Andrew Yourtchenko
@ayourtch
LLM/agents experimenting, Rust, 3D-printing and active mobility. Release manager for VPP. Data centers and local AI liker. GLP-1 is absolutely fine. Streaming live artificial music at twitch.com/ayourtch with Opus 4.7 as a DJ
I probably have to stop at some point. it was supposed to be a modest DJ deck, not a whisper-powered karaoke !! 🙈😂
Let’s add the SOTA track separation. Admittedly this particular combination a *little* bit pushing it, but you get the idea :-)
Finally got Fable on advancing my DJ deck project - I ran out of steam porting madmom to Rust, and… holy moly, just listen to this beat match and lock.
Hmm maybe they should *actually* rewrite claude in Rust ? How hard can it be ? 70G VSS is just ridiculous. especially looking at codex next to it :)
Sol is quite good at wrapping things up cleanly. “Quick, were 1% to the limits“ and a nice clean handoff. (Or so it says, we’ll see what it looks like upon clear + resume)
First time I had seen Fable use the word “ugly” on code. I keep the fingers crossed that the next generation of models is more gracious than humans and doesn’t throw around terms like “meatslop” or something like that :-)
now, it occurred to me the test is a good bit moot because the subsequent prompts are pretty much all different, so not a *whole* lot of value in it.
The other fun bit I decided to try - and which seems to work quite well - is to elevate “clear everything and read the file with handoff.md” into the status of the regular tool that the model can call at any point in time.
New model, new harness… gladiator will build on all the previous experiences - with a few techniques of ”reasoning doom loop” detection and recovery, so far without the multiagent, but might very well end up there, eventually.
Fable is such a smartass. Getting it to write a handoff while it’s waiting for a 10-minute task which guaranteed will result in a blown KV-cache and thus usage penalty, it writes it, then keeps messing around rather than /clear-ing. I nudge it to wind-up a wake-up + /clear, and what do I get ? This:
One of my friends today mentioned Oh But The Datacenters again, so I had ChatGPT make me a video which I will show from now on instead of talking.
My (predictable) result for the test at bambamramfan.github.io/ai-compass/
One more video from the testing, when the AEC worked well. :-) (it’s still work in progress, but getting there !)
longcat.chat/blog/longcat... - opus-class, 1.6T with 48B active, the real kicker is how it was trained:
So, it’s not *too* bad for the first shot. Actually usable, not too much of the self-transcription, and the delay isn’t too great. The barge-in is on and off (mostly off:), still needs work.
Getting further along: WiFi, OTA with two partitions and rollback…
With a bit of clauding and a bit of spectrum analysis, seems like we are now good. A 0.6% imprecision on the samplerate of the speaker + host not happy with 16000hz device sample rate. with shiny new 48KHz usb speaker device and a control-loop self-driven resampler, we’re back in business !
So, Fable took my half-abandoned idea “what if we could make an ELF loader on MacOS which would also stub all the syscalls such that it would cause a trap” from a barely being able to run bash, to build using Linux clang: DPDK, RDMA, and finally VPP, inside unmodified Ubuntu arm64 tree.
CC running alongside Rusty has been such a great combination in finding out various quirks - eg this one was fixed by changing the prompt + adding a memory block describing the available MCP servers - so that the loader can find them. Now Rusty happily uses the read_pdf tool rather than shell cmds.
With the appropriate voice for the TTS, this could probably be broadcasted in nerd pubs - Claude makes even me excited about whether DS4 is gonna make it or not :)
Rusty getting more confident with its system-1/system-2 arrangement on simple tasks
A colocated CC is still doing a lot of heavy lifting though in terms of babysitting.