Thinking about Satoru Tsuda who probably lost his job as a cat dresser because of AI
If Google put as much energy into improving its AI models as they do into UI redesigns nobody asked for, we’d have AGI by now
I usually scoff at radical takes, but it turns out I have one: I don't trust recommendation feeds On YouTube I only check Subscriptions. On X only Following. The rest is newsletters and direct visits It takes an initial investment of time but it protects independent thinking
Berkeley tested 21 model-harness pairs: minimal harnesses cut token costs in half with zero drop in accuracy. Interesting results, but take them with a grain of salt. A benchmark test is very different from a real dev workflow.
I don't get this "AI guilt": building a tool with Claude in 3h that saves a company $2k/mo, then charging $20 because you didn't work hard enough is crazy Customers pay for the headache you remove or value you deliver, they don't care if you built from your phone at the beach
running a task on Fable basically burns 10% of weekly credits in a jiffy, with barely any moat over other models and harnesses. i don't see how this is sustainable: if open-weights keep closing the gap, i don't think they can keep these prices up. long live the competition
hearing a lot about these elusive software factories. maybe i'm missing something? what's working for me: • parallel workers over linear chains • a shared roadmap file to log decisions + state • tiered adversarial reviewers (code, UX/UI, customer personas..) to check the output
there's been a lot of debate around Dario Amodei's essay on pacing frontier AI, so i decided to join the conversation. i wrote down my thoughts on my blog: https://christianvismara.com/blog/the-prisoner-s-dilemma-of-frontier-ai
The Prisoner
When I saw the essay that Dario Amodei published about pacing frontier AI, my first reaction was to reject it, but not based on the content but based on a sense of distrust for the person that wrote it himself.
christianvismara.com
knowing the character my first reaction is to distrust Amodei, but pre-release review of frontier models isn't a bad idea. we're in a prisoner's dilemma and everyone protects their own interests. until a lab eats into its own profit to slow down, calls to pace AI are just PR.
Frontier labs claiming they’re solving math, replacing software engineers, and building tech that could end the world yet they still haven’t solved the hardest problem of the 2000s: how not to get copied by China.
Watching @karpathy use Opus to turn Tolkien into a 3D scene is weirdly cute, but now imagine listening to an audiobook while exploring the scene, moving the camera, checking next door. Cheap 3D can turn books into a medium built for the overstimulated brains of the late 2020s
people ask me how i review code these days, the thing is, i don't. i haven't opened VSCode in months, human review catches 15-30% of bugs at best, a proper LLM pipeline matches or beats that, and i move 10x faster. i'd be a huge bottleneck
this is what it looks like, weirdly mesmerizing
Christian V@imcbv.bsky.social · 2mo agoagent swarms just built a walkable 3D Kyoto simulation from open data. this is basically one step closer to predicting the future: computing infinite parallel worlds from a single point in time to bruteforce test all possibilities https://github.com/PhiloLabs/fable51-worlds
agent swarms just built a walkable 3D Kyoto simulation from open data. this is basically one step closer to predicting the future: computing infinite parallel worlds from a single point in time to bruteforce test all possibilities https://github.com/PhiloLabs/fable51-worlds
GitHub - PhiloLabs/fable51-worlds: worlds via code, from fable 5.1
worlds via code, from fable 5.1. Contribute to PhiloLabs/fable51-worlds development by creating an account on GitHub.
github.com
Most of my day is in Codex and Claude Code. @simonw found Codex ships a full copy of LibreOffice, Python, Node, git in its cache: 1.7GB of binaries. The harness is becoming THE workstation, give it two years and macOS is just the bootloader
If AI music were actually bad, you wouldn't need to ban it from the charts, you'd just outperform it. Australia banning it instead tells me either the AI is better than musicians want to admit, or the musicians are worse than they want to admit. What's the truth here
They trained an AI to paint by writing JavaScript and the results are quite beautiful. (This isn't diffusion, just brushstroke math and code)
The author isn't whoever typed the words, the author is whoever takes responsibility for them. A machine never will.
Imagine selling a technology as something that might end humanity and displace workers and then being surprised most Americans are skeptical. Amodei calls it a crisis of trust, I call it cause and effect.
my approach to LLMs: a full on stream of consciousness, regurgitating the plan for ten minutes. while everyone's selling over-optimized workflows, the old man yelling at clouds approach remains best. even @karpathy.bsky.social uses it and calls it a ramble session, nice to be validated 👴🏻☁️
so @anthropic.com marketing finally went from "our top model is gonna destroy the world" to "our most powerful model can do math." that's a step up
the irony: schools worried about students using AI are using AI to judge them, although even Turnitin says don't use them for grades. it's a lazy fix to laziness: much better to teach AI-augmented thinking and test the understanding of what students produce
Ok we get it you guys don’t like the new Ferrari but most of you can’t afford it either, so what are you even talking about?