OpenAI's roon argues faster software and UIs are required to prevent Amdahl's law from bottlenecking rapid AI models
Garry Tan asks if this will drive Rust adoption.
Many users stress the need for faster builds and interfaces to match accelerating AI models because speed improves usability for agents and enables future applications, while others reply with personal insults and hostility.
Most Activity
@tszzl So it’s gonna be like all Rust and Erlang from here?
people have to build much faster software and interfaces so as to not Amdahls law the increasingly powerful and increasingly fast models
@garrytan we’re gonna a need a whole lot of cacheing
@tszzl We need to stop building Software as a Service. We should build Software as a Servant.
APIs should evolve into Agent-Powered Integration.
@tszzl what are the biggest areas of improvement?
there's the physical piece of networking, topology, etc.
then I know less about the software/harness layer you're talking about. educate us pls!
@tszzl i think we need to start doing it at the foundational pieces.
operating systems, compilers, browsers, etc
@garrytan @tszzl Elixir and Rust together!
https://fly.io/phoenix-files/elixir-and-rust-is-a-good-mix/
@tszzl All that foom theorizing and they forgot about Amdahls law
@tszzl We should also aim to turn more traditional domains neural since models perf & power improves much faster than trad code. FSD didn't end up as 300M lines of LLM-written C++
@tszzl The model is rarely the bottleneck anymore. It's file I/O, shell execution, browser automation, tool calls. The interfaces around the model are what Amdahl's law actually hits.
@doodlestein @tszzl Franken-transformers lib when @doodlestein ? 👿
@tszzl @garrytan
@tszzl Rust compiler needs to be faster or else we’re gonna have to switch to go, which later models will surely punish us for
@tszzl Honestly, I don’t think most people will need apps.
They’ll tell the chatbot what they want and it’ll just happen.
That’s why I’m 42, staying in corporate, and waiting for your AI abundance. Cant wait to have my own robot.
@tszzl Yup. Hence my FrankenSuite efforts. No more Python.
@JungleSilicon @tszzl Databases
@tszzl people have to do faster model so as not Amdahls law the autoresearch agent running on 8xB200 gpus
@tszzl @grok please elaborate on what Amdahl’s law is and how it applies to LLM model performance and software
@Jeremy_AI_ @tszzl Oh shut up jeremy, I am working using codex. And codex is down so bad fr right now
@garrytan @tszzl i wish :( subjective experience is coding models are ~6mo behind in Rust vs. Python. That and the fact that compilation times are weirdly a bottleneck now
@tszzl Traditional software still cannot escape the bottleneck even with AI. We need to design better software instead of just faster.