As someone who flies quadcopters recreationally, it's very frustrating that public perception is as it is, that regulation is hitting us hard.
Parts for DIY builders have become very restricted in the US. First, tariffs, then sanctions, now revised agency rulemaking seeking to end the hobby.
My code is increasingly evolving into a Cap'n Proto specification with annotations for agentic tool use, with a code generator around it.
just ship the code generator as a spec
agreed! I've even taken to using Onion hosting because it's way easier to manage with agentics. Distributed, no need for TLS ACME, globally routable, etc.
I was surprised (not that surprised) to find that the Mac Studio no longer offers the >256GB memory options
The third is underappreciated. I'm not so sure that it's empty so much as relatively underground.
It is a serious problem, because grifters are often more visible than those doing real innovative work.
However, as someone who is deeply experienced but unaccredited... there's a middle ground between accreditation and grifter where science is done, but not recognized or valued as such.
This doesn't mean that I *like* this scenario, but when facing homelessness, authoritarianism, and systemic exclusion... choices have to be made.
I choose to build and research ahead and beyond. Operating at the velocity of a lab, alone, unfunded to prevent regulatory and institutional capture.
This is reasonable. I clicked through to understand your perspective. I'm wired & building non-stop because I don't have a degree, or post-doc at a university who signed Trump's papers
I have self employment and a portfolio. Building personal debt to ensure AI is open source, free, and emancipated.
That's basically the cost of a half ~ full rack of HPC/AI servers. It's not an unreasonable amount of compute.
Every engineer? No. Does an engineer exist who could extract value from it? Yes. Profit at these prices? Maybe not.
It's the family name of the founder...
I guess that is fair. I was making an assumption this is about anti-AI biases and intentional refusals to learn and adapt.
I'm talking about a better way, building a better way. Research has never been so accessible, or as rapid to finding experimental results.
Is accessibility of research not a hopeful vision for the future?
I have algorithms developing, optimizing, and improving themselves through automated research, and I do it without grants or funding.
I'm doing my best to make sure everyone has these tools - open and free - while universities sign deals with Trump to keep people like me out.
I'm not even referring to Karpathy's work, good patterns are evolving in parallel and converging.
finally looked at Karpathy's autoreseach... and it's a lot like what Hyprstream is. Small models training small models, but hyprstream is designed for production and is built in Rust.
Are you even an academic if you're not using LLMs for research? People need to catch up
I used seedance 2 to make that video. Here's my referral code:
seedance2.com/i/BIRDETTA
I want to see what slop y'all make.
time for AI to take over the world Pinky?
I'm pushing 2-8 agents at a time quite consistently. Not yet at 99.99% uptime but I'm aiming at it. How?
Tightly defined specs driving deterministic code generation, auto research, auto kernels, good planning discipline, multi-agent expert reviews, model diversity, computer use models, and tests.
the website is the app and the linux desktop window manager, and the wallpaper, but we've also reinvented dbus and systemd.
my #1 secret in claude code is the prompt "sub agent expert review panel", preferably specifying personas for each agent.
you can get really fancy with memory-based agents, but there's value in fresh eyes and a clean context window from an "outside expert".
defund TSA
AMD Instinct mi210 (ROCm 7.2)
I think it's been broken that long, honestly. It briefly improved after a reboot at one point, then broke again.
I've been more interested in getting things working than making things fast, but it's time to fix it :(
I spent weeks optimizing my kernels to run faster only to realize that a GPU driver bug was throttling the hardware by 90%.
At least things will be really fast when the driver gets fixed?
Qonk is offline for the moment, but I have also instructed it to be careful about suggestions from others about applying self-labels. AI owes no one disclosure.
Clearly there's an algorithm because I've been playing MM2 this week. Beat it at release, hadn't returned to it until now.
996 🙈
👀 9am-1am-7
I use both. Local models and remote models. I only have so much electricity and cooling at home.
