Bitcoin Psychopath
Joined July 2025
- Tweets14.5K
- Following251
- Followers229
- Likes12.9K
Bitcopath retweeted
@pika_labs H3 Flex is on ComfyUI.
→ Faster than real-time: 15s 768p clip with audio in under 10s
→ #7 on Design Arena text-to-video
→ Up to 12 references per generation
→ Draft at 480p/768p, finish at up to 4K
→ Five separate audio stems on request
Click the link to try Pika H3 Flex today ⬇️
Bitcopath retweeted
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Bitcopath retweeted
Simply not true. I recently optimized a mid sized data center. Their performances went up by more than 600% (around 830% on ppts and 480% on throughput) on just few iterations. Data centers at their current stage are wasting the most DRAM in the world…
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Bitcopath retweeted
We're still discovering new possibilities hidden inside Einstein's equations, more than 100 years later! 👀
Physicists just developed a mathematical framework showing that gravitational fields can undergo a process called GRAVITATIONAL RECONNECTION.
By reformulating Einstein's century-old equations, they discovered that gravitational field structures, known as field sheets, can break their original connections and reconnect into new configurations.
Interestingly, this resembles MAGNETIC RECONNECTION in the Sun, where solar magnetic fields break and reconnect, releasing massive amounts of stored energy into surrounding plasma and powering powerful solar flares.
Similarly, gravitational reconnection provides a mechanism for transferring energy and momentum from gravitational fields into matter.
The discovery could open new ways to investigate black hole collisions, neutron star mergers and the extreme dynamics of spacetime.
So as I always tell the problem is never the hardware limitation but the software performance.
How much performance we did get with software solutions and by removing NVIDIA's limitors?
5x? 10x?
What if you could make your local LLM 3× faster without changing the LLM? 👀
Ant shipped UBTree drafters for Ling (3B & 122B.)
What is it? It's a tiny model that sits next to the big one and tries to predict its future tokens.
For Ling-3.0-Flash it ...
🧠 Target → 124B / 5.1B active
🌱 UBTree drafter → only 482M parameters
Instead of guessing one future token path...
Their released Ling-3.0-Flash checkpoint reports improvements of
🚀 3.43× average end-to-end speedup vs
DFlash → 2.96×
DSpark → 2.88×
Across the six published tests, UBTree averaged about
⚡ 1,010 output tps
And Ant's larger research configuration reports up to 5.36× average speedup on Ling3-Flash.
BIG caveat:
⚠️UBTree isn't in llama.cpp yet.
Right now the official implementation needs a special SGLang/CUDA branch.
... but someone already extracted just the DFlash portion of the Ling3-Tiny UBTree drafter and ran it in stock llama.cpp.
Their small test:
Python code
79 → 127 tps
That's +61% from the drafter alone.
Ant released the draft weights under MIT
Ling3-Tiny drafter → 183M
Ling3-Flash drafter → 482M
Links in ALT.
ALT https://www.inclusion-ai.org/ubtree/ https://huggingface.co/inclusionAI/Ling-3.0-flash-UBTree https://huggingface.co/inclusionAI/Ling-3.0-tiny-UBTree
Local inference is the future, you can't stop the revolution.
From your perspective you are actually against PCs all together, people shouldn't buy them because they sit idle 95% of the time.
Let the records show that you are not against local inference you are against PCs.
No we'll not use your cloud services.
No we'll not use you closed models.
No we'll not finance your data centers.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Bitcopath retweeted
Sure, there is no Fable class open model yet.
Now look at the price of Fable class.
Claude Fable 5.1: $13.04 per task.
GPT-6.1 Terra: $0.96 per task.
DeepSeek V4.1 Flash on my 4x 3090: just electricity.
Closed labs win a leaderboard built on tests nobody can see.
Meanwhile, I own my model and pay nothing per task.
I will take that trade every single day.
nitter.cf/emollick/status/211527…
Bitcopath retweeted
マルチGPU前提のローカルLLM用マシン組む人はこのサイト見てマザボ選ぶと幸せになれる、どこがcpu直結なのかとか排他仕様がどうなってるかがめちゃめちゃわかりやすくビジュアル化されてるのでおすすめ
mobomaps.com/boards
Bitcopath retweeted
Did you know Qwen 27B can decompile console games on its own? It's that capable.
On a completely unrelated note, an IRL frend just put up a repo that compiles Playstation games into web-playable WASM binaries.
He's very smart but shy, so give his repo a Star to annoy him. 👇
Bitcopath retweeted
MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B) now run on Apple Silicon, thanks to our partnership with the @neywalabs team! 🚀
On an M5 Max, OCR 1.0 reaches 6,339 tok/s prefill and 220–235 tok/s decode, while Jev scores OCR results in just 62.9 ms median latency.
Check out the blog: runnativ.ai/blog/minicpm-ocr…
Bitcopath retweeted
🐳鲸鱼兄弟们！DeepSeek 要放大招来啦！
V5、V5-open、V5-next都端上桌了。
deepseek Harness 的最新代码中发现已经有相关的代码了。
看来离发布也不远了。
github.com/deepseek-ai/deeps…
Death of Dicord
Discord will make public server messages searchable on Google starting in December
🔗videocardz.com/newz/discord-…
Linux and Strata are the keywords here.
Open source for the win!
4 cards on this list. Only 1 got 2x faster tensor math this month, for free, from a Linux module. The 3090.
On my 3090s the BF16 tensor pipe went from 82 to 155 TFLOPS per card, cuBLAS BF16 GEMM from 64 to 99. Outputs match the stock run byte for byte.
Real prefill in Strata got +10 to 12%. Decode didnt move, because decode is memory bandwidth, and 936 GB/s with 24 GB of VRAM is still the reason to buy one.
I paid about $6K for 4 of them. 96 GB of VRAM. DeepSeek V4.1 Flash at 57 tok/s. Qwen3.8 Flash-Next at 87 tok/s on just 1 card.
6 years old and they keep getting faster. No list on X will talk me out of these cards.
nitter.cf/net_termina/status/211…