puffybsd (@puffybsd) header
puffybsd (@puffybsd) profile picture

Protein Language Models, Functional Programming, GPUs, HW Tinkering, Arch Linux, Mechanical Keyboards.

Joined May 2008
Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)Photo by puffybsd (@puffybsd)
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?
4
1
1
20
17,750
Has anyone created a derivative programming language by using an llm tool to remove or constrain features to eliminate ambiguity? Subtractive language design. Strict subset. Contrastive loss? Inverse reconstruction loss? GAN style loss functions?
3
3
80
$ANT to zero
Woaaa, faster than I anticipated. LOL! Remember yesterday when Silicon Valley access journalists were bragging “demand was huge” for Anthropic’s IPO? Barely 24 hours later, $ANT has dumped from its $111.50 opening print to $85, and Morgan Stanley already burned through its entire 147M-share greenshoe option on Day 2 instead of holding the 30-day window to defend the $83 floor. Underwriter got paid, Anthropic got its cash, and retail FOMO buyers got left holding a 23% bag with the safety net ripped out. In plain terms, the safety net under the stock is gone. Anyone who bought near Thursday's $111.50 opening print (or $115.10 peak) is already sitting on a roughly 23% to 26% loss, and if $ANT tests or breaks below its $83 IPO floor next week, Morgan Stanley will not be stepping in with stabilization money to catch it.
1
2
104
"POSIX sh, not bash — no bashisms needed." -Qwen Flash Next (em dash, because it keeps it real)
2
51
qwen flash next fixed a segfaulting bambu-studio cert issue and login issue with some computer use and other goodness.
2
28
qwen flash next is great at git-fu.
1
44
So bad at writing cuda kernels figured I'd try writing some hip ones.
29
rocm.docs.amd.com 's configurator for torch that mimic's torch's page works well, just replace python -m with uv and it's good go to go. Picking specific processors is great. rocm has come a long way since 2021, thankfully.
34
Interesting that there's so much advanced intelligence and Amazon is still like: 'You wanted a 14" privacy screen, how about this 16" one instead?'
2
4
201
Setting up rocm on strix point and seems like uv pip install is the way for uv users. Frustrating that there were no guides that show this, given that most projects are uv based these days.
1
34
puffybsd retweeted
Detaining a PhD student on her way back to Princeton from presenting scientific work at a conference is not how we get the best and brightest scientists in the world to want to come to the US. I hope this is resolved quickly! More details: dailyprincetonian.com/articl…
Our colleague Eva Yi Xie, a Princeton neuroscience PhD student, was detained by ICE at Reagan National Airport on her way home from SfN, where she’d just presented her poster. She’s being held in Chantilly, VA. Please help us bring her home. 📷
404
664
53
6,907
547,080
These ai celebrity dance videos have become vomit-inducing. "Nauseating" is not visceral enough.
35
Started reviewing RL again, just in time to see HF putting out some awesome RL fine tuning models today.
39
brave and librewolf are pretty good browsers.
50
It feels like the thing that happened to deep learning people is happening to the formal methods folks.
2
171
People forget nvidia was at $11.20 in Oct '22.
1
1
98
ROG z13 flow is tempting. 64GB is on sale, but nobody want that. 128GB needs to go on sale.
1
1
55
A pair of tenstorrent p150a blackholes with a qsfp cable sounds at least as much fun. Would need a home though. Have a few options that might not involve growing the armada.
1
54
Showing a friend some jupyter notebooks from early deep learning edu and research stuff. thinking: "damn I wrote some clean and organized code, with pride of ownership and discipline. Well, kind of sad these will no longer exist except as a form of meditation or katas" (there's a fugazi song about this, pretty certain)
2
54
From the perspective of someone who always hated reading their own, awful code. In hindsight, not as terrible.
11
Working on discipline and security in the homelab. Proud of some of the improvements I've made lately, though really far to go. It's also amazingly scary how poor security posture is across modern tooling and components, most of which are backed by large orgs and hyperscalers.
1
32