eurekin

6 ms·

1.
by eurekin 9d ago
Brain fog? This might be a legitimate symptom worth seeing a doctor for
2.
by eurekin 24d ago
Supposedly having over 125 (130-ish) can affect "normal functioning", it's very counterintuitive, but I'd recommend to just go through common psychology checks for that case.
3.
by eurekin 26d ago
Have you ever gotten an IQ test?
4.
by eurekin 1mo ago
Care to share any details? I'm about to check the 2.6b lfm on document editing.
5.
by eurekin 1mo ago
It does? I wonder, what it does to the gut biome then
6.
by eurekin 2mo ago
The needle benchmarks show, that models extended context works for the part, that can be explained as: "I can access/adress that part of the input". I have no idea, why in that context, the number of attention heads isn'
7.
by eurekin 2mo ago
I have a feeling this comment will make history
8.
by eurekin 2mo ago
It's a mcp, so connects quite easily to agents. With mcpo, I also connected it to open-webui (which has better support for OpenAI style tools/functions). Used it in claude code with that mcp plugin set-up too. Only ever used it fo
9.
by eurekin 2mo ago
With 3.6 27b, I just stopped changing local models and started tinkering with things on top (like mem0). Feels genuinely useful and more than a toy
10.
by eurekin 2mo ago
Also explains why we like music with two simultaneous distinct sections (bass + the rest). One without the other doesn't feel as complete
11.
by eurekin 2mo ago
Disclaimer, I only use it to grow the "knowledge hub". It's a single git project at my $USER home, that is referenced in global memory. It contains as much information about work things, as possible, to be productive. I found
12.
by eurekin 2mo ago
That's where the annotation in plannotator helps. I'm asking to scan projects on gitlab, go through some docs to find more grounding material, write a subarticle (in the same style), scan logs on the test env, issue some curls, et
13.
by eurekin 2mo ago
Oh, I know exactly what you mean. I use plannotator with claude a lot and have much better time, since I asked for a specific styleguide. I used "CD era MSDN reference and Raymond Chen blogging style" as a starting prompt for the
14.
by eurekin 3mo ago
Similarly, I never try to imagine a bright sunshine. It can wake me up
15.
by eurekin 3mo ago
There never are. Those are going to be viewed as two discreet successful interventions. One for lay-offs, because it was the best move at the time with the knowledge they had. Second for quick correction, ability to pivot and execute quickl
16.
by eurekin 3mo ago
I remember seeing this popping up in discussions the first time, but never noticed any resolution (other than to train both sides). Has SOTA advanced?
17.
by eurekin 3mo ago
> A shift is happening among major AI labs, who are becoming increasingly skeptical of endless parameter count and training data scaling I'm pretty sure it's mostly due to the training data quality. No idea, why this never gets
18.
by eurekin 3mo ago
Noticed few cliffs. Sometimes it was a spurious stop (had to write "go on" or "continue" to restart), othertimes it was randomly saying: "Oh the user wants [the thing we already resolved]" and goes back in hist
19.
by eurekin 3mo ago
If I started today, with building a server, I'd jump right into verified set-ups and writeups, like this one: https://github.com/noonghunna/club-3090 You can find info about running a patched version of vllm for 1
20.
by eurekin 3mo ago
> The model is running so hot, that it shoots past the goal and starts looping later: > My latest experiment was setting up vLLM (the gold standard for production and concurrent serving) and even with an NVLink (175GBP) and tensor par
21.
by eurekin 3mo ago
I'm still sour they had only one toast in, in a two slot toaster
22.
by eurekin 3mo ago
Also, supposedly biggest MMA event Rogan doesn't want to be on. Something is up
23.
by eurekin 3mo ago
I kept getting recipes with "that one ingredient", which was either a major PITA to source or produced too much waste, even from a real world dietician consultation. Example, use 1/4th of a pumpkin for something. Those were g
24.
by eurekin 3mo ago
Chrome driven by the OS accessibility API
25.
by eurekin 3mo ago
I keep finding more and more usecases for Q3.6 27b (same league) and the best performance is, when answers to my question is already in the context. The moment I'm trying something open-ended or ambitious, Claude/ChatGPT clearly t
26.
by eurekin 3mo ago
Yeah, it won't. SWE here with a sidegoal to tackle the deployment side through various means (homelabbing, grabbing sre/cloud/observability tasks at work). The biggest observable improvement in my post and pre ai development
27.
by eurekin 4mo ago
It feels now like an alternative timeline, one which performance optimisations were first and foremost still. Sometimes I fantasize, thinking how would our current development ecosystem look like, if we never abandoned the "be very vig
28.
by eurekin 4mo ago
Batching lowers that, since the model is read once from memory. Activation accumulation doesn't scale as nicely
29.
by eurekin 5mo ago
On my homelab. It really feels like a dream come true for my usecase. No more puppet agents. No more declarative syntax, that you have to work around to do basic imperative ways. Or use a module, that stopped being maintained 3 years ago. J
30.
by eurekin 5mo ago
Yes, my goes either-or quite abruptly and, while that's not really annoying, I notice it doing way too often, than I'd like to. It shouldn't be done in a binary fashion as well.

More ›