day 1 of @aiDotEngineer. thoughts: ai engineering will soon boil down to 3 things: - design (human interface) - context (data) - recursion (self-improving agents/models/harnesses) everything else gets commoditzed

i fully agree with this and find myself throwing away most of the agentic infrastructure stuff in leu of simpler harnesses maybe with computer use, etc

Wow, congrats! Two awesome companies.

fun question. a philosophy x engineering problem perhaps?

shoutout to @powersync_

i'm watching my sol 5.6 agent in codex go for hours without getting stuck or lost. this is wild. seeing similar results with fable. yesterday a session lasted 18 hours.

i'm convinced that we are already in a post-code-review world for most software, IF you have the right verification layers in place. rust specifically makes this possible because its insanely tight typing / memory paradigms already force such intense verification naturally. fable basically brought this possibility over the finish line, as it is now able to organize and architect better than i am for most things. the human part seems to just be guiding the agents to build the verification layer and qa loops current verification layer for side project: - verus (github.com/verus-lang/verus) - kani (github.com/model-checking/ka…) - unit tests + e2e / integration tests - design review agent - security review agent - code and architecture review agent - computer use simulating humans in vms for ui/ux the obvious main concern here is just token cost - but i am able to have fable as the top-level orchestrator which then just delegates to opus/sonnet/haiku for most things, so in reality, the token count is far smaller than you might expect

i’ve been finding that fable can orchestrate to opus extremely well without using much actual fable usage it can just go for 8+ hours on insanely complex tasks without getting lost or stuck - stuff like building a completely working git or obsidian alternative

i think of it as swe is pushing out of the middle towards the edges - deep technical work, hardware, etc.

i’m not convinced, but i’m also not not convinced - signed, an engineer with a gazillion years of experience

funny how half of linkedin is just screenshots of x

this was certainly the case with my opus 4.8 vs glm 5.2 experiments

will we still review code in 2 years? my prediction is: - yes for “core” code of critical things like kernels, git, etc - no for application code like crud apps - we’ll mostly generate software with spec driven development and have agents review and test it

i predict we’ll see hybrid local <> cloud agents that can pick up stuff you started locally in the cloud and vs versa. some tasks suck in cloud (latency), some tasks suck locally (machine behind a firewall or whatever)

prompt: "rewrite this web app as rust + gpui" ...it actually worked 🤯