@dexhorthy@walden_yan@tobi@karpathy Agent swarm is a terrible term for two reasons.
Scary sounding and implies a very stupid approach is actually the goal.
Love agent soup term.
Workflows of inner agent loops is exactly how I do my stuff. Thanks for the two word shutdown for counter arguments.
I finally found *the* solution I wanted to the old/new editing problem. And it is a solution that at the same time works extremely well, is quite elegant I believe, and can't be implemented if you don't build something like DwarfStar. Thread (but check [upto] in the screenshot).
“Flash-level” local frontier at ~27 t/s (~500 prefill) thanks to @antirez on my m5 128gb.
However, integrated harness/inference/sampling loop might be the true unlock.
It lightens the load on the model and corrects its work much more efficiently.
Check out the [upto] injected control op in the replies. I can imagine many more features to lighten the burden on the model.
Eliminating http overhead and inspectable/steerable tool calls (basically intellisense for LLMs) might make these models smarter and faster at doing real work than much bigger/better models.
I'll tell you one thing that shows a fundamental advantage of the "agent handles the LLM inference" in the specific case of local single user agents/LLMs. You can stop the tool and report an error as soon as in the first parameters of the tool calling you detect a problem. No need to wait for a long "content" generation that will fail immediately after.
@antirez Okay, this is getting interesting now. I've spent a lot of my life abusing this algorithm. This might be the unlock I've been looking for. https://t.co/2eefZNSvXS
@TheT8or@LLMJunky Follow @antirez and @ivanfioravanti … This is what you want to be following for your m5 128 … or at least that’s what I am doing :) https://t.co/HynPyHdPpR
Translating this to @badlogicgames's Pi architecture:
- antigravity CLI is "pi-coding-agent: Interactive coding agent CLI"
- antigravity harness is "pi-agent-core: Agent runtime with tool calling and state management"
- gemini LLM is "pi-ai: Unified multi-provider LLM API (OpenAI, Anthropic, Google, …)"
https://t.co/w4HfZFLFx5
@badlogicgames I asked my clanker to assess how my code is going to have to change based on the WIP SDK code inside of Pi.
Apologies if this isn't useful slop.
hey surprise - you can just launch interactive in tmux and then tail the jsonl - shipped a small wrapper...ralph loop iterating to full parity rn https://t.co/3N4klSSEwd
@hk_net@GaryMarcus I thought this was neurosymbolic then and thought Gary was right about what was needed. The neuro part wasn’t smart enough to do what it can do today, but building on the fly bash tools definitely seemed inevitable.
@hk_net@GaryMarcus I saw it do something harder… use wolfram alpha with a ChatGPT plugin (effectively a “skill”) in April 23 with plain English “integration code” https://t.co/KKhmH8tvyF
I think it comes down to this quote applied to all the sv technorati from the show Silicon Valley :
Gavin Belson: "I don't know about you people, but I don't want to live in a world where someone else makes the world a better place better than we do."
That and Demis told Elon that his AI would follow him to Mars