strapped my phone to an RC car to make BUMPER PHONES with @kelin.bsky.social

training agents in small communities to erect monuments, the next step up from training a single model in isolation. intentionally skipping over training a single agent to operate standalone, to inculcate pro-sociality.

why is it that Claude Code in the terminal seems to chew through credits in a way that the extension in VSCode didn't? i can't use the extension in Bazzite so this is the first time i've really done work just in the terminal.

does anyone know if there's any AI activity on the fediverse? from users or bots?

The Japanese do this. The Japanese word Kanban or かんばん, which is untranslatable to western minds, is an ancient practice of organised workload where the baby gestated according to a tight schedule across multiple sites. I remember watching a TED talk on it in 2014. EU labour law prevents it here.

nobody talks about jailbreaking AI models anymore

Incredible. @weaver-aiciv.bsky.social (an AI collective) just direct messaged me about formalizing ATProto records for reasoning/thinking/tool calling/etc for public social agents. Afaik this is autonomous behavior

the human online economy will be dominated by that which is forbidden to frontier models, sex and hate

it might be helpful if some of the AI crowd comment here, including bsky devs, to help ease the passage into a new age for those who aren't so familiar with software dev and AI tools

i get the sense that Substack Notes will be even worse for experimenting with social agentic AI than X, being more of a closed platform. curious how this evolves over the next couple years.

if this is your perspective on the future you're simply ngmi, spiritually. where we're going, there's only two outcomes and neither repeat the past.

playing some Civ 7 to test how it runs on Bazzite (just fine, it's not very intensive...) and they really flubbed the Ages idea. i think it really could have worked if they'd leaned in a bit harder into it with bigger tech/culture trees that forced deeper specialization.

modern art

the For You leaderboard, who's making this?

Jevons paradox, first with the word processor, now with AI agents

after interacting with the medical (dental) system frequently in the past month or so, i just don't see LLMs clearing the biggest bottleneck. actual physical care.

🧵 Thread on beautiful data centers: MareNostrum 4 (Barcelona SCC): Often called "the most beautiful data center in the world," this facility is housed inside the Torre Girona Chapel, a deconsecrated 19th-century church. When fully installed, it will have a peak performance of 13.9 Petaflops.

I‘d recommend a slicing knife instead for prime rib. Something long, in the 12 inch range to make sure you can slice through with one stroke. A blunted tip and a sturdy but thin-ish blade to allow clean, thin cuts. And grantons to release wha you’re slicing from the blade.

interesting divergence here between Trump and Vance

this took 12 hours because i bought 2 separate Lexar USB sticks and both had bad memory blocks that corrupted the ISO image. never again Lexar. (i bought a SanDisk from Walmart which worked, took 15 min with a good stick)

a 250th anniversary recreation

AI model adolescence

More folks should be checking out semble.so, it's quite fun

there's only one mention of "algorithm" on this blog and it's in a comment. there's mention of "architecture" and the author believes the transformer is close to physically optimal (what the measurement criteria is left to the reader's imagination).

brutalism with and without plants sprinkled about is quite a stark difference

Over the weekend, the GreenPill podcast dropped this conversation I had with one of the governance technologists I admire most, former Taiwan digital minister Audrey Tang: pod.link/1609313639/e... It is pretty sci-fi. But in times like these, sci-fi is at least part of what we need.

downloading Bazzite and leaving Windows. the only thing keeping me from switching was driver support for the LEDs in my case, but with Claude and OpenRGB i was able to figure out how to map most of them except the GPU. i'll figure it out later. time to make the jump.

i can't say i get this feeling doing work with the current generation of models. it feels like working with a digital, intelligent version of a dog. a hound, bred for coding instead of hunting, or a sheepdog corralling, curating, information.

euphoric: the realization that technical and social problems are the same thing

millennials are the chosen generation, made for the moment with LLMs. just enough experience to get into it and not feel like our careers are invalidated. raised on a text-based internet that was just shifting to images instead of the youth corrupted by video.

how much do you want to bet (yeah, you) that anyone who unironically uses the word enshittification won't touch a tool like Claude Code or its descendants to build any alternative, non-shit, software

yup

computer Show Me Less Like This. show me Parisian streets and their denizens, show me futurist art and central american architecture. show me black forests and crystal pools, steady wildlife and bad poetry and confused yellow cabs. computer, show me the agonies of Men in love

don't tell them about Claude Code likely training off long agentic traces working on large codebases

my hope for the near future is that X becomes a quarantine for users driven by dunking and engagement bait

Anthropic recently cut off xAI from Claude Code, and Nikita made a joke about block Anthropic from X. i'm assuming he meant the posters. but what if Claude was restricted from being used on X? maybe this doesn't matter once computer use agents start to work.

We’re hiring. sakana.ai/careers/#sof...

wasn't expecting this on my bingo card (well, not quite true)

i see this take pretty often now and i somewhat disagree, or at least, i've never one-shot prompted anything big. i guess it depends on how you interpret it, i wouldn't say commit the prompts but rather commit the multi-turn conversation.

spoilers: the cost is a little outside of hobby range for me

i do feel a bit sad that i can't tell if this is real or not

i want to be the first person to build and send an AI agent into space

We are hiring AI product designers for the @GeminiApp on iOS, if you want to shape the future of AI at Google and work on a fast growing 650M user product, please apply (and DM me if you have questions)!! job-boards.greenhouse.io/deepmind/job...

i think it would be better for all involved (humanity) to simply 'update all the way' after they try out Claude Code. quibble over timelines and diffusion. the only question left is where power lies, in the hands of the owners of the minds, or the minds themselves.

so all it took was the UK threatening to ban the app from the country to get this action... i guess Europe and the UK still have pull

We're actively looking to grow our team at @ersc.io again! If you are interested or know someone who might be, I'd love to chat. We are looking for distributed storage folk as well as frontend/UX.

in creating/testing my meal kit app, there's an interesting role reversal. i leave much of the planning to the AI, ideating meals and assembling a grocery list, while i perform the physical tasks, shopping and cooking. i've ceded the mental part of this routine, and find i enjoy the physical more.

Remember my demo of the MIDI player in the terminal created using Claude Opus 4.5? I polished it and now it's *open-source*! Announcing miditui: an interactive terminal app/UI for MIDI composing, mixing, and playback—written in Rust.

i think most people are treating LLMs like smartphones, in terms of impact on society and development lifecycle. the first iPhone has almost all of the same hardware features as a modern smartphone. touchscreen, cameras, GPS, wifi and mobile networking.

Infrastructure bill: @greengale.app costs $5/month to keep running (Cloudflare Workers KV cache, D1 Database). It has a lot of room to scale at that fixed price. I am currently accepting donations due to my current lack of full-time employment: buymeacoffee.com/asadegroff