Pinned Tweet
I’m building Handbeam, an experimental open-source AI agent designed for phones—not just a remote control for a desktop agent.
Built with Elixir/OTP. The agent loop and tools run on-device; LLMs via API.
Coding is just one use case. github.com/youfun/Handbeam
对于不影响交易的服务/系统来说，，完全无法使用的bug其实还好。尤其是agent/创意制作类。但如果涉及到交易，比如crm/电商。系统半个小时不能用，经济损失可就不好说了，可能会导致客户失去信息而切换产品。
25
We benchmarked six AI models on their Elixir programming capabilities using Droid CLI. The test used the
open-source project tailwind_formatter (a Tailwind CSS class sorter for HEEx templates) as the baseline, with
two coding tasks: implementing a validate_class/1 function for class name validation, and extending the
formatter to support data-class attribute sorting. Each model worked on an isolated copy of the project, and
correctness was objectively verified by running mix test after each model's changes. We recorded test pass
counts, number of new test cases, execution time, and token cost.
The six models tested were muse-spark-open, beam-1.0, mistral-large-4, gpt-6.1-sol, glm-5.3-flash, and
gemini-4-argon. Results showed that mistral-large-4, muse-spark-open, and gpt-6.1-sol tied for the highest test
pass count (81 each), with muse-spark-open and gpt-6.1-sol demonstrating the deepest test coverage. Notably,
glm-5.3-flash achieved a near-top score at a fraction of the cost (7,108 credits, only 3.9% of the most
expensive model), making it the clear winner in cost-efficiency. In contrast, gemini-4-argon was the most
expensive yet performed at the bottom, and beam-1.0 had the narrowest test coverage. All models successfully
completed the Elixir tasks with zero test failures, but they differed significantly in code robustness, test
completeness, and cost efficiency.
Mermaid 支持改成 Pele实现teapawt.net/kepano/status/21114720…
I’m building Handbeam, an experimental open-source AI agent designed for phones—not just a remote control for a desktop agent.
Built with Elixir/OTP. The agent loop and tools run on-device; LLMs via API.
Coding is just one use case. github.com/youfun/Handbeam
22
Yoyfun retweeted
I'm writing a book: Native Mobile Apps With Elixir.
From your first screen to a production app with Mob, the framework that runs the BEAM on your phone.
Register for early access: forms.gle/ksjRXyEqNBmFwLhn6
#ElixirLang #MyElixirStatus
Yoyfun retweeted
I took a stab at improving the performance of Elixir. I used the PR for Elixir that @ZachSDaniel1 posted and worked some more on it. In the benchmark Elixir is now faster than Rust (ccece30 newest commit) for the Sidebar and Search benchmark.
Last conversion of the Campfire from my token stash goes to JavaScript/Express. As with the others, I'm sure you can find more to optimize. But this gives you a sense of what order of magnitude these options exist within with agent-led development. github.com/basecamp/once-cam…
经过优化后的 rust 肯定比 elixir 版本快，但迭代一次和多次会拉大这种差距。想起来之前说如果提示里面没有说遵循最佳实践，部分 llm 会把 elixir 写的像是其他语言。
Oh FFS. It turns out the req/s in the initial vibed Rust version is within the same order of magnitude as the initial vibed Elixir/Go versions, and even performs worse in some of them.
The history is in the repository! The Rust version only got faster with a bunch of additional work. The initial Elixir/Go versions presented in the tweet below saw none of this additional work. It was comparing a 300 commits Rust version with a 2-3 commits Elixir/Go version.
Source: github.com/basecamp/once-cam…
46
DHH says a language agents can't write well from a one-line prompt is in trouble, while José says the Elixir port was slow because the agent copied Rails instead of being told to use Elixir idioms, and I side with José.
The prompt was just something like "Rewrite this in Elixir. Review with Codex. Optimize it." Same as what was used for Go, Laravel, Rust, etc. If the agents can't write good Elixir without good steering, I think that's worth attempting to fix. But I'm also happy to take PRs!
37
