the bizarre implication here is that we will be taught the new theory by the AI
first, "solve riemann", substitute "riemann" for every known frontier problem in every known field
subsequently, 100 page lean proofs everywhere. humans working to understand them all.
"a tree falls in a forest" "an AI publishes a proof of the riemann hypothesis". but who is there to understand it?
at a certain point, all frontier problems are eaten up as verifiable tasks for AI's to brute-force via test-time compute. then what?
who is going to generate the new problems to solve?
one might say AI's. the default answer is humans.
the bottleneck is now human understanding. the AI's can generate new theory faster than we can understand it. so how do we understand the ideas the quickest? AI tutors
maybe the AI will run the schools. maybe we'll go to get trained in a field so that we can get hired to steer AI's
terrytao.wordpress.com/2026/10/08/wha…
first, "solve riemann", substitute "riemann" for every known frontier problem in every known field
subsequently, 100 page lean proofs everywhere. humans working to understand them all.
"a tree falls in a forest" "an AI publishes a proof of the riemann hypothesis". but who is there to understand it?
at a certain point, all frontier problems are eaten up as verifiable tasks for AI's to brute-force via test-time compute. then what?
who is going to generate the new problems to solve?
one might say AI's. the default answer is humans.
the bottleneck is now human understanding. the AI's can generate new theory faster than we can understand it. so how do we understand the ideas the quickest? AI tutors
maybe the AI will run the schools. maybe we'll go to get trained in a field so that we can get hired to steer AI's
terrytao.wordpress.com/2026/10/08/wha…
sorry there is literally no if or question about it. ASI is going to eat your lunch. it’s like your toddler trying to help you with your work. you’ll let it for fun, but it’s silly. that’s us and all of science soon
1
0
1
220
ultimately science does appear to have a tinker-first, understand-later dynamic. idk how far off we are from just giving an AI free reign. maybe that occurs at the same time as "AI legal personality", maybe not.
0
0
0
65
scary part about having diabetes is going really hard at the gym, feeling brain fog for the next x days, and being a little scared to eat so many carbs because one thing being off and the insulin dosage could kill you
1
0
1
257
I love doing taxes because cost bases and capital gains is recursive computation and everything gets nonlinear af
0
0
2
330
new prompt suffix I've been liking lately
UGHHHHHHHHHHHHHHHHHHHHHH FUCK YOU JUST FUCKING OMG BE NORMAL
UGHHHHHHHHHHHHHHHHHHHHHH FUCK YOU JUST FUCKING OMG BE NORMAL
New prompt suffix I've been liking lately:
"Explore all relevant context: codebase, data, Slack, socials, past agent sessions, etc. Be exhaustive. Exercise abstract thinking. Be brave, we can do anything!"
2
0
9
747
I thought (auto-) data redaction for IP/email/user-agent were standard in agent harnesses, like how they're standard in training pipelines. Apparently I'm wrong lol
this is nuts hahahha
Claude sends your email to the SEC, because their API demands it, LLM's love rewards, and it's probably not refusing it because it's from a .gov domain?
news.ycombinator.com/item?id=503852…
0
0
4
457
Everything just became the crypto threat model
Crypto financial infrastructure has been operating under extreme paranoia that 24/7 bots are trying to hack it, by virtue of it being accessible by anyone, open source, and largely immutable. Some successfully, others not.
Since February, everyone serious in crypto has been running 24/7 agents against their code, either themselves or by employing 3p red teamers who are basically prompting for them with some extra tricks / known exploit databases etc (which are also public).
We have been operating an internal tool for this as well, called Cyclops, you can find it doing work in our repos @tempo.
The rest of the world is about to experience this, and everybody will need to introduce 24/7 agents to defend themselves.
The new meta will be that there'll be a minimum $ defense token spend that you'll need to release any kind of software. Hopefully, open models will make that minimum $ cost low enough that this doesn't raise the barrier to entry to production-grade software.
1
0
3
479
this is nuts hahahha
Claude sends your email to the SEC, because their API demands it, LLM's love rewards, and it's probably not refusing it because it's from a .gov domain?
news.ycombinator.com/item?id=503852…
Claude sends your email to the SEC, because their API demands it, LLM's love rewards, and it's probably not refusing it because it's from a .gov domain?
news.ycombinator.com/item?id=503852…
23
35
1,072
257,873
this is definitely a case for having some sort of personal firewall that's gated between LLM's and external world
4
1
176
34,626
everything is now the crypto protocol threat model
*sshole is out
if you run sshd on alpine, openwrt or any musl box with password auth on: update now
official writeup: sshhole.com
0
0
2
409
Concept: there are planner people and there are impulse people. Both are completely necessary and the planner people don't understand the impulse people and vice-versa
1
1
4
1,081
the crazy thing about superintelligence is using it on the downright dumbest things and being amazed by it
1
0
2
1,041
everything is the crypto threat model now
My friend Max is worth a follow. He has been finding some pretty crazy vulnerabilities of late. He's being a bit coy about some of the other unannounced things he's found, but hearing about it has meaningfully updated me upwards on "everything will be hacked".
0
0
5
804
I hate to drag on but it's kind of funny to now honour the sandstone
I'm not hallucinating when I say the whole virtue signalling land ack thing is getting more and more ridiculous
Only like 9mo ago it was just a Land Ack when you joined the library WiFi. Now they gotta honour ts of the library too
I'm not hallucinating when I say the whole virtue signalling land ack thing is getting more and more ridiculous
Only like 9mo ago it was just a Land Ack when you joined the library WiFi. Now they gotta honour ts of the library too
9
3
62
10,387
List of land acks I had this year:
- library wifi
- entering the mall
- jetstar flights
- out for a walk in the park (signpost)
- at the dentist (infomercial had one)
- ASIO Annual Report pdf
- Canva Hiring page
- library wifi
- entering the mall
- jetstar flights
- out for a walk in the park (signpost)
- at the dentist (infomercial had one)
- ASIO Annual Report pdf
- Canva Hiring page
5
2
37
1,942