what looks like a grind to others and and what feels like a grind to you. More nuanced take: https://t.co/M2UZutmpit
A bit over a decade ago, we got fuzzers. A fuzzer is an automated vulnerability-finder that repeatedly runs a target program with semi-random inputs. One particular fuzzer, American Fuzzy Lop, was notable for being really good at searching the space of all possible branches in code in order to find the buggy ones. @BenLaurie found some security bugs in my own Cap'n Proto using AFL -- the first vulnerabilities reported in my code. And honestly, I thought that was really cool.
Today projects like Chromium and V8 have extensive fuzzing infrastructure that find tons of bugs. Most V8 security bugs are found by their own fuzzing, often before the bug is even released. And, you know, that's pretty great!
If you point a fuzzer at a project that hasn't previously been fuzzed, you will probably find a bunch of security bugs. It's not that hard.
And of course, bad guys can use fuzzers too.
But all the interesting targets have already been fuzzed. So. It's not really that useful to bad guys. On the contrary, fuzzing likely made it a lot harder for bad guys to find vulns.
My take on the latest developments in Iran:
The MEUs are a backup option, for now. Plan A is still to make a deal. But there’s no deal while the IRGC has anything resembling a unified command. Watch the target list.
Good essay. One question it raises for me: why are they spending so much money on data centers if the fundamental algorithms for continual learning aren’t yet understood?
3 years ago I bought ~15 T Shirt brands to see which was best. I now feel I can select a winner, and loser, based on Fit, Wear and Price
Surprisingly to me, Vince Slub Cotton T Shirt won, easily. Not at $88 list - wait for $45 on sale
The worst was Buck Mason. By a lot