Rich Rican

ai musings ￨ impossible → hard → easy → elegant ￨ ≪†✝️†≫

[Rich Rican @poor__rican]

Even if we don’t follow each other, I admire you all.

Y’all keep me motivated, and in a certain sense that kinda makes us friends.

It’s easy to get depressed on social media but the TL you take part in helps me realize the cup is half full.

Gm anon.

[Rich Rican @poor__rican]

@TheAhmadOsman How do you get multiagent structures with Hermes??

[Rich Rican @poor__rican]

@josepha_mayo Good choice with hrm.

I would like to donate. DM?

[Rich Rican @poor__rican]

@josepha_mayo What model architecture?

[Rich Rican @poor__rican]

@Tech2Wild These are the correct questions.

[Rich Rican @poor__rican]

PSA: just propagating the idea that Mythos class models are likely loop, mixture-of-depth, or another form of recurrent models.

Given the recent actions of Anthropic, significant attention of the open source LM community should be diverted.

[Rich Rican @poor__rican]

@MikeWingerii Not simple, but Logos.

[Rich Rican @poor__rican]

@doodlestein You're probably good. Looked at the repo and the readme felt more like describing a metabolic process than ML. The neurotransmitter ontology probably'll obfuscate.

now, as for training... idk

[Rich Rican @poor__rican]

@eliebakouch The real question is: what do they consider "frontier LLMs"?

I hope it has a low false-positive rate.

[Rich Rican @poor__rican]

@TheAhmadOsman Many things can be said about a company that simultaneously says "we should pause AI" while releasing a model "unlike anything before".

Also didn't know ML accelerators were against ToS.

[Rich Rican @poor__rican]

I've been forced to rice out my GPU server as I transition from macOS to Linux.

Don't know if I should be delighted or shameful. Am I in relapse?

Arch Linux btw.

[Rich Rican @poor__rican]

PSA: Given the current discourse about loops, a helpful reminder to not give into low-effort techniques. Low effort techniques will consistently yield mediocre code.

These larger models can be pushed to output fairly novel, elegant, effective code while being better / faster than handrolling. However, I can assure you, it's not as trivial as "just write a loop."

If you're looking for durable code that serves a novel purpose, a loop is a recipe for an over-engineered and fragile codebase.

If you haven't realized, these tools are multipliers and what you put in is what you get out.

[Rich Rican @poor__rican]

hot take: all coding tools suck. Each for various reasons.

However, Claude Code is the one with the least amount of issues and the highest amount of features.

If you disagree, either A, you hate Anthropic, in which case your assessment of a CLI tool isn’t sound. Or B, skill/trust/plan-reading issue.

[Rich Rican @poor__rican]

Recently laid off, so I'm taking the opportunity to recalibrate and leverage ML / AI Engineer Xitter networking.

Forcing myself to post what I'm working on by Friday...

[Rich Rican @poor__rican]

@davidzmorris @mil000 The real skill that needs to be taught is having people output specific enough natural language.

Obviously that doesn't sell as well, so everyone just perpetuates the FOMO.

[Rich Rican @poor__rican]

If you're vibe-coding, were you really a SWE to begin with?

[Dmitrii Kovanikov @ChShersh]

Vibe-coding is just a gambling addiction for SWEs

[Rich Rican @poor__rican]

@Hesamation "Is the bug-free app in the room with us right now?"

[Rich Rican @poor__rican]

The people who rely on /loops for all coding are either not mentioning in-depth planning, not writing heavy/complex/performant code, or they expect to never actually maintain their codebase themselves again.

No way there's no hours long planning.

Or it could be all of the above....

[Peter Steinberger 🦞 @steipete]

Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.

You should be designing loops that prompt your agents.

[Rich Rican @poor__rican]

@doodlestein YouTube’s anti-bot measures are very complex.

If you standup a Docker container with a real browser, avoid using tools (ie: playwright, selenium, etc) and inject JS manually, you’ll avoid 99% of problems.

Also, it defers any FFI into static function calls.

[Rich Rican @poor__rican]

@PhreeStyleBTC @bindureddy You’d actually be surprised how well their smaller models do on non-typical / possibly OOD vision tasks.

Definitely poor reasoning depth and weak logical reasoning, but very good generalization in multimodal tasks.

No point in relying on benchmarks these days.

[Rich Rican @poor__rican]

@CryptoCyberia @banteg It’s very easy to write sloppy rust.

It gets the basics but complex functionality causes it to breakdown