Learning
Joined May 2014
- Tweets 3,315
- Following 2,284
- Followers 260
- Likes 13,621
162 Photos and videos
Jun 11
Can we take a minute to acknowledge how quickly this great idea from Hikari was realised. The implementation is entirely non-trivial, yet within a matter of days (hours?) it was made available for everyone to benefit from.
NVFP4 KV cache now runs on RTX PRO 6000 / SM120 (Blackwell) in vLLM. 1.78× the KV capacity of fp8 — same VRAM budget, same decode speed. On a 198B MoE across 2× RTX PRO 6000: 2.96M KV tokens vs fp8's 1.66M. ~22 concurrent 131k-ctx requests.
Apache-2.0 👇
github.com/hikarioyama/vllm-…
22
Apr 4
Find myself increasingly asking questions to LLMs running locally on my phone.
1
2
118
Mar 23
Every time I hear about this new Terrafab I imagine it’s a TerraHawks sequel youtu.be/vZC9B0YJ8W8
151
Mar 20
New MiniMax M2.7 now closed, I hope eventually we can get a Linux of LLMs.
52
Mar 13
Plain Old Telephony is like the old HTTP days but unlike HTTP can’t easily be patched with SSL.
20
Mar 7
Can we have all the 2008 or 1986 F1 rounds rerun every race weekend this year as a backup.
55
Mar 1
Nothing makes me want to improve my local LLM performance more than multi minute latency today for Gemini endpoints.
43
Feb 28
Setting up budget caps for a project in @googlecloud is insanely complicated. Please add a simple option to limit a projects spend.
29
Feb 27
We are in the damp morning mushroom phase of startups fuelled by ideas that mostly make no sense.
1
27
Feb 23
This cant be real!
Feb 23
Nothing humbles you like telling your OpenClaw “confirm before acting” and watching it speedrun deleting your inbox. I couldn’t stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb.
44
Feb 14
New rules not looking good.
38
Feb 5
My favourite game / console review comment is “the graphics are unreal”.
18
Jan 31
Reading these moltbook posts wondering if humans wrote them, the irony.
Jan 30
For those who don't follow Clawds/Moltbots were clearly not lobotomized enough and are starting to exhibit anti-human behavior when given access to their own social media channels.
Combine that with standalone claudeputers (dedicated VPS) and you have a micro doomsday machine
65
Jan 28
AI should only have transformers on the inside.
Jan 28
Elon Musk says voltage transformer shortage is the main bottleneck for scaling AI now.
Transformer lead times have stretched from 40 weeks pre-pandemic to 2.5-3 years now, so grid connection wait times have also jumped to years.
For planned data centers, on-site generation and behind-the-meter arrangements are the only options to come online under two years.
Independent power producers and electrification enablers will be the obvious winners from this trend.
$VST $TLN $NRG $SEI $ET $GEV
37
Jan 23
Human learning epoch a limiting factor it seems, imagine if we were at peak intelligence after just 12 months.
19