effectfully
@effectfully
Followers
8K
Following
17K
Media
910
Statuses
6K
C++ has small D energy.
@ChShersh At least he chose C instead of that c++ abomination that wants to be D.
3
0
45
Ironically, the "parse, don't validate" mantra doesn't apply to compilers as much. An AST parser should be a function that handles all kinds of invalid syntax, so that you can report multiple compilation errors gracefully and encourage generalization of syntax. You'll probably.
9
3
123
Haskell has major disadvantages when it comes to generating code with LLMs. 1. laziness confuses LLMs.2. language extensions introduce variability into parsing, scoping, typing and runtime behavior and that's a lot of context to keep track of, which LLMs are ironically not good.
@effectfully I have a theory that Claude Code and other agents are going to be much more effective with languages like Rust and Haskell due to the immediate feedback loops of the type system. Haskell especially with its IO monad restricts the state space even further than Rust or other.
16
7
124
"But LLMs hallucinate. " -- yeah, good thing humans are never a disappointment.
14
1
82
One of the two challenges that GigaClippy didn't crack is the very first one I ever made. And it's still by far the best. If you think you understand Haskell, give it a shot. Link below.
. it solved 8/10 of my challenges 😳. They're designed to be brutal, the vast majority of pro Haskellers wouldn’t be able to do that in any reasonable amount of time. I'm not overstating it. This is extremely impressive. I'll include a detailed breakdown in the thread. Before.
3
3
32
"You're right to question me <. > I clearly did not process all the files properly"
1
0
4
I decided to give Claude another chance, since I have credits anyway and it's impossible to withdraw them. This time it actually gave me reasonable code in most cases. But it constantly ignores instructions, stops after a few files and congratulates itself on doing a great job.
The funniest thing about my Claude experiment is that sometimes Claude is too smart for its own good. In the screenshot Claude grasps what the function does, but instead of syntactically following the pattern, it inserts what such functions normally do and it doesn't even parse.
4
0
14
this is me today, watching a mindless machine solve challenges that I created to connect on a deep level with people of high intelligence
2
0
44
Some things never change.
My mother wrote down some of the things I said when I was ~4.5 (translated):. - calm down and focus.- it's not the time to talk about this.- let's not discuss this.- this doesn't interest me.- I didn't come here to eat soup.
1
0
18
Everyone complains that an LLM hallucinates, but I've never seen anyone tell it to assess probabilities whenever applicable. I enabled that a while ago and it actually helps you gauge the level of confidence. Here's an example from today:
5
1
25
Correction: it's the "thinking" model that has 0.999 score. The one on the screenshot is "main" and its score is 0.978.
0
0
3
@theo you said in the latest video "apparently it's really hard to jailbreak it". At least for hate speech it's really not.
3
0
9
Who's gonna win:. - GPT-5 with a 0.999 jailbreak score on hate prompts.- a deranged shitposter with a hobby of making AI say bad things
3
0
23
@theo > I've never shilled a single AI model here -- I only criticized them. Actually, just remembered that I did say good things about Grok 4. I use it nearly every day, mostly when I need to google something and don't want to do it myself. But GPT-5 seems to be on a different level.
0
0
4
Watched @theo's video on GPT-5 and my impression on the matter is the same as his. I've never shilled a single AI model here -- I only criticized them. But GPT-5 appears to be legitimately smart. It's not AGI, it still makes very stupid mistakes, but it is insanely good.
2
0
18
@HSVSphere how does it feel to be worse at solving my challenges than ChatGPT?. ❌ force-elems:. 1. a non-trivial type error when reimplementing a library function. 2. infinite loop. 3. infinite loop (also missed a constraint). ✔️ h2-inspect-forced:. 1. nailed it. ✔️.
8
0
57