maxignol

5 ms·

1.
by maxignol 27d ago
Author here. Two things that are worth repeating up front: corpus A is 84% of the pooled bill, so the pooled −0.01% is pretty much "corpus A plus noise" so read it per-rows instead. And p_fire is a lower bound because of the detec
2.
Caveman prompting saves tokens, until you run it in real sessions(jayn.app)
3 points by maxignol 27d ago|1 comments
3.
by maxignol 1mo ago
> I don't know if engaging with that button means I'm less likely to be served such content, or whether it's cleaning up everyone's feeds. I would guess it's both cause I've been having less AI slop on Linke
4.
by maxignol 1mo ago
Been laughing for the past 10 minutes, thanks for this one. (I tried the death of my cat by car in front of my house in a Go Girl manner, absolute cinema)
5.
by maxignol 1mo ago
Optimizing speed is really the way to go. Yet 24GB is not what everyone can afford. Maybe we could take some of those 56tk/s and transfer into some free RAM space using MoE loading ? I'd be glad with a less than 10GB and more than
6.
by maxignol 2mo ago
Is the model picked through the router only for the first user turn or is there multi-turn routing (or planned to be added) ?
7.
by maxignol 2mo ago
I'm really excited about what's been happening couple last weeks for local inference. I feel like it all started after colibri [1] was released. Great work ! Anyone got recommendation about what local model to use for what purpose
8.
by maxignol 2mo ago
I dont quite understand why GGUF is better optimized. Are the performances better for the same amount of VRAM ?
9.
by maxignol 2mo ago
> I’d suspect the harness to massively affect token use and optimisation Yes they do according to databricks -> https://www.databricks.com/blog/benchmarking-coding-agents-d... That's why I find comparing mod
10.
by maxignol 2mo ago
What is your harness with every model ?
11.
by maxignol 2mo ago
> my use cases stop aligning to swebench pro around 50% accuracy, and more closely align with DeepSWE. What do you mean by that ? If the model is higher than 50% on swebench pro then it tends to drift from what you like it to do, like De
12.
by maxignol 2mo ago
Shouldn’t we fear they start doing only close source like most us labs once they catch up in market shares ?
13.
by maxignol 2mo ago
Well, I had not heard of RLM before, just read the paper, thank you for introducing me to your lazy version !
14.
by maxignol 2mo ago
I’ve never seen the Id approach before, that’s a good idea ! Though I was wondering how do you manage to keep costs low within the 7 agents ?
15.
by maxignol 2mo ago
Giving the accuracy-token graph and not the accuracy-cost graph, thus we cannot easily compare costs with other models, is not a way to gain my trust
16.
by maxignol 2mo ago
I’m truly impressed by your work ! I don’t know if this is planned for near future, but how about adding energy efficiency benchmarks ? Because running locally is a great feeling, but the electricity bill should not be forgotten
17.
by maxignol 2mo ago
Well, had a meeting with a VC that literally brought his note taking 3rd person to the meeting, did not even say hello. Tbh, I still find it creepier than an AI transcriber.
18.
by maxignol 3mo ago
This seems really bad…
19.
by maxignol 3mo ago
I guess the technology used here must be ground-breaking lol
20.
by maxignol 3mo ago
Did not seem to find how much tokens per second he achieved with this setup ?
21.
by maxignol 3mo ago
Thus the prices in switzerland are higher than anywhere else. I don’t know about you, but I’d have no use of 25gbits/s
22.
by maxignol 3mo ago
Lol next time I’ll just apply with 4 accounts and maybe get in once.
23.
by maxignol 3mo ago
Are many people using HackerRank ATS ?
24.
by maxignol 3mo ago
Have you tried opencode go ?
25.
by maxignol 3mo ago
Would you recommand some ressources about how multiple neural engines are used in data centers ?
26.
by maxignol 3mo ago
I guess it was inevitable. Is it only RAM related ?
27.
by maxignol 3mo ago
Funny one x) Though I ain’t sure if even more data is useful on hackernews
28.
by maxignol 3mo ago
Great way to honor Tony and his work
29.
by maxignol 3mo ago
In the end, sorting prs and vulnerabilities has been the same for open source maintainers. How about adding a credibility score to every github account ? Couldn’t that cut sorting times ?
30.
by maxignol 3mo ago
3B param on par with opus 4.5 sounds interesting. Will read the full article before making my mind

More ›