mirekrusin

7 ms·

91.
by mirekrusin 4mo ago
Mythos and other models are not brute-forcing passwords (and with this analogy passwords, ie. systems are the same). We're not talking about dogs, but LLM systems. Mythos is not exploring entire solution space either. Usually looping i
92.
by mirekrusin 4mo ago
If mythos can break into almost all of their classified systems in hours then other models including opus, gpt, gemini and large open weight models can do so as well, maybe you'll have to double hours or it may become days, but they al
93.
by mirekrusin 4mo ago
This suggestion fails for values that can be null, need to be mutable or need references from multiple places etc – it's not "just performance penalty". Go has a problem, "just remember to always do X, never Y" patt
94.
by mirekrusin 4mo ago
Did they rewrote WebKit in Rust or not yet?
95.
by mirekrusin 4mo ago
That reply was AI written.
96.
by mirekrusin 4mo ago
Did I get it wrong or the first link has dataset with 30 entries only?
97.
by mirekrusin 4mo ago
Wikipedia is a lot of data as well but we manage to do it, no?
98.
by mirekrusin 4mo ago
I don’t understand why there isn’t public dataset for reasoning that can be improved by humans/llms like Wikipedia (ie with auto judging contributions etc).
99.
by mirekrusin 4mo ago
So the goal is to evaporate it with minimum number of shots?
100.
by mirekrusin 4mo ago
It's huge force multiplier for people who have hands on, detailed technical knowledge as well. 3 * N < 42 * N 42 - 3 < N * (42 - 3) It helps to know layer you're working on. People seem to make mistake of thinking how good L
101.
by mirekrusin 4mo ago
For comparison couple of months ago I acquired zed, pi, opencode etc. for $0.
102.
by mirekrusin 4mo ago
„Coders being obsolete” is not a measure of AI capabilities. I see coders being more busy than ever before. I see people without coding knowledge getting more behind. The gap is widening, not shrinking.
103.
by mirekrusin 4mo ago
You’re adding extra gpu for more vram, not speed.
104.
Ask HN: Year of Linux Desktop is fun with LLMs
6 points by mirekrusin 4mo ago|3 comments
105.
by mirekrusin 4mo ago
on 2x 4090: 90 t/s for 27B Q8 256k context 260 t/s for 35B-A3B Q8 256k context
106.
by mirekrusin 4mo ago
There were bugs at the beginning (imho worst ones where it kind of works but sucks), you should re-try with latest llama.cpp/quants/whatever you're using. Stuff like repeated nonsense, endless ???????? output, bogus code, loo
107.
by mirekrusin 4mo ago
You can definitely go above 30B on consumer hardware – 2x gpus, spark, mac, half byte quants etc.
108.
by mirekrusin 4mo ago
Suggest to him to automate what he's doing.
109.
by mirekrusin 4mo ago
They should use siri to code it quickly, no?
110.
by mirekrusin 4mo ago
"Some" of the cost? More like 120%-200% recovery during shortage and it's still going to be an asset after that period.
111.
by mirekrusin 5mo ago
It works very well on dense models, imho great alternative to MoE. As verification is cheaper than generation it could be fundamental, first class primitive, maybe even to recurse on it, do live distillation during inference etc. MoE is mor
112.
by mirekrusin 5mo ago
If “speculative” approach works so well in different contexts why not make it first class and use everywhere, possibly recursively?
113.
by mirekrusin 5mo ago
But auto merge would have to decide the order when resolving and it can't do it without custom merge.
114.
by mirekrusin 5mo ago
Won't appending to .jsonl keep creating conflicts?
115.
by mirekrusin 5mo ago
Go public so everybody can benefit?
116.
by mirekrusin 5mo ago
Just like cloud is "cheaper" than colo/metal, right?
117.
by mirekrusin 5mo ago
It starts to be interesting when latency is better than average website.
118.
by mirekrusin 5mo ago
Julia sits somewhere between jit and aot, there is no interpreter part, if something is to be exectued, it's compiled to machine code. The point is that you can have dynamic language that executes natively. You can also compile whole p
119.
by mirekrusin 5mo ago
They're going to start seeing real bill on Monday from Microsoft/Github.
120.
by mirekrusin 5mo ago
Julia?

More ›