Angehefteter Tweet
Devon James ☀️
19.8K posts
Devon James ☀️
@DevonRJames
Co-Inventor @OpenIndexProto | CTO @Alexandria | formerly sales @Apple, VFX artist @hensoncompany @Sony & @wbpictures, Infantry @USMC & Technical Dir @web3wg
California Beigetreten Haziran 2007
6K Folgt5K Follower
Devon James ☀️ retweetet
Solo dev reverse-engineered Google's billion-dollar algorithm in 7 days
Google published the paper that crashed memory stocks worldwide. Then shipped zero code.
Tom Turney read the math, opened his terminal, and built the whole thing with Claude - then made it faster than Google promised.
Day 1-3: Core algorithms, 141 tests, Python prototype
Day 3-5: C port into llama.cpp, Metal GPU kernels
Day 5-7: Speed optimization from 739 to 2747 tok/s
That's a 3.7x speedup through pure engineering:
> fp32 → fp16 WHT
> half4 vectorized butterfly ops
> graph-side rotation
> block-32 storage layout
Then he added his own research on top:
> Sparse V: skip 90% of value decompressions at long context
> Asymmetric K/V: keep keys precise, compress values harder
> Temporal decay: old tokens get lower precision automatically
Result: 35B model running on a MacBook with 4.6x compressed cache.
613 GitHub stars in a week. Google still hasn't released their own code.
BuBBliK@k1rallik
English
Devon James ☀️ retweetet
🚨SOMEONE REINVENTED HOW TEXT RENDERS ON THE WEB AND ITS ABSOLUTELY INSANE.
the goated dev behind react, reasonML, and midjourney’s frontend, just dropped Pretext. a tiny typescript library that measures and lays out text 500x faster than the DOM.
he trained models against real browser rendering for weeks until the output matched safari, chrome, and firefox exactly.
the demos are insane!! hundreds of thousands of text boxes at 120fps. magazine layouts and chat bubbles that actually wrap right.
engineers from Vercel, Remix, Figma, and shadcn all cosigned. this is the kind of open source that makes you want to be a better dev.
here are some cool demos in the past 24hrs👇
Cheng Lou@_chenglou
My dear front-end developers (and anyone who’s interested in the future of interfaces): I have crawled through depths of hell to bring you, for the foreseeable years, one of the more important foundational pieces of UI engineering (if not in implementation then certainly at least in concept): Fast, accurate and comprehensive userland text measurement algorithm in pure TypeScript, usable for laying out entire web pages without CSS, bypassing DOM measurements and reflow
English
English
@DevonRJames @xumas_iq April 9 2003 ( 15 mins away from American )
English
Devon James ☀️ retweetet
👀 Could burning the entire LLM (weights, attention layers, and everything else) straight onto a chip and board lower cost and speed up inferencing by hardwiring LLMs?
YES ✅ — and it’s already being done.
Taalas HC1 is using these ASIC “LLM burners” right now. 17k+ tokens/sec on Llama 3.1 8B, ultra-low power, rumored cost ~$300–400 PCIe card, 100% offline. Medium models (such as Qwen 3.5-27B) dropping to lab for testing Spring ’26.
If sold to public could bring local hyper-token AI from sci-fi to your desktop. ⚡🪪🚀
David Hendrickson@TeksEdge
🎗️ "Medium-Sized" LLM Burners Coming Soon! 🔥 This Could Make Local HyperToken Generation a Reality. ⚡️ NVIDIA’s worst nightmare? 😱 ⚙️ Application-Specific Hardware Taalas new PCIe ASIC board would burn the entire medium-sized Qwen 3.5-27B LLM straight into silicon 🤯 (already doing it with small models) Taalos said medium models on ASIC would be available in their lab by Spring '26. 💭Imagine: 🚫 No more loading weights 🚀 ~10,000 Tokens Per Second locally (Llama 3.1 8B already @ 17,000 tps) 💻 Standard PC slot, ultra-low power (10x less) 🔋 🌍 100% offline with no cloud, no GPU farm 💰 Reddit unit cost rumor $300 to $400 🖥️ Imagine HyperToken generation on your desktop. 🤖 AI agents that think at light speed. ⚡️ Are you ready? 👀
English
- Drafted a blog post
- Used an LLM to meticulously improve the argument over 4 hours.
- Wow, feeling great, it’s so convincing!
- Fun idea let’s ask it to argue the opposite.
- LLM demolishes the entire argument and convinces me that the opposite is in fact true.
- lol
The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.
English
@0xSero you skipped from 24 to 64, what would you recommend for 48 (across 2 4090s, so prolly not a fully usable 48)?
English
Best models to run on your hardware:
—— 64 GB ——
- Qwen3-coder-next-80B-4bit (coding, Claude code, general agent)
- Qwen3.5-122B-reap: (browser use, multimodal, tool calling, general agent)
—— 96 GB ——
- GLM-4.6V (multimodal and tool calls)
- Hermes-70B (Jailbroken)
- Nemotron-120B-Super: (openclaw)
- Mistral-4-Small (general agent)
—— 192 GB ——
All these are excellent top tier LLMs and approach sonnet in capabilities
- Step-3.5-Flash
- Qwen3.5-397B-REAP
- MiniMax-M2.5 (soon M2.7)
- GLM-4.7-Reap
0xSero@0xSero
Best models to run on your hardware level I'll be doing this every week, I hope you guys enjoy. ---- 8 GB ---- Autocomplete for coding (like Cursor Tab) - huggingface.co/NexVeridian/ze… - huggingface.co/bartowski/zed-… Tool calling, assistant style - huggingface.co/nvidia/NVIDIA-… ---- 16 Gb ---- Here things get better: Multimodal - huggingface.co/Qwen/Qwen3.5-9B - huggingface.co/Tesslate/OmniC… - huggingface.co/unsloth/Qwen3.… ---- 24 GB ---- - The best model you can get (thanks Qwen) huggingface.co/Qwen/Qwen3.5-2… - Great model (strong agents) huggingface.co/nvidia/Nemotro… - Mine hehe huggingface.co/0xSero/Qwen-3.… I'm doing a weekly series
English
Devon James ☀️ retweetet
Best models to run on your hardware level
I'll be doing this every week, I hope you guys enjoy.
---- 8 GB ----
Autocomplete for coding (like Cursor Tab)
- huggingface.co/NexVeridian/ze…
- huggingface.co/bartowski/zed-…
Tool calling, assistant style
- huggingface.co/nvidia/NVIDIA-…
---- 16 Gb ----
Here things get better:
Multimodal
- huggingface.co/Qwen/Qwen3.5-9B
- huggingface.co/Tesslate/OmniC…
- huggingface.co/unsloth/Qwen3.…
---- 24 GB ----
- The best model you can get (thanks Qwen) huggingface.co/Qwen/Qwen3.5-2…
- Great model (strong agents) huggingface.co/nvidia/Nemotro…
- Mine hehe huggingface.co/0xSero/Qwen-3.…
I'm doing a weekly series
English
English
English
Speaking for the pagans, no it's not.
It's obviously Catholic.
Human Events@HumanEvents
.@JackPosobiec: Lord of the Rings is overtly pagan.
English
@JackPosobiec @VDAREJamesK You need to try harder to understand the concept of an "allegory"
youtube.com/shorts/ICNVgTE…
YouTube
English
@VDAREJamesK With no Christ figure, no religion, not even prayer. Ok lol
English
On April 1, 2003, Iraqs information minister went on TV and said "they are nowhere near the airport …They are lost in the desert... they can not read a compass. they are no where near Baghdad! This is silly!"
We found this pretty funny to hear. The next day the airport was taken. My platoon got there on April 4 I think. We crossed the Diyala river into Sadr city the night of April 7. The statues in Baghdad started getting pulled down on April 8.
It's pretty standard for the losing side when facing overwhelming defeat to just lie through their teeth to give themselves as much time as possible so they're ready to go into hiding the moment the regime collapses.
English
Let's hope the balance of power in the actual war is the exact inverse of the meme war, because Iran appears to be wiping the floor with us in Legoland.
Did the Six Eyes somehow fail to anticipate having to fight on this front?
Patricia Marins@pati_marins64
Iran doesn't seem intimidated at all and has just released another Lego video mocking the coalition.
English
Devon James ☀️ retweetet
Devon James ☀️ retweetet
Upgrading your RAM is now unnecessary.
Introducing our new ComfyUI Dynamic VRAM optimization. Running local models is now possible on even the most memory constrained hardware.
Read more here:
blog.comfy.org/p/dynamic-vram…
English
Devon James ☀️ retweetet
Two incredible innovations in the local AI space in a span of three days. I am so excited.
ComfyUI just shipped "Dynamic VRAM" and it seems like a big deal for anyone running models locally.
The problem: large AI models can have many GB of weights. If your system lacks the necessary RAM, you'd normally hit memory crashes or grind to a halt on the page file.
Instead of loading the entire model into memory at once, ComfyUI now reads the model file piece by piece directly from your SSD. Only the specific parts needed for the current step get pulled into memory. Everything else stays on disk until it's actually called for.
On the GPU side, they built a smart system that loads weight data at the exact moment it's needed. If your GPU runs out of space, it doesn't crash.
It uses a temporary workaround to finish the calculation, then cleans up after itself. It also keeps track of what didn't fit so it doesn't waste time trying to reload things that won't fit again.
The other big improvement is for workflows that use multiple models. Previously, swapping between models would pile everything into system memory and bog your machine down.
Now when a model gets swapped out of the GPU, it just goes back to the "read from disk when needed" state instead of sitting in RAM.
The result: a 56GB model can now run on a machine with only 32GB of memory. No crashes, no slowdowns from swap.
Available now for Nvidia GPUs on Windows and Linux, with AMD support on the way.
No idea how fast this is, but this seems incredible. Cannot wait to get my workstation going.
ComfyUI@ComfyUI
Upgrading your RAM is now unnecessary. Introducing our new ComfyUI Dynamic VRAM optimization. Running local models is now possible on even the most memory constrained hardware. Read more here: blog.comfy.org/p/dynamic-vram…
English
Devon James ☀️ retweetet
🎗️ "Medium-Sized" LLM Burners Coming Soon! 🔥
This Could Make Local HyperToken Generation a Reality. ⚡️ NVIDIA’s worst nightmare? 😱
⚙️ Application-Specific Hardware
Taalas new PCIe ASIC board would burn the entire medium-sized Qwen 3.5-27B LLM straight into silicon 🤯 (already doing it with small models)
Taalos said medium models on ASIC would be available in their lab by Spring '26.
💭Imagine:
🚫 No more loading weights
🚀 ~10,000 Tokens Per Second locally (Llama 3.1 8B already @ 17,000 tps)
💻 Standard PC slot, ultra-low power (10x less) 🔋
🌍 100% offline with no cloud, no GPU farm
💰 Reddit unit cost rumor $300 to $400
🖥️ Imagine HyperToken generation on your desktop.
🤖 AI agents that think at light speed. ⚡️ Are you ready? 👀
English
@Didicoy_Tonttu @johnnymaga look at the backs of the chairs. rubios is clearly a different style of chair, it’s possible that the seat is higher.
English
@johnnymaga Trump - 6'3"
Rubio - 5'9"
Burgum - 6'1"
Why is Trump the shortest person in frame?
English
Burgum on Venezuela: I literally think they’re going to put up a statue of President Trump
Trump: That would be a great honor
*2 mins of updates later*
Burgum: Their oil now flows to our refineries
Trump: Forget that. When are they going to do the statue? 😭
English
@MichelleWth6 @DataRepublican names that are part of current active investigations were not released 🤷♂️
English
@9to5mac it's really great news because finally we're gonna be able to force all of these idiots who insist on using Photoshop on a Mac rather than a PC onto a PC so that they learn how to use a real computer and stop costing IT departments tens of thousands of dollars a year
English
Devon James ☀️ retweetet
Training distributed DiLoCo / SparseLoCo over eduroam wifi, awesome!
Swarnim Jain@swar_ja
I trained models across MacBooks using Apple's AirDrop protocol. grove is a distributed training library for Apple Silicon. Devices discover each other over AWDL, a direct radio link. If there's a shared WiFi network it upgrades to that for speed, otherwise everything goes over the direct link. No router, no cloud, no setup. grove start