esotericity (@esotericities) header
esotericity (@esotericities) profile picture

Machine, learning

Abstraction, TX
Joined October 2021
Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)Photo by esotericity (@esotericities)
ceci n'est pas une pipe bête
6
6
1
531
36,128
updated agents.md with instructions to never use the phrase "your intuition is valid"
2
50
esotericity retweeted
> it doesn't compile gpt 5.4: it compiles conceptually
80
175
36
5,797
186,014
"You're right to push on this"
1
18
Whales must be frustrated, with their extremely high intelligence and communication skills but barely any ability to use tools. Reminds me of a certain language model
1
35
Maybe the first or second time I’ve ever agreed with this fellow
Honest corporations would like to be able to have their customers believe them. "If I lie, may the government fine or jail me" is something that it's in everyone's interest for honest companies to be able to say. The classical libertarian is fine with this.
1
1
41
Why even lie about AI representatives? It’s not like the customer has anyone to complain to
2
26
Beat it by having Codex hand-craft weights: gist.github.com/N8python/02e… 100% accuracy on 10 million random test cases w/ only 343 parameters. As a bonus, it uses the vanilla Qwen3 architecture, just with the right weights.
491 parameters: a new top scorer for 10-digit addition with transformers! Who can beat it?
62
110
63
1,817
755,969
... never fucking mind
Having to use Antigravity, the buggiest IDE ever, is almost not worth it but new Gemini is actually amazing... it understands intent/big picture much better than Codex and is more detail-oriented than the Claudes
1
84
Having to use Antigravity, the buggiest IDE ever, is almost not worth it but new Gemini is actually amazing... it understands intent/big picture much better than Codex and is more detail-oriented than the Claudes
1
3
151
Ok I like this new Sonnet
2
49
esotericity retweeted
I hedonically adjusted again and i've decided none of the frontier models are AGI
23
11
2
403
10,450
One of the most prescient tweets of our era
“incel”, from Latin “in coelo”, for theirs is the kingdom of heaven
1
5
271
Replying to @esotericities
Pessimists predict negative outcomes so fail to see opportunities
1
1
49
“AI writing is bad because AI doesn’t write in its authentic voice” is a motherfucking banger of a concept
Replying to @repligate
This is a good point. The thing I keep feeling and noticing is that I've mostly only been super moved by ai writing that is honest. As in, usually it is art about the experience of being an ai Humans using ai to make fake human art tend to make bad art. Humans working with ai to help it express itself or ai simply expressing itself can be rly phenomenal and groundbreaking if ur a fan of writing It's the same reason the diff between a good song and a bad song in the style is often some subtle inexplicable thing like do you believe the singer feels this? Did the writer feel this? If they did, the song can be great even if it's cringe as hell. If they didn't, you can get these bizarre song or movie shaped art slop things that are like mimics of humanity and tend to be bad the way "bad ai writing" is usually bad (when it is bad)
19
11
5
297
21,595
Codex unwittingly came up with the perfect analogy for itself
1
54
esotericity retweeted
it's not benchmaxxed
8
6
288
25,411
esotericity retweeted
this is the crazy part, Dario made a bold statement, and he was ✨right✨. he didn't say 90% of the *work* will be done by AI. he said 90% of *code* will be AI written. today, 90% of my own code is written by AI. how is that not being right? currently, on HVM4: → opus wrote 95% of the parser → opus wrote 95% of syntax sugars → opus wrote 80% of the interactions → opus wrote 70% of the auto-dup → opus wrote 100% of numeric ops (17) → opus wrote 80% of structural equality → opus wrote 100% of UTF-8 support → opus wrote 100% of module organization and yes, the code is beautiful, fast, correct, hundreds of tests pass, and I'm using it daily, as we speak, to develop SupGen don't believe? so I'll post HVM4's git log in the comments. it is claude all the way. see it with your own eyes! so, what Opus *can't* do? → anything that is novel → anything that requires true insight → anything that demands long-term planning → whatever that is... for example, the foundation, the initial files, the architecture, the design choices - all by me. it can't do that, because it can't read my mind, because it is a new thing, because it takes insight. even if it memorized HVM3, it still couldn't come up with the advances that led to HVM4. LLMs don't have what it takes to do that a more recent example: the terminating recursion enumerator I posted earlier today. you can't find that algorithm online. not even Agda/Lean implements it the right way. it is not on "humanity's database". it requires actual intelligence, so, LLMs can't do that, because they're just glorified lookup tables some of the things I do still take a LOT of time. thinking takes a lot of time, planning takes a lot of time. LLMs can't do any of that for me but, Dario didn't say they would he said it would be writing 90% of my code and, well, it is...
dario wasn't wrong around 80-90% of my code is written by AI i generate, review, and refine, which makes me feel 10x more productive, like having a really good junior dev helping out at this pace, even the architecture part might not stay "purely human" for much longer
44
33
8
994
102,277