Dayanne Fernandes

local models | open weights | fine-tuning | small models | RAG
.
building: @reopenly_ai 🚀
prev: @LandingAI

[Dayanne Fernandes @dayoffdev]

A fine-tuned 8B model beat Claude Sonnet 5 on 437 contract clauses.

73.7% vs 70.0% accuracy
250 ms vs 2,230 ms per answer
$0.125 vs about $2.00 per 1,000 answers

Sonnet 5 still won the rare categories, 67.4% vs 60.4%.

On a narrow task, the specialist wins.

[Dayanne Fernandes @dayoffdev]

That is again one of a thousand reasons to use local models

[Pirat_Nation 🔴 @Pirat_Nation]

Anthropic is cracking down on AI-generated video game clones through Claude Code.

A new instruction added tells Claude to build original games or use ideas from existing genres, but not recreate specific games, even if they’re renamed.

The move could affect people using AI to recreate games like Terraria, Minecraft, Halo or GTA as browser games or unofficial ports.

[Dayanne Fernandes @dayoffdev]

@rohanpaul_ai I wonder what anthopic and openai are using to do the same, I felt that the latest models are becoming quite good at double checking any push back

[Dayanne Fernandes @dayoffdev]

@lvwerra why not? maybe another form of hands? like the industrial ones?

[Dayanne Fernandes @dayoffdev]

Quickstart for embeddings: https://t.co/UAYIbVL2L7

[Dayanne Fernandes @dayoffdev]

Once it's trained, we host it for you and you call it with the OpenAI SDK. Change the base URL and the model name, and that's it.

[Day~ @daayoff]

@LulaOficial investe na https://t.co/TfgebWdOog pra termos uma potencia de IA no Brasil tb

[Dayanne Fernandes @dayoffdev]

We didn't pay for it or see it first. The full review, with the video and what to watch out for: https://t.co/xxD6TcuXkM

[Dayanne Fernandes @dayoffdev]

An independent review of ReOpenly is out. Present of AI signed up, used the free credits and fine-tuned a small model on their own examples, twice.

From what they found: "Round 2 trained in 6 min 6 s, and all 5 held-out answers came in within 20 words."

[Dayanne Fernandes @dayoffdev]

@MaziyarPanahi I'm building a platform to help curate data and be able to create specialized models: https://t.co/3DL1z4KE6r
so far we have text and embeddings models; so for healthcare case, it can improve notes, triages, cases retrieval, etc

[Dayanne Fernandes @dayoffdev]

@MaziyarPanahi yes! we need more improvements in the feedback loop and evaluations

[Dayanne Fernandes @dayoffdev]

@MaziyarPanahi Indeed! Me too, I only have the feeling that people are trusting too much nowadays, they should definitely reason on the response and give a feedback when things are wrong instead of blinding believing in it

[Dayanne Fernandes @dayoffdev]

@rohanpaul_ai Interesting work, but this does not seem too efficient, why training 1 lora adapter per mistake instead of updating one adapter only?

[Dayanne Fernandes @dayoffdev]

@TheAhmadOsman Indeed, and this also affects the gaming industry

[Dayanne Fernandes @dayoffdev]

https://t.co/Wmsq7u2F4q

The README also points to ReOpenly skills for your coding agent: upload Q&A data, fine-tune and deploy text models, and call them. https://t.co/FgXIdnQH0E

[Dayanne Fernandes @dayoffdev]

Our embedding examples are public on GitHub. Rank passages for a question, plot six texts on a chart, report searches for review, compare a trained model with the base. One command each, plus a notebook with all four. Nothing in there starts a training run, so play away.

[Dayanne Fernandes @dayoffdev]

The guide: https://t.co/9oPirxL9hy

[Dayanne Fernandes @dayoffdev]

Fine-tune an embedding model on your own searches: the results that helped, plus the near misses that looked right but weren't. You only pay for the GPU time while it trains.

[Dayanne Fernandes @dayoffdev]

I get it, we're not really reviewing code line by line anymore. Agents write it, agents review it.

But how are you reviewing the changes? What changed in behavior, what's risky, what the checks actually proved. GitHub Universe 2026 didn't touch that at all.

https://t.co/s1S3WWEtsq

[Dayanne Fernandes @dayoffdev]

I was just watching Kent Beck talk in a event, and I loved that he called the LLM solutions as "Genies" since it grants wishes but not as you would like