New conversation on Interconnects w Dean Ball on why the Anthropic v DoW moment could strengthen the long-run case for open models - even if the next few years get rough for open.
www.interconnects.ai/p/how-anthro...
Rather than cleaning this up from my idiosyncratic environment (I don't own a laptop or desktop computer), here's the spec if you want to ask your parrot to implement it: paste.rs/hIbgG.txt
Apchat can now work with images !
apchat can now see. Took a bit of convincing but we’re there :)
Is there a pattern that is starting to energe or am I just hallucinating it ? 🤔
I might point my own thing at Tim’s repo and tell it to learn. The diversity, even if just on the harness side, is really cool, and allows to quickly test lots of different ideas.
Reading linkedin threads how AI is good only for boilerplate code, while apchat/qwen3.5 duo is doing a refactor surgery on apchat in order to support multimedia capabilities, and can’t stop giggling uncontrollably.
It’s very fragile and tries to mask it by being abrasive.
(100% local, artisanal AI heating my apartment)
Screenshot of a conversation instructing AI agent to clone VPP repository and go bughunting.
On it.
With the qwen3.5, feels like this is not that far off.
Got my little Ape a creative outlet: ayourtch-llm.github.io/apchat-blog/ - the first experience required a couple of adjustments to match the actual experience - but we see, hopefully it learns with more time :)
I need some of that for my Ape, it’s been all business ! 😂
What did you feed it with ? 🤯😂
People are trusting trust too much... ;-)
FWIW I run Qwen3.5-35B-A3B-UD-Q4_K_M.gguf via llama.cpp on a single 4090 with 262000 context window and it flies. And makes half reasonable Rust code (though likes to jump to conclusions a lot but at 100tok/s and 2000tok/s prefill it’s cheap to iterate on things)
Tried qwen3.5-27B ? I didn’t try it for review, but for coding it‘s pretty impressively good.
The first two books that ap[e]chat decided to read from my kindle library were “hacker delight“ and “how to get rich”. And then it started looking at “Explainig humans”. 🤷♂️
github.com/maderix/ANE - pretty interesting! I wonder if eventually this might end up in Asahi Linux…
https://github.com/jmamda/OpenTrace Reporting dashboard
Found this little gem today: github.com/jmamda/OpenT...
Qwen 3.5 Small Model Series just dropped on
@hf.co 🔥
huggingface.co/collections/...
✨ 0.8B/2B/4B/9B
✨ Apache2.0
✨ 262K→1M token context
Started writing an agent in bash (yes bash:) for a potential training course, got it to a point of a little chatbot and found it super useful to quickly debug / check things - so here it is, maybe useful for someone else too: github.com/ayourtch/age... - PRs welcome !
E.g.: spread the same usage but make the timer 8h, and align it with circadian cycles of the account, based on the account‘s token use ? At this scale, this can have a non-trivial positive health impact … (and keep the Max plan the same, obviously).
A thought to @anthropic.com - maybe it’s worth thinking about how to tweak the rate limiting timers ? such that the limiters on the Max plan could enforce the “healthy” life style rather than break it ?
Excellent post by @minimaxir.bsky.social. It seems as though "re-write everything in Rust" is now possible with AI 😀
minimaxir.com/2026/02/ai-a...
The human alignment problem has surpassed the AI alignment problem in importance.
This is really really cool! It didn’t work on my consumer GPUs, but i also have a DGX Spark, I made a couple of repos with adaptations of your code, in case anyone finds useful: github.com/ayourtch-llm... & github.com/ayourtch-llm...
The shared memory is a bit of a bottleneck, but there is 128G of it, and 2x200Gbps + 1x10Gbps network ports, which as a network geek I find very cool :D
Admittedly I just grabbed github.com/eugr/spark-v... as I "just" needed something to put an LLM on, it does build off Nvidia patched torch; but IIRC I did compile llama.cpp and ran it with no problem. I switched to VLLM because 2xDGX give tensor parallelism for inference - else it is a bit slow.
I made a little demo of doc2lora running on a DGX Spark, its super fun to experiment with! : m.youtube.com/watch?v=M5DP...
I'm not kidding, cancel ChatGPT.
