Ruemic retweeted
The latest episode of the Pingback Podcast with @joelga is out!
We had a special guest this week! @samruby has more years of programming under his belt than I have lived!
His recent project Roundhouse is having a good momentum. Go and give a listen!
pingbackpodcast.com/episodes…
do it @josevalim !
I am publishing archives of agents one shotting a spec app at github.com/sheetgenius/agent… and I would be super honored if you could make a PR for the elixir prompt. Basically I’m just trying to have an even playing field where there’s a expert level prompt for each language to make it as terse and idiomatic as possible, and a shared spec they need to hit.
Do you have the prompts or threads publicly available?
The agent clearly chose to follow Rails compatibilty more closely in the Elixir version. So a few of Rails limitations got ported to the Elixir one. Here is an example from the README itself:
> Elixir retains Redis and Resque-compatible jobs, while Rust uses integrated queues and a different frontend/server implementation.
The agent knows, it just chose to prioritize something else, which will always happen if you are not steering.
I am certain coding agents can spot those issues and bring the numbers up for both Elixir and Go with no troubles whatsoever (although I don't expect them to beat Rust).
Initial explorations with Spinel are very promising! Seeing around 6x+ perf And request throughput on the canonical github.com/sheetgenius/agent… app! It wasn't easy, but it works! 2 PR's landed upstream, lots more incoming.
Best thing I’ve read all year
Everybody wondering why Omarchy is succeeding should read this classic Joel on Software essay:
joelonsoftware.com/2006/04/1…
I can't wait for assembly on rails. I don't want my agents wasting tokens on bike shedding their assembly code
we have continual learning, it just works in epochs!
so opus 5.5 did substantially better than fable 5.1 on all my prompts. Anthropic says it is "as good as Fable 5.1" but, in my experience, that does no justice to it
now, before you call me a paid shill, remember Anthropic is actively training on my domain. this is easy to verify. ask Opus 5.5 to write the grammar and rules of the Interaction Calculus. it will nail them, to small details (see it on VictorTaelin/Interaction-Calculus), while GPT models barely know it. so I think I might be experiencing Claude's jumps more profoundly than most, and that explains my preference
needless to say, I'm having a fields day and feeling an enormous urge to lock in and (finally) write Bend in itself. that'd unlock so many incredible things. but I have PRs and a raise to handle :|
(if you're from Anthropic, I'm very thankful you guys are doing that and I hope you keep it up...)
@doodlestein i’m wondering if you’re seeing agent flywheel familiarity showing up in the newer models, even though you have a clause to not train on it in your terms. Also, have you considered removing to activate this dimension of the flywheel?
I had Fabel and Astra tag teaming 3.8 flash to try to get useful out of it and after hours of it, it was still largely unproductive. The task wasn’t even hard. I was gonna see if I can use it to do basic js application refractoring, and it was basically just too sloppy, nothing it suggested passed muster
