Edge0-145B: Frontier AI Capabilities on Your Own Computer
A commenter notes that the 145B parameter model can be run with only 29GB peak memory usage, which is impressive.
30 stories
Get a notification every evening
You hear about the digest when it is ready. Turn it off any time.
Notifications are on
You will hear about each evening’s digest when it is ready.
To get notifications on iPhone or iPad, add this site to your Home Screen first: tap Share, then Add to Home Screen. Open it from there to turn notifications on.
Notifications are blocked in this browser. Allow them in site settings and reload.
New notification sign-ups are closed for now. You can follow along with RSS.
Notifications could not be turned on. Check your connection and try again.
Notifications could not be turned off and are still on. Check your connection and try again.
The day’s top discussions
A commenter notes that the 145B parameter model can be run with only 29GB peak memory usage, which is impressive.
A commenter says that Nvidia knows it will sell all its inventory due to supply chain bottlenecks, making circular financing a normal business strategy.
A commenter notes that the new quantization (v3.1) of the Qwen3.8-27B model, while slightly worse at debugging, retains its overall advantage and significantly reduces token consumption.
A commenter claims that 607.8 tok/s on $200k hardware is not impressive and higher throughput can be achieved cheaper with 8xB6000 cards.
The project author explains that they rebuilt three core ideas of DeepSeek V3 in about 190 lines of pure JAX, training a 36.2M parameter model on a MacBook Air M5.
A commenter says that using parallel agents in Claude Code has become common and projects are his main way of coding.
A commenter states that Gödel was a perfectionist and while his notes likely contain sharp insights, it's unlikely he produced illuminating new theorems by chance.
A commenter says ggml and llama.cpp provide 5-10% speed improvements on the same hardware with each update, and verifies this with his own benchmark.
A commenter questions why every factory game revolves around mining/smelting iron and copper and asks if there are more interesting industrial processes.
A commenter expresses confusion about the heavy reliance on demonstration examples in JEPA world models and argues that the real task is to move between any two arbitrary poses.
A commenter states there is a sharp distinction between abstract mathematical correctness and real-world functionality in software, and it's important to determine when formal correctness is useful.
A commenter notes that Oxide's Propolis also worked on live migration with bhyve.
A commenter finds it impressive to implement out-of-order execution (OOO) in a one-person project.
Commenters discuss the duration of memory billing after last request and the cold-start sample size.
A commenter points to a more detailed Bleeping Computer article about the Apify security breach.
Commenters share an orbit simulation and note it was found in 1937 photographic plates.
Commenters discuss the article about bypassing Android clipboard read security, noting GrapheneOS's secure paste and recommending password managers.
Commenters question why Haskell isn't popular in AI and express wishes for Haskell to be more universal.
Commenters note that Excalidraw can run fully locally via its NPM package and that a similar solution exists in Emacs.
Commenters discuss that general models run robots poorly and that the disbanding might be cost-cutting before an IPO.
The author and commenters discuss that AI-generated apps are vulnerable to data leaks and users are unaware of risks.
One commenter states that Kapso efficiently calibrated Relflow's hyperparameters and outperformed TabPFN.
Commenters discuss that such tools are good for simple answers but need more structure for effective self-teaching.
Commenters note that tokenization inefficiency in non-English languages causes more energy use and that methods like H-Nets are suggested.
A smaller, more technical community
A commenter admires this setup, mentions being just before the bar exam in Germany, finds the tools interesting but is in the Emacs camp.
While a commenter notes Kildall is better known for his work on APL, a project describes working on Cleavir, a new Common Lisp compiler.
The week’s most-starred new repositories