social media influencee
- Tweets 32,952
- Following 1,054
- Followers 12,956
- Likes 244,658
Jevons paradox for mathematicians would be the funniest outcome
More mathematicians needed
terrytao.wordpress.com/2026/…
23
995
Would have been nice to know in 2020 that you had 2 years to produce writing or code artifacts that prove you can do stuff, before everyone has a GPT to produce for them.
Maybe we are in a similar situation now with Pangram.
Last call to produce work Pangram can tell is human
6
2
100
2,652
It seems clear that Pangram doesn't hold up against arbitrarily high capabilities. Neural networks are universal function approximators, so they should be able to model the human distribution of text, given enough params and the right training setup. When it happens -- no idea
1
20
313
If you are tempted to give a probability range, instead provide a model!
e.g. "Either ASI is possible or not (50%). Also maybe alignment is easy (10%). Finally conditional doom chances: ASI+hard 90%, ASI+easy 40%, AGI only 10%"
Way more informative & still a valid distribution
I think the temptation to give probability ranges should just be really strictly avoided
The reason it feels good (not committing yourself to any one thing) is the reason it's bad (probabilities that don't commit you are noise/meaningless)
4
3
30
1,447
I think this illustrates how 95% of the AI debate is mistake theory. Most believers in unaligned superhuman agents think we shouldn't make them; most others don't think that is really going to happen. The believers in the labs think they have a shot at alignment.
Interesting back and forth. Ezra describes the HF incident, Jensen says "well they shouldn't release the product." Ezra says "this product wasn't released," and Jensen's response is that if they say they can't contain their experiments then "we have to shut the labs down"
2
31
1,160
TPOT was just the MOPs of the twitter postrat scene.
I was an early MOP and picked "prerat" as my handle bc I didn't know as much about LW stuff. (Well I have read the sequences but I felt like the postrats knew a lot more especially culturally about the rat scene.)
I think there are a lot of cool things about TPOT that are less common among the rats, but on the whole I think it's more true to say that TPOT is "pre-rat" or "not familiar with the sequences" than to say that they're "post-rat" or "moved on past the sequences".
6
3
125
9,978
Look we've tried Uranium 235 with k_eff of 0.1, 0.2, 0.3, and none of them have had a "recursive self-propagating neutron chain reaction".
If you were right, we would see some smaller explosions before 0.9.
Your prediction of explosion past the critical point is unfalsifiable!
18
39
461
10,541
The machines didn't create the Matrix for electricity.
They did it to benevolently maximize hedonistic utility.
Canonically their first attempt was a literal wirehead utopia sim; only changed bc humans kept rejecting it. So they did the minimum acceptable utopia (the 90s).
64
53
1,683
41,751
AI seems paradoxically both easier & harder than I expected
Easier bc we just keep scaling compute & data & "merely incremental" algo progress & it keeps getting smarter
Harder bc even now, even tho its in many ways superhuman at coding, theres no intelligence explosion (yet),
4
1
20
774
Why did the ozone thing work out so well? Was it just cheap to switch to other hairsprays or whatever so there's no pushback?
For AI, it seems like "don't do an intelligence explosion" is cheap to most ppl, esp. since most don't even think it's possible. Main concern is side fx
If you like not having skin cancer, you may appreciate knowing that today is the 39th anniversary of one of the most impressive feats of coordination humanity ever pulled off: signing the Montreal Protocol, to ban CFCs.
The treaty was signed just *18 months* after scientists raised the alarm about the hole in the ozone layer. It is the only treaty to have been ratified by every single member of the United Nations.
Humanity can do great things sometimes, when the stakes are high.
2
1
33
791