Ox Alpha - a new model from an anonymous creator

Just hype.

I burned around 3B tokens on this model. It’s quite good (except for long tasks—it easily gets stuck in a loop). Better than OW Alpha before.”

I can second on this — been burning some tokens with it since yesterday, and it does get stuck mid-sessions. Quite annoying.

But I think it’s quite good, I been tackling on some tasks with it, stressing with questions on specifics about the tools I’m working with (specially, Phoenix and its ecosystem + Kanidm and its architecture/ecosystem/etc.). He ended up cloning Kanidm repo to read it all and answer specific questions, explore alternatives, and write a spec the way I wanted. Not bad!

Whoever made this, be OpenAI, Anthropic, Alibaba, or some random not-so-known AI lab from US China France Germany India etc., my honest Thank You for the tokenmaxxing weekend My Saturday night will be fun!

I have same issue when I’m using Zed editor + OpenRouter provider. After I move to OpenCode agent + Caveman + add Elixir skills + OpenCode provider the issue is gone, I can run long session without stop.

You said “Elixir skills”?

Tell me more about this! Which ones you’re using? And which ones you’d recommend?

I got some skills from people in this forum a while ago. Unfortunately, I didn’t bookmark the link.

I used the following EEF Security WG to create skill to publish and maintain library, it helped a lot especially keeping the deps updated. You should adapt to your project and needs

You need an LLM to run mix hex.audit and mix deps.update --all?

There is a bit more involved than running these 2 commands and merge them blindly.
See below some additional links
User tooling and Supply chain
For example, the EEF Security WG recommends checking the actual diff of dependency updates and making sure the changes do not introduce malicious code.
mix hex.outdated, mix hex.audit, mix deps.audit and especially mix hex.diff are all part of that picture. Then you still need to run the project tests, checks and review the PR.

I personally use Dependabot, but I still treat every dependency PR as something that needs to be manually reviewed and tested.

The LLM Skill is useful (to me) for encoding that workflow, not to remember the mix tasks.

Im using it right now and its going to take me out of the permanent underclass! Jokes aside, I notice that non standard models seem to be good at writing elixir, although they ignore the concept of DRY and write the most verbose elixir code you will ever see, which makes sense giving how elixir can be written and most models arent post trained + fine tuned on elixir, but it sort of gets the job done and right now 0xalpha is allowing me to burn tokens to get an mvp up while I have gpt sol and luna coming in to fix out bugs without wasting precious frontier tokens. Oxalpha seems smart enough