yzma v1.26 is here and it is the big one: local inference in the browser! Write @golang.org, compile with @tinygo.org, run GGUF models 100% locally on WebAssembly. Uses WebGPU/CPU Plus @hf.co llama.cpp v0.4.0 support. Go get it! github.com/hybridgroup/yzma #golang #tinygo #wasm #webgpu #yzma

GitHub - hybridgroup/yzma: Go with your own intelligence - Go applications that directly integrate llama.cpp for local inference using hardware acceleration.

Go with your own intelligence - Go applications that directly integrate llama.cpp for local inference using hardware acceleration. - hybridgroup/yzma

github.com

Specialization means performance, accuracy and lower costs. It's still the time for general models, but examples like this make clear what the future would be. Great work 👏

Daniel van Strien@danielvanstrien.bsky.social · last wk.

A 15.9M-parameter specialist OCR model outperforms many VLMs hundreds of times larger. On a 2,165-page historical eval dataset, Kraken PP-OCRv6 ranks #4 on reading CER and #1 when long-s, ligatures and case are preserved. huggingface.co/spaces/fineb...

I like to see how AI changes the access to knowledge and the knowledge itself

Daniel van Strien@danielvanstrien.bsky.social · 3w ago

Every illustration in the British Library's 19th-century book images dataset now has an instance mask: 1.02M clean cutouts of public-domain engravings, no image generation involved! One small segmentation model, $3.24 of compute. Model, masks and a search Space with cutout view all open.

I'm still trying to understand how and why platforms like @bsky.app ignore developers and scientists needs. These platforms are built by developers (and programming is a Math), but they don't have code highlighting, nor they have Markdown or LaTeX support Who's making product decisions there?

Hi, @feross.bsky.social what is @socket.dev's position on this? By default inference APIs don't escape special tokens which split chat into system and user messages. What let user to send system message github.com/ollama/ollam... github.com/huggingface/... github.com/google-deepm...

Special token in user input is tokenized as-is for Gemma4:26b · Issue #15931 · ollama/ollama

What is the issue? When calling: curl http://localhost:11434/api/chat -d '{ "model": "gemma4:26b", "messages": [ { "role": "user", "content": "say \"<|channel>\"" } ], "stream": false }' It appear ...

github.com

🦥 Unsloth turns SOTA open-weight models optimization into something usual

Unsloth AI@unsloth.ai · last mo.

DeepSeek-V4-Flash can now run 2× faster locally with DSpark! ⚡️ DSpark enables V4-Flash-0731 GGUFs to generate ~1.4–2× faster with no accuracy change. DeepSeek-V4-Flash-0731 can reach at 120 tokens/s. GGUFs: huggingface.co/unsloth/Deep... Guide: unsloth.ai/docs/models/...

Train a small language model on $8 ESP32-S3? This model speaks Klingon (poorly) and is not capable of meaningful conversation. It's 319K parameters only 🤷‍♂️ But the whole learning process has been running on the ESP32 chip 🤯 📦 Project: github.com/Carloscodix/...

GitHub - Carloscodix/qapla: A char-level transformer trained from scratch on an $8 ESP32-S3. Not inference: the chip runs the full training loop, with backprop written by hand in C.

A char-level transformer trained from scratch on an $8 ESP32-S3. Not inference: the chip runs the full training loop, with backprop written by hand in C. - Carloscodix/qapla

github.com

Gemma4 on iPhone using only 500MB of RAM? Yep! In the demo the new model controls Calendar app from the device huggingface.co/fraQtl/Gemma...

fraQtl/Gemma-4-E2B-it-Hi-Fi-GGUF · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

It would take time for a wider support and adoption of diffusion text models, but the beginning is promissing

Sung Kim@sungkim.bsky.social · last mo.

Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of

🦥 Unsloth quiantized Kimi K3 🤩

Unsloth AI@unsloth.ai · 2mo ago

Kimi K3 can now be run locally! ✨ The 1-bit model retains ~78.9% accuracy after we shrunk it from 1.56TB to 594GB (-62% size). Run on a Mac Studio + 128GB RAM device. Kimi K3 is the strongest open model to date. Guide: unsloth.ai/docs/models/... GGUF: huggingface.co/unsloth/Kimi...

A guy is running Kimi-K3 on a M1 MacBook with 5 tokens/minute speed Near Fable-5 model on a MacBook 😅 github.com/gavamedia/de... #KimiK3 #dev #llm #ai

GitHub - gavamedia/deltafin: Run Kimi K3, a 2.8T-parameter Mixture-of-Experts LLM, on a single Apple Silicon Mac. Streams MXFP4 experts on demand over HTTP into a local disk cache — fused NEON kernels...

Run Kimi K3, a 2.8T-parameter Mixture-of-Experts LLM, on a single Apple Silicon Mac. Streams MXFP4 experts on demand over HTTP into a local disk cache — fused NEON kernels, Metal/MPS compute, exact...

github.com

Programming is making math work for humans 🖖

iamwil@interjectedfuture.com · 2mo ago

side effects and state management. x.com/tsoding/sta...

Things like DeepWiki.com and Repowise.dev are the future But I can not imagine my work without Github1s.com, it provides all the VSCode features for code navigation and search right in a browser

Mike Schilling 🦀@oxidebyte.bsky.social · 2mo ago

I found a new way to read GitHub repos and I like it 😍 deepwiki.com

It has been discussed bsky.app/profile/pave... It's a two days old trick. It should be something breathtaking now

mr. TIM@timkellogg.me · 2mo ago

the Dinitz-Garg-Goemans conjecture, a graph theory problem that’s gone unsolved for 30 years, was solved with this prompt

It's the biggest opportunity window this year We are on a plato for many capabilities. Smaller players mostly can only adopt what providers give them, and providers are fighting for the best coding agent

Ethan Mollick@emollick.bsky.social · 2mo ago

I still believe that everyone is too fixated on the state of play in AI right now (which labs are ahead, how to manage costs, etc.) and not focused enough on the continued steepness of the capability curve for AI At higher capabilities (like the ones expected in the near term), a lot changes fast.

"We introduce requential coding, where a teacher model selects training samples drawn from the student's own distribution" 👏

Andrew Gordon Wilson@andrewgwils.bsky.social · 2mo ago

How far can we compress billion-parameter LLMs? We introduce requential coding, which achieves < 1-bit per param compression, and explains why scaling doesn't hit a generalization wall! arxiv.org/pdf/2607.11883 w/ Shikai Qiu, Marc Finzi, Yujia Zheng, Kun Zhang 1/🧵

ThinkingMachineLab has introduced their open-weight model Inkling (Apache 2.0) with near Fable 5 performance on IFBench (instruction following) and AIME 2026 (math problem solving) with only 41B active params thinkingmachines.ai/news/introdu... #ai #llm #release

@opensource.google EmbeddingGemma has been trained on Gemma3: ai.google.dev/gemma/docs/e... Do you have plans to train it on Gemma4? It should increase relevance and can reduce model size

🚶‍♂️ Job interviews today: You're not allowed to use AI on interview! 🏎️ Job interviews in the future: What prompt would you use to do your job?

Hey, @deadprogram.com it seems like yzma.ai website doesn't work. Did it work in the past?

It's not a question of opinions. This is math! Some people believe that 2⁴ is the same as 2 + 4 just because they saw 2 + 4 before, and think it's 6. AI automates processes which require years of studying and hard work and turns these years into hours and even minutes

And more autonomy is coming...

Simon Willison@simonwillison.net · 2mo ago

The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving tokens simonwillison.net/2026/Jul/3/j...

How to make models be more creative without telling them to be more creative:

Kingson Man@kingson.bsky.social · 2mo ago

Hola from #ASSC29 in Santiago, Chile! We are presenting early work here called "Skin in the Game": What kind of mechanism is needed for feeling-like behavior to appear in machines? 🧵...

I feel disappointed after receiving AI-generated reply to my supportive comment in a dev channel. It was just a basic "exactly that. thanks for noting this" LLM-styled reply! I'd like platforms to react faster to behavior like this with better reporting. Either way we would quit them

$333/month is the cost of owning new AMD's Halo for one year If the capability of the open-weight models would reach Fable-5's in the next 6 months it would become economically effective to have one. Especially for small companies. But I'd like to see demo on real use cases #AI #AIHardware #LLMs

You'd better let AI to rewrite your abstracts and probably this message too 😁

@patqdasilva.bsky.social · 3mo ago

🚩 AI-assisted scientific writing faces increasing scrutiny Can human or LLM editing fix it? 👇 240k edits on scientific abstracts expose LLM writing flaws, and show that human editing fails to fix them cc Sanchaita Hazra, Doeun Lee, @shocheen.bsky.social, Bodhisattwa Prasad Majumder 🧵

Hey guys, ASI is coming closer. Seems like Sakana has beaten Mythos 5 in some tasks 🤯 Congrats to the team!

hardmaru@hardmaru.bsky.social · 3mo ago

Human intelligence is fundamentally a collective intelligence. We solve complex problems by participating in a vast cultural network that builds upon ideas across generations. I believe the strongest AI systems will become a collective intelligence, too.

Would AI era end Git's 20 years of dominance? There were two new version control systems announced by big players this June. DeltaDB by @zed.dev and Lore by Epic Games. It's the first sign of an AI shift in this direction. Curios would it be enough? #AIDev #AICoding #dev #git

Stanford analyzed the state of AI R&D and found 5.6M projects on Github in 2026. The number has almost doubled since the previous year and keeps growing hai.stanford.edu/ai-index/202... #ai #dev #opensource

After thoughtful consideration, I believe the solution to this is a reasonable limit on the contribution size for non-members and for overall number of contributions over time period. It would help to protect maintainers without sacrificing contribution activity #git #ai #dev #aislop

Pavel AI@pavelai.bsky.social · 3mo ago

After reaching a new equilibrium we will get from "stop ai slop" to the usual "probably would be implemented" state of contributions. Just give it some time