Melvin Vivas
@melvindvivas
Local models, Agents, AI Engineering
Coworker
Astra Orchestrator
pylib
976 Following 4.1K Followers
We can now ssh into Devin's computer 🤯
omg this is awesome
I think you need to spend $200 minimum on coding plans if you do a lot of stuff
Devin is the most underrated agent right now
Codex Pro 20x wins even limits just last 2 days
Let's settle this Codex, Devin, Cursor or Claude
If you had $200 for a sub, which would you choose?
Best way to perfect model routing is to finetune your own model with your own prompts
Either to Astra, Luna or whatever model you want
Try this for a start
model routing with jev? not worth it.
At least not for my workflow. The routed "team" cost almost as much as Astra alone and hit the time limit before it was finished.
Who told you that you must use Linux to do AI, I'm on windows now lol
Working on something to serve local models
Adding support for ONNX too for classification models which Jev made popular lol
I'm back to using the IDE with @devindesktop
I missed this
I asked to take a break from a project a few weeks ago because I got sick. Now, they don't need me anymore
So I'm doing open source full-time for now until the next paid gig
That's why I'm able to spend more time here in X too
Show more
I think I'll cancel Codex Pro 5x and just wait for 20x to be available again
Codex is great but if you gotta save, you just have Luna
No other choice
FYI guys
SWE-2 is free in Devin Cloud Agents, Devin CLI, and Devin Desktop until October 8 with all plans.
Great time to try cloud agents for the first time, at no extra cost!
Are you happy with your Codex
A language model can perform well on a clean benchmark and still struggle with the cases that matter in real-world use.
Here are the evaluation practices that helped us move from promising prototype results to production. ✅
Show more
Banger paper from MIT and Sakana AI.
They show that self-improving coding agents work.
The best part is that their approach, Self-Improvement via Fast Tree-search (SIFT), runs at a tenth of the CPU hours of DGM.
They reach 35.1 percent on Polyglot with o3-mini after 30 expansions. DGM reaches 30.7 percent after 80 nodes of tree search.
SIFT does it in under 50 CPU hours and under 5 hours of wall clock. The Qwen3-30B configuration runs its full search at 224 CPU hours and $34 of API spend, a tenth of the DGM baseline.
The saving comes from where the money goes.
Benchmark evaluation is the runtime bottleneck, so an LLM judge ranks candidate self-modifications first and only promising candidates get evaluated.
Judge quality decides the run.
On TerminalBench, gpt-5.4-high as the pairwise judge finds a 36.7 percent agent against a 29.2 percent starting point. gpt-5 finds 34.5 percent, and its top-ranked candidate is not the best agent its search produced.
Paper:
Show more
Timing, luck and adoption is very important
First-movers don’t necessarily become the successful ones
Execution is key and how Jev was launched was epic
poor guy claim to have built Jev a year ago but no one cared, and now Jev stole all the thunder
many people are saying “you gotta tell your story” or “marketing is important”, and they just completely missed what actually made the difference here
i just looked into this laya model and:
- it only supports 512-1k context… a lot of use cases won’t fit at all
- evaluating the model directly shows its accuracy is as good as a coin flip. in order to get good results, you need to first fine tune it
i’m sorry, but that’s not Jev
there’s a massive gap between an interesting research and a useful product
you can “tell your story” all you like, but you can’t blame Jev for stealing your thunder when Jev did all the work to make a well-packaged solution anyone can just grab and go
Jev is not completely new from an academic sense, just like how ChatGPT was not the first LLM
don’t underestimate the effort and value in putting together something that’s actually good enough for adoption - it makes all the difference
Show more
I can already smell Grok 4.7
Never use Astra in Codex Plus, you’ll hit your limits faster
@Beareka Don’t use Astra on the plus plan except for exceptional queries. Stick to Luna and sol
Mission accomplished @DevinAI
I have 2 hours to touch grass
