Microsoft just opensourced a blazing fast inference framework for 1-bit LLMS. Now you can run up to 100B LLMS locally on your CPU (GPU and NPU support coming in the future) and get 5-7 tokens/second.

https://github.com/microsoft/BitNet

Brilliant article on Arm's approach for the future, challenges (AI, Qualcomm and Apple turning into competitors, AMD and Intel joining forces to improve x86) and mini-interview to Christopher Bergey, senior VP and GM of the Client Line of Business at Arm. A very good read.

https://www.theregister.com/AMP/2024/10/22/arm_custom_silicon_interview/

Thomas Eugene Kurtz, father of BASIC, has died peacefully today in his retirement home. He was 96. Thank you, Thomas, for enabling at least two generations of human beings to get to learn and use a personal computer.

This is truly the best time for retro gaming and emulation. Ten years ago I would have sworn we'd never see a working emulator for extremely complex systems such as the PS3 and its IBM Cell, a PowerPC-based CPU. Now RPCS3 runs even on ARM64 CPUs, including Apple Silicon, thus opening a new landscape for PS3 emulation even on extremely portable devices.

For more information on the specific RPCS3 case, read this blog post on their website.

Interesting article on AMD's quite incredible feat to run crude C++ code on their GPUs for AI workloads, thus threatening the de-facto monopoly Nvidia has with CUDA and their GPUs (biggest market cap in the world).

https://www.phoronix.com/news/AMD-Standard-C-Code-GPUs

Find your favourite telegram channel admin at FOSDEM 2025 (01-02/February)!

The research paper "All You Need Is Attention" from the Google Research Team that came up with transformers (which radically changed and improved artificial intelligence and gave rise to genAI as we know it) might have found its successor.

https://arxiv.org/abs/2501.00663

Some good discussion and interpretation of results can be found in the comments section here.

Imagine playing a simple game of Hangman with an AI.

You'd think that explaining the game mechanics step-by-step would be enough, but my favorite benchmark test for Large Language Models (LLMs) has consistently highlighted their limitations.

Most LLMs can't successfully play Hangman—even after a detailed explanation. They end up accepting all guesses as correct, fail to understand the turn-based nature of the game, and often mix up who should come up with the blank word or provide guesses. In my recent tests, between the commercially available LLMs, only ChatGPT 4o has been able to play Hangman correctly.

This quirky but telling failure shows that while these models can seem incredibly smart, there's still a big gap between surface-level understanding and actually grasping logical, step-by-step processes. Read more about it in my latest post on Substack.

New post on Substack. Just in time for a good lunch read, I suppose. Don't miss it!

https://atsetilam.substack.com/p/the-fabled-ai-on-the-edge

Asahi Lina left the Asahi Linux project for “personal reasons”. We lost a fantastic developer who was focused on an incredibly difficult and extremely important work - bringing Linux onto the Apple Silicon platform. Such a pity. She says she is “safe”, luckily.

We got GIMP 3.0 before GTA 6

"Model Context Protocol (MCP) is an AI tool calling standard that has been rapidly gaining adoption over the past few months. MCP tools give LLMs a standardized way to call functions, look up data, and interact with the world. Anthropic created the protocol and built the first GitHub MCP server, which grew to be one of the most popular MCP servers in the expanding ecosystem. We are excited to take ownership of the server and continue its development."

https://github.blog/changelog/2025-04-04-github-mcp-server-public-preview/

Microsoft has released a tool under MIT license to convert your docx into markdown.

https://github.com/microsoft/markitdown

OpenAI has just released a FOSS CLI tool for “developers who already live in the terminal and want ChatGPT‑level reasoning plus the power to actually run code, manipulate files, and iterate – all under version control”.

https://github.com/openai/codex

Babe wake up, Microsoft released a 1-bit LLM under MIT that is optimized for running on CPUs: microsoft/bitnet-b1.58-2B-4T

Salvatore Sanfilippo (aka antirez) is back!
After stepping away from Redis for some time, he's returned with a major contribution: a brand-new data type called vector sets. This addition brings semantic similarity search to Redis, making it possible to query based on meaning rather than exact matches.

Check it out: https://redis.io/blog/announcing-vector-sets-a-new-redis-data-type-for-vector-similarity

A must-watch for anyone interested in the future of #AI. In this interview for NVIDIA Developer, Yann LeCun - Turing Award winner and Chief AI Scientist at Meta - shares his take (and a very robust contrarian opinion, IMHO) on the limits of today’s language models.

In the interview, he argues that LLMs (like OpenAI's GPT or Meta's LLaMA) are not the path to true artificial general intelligence. They're impressive, yes, but fundamentally constrained by the Transformer architecture. Scaling up won’t solve this. Why? Because human intelligence isn’t just about language or token prediction - it’s about understanding, reasoning, and interacting with the physical world through systems more akin to what psychologists call System 1 and System 2.

Think about a cat: it can leap with precision without any concept of physics, nor of the languages (mathematical and linguistic) which can explicitly explain physics. Ask your local orange alley cat about it, if you don't believe me.

LeCun also shares some striking numbers to highlight just how limited language-based learning really is:

• Human language processing has a very low data rate - roughly 12 bytes per second. That’s about 4.5 words per second, assuming each word is encoded in about 2 bytes.

• Vision operates on an entirely different scale. Our two optical nerves transmit a combined stream of roughly 20 megabytes per second, based on the million fibers in each nerve sending about 10 bytes per second.

• Over just four years of being awake, a child accumulates around a petabyte of visual experience - far more than the total training data of even the largest language models.


To put it plainly:

Visual perception carries around 16 million times more data than reading or listening to language, and a preschooler has already taken in 50 times more data than what goes into the largest text-trained LLMs.


LeCun’s takeaway is that the infrastructure modeling the data and the structure of data matter just as much as the quantity of data. Self-Supervised Learning thrives on redundancy, and sensory inputs (especially vision) are packed with the kind of statistical richness that language alone can’t provide.

Most of what we know - and certainly what animals know - comes from experience, not explanation. Language is a brilliant tool, but it's the final layer, not the foundation.

I explored this in my own way in Hangman and Circles, where I reported some results of a research by the Apple AI Team, reflected on the limits of linguistic abstraction, and talked about why it may be misleading us in the pursuit of AGI:

https://t.me/bytebaibyte/19

This interview is amazing, and super-fun too - definitely worth your time:

https://youtu.be/eyrDM3A_YFc?si=oMiDKJAXUYIjfjIu

PS. I also found watching the ~2h interview on the Lex Fridman Podcast extremely interesting - it dives deeper into the topics mentioned during the Nvidia Developer interview.

also relevant: https://arxiv.org/abs/2503.21934

Who would have thought we'd see PewDiePie advocating for Linux on the desktop before GTA 6?

In the upcoming Ubuntu 25.10, Canonical plans to use an alternative to sudo that's being developed by the sudo-rs project and written in Rust. In March, a similar decision was made to replace GNU Coreutils with uutils, which is also written in Rust. There are currently initiatives under consideration to replace zlib and ntpd with zlib-rs and ntpd-rs.

https://www.phoronix.com/news/Ubuntu-25.10-sudo-rs-Default

via Valentina Lenarduzzi on LinkedIn:

"Our paper "Does #Microservices Adoption Impact the Velocity? A #Cohort Study" has been accepted at Empirical Software Engineering Journal

Microservices are often praised for improving development speed thanks to their modular and independent nature. But do they actually lead to faster feature delivery and bug fixing? In our latest study, we explored this question using a retrospective #Cohort design - a methodology widely used in medical research but still rare in software engineering.

What we did: We conducted the first large-scale empirical study comparing GitHub projects built with #Microservices from the start against similar monolithic projects, using a #Cohort study to assess causality-not just correlation.

What we found: Surprisingly, no statistically significant difference in development velocity was observed. Even after controlling for confounding variables, #Microservices adoption didn't show a measurable impact on how quickly projects deliver features or fix bugs.

Why it matters: This study not only challenges assumptions about #Microservices and velocity, but also introduces a powerful empirical methodology to our field. We're excited to contribute one of the first works applying cohort studies in software engineering research.

https://www.researchgate.net/publication/391482952_Does_Microservice_Adoption_Impact_the_Velocity_A_Cohort_Study