Pinned Tweet
dropping github.com/say4n/pi-sync, take your @pidotdev config across all your hosts without any fuss
sayan retweeted
So one of our ambassadors @divagr1925 has been working on something cool
He found a way to use Qwen's Gated DeltaNet state as memory, and built a 32-bit RISC-V CPU directly into the model's weights. So instead of relying on predicting the answer to a calculation, the model can execute instructions internally. It can also run C and MicroPython programs.
He calls it Qwenputer, love the name btw. The results so far are pretty impressive. On one arithmetic benchmark, accuracy jumped from 48.6% to 100%.
Pretty cool to see developers experimenting with Qwen's architecture. Great work, my friend!
sayan retweeted
Using a MacBook, you can run:
LFM3 (1.7B): 620 tok/s (3x faster than mlx-lm) on M2 16 GB.
Flint (40B): 319 tok/s (1.6x faster than mlx-lm) on M4 24 GB.
New in Locals: LFM3 + Flint. This speed is with 8 agents. Our custom harness splits problems into subproblems, so you get multi-agent speed with a single-stream experience.
Open source: github.com/Generated-Compute…
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
sayan retweeted
It is pretty wild.
qwen-neurons-site.vercel.app…
sayan retweeted
AI-assisted PCB design for KiCAD 10 using MCP
blog.adafruit.com/2026/10/26…
sayan retweeted
You can now run Nintendo Switch games natively on iPhones and Macs
blog.adafruit.com/2026/10/27…
hell yeah!
I've been doing this with basically everything I own
here's my liene pixcut S1 printer / cutter from browser
brencut.brennan.computer
Tears of the Kingdom, running native on iPhone Duo!
I'm continuing to cook on these "ports" so make sure to follow for updates in the future. so
Just like the last one for Super Mario Odyssey, I’ve added:
- Controller and On-screen controls w/ native device gyro.
- PS5 DualSense controller support for the unique haptics and adaptive trigger engine. Drawing a bow gives some resistance, and shield surfing feels incredible.
- Haptics on device itself when using without a controller.
- Play in folded or unfolded mode on the Duo, keeping your gameplay seamless in whatever format you want to enjoy it in.
Github repo in the comments so you can try this out on your own legally obtained game files.
sayan retweeted
A configuration mismatch in GLM 5.3 on Workers AI triggered an SGLang bug that exposed KV cache pages across requests. Here is how we addressed the issue. cfl.re/4srMk2v
sayan retweeted
My first open model lecture, covering all the many pieces that build the ecosystem, and the open questions/debates determining where it goes next.
How does distillation fit in?
How do we manage risks?
How do we see access to models changing in the future?
I'm excited to revisit these questions about every 6 months, to give a state of the union on the field.
Chapters
00:00 Introduction
01:36 What is an open model?
08:47 Why should we care about open models?
18:52 How competitive are the best open models today?
25:28 How are open models built? The role of distillation
34:52 Are people actually using open models?
41:47 What comes next with open models?
47:32 Conclusions
What questions did I not answer that you have?
sayan retweeted
a couple weeks ago I posted about how bone grows into titanium implants because cells can't find the boundary
turns out there's a catch engineers have been trying to solve for years:
titanium is too stiff (~110 GPa vs spongy bone's 0.02–3.7 GPa)
so the metal takes all the load, and the real bone around it actually weakens because physics says "why maintain structure that isn't carrying stress?" (stress shielding)
enter selective laser melted Zn-1Mg-0.1Sr gyroid scaffolds:
→ mathematically generated TPMS gyroid lattice (70% porous, 550 μm pores)
→ elastic modulus tuned right down to 2.3 GPa so it flexes like real cancellous bone
→ alloying zinc with 1% Mg + 0.1% Sr precipitates Mg2Zn11 & SrZn13 intermetallics, more than doubling strength over pure zinc (33.7 MPa)
→ degrades at 0.36 mm/year with zero hydrogen gas pockets
and the wildest part
as it dissolves, it turns its own surface into calcium-phosphate & zinc-phosphate minerals while feeding Zn2+ and Sr2+ ions that make bone-forming cells grow faster
titanium tricks the skeleton into adopting permanent metal
this thing holds the load, feeds the bone to rebuild itself, and vanishes when the job is done.
radical engineering.
today while talkin to my friend (who’s a doctor) I learned something really crazy
That the titanium implants don't just sit in bone they actually bone grows into them
the oxide layer on titanium surface is chemically close enough to biological tissue that cells can't find the boundary so they erase it
osseointegration isn't the implant holding on it's biology deciding the metal belongs there
the skeleton adopts the foreign object as its own
Cyborg vibes
sayan retweeted
My take: SaaS is dead. Software is charity now, make money elsewhere.
No one can make $$ selling the same features for years anymore.
Instead:
- Use good software to collect user data and sell data.
- Use good software to deliver outcome and sell outcome.
- Arbitrage tokens under the guise of selling software - buy bulk, sell retail.
- Use good software internally to drive your own margin higher for your services/products
- Build RSI agent in a domain that has the fastest learning loop so it has the most, latest expertise: e.g. a marketing agent that knows the latest tricks to get viral in X coz it's doing it for others every day and learn and evolve.
- Give away all your features for free, it's futile anyways. Open source them. Make money elsewhere.
occam’s razor
Replying to @cohere
Simplest explanation is just a narrow RL training reliant on their internal infra (which isn't good design for an open weight model)
im about to get 5 press requests about this 🤦♂️
sayan retweeted
We just launched OpenUSD support on the @huggingface Hub 🎉
Yesterday, @nvidiaomniverse put 2,500+ SimReady assets on the Hub. Forklifts, robot arms, engines, hospital equipment, all OpenUSD. You can also search the whole library in plain English
huggingface.co/spaces/nvidia…
huggingface.co/datasets/nvid…
sayan retweeted
Introducing ☀️ Sunlight: The first infinite canvas window manager for Mac
- 14″? How about 355″?
- 100% native
- 120fps multi-touch interactions
- No window reflows while zooming
- 100% free and open source
- Coming next month
sayan retweeted
If you're wondering why GPU prices are going up ... algorithms have become wildly more efficient while prices haven't come down nearly as much.
There is a major "arbitrage" right now between token prices and GPU rental rates. This means that the price of tokens goes down (supply up) while the price of GPUs goes up (demand up).
@opencode is one of the top providers of tokens outside of the frontier labs. Let's suppose they're using a GB300 NVL72 and ballpark the price at $5 million.
DeepSeek v4.1 Flash 1017:
• 98% input tokens at effective in of 0.82c per million.
• 2% output tokens at 60c per million.
• Weighted average = 2c per million.
Revenue = 6.77T x 2c / million = $136,000 / day. This would be circa double the $38 / hour that DeepSeek reported.
If the GPU Rack was $5 million then then the payback for "just" the GB300 is 37 days.
Obviously there are other costs but this gives you an idea.
sayan retweeted
🥐 Hi everyone! FlashAttention's backward pass has the same zero-sum bug
@jon_barron
just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.
































