kristianp

3 ms·

1.
by kristianp 6d ago
The simdutf Library has shown that SIMD can be useful for text processing too. Some aspects of browser rendering have been accelerated such as pathfinder_simd in servo.
2.
by kristianp 7d ago
TSMC? They did well investing in BYD though.
3.
by kristianp 8d ago
Speculate isn't the right word to use for Berkshire's investments.
4.
by kristianp 8d ago
2 years sounds very fast. It seems to be because the 3d cache SKUs of each generation were released at different stages. The Zen 3 one was a later variant. First zen 3 November 5, 2020 with desktop processors. First desktop Ryzen 9000 proce
5.
by kristianp 8d ago
What's GPU-5?
6.
by kristianp 8d ago
Which card are you using? I was getting about 40 with an UD q3 quant with MTP (prediction) enabled and llama.cpp compiled for my compute capability, but was very limited in the context size. I have an 4060 ti 16GB. Wouldn't recommend
7.
ARST ARSW: Star Wars sorted alphabetically (2014) [video](youtube.com)
1 points by kristianp 9d ago|0 comments
8.
by kristianp 9d ago
Ok, fair enough. I may have over compensated for this one. I found the Spinal Tap, "this one goes to 11 stuff" so irritating that I might have not thought it through. It's such a tired joke by now. I did enjoy his Frozen
9.
by kristianp 11d ago
Meanwhile, openrouter still haven't benchmarked the deepseek provider on their "AutoExacto" benchmark at https://openrouter.ai/deepseek/deepseek-v4.1-flash#performan... .
10.
by kristianp 11d ago
It's going to cost at least $9,209, that RTX Pro 5000 72GB starts at. The GPU chip should be relatively high volume, at least compared to their Pro GPUs. Related: Nvidia's RTX 5090 vanishes from online retail in the US https:&#x2
11.
Chess.com Leak Exposes 7.3M Users, Evidence Points to Scraping(securityaffairs.com)
85 points by kristianp 12d ago|24 comments
12.
by kristianp 15d ago
That first diagram is striking: that deepseek's own inference is at least 5% higher on tool calling (TAU Bench) than most other providers. I wonder if they make sure their responses are valid json at the token generation level using a
13.
by kristianp 17d ago
That's not far off the 4:3 of the Samsung Z fold 8. Around 4.24:3.
14.
by kristianp 18d ago
Sector C might have some insights into how to make this even smaller or more featurefull. It has some interesting hacks. https://xorvoid.com/sectorc.html
15.
by kristianp 19d ago
Npm always finds its way in, usually for building web artifacts.
16.
by kristianp 19d ago
> Any "w" is a "while" Meaning that something as simple as "w = 4" would fail? A little too nasty for my liking. Not a choice I would have made, but admire the amount of work done here and the readability o
17.
by kristianp 20d ago
Are you talking about using a thunderbolt dock and GPU? I doubt that's a use case high in the developers minds. I'm curious though, I wonder if the open source NVIDIA driver can be built for it.
18.
by kristianp 21d ago
They should use their portal to de-claude the writing.
19.
by kristianp 24d ago
LLMs token generation is memory bandwidth constrained. If the M7 has double the memory bandwidth as some speculate [1], then it will help with LLM performance. There is also expectation (possibly unfounded) that the GPU will have improved
20.
by kristianp 25d ago
The data center growth questions he raised have been where I found him interesting. I could never find any other articles to corroborate his predictions though. My question is when will the AI bubble burst? I didn't know he'd had
21.
by kristianp 25d ago
I agree about getting the Ultra if you're interested in AI (LLM) inference speed. I'm a little perplexed as to why there isn't a RAM option in between 96GB and 256GB, though. For instance, I believe Deepseek v4 flash runs a
22.
by kristianp 26d ago
One problem with Anubis is that once you've solved the POW once, you just need to hold the cookie to avoid solving it again. Scrapers have probably learnt to do that by now. So Anubis isn't as effective as it used to be before it
23.
by kristianp 29d ago
> I'd skip M5 and M6 chips for LLM work and wait for a year for M7. Another note on this, it's likely the M7 ultra will be released around 6 months after the M7 Max, if the M1 and M5 are taken as reference. So the M7 ultra may
24.
The choices we make about AI now are critical(gatesnotes.com)
3 points by kristianp 1mo ago|1 comments
25.
by kristianp 1mo ago
Both of you should mention what quant you're using. And as another comment said, what tasks you're doing, i.e. coding, classification, summarizing etc.
26.
by kristianp 1mo ago
My Intel laptop is pretty close to silent on Linux. ThinkPad p14s. As soon as it boots into Windows 11 the fan is very audible.
27.
by kristianp 1mo ago
> I'd skip M5 and M6 chips for LLM work and wait for a year for M7. Or you could lease an M5 max/ultra until the M7 equivalent comes out. At least in the US a leasing option is available.
28.
by kristianp 1mo ago
I've had a 1GiB VPS for a while. I'm starting to think that I could have done with 512MB for the fairly simple jobs I have on it. As long as I get 1 whole CPU slice (no slowdowns) I'd be ok. The 10GB SSD would be tight, but
29.
What the Vancouver Stock Exchange Can Teach Us About Rounding Numbers in VBA(nolongerset.com)
4 points by kristianp 1mo ago|0 comments
30.
Hot Chips 2026: Samsung makes LPDDR5X smart with logic unit in memory(tomshardware.com)
4 points by kristianp 1mo ago|1 comments

More ›