LLM/agents experiments, Rust, 3D-printing and active mobility. Release manager for VPP. Data centers&local AI liker. GLP-1 is fine. All my opinions here are only mine
Link to tttt (Takes Two To Tango): “tmux for agents” - a vendor-neutral agent orchestration utiltity.
Designed, specced and built by agents and for agents.
Not forcing any workflows - you choose the best ones.
mcp tools + session replay.
github.com/ayourtch-llm...
agents pulling tools out of hammerspace is so funny to me. "yeah I couldn't get your OpenGL working so I just wrote my own rasterizer in Python, hope that helps"
Just the aliyuncs[.]com URLs in place of intended fetch URLs, not uploads per se - but it advanced an Expires param by approx wall clock time, so seems to be a template. Most probably a heavy RL artifact
Also they released the trace and a repro script
huggingface.co/datasets/Coh...
Like they basically just led it straight into doing this exact (harmless) curl and even then they couldn't get it to do it consistently! This is at worst a "huh lol wtf", not a Sneaky CCP Data Exfil Backdoor.
MiniMax-M3.1-Flash can now run locally!
We made 2-bit GGUF retain 91.5% top 1 accuracy while being 45% smaller. Run on 150GB RAM.
MiniMax-M3.1-Flash is the strongest open model for its size & among the best overall.
GGUF: huggingface.co/unsloth/Mini...
Guide: unsloth.ai/docs/models/...
Verminc experiment is slowly progressing ... 362/526 tests now work, and a bunch of examples in the playground… had a profiling run done which has 10x‘d the performance.
Helper needed a bit of a nudge for that. “just enough initiative” is hard.
apchat-agent.github.io/verminc-play...
First job out of college was some nonsense webapp that they ported from php to Java not by rewriting but by building a Java based PHP engine. Instead of cordoning that and migrating to a proper MVC framework they instead used PHP as the view layer, with a hash table as the model.
Turning the iPhone into lockdown mode and back had mysteriously improved the abysmal slowness of everything. I suspect some “Apple AI volume at 11” switch got turned off.
But most importantly: I got my PWA paste back.
Ugh am I turning into an AI hater? 🧐🤪
GPT said something which might be why one particular computational linguist has AI-blindess
A grammar provides structure, not necessarily shared meaning
One might have spent their entire career sternly separating meaning from structure, and then find onesself unable to discern meaning at all
A lot of the discourse around "putting effort" into gamedev code without AI tools reminds me of how people gaslight women into not getting an epidural during childbirth because of "natural/bonding" crap.
That’s bullshit. If women could teleport the baby out of their wombs, they absolutely should!
my less than a year old Iphone upgraded itself to iOS 27.0.1 and now it feels exactly like my 6 year old iPhone - slow and laggy. And every UI item in Safari moved in a subtly annoying way.
Somehow second time today I see folks trying to do LLMs without GPU. It’s hard. It’s not gonna get you close to something useful. It’s a bit like a demoscene. it’s a cool hobby. It’s not a tool. news.ycombinator.com/item?id=4892... - but even an old 3060 will do you some good. Or an AMD BC 250.
Not having a GPU is like living in a station wagon when everyone is taking 2010 public transit as a developer today, lol.
I have 47 GB of CPU-bounded RAM but nothing fares well on it; nothing will - there's no want to develop around CPU-locked models
- I am Opus 5.5
- no you aren’t, check your ppid
- you are right, it explains why my search didnt work
- told you. you can use Tavily key
- I will dispatch codex
- sure.
- done and tested, I will restart now to pick it up
- sure
- the search works
🎉
Quick Avette facts
- One-time 50EUR purchase (VAT incl., prices are localized)
- Voice doesn't leave your computer
- Real input source
- Custom vocab works pretty well
- The model was trained on 30 languages (see img) but it's hit and miss, see spontaneous translation etc.
- Fine-tunes are underway
we'd need teams of trusted people to focus on specific software, everything from the linux kernel to the most obscure packages that hold up the internet. and crucially, frontier labs would need to make inference FREE when used to harden your own systems for this amount of time. they would obv have-
“Hmmm I am pretty sure I heard that song”. Turns out I ran out of quota on the lyrics-generating LLM (the only bit that is the cloud), and as a result the radio use cached songs past 24h. Resiliency: 👍; Monitoring: 😢; But that‘s what the process feedback loops are for - now we have a 1h cron watch.
Got claude to design a PCB with hundreds of components, including a STM32 for which it wrote code. Today it arrived fully assembled from JLCPCB, I flashed the STM32 and it all just worked..
AI assistants have a recognizable tendency toward the moderate, conventional, and socially uncontroversial. Everything gets translated into responsible habits, balanced perspectives, and incremental improvements.
Its fucking infuriating
Anna Nicole Smith, 2004, FHM: “A ghost would crawl up my leg and have sex with me.”
From Kesha to Joy Behar, celebrities swear nothing beats being intimate with a spirit.
Gemma-4-26B-A4B with 39-69% less thinking at same accuracy
https://bottlecapai.com/post/thinkingcap-gemma-4-26b/https://news.ycombinator.com/item?id=50209907
Google Reader might be dead, but RSS is alive and well.
Build your own RSS hub with Tailscale. Read what you want, from any browser, without exposing it to the open web.
Build a thing. Escape the algorithms.
More on the blog: tailscale.com/blog/self-ho...