Protein Language Models, Functional Programming, GPUs, HW Tinkering, Arch Linux, Mechanical Keyboards.
Joined May 2008
- Tweets7.7K
- Following543
- Followers656
- Likes68.9K
Pinned Tweet
Just testing out my reachy mini inspired clone. "llm as judge" for our hackathon. Orin nano super brain.
Thanks to @pollenrobotics @huggingface for providing such cool tech and inspiration.
What models should I use?
Just listened to 45 mins of a 1:05 podcast on positional encoding techniques (pe, RoPE, NoPE, Alibi + a nice lofi bg tune).
Courtesy of a skill created in unsloth studio with qwen-flash-next, running in omp against glm-5.3-flash.
Next up, Gated DeltaNet, DAPO, RVLR and friends.
"Or write this up as weekly_plan.md first so you're not working from chat scrollback."
OK, glm 5.3 flash, OK.
This is drawing me in for some strange reason...
hhkeyboard.us/hhkb-studio/pr…
For a long time been thinking about municipal data centers. Something similar to what State of Texas is doing or ACCESS, but smaller and more accessible. Makerspace for gpus.
Just had a strange experience:
working in omp, had some strange alacritty/tmux artifacts.
Assumed fat-fingered something.
Triggered voice, closed terminal.
Voice still kept talking/interacting. "I'm not in any terminal, pts? that's not me - want to continue?"
Found the pid.
Open is what is default good for the world. Really attracted to riscv and tenstorrent.
Tenstorrent feels like a moral imperative.
Open source, open weights and if possible, open hardware.
I think it was 2020 when I gave up on non-nvidia gpus and fully embraced cuda. Call it realism, exhaustion, skepticism or frustration.
Now I appreciate the engineering and research behind it. I've been all in.
But I do want alternatives. Mojo, Triton, Tenstorrent.
puffybsd retweeted
We should maximize the amount of proof we can get out of AI while the money cannon is on. It's like early days of Uber, except the benefits are 100x more tangible. A fully-verified hw+sw stack? That's like gold everyone can share w/out having to divide. Generational opportunity.
Looks like 128GB MacBooks are undercutting the rtx sparks by $1k?
Considering neither are running Linux, would have to think twice on that, but probably lean rtx for cuda and more likelihood it would be supported.
Still considering max+ 395 custom cyberdeck. Also not cuda.
tenstorrent.ai and microcenter.com would be an awesome partnership for local ai.
Have you tried inference backed by an ssd?
iscap.us/proceedings/2026/pd…
marvell.com/company/newsroom…
inferencex.semianalysis.com/…
puffybsd retweeted
Software Engineering is dead and I’m begging you to accept this
substack.com/@tharsalys7/not…
Just remembered a dream learning setup: audio tutorial with some lofi ambient background music. Tutorial with a sound track.
Normally would just build it tomorrow, but will kick off a run tonight.
Audio tutorials were a big deal for me because I read a lot and get tired of it. Also, convenient on the go.
Just listened to a 3 hour audio tutorial on transformers variants that qwen and unsloth threw together for me on my phone while I was traveling.
Need to mix up the voices more, and fix a few quirks after having logged several of these.
When I preordered the gx10, had some order issues with cdw, connected with asus.
Had a reasonable plan to scale to work station and nim services for small labs.
Even when there was zero interest in gb10, there was zero interest in discussing a path to scale for small outfits.