@tritlo

uncertified haskell engineer | certified researcher

 Gothenburg, Sweden
mpg.is
Joined January 2010
  • Tweets28.1K
  • Following2.1K
  • Followers5.5K
2.4KMedia
Pinned Tweet
Matti Palli retweeted
Matti Palli
@tritlo7 Mar 2021
Friends don't let friends use dynamic types – at Iceland
22
312
13
1,891
wouldn’t it be quicker to inject the microplastics directly?
Dave F.
@DaveF937Oct 25
I gotta hand it to those hippies... They're dedicated, lol.
6
476
sometimes one of my neurons gets stuck like this, and i ramble on about [current interest] thankfully i can update my weights on the fly
Sauers
@Sauers_16h
I found a neuron that makes the model green.
1
8
413
I see Hylang* has a new competitor *the language that combines the readability of LISP with the speed of Python
Dominik Tornow
@DominikTornowOct 25
Postgres on S3?! No, S3 on Postgres. Introducing pgS3. The scalability of a Postgres box. The query power of S3.
2
1
1
26
1,665
reminds me of this excellent project, a webframework for Cobol github.com/azac/cobol-on-whe…

GitHub - azac/cobol-on-wheelchair: Micro web-framework for COBOL

Micro web-framework for COBOL. Contribute to azac/cobol-on-wheelchair development by creating an account on GitHub.

github.com
5
188
sadly they will not be prioritized. we must focus on shrimp welfare if we are to be effective
roon
@tszzlOct 25
it will take hundreds of years for the blue whale population to return to pre-industrial levels
1
13
573
we’re going to get a lot of “good idea but tedious to implement” papers in the next few years thanks to agents finally getting good
Kiran@kirancodesOct 25
It's the summer of rebooting dead PL ideas babyyyyy
Eight ideas went silent for 11 to 27 years (each comeback cites the earlier work)

ALT Eight ideas went silent for 11 to 27 years (each comeback cites the earlier work)

1
6
1
48
1,543
the nice thing about improving performance is that every fix makes the next iteration faster!
2
15
587
agents are also great for setting up some spikes, running them, acting on minor details and i don’t have to babysit a 4 hour run!
187
> “Types are a bad way to formalize mathematics” > Dependent types (Lean) currently being used to push the frontier of math make it make sense??
Kiran@kirancodesOct 24
> be Leslie Lamport > literally wrote THE book on Paxos > specifications shouldn't be math > specifications should be text > "Claude, make no mistakes" Preempted LLMs by 30 decades. That's why he's the GOAT.
9
3
2
113
9,029
“but we can do it in natural language this time!” i’ve been using natural language to communicate with colleagues for years, it doesn’t solve the problem. you have to learn how to express yourself. changing the method of communication isn’t a magic bullet
Matti Palli
@tritloOct 23
1950s: you write programs on punch cards. doesn't scale. 1960s: you write programs in assembly. doesn't scale. 1970s: You write programs in C. doesn't scale. 1990s: You write programs in Python. doesn't scale. 2026: You write programs in a spec language. doesn't scale.
1
2
41
1,296
Ok wow I can run v4.1 on my 4070 super via strata/maya, offload into ram and get 8 t/s it's not much but let it run overnight and you can get a lot of work done "for free" exciting times for local models
Gabu
@gabu3d_plOct 23
😍🥰🥳 huggingface.co/ISTA-DASLab/D…
3
10
928
More details: xitter.cf/BUTA_no_rakuen/status/…
どすを (DOSUO) | 📈PCパーツ価格動向📊 / AI・LLM
@BUTA_no_rakuenOct 23
来ましたよ！間に合わなくて、今日はないのかと思ってた💦VRAM 12GBで動く模様！ 🔥 763BモデルがVRAM 12GBで動く時代が来た ISTA-DASLabのDeepSeek-V4.1-Flash-1017 GSQ-RCO量子化版が公開されました。 📦 容量感 （全部で230GB ~ 265GB) ・Engram：約110GBまで圧縮 ・バックボーンは120GB ~ 155GB 💻 ローカル動作目安（Strata使用時） ・VRAM 12GB + RAM 64GB → 動作可能 ・VRAM 12GB + RAM 128GB → 快適ゾーン ✅ 品質はAIME25 / GPQA / LiveCodeBenchなどでベースモデルにほぼ匹敵。 フロンティア級をコンシューマ環境で実用的に動かせるラインまで来てます。 📊GSQ-RCOって何？ 既存のGGUF系とは以下の点で異なる。 ・動的ミックス精度（重要度に応じてビット数を変える） ・RCO（Routing-aware Compression）でMoEのエキスパートを効率的に圧縮 → 通常の量子化より低ビットでも精度落ちが少ないのが特徴。 かなり圧縮されてるとはいえ、763Bのモデルなので、RAM 64GB勢の実測がどのくらいになるかで、話題の爆発度が変わりそうだ！ ※実測は使ってみないとわからないので、参考程度に！ #地元のLLM #DeepSeek
331
1950s: you write programs on punch cards. doesn't scale. 1960s: you write programs in assembly. doesn't scale. 1970s: You write programs in C. doesn't scale. 1990s: You write programs in Python. doesn't scale. 2026: You write programs in a spec language. doesn't scale.
Matti Palli
@tritloOct 23
as happy as I am with the explosion of formal verification, we’re immediately running into the classic problem: writing down what the program is supposed to do proving your program “correct” only works if you know exactly what “correct” means!!
4
4
2
130
7,357
at some point we have to realize that the problem is not the language, it's the fact that we have no idea what we want
4
2
24
1,081
as happy as I am with the explosion of formal verification, we’re immediately running into the classic problem: writing down what the program is supposed to do proving your program “correct” only works if you know exactly what “correct” means!!
Shriram Krishnamurthi (primary: Bluesky)@ShriramKMurthiOct 23
Replying to @rajashree
Up to the quality of the specs. Not the same thing.
13
17
2
205
14,768
formal verification is like number theory super useful once you know how to apply it
Talia Ringer 🕊🪬@TaliaRingerOct 23
Where we started (left, 1979) vs. where we are now (right, 2026)
Social Processes paper arguing formal proof will never be important to CS and softwate engineering

ALT Social Processes paper arguing formal proof will never be important to CS and softwate engineering

Conclusion from the Anthropic post from yesterday arguing verification is about to become as practical as fuzzing

ALT Conclusion from the Anthropic post from yesterday arguing verification is about to become as practical as fuzzing

1
43
1,305
agentic world problem: I want to play a game, but my system is fully saturated running spikes, benchmarks and local models such is the local life
2
7
792
realization just hit: the next gen models will be RLed on dependent types and formal verification specifically exciting times ahead !
Matti Palli
@tritloOct 23
it’s incredibly gratifying to see dependent types finally escaping the bubble and having their moment having the formal specification and implementation go hand in hand was always the way to go, and now LLMs have unlocked that for the non-experts
6
6
1
86
2,118
update: there’s a woman who beat her parents for years and ended up murdering her dad. story: she’s *uncomfortable* in prison! yeah that’s the whole point lady
Matti Palli
@tritloOct 21
In Iceland the current debate is if it is morally permissible to give *a child* a 3.0 grade out of 10 in their schoolwork. This is seen as "cruel", it might make them feel bad (even if they didn't put in any work). The empathy has gone too far
5
559
“We choose to go to the moon in this decade and do the other things, not because they are easy, but because they are hard.” - JFK you can just do (hard) things!!
Talia Ringer 🕊🪬@TaliaRingerOct 23
You won't believe how many people told me over the past 10+ years that proof assistants would never achieve widespread use, and that my work was therefore not useful. Ignore people who discount entire fields of research because they are hard. All the more motivation for research
2
2
15
737
Load more