Fatha

An engineer chasing fleeting ML waves | A believer that a good cup of chai can solve colossal problems

[Fatha @fathaa_s]

I need a new code editor, I want to have no barriers between my code and me. Cursor is cooked VS code as well!

[Fatha @fathaa_s]

Lets check out

[Fatha @fathaa_s]

Is there a benchmark that measures visual creativity/artistic sense of LLMs. I want to figure which LLM can be my UX expert?

[Fatha @fathaa_s]

It's mid week and I might have trained around 50 yolo models. Not sure if people train that many detection model. And, I am not tuning the model... I am tuning the data. Each run had entirely new set of data... Crazy times we live in.

[Fatha @fathaa_s]

@brancasi What are you cooking?

[Fatha @fathaa_s]

Very excited about this - https://t.co/HwGbh4QpTU

[Fatha @fathaa_s]

@brancasi What did you learn today?

[Fatha @fathaa_s]

@brancasi I hear they are hiring vision engineers now. Not sure if it's for multimodal or image generation.

[Fatha @fathaa_s]

@DavidSHolz Wild! We need more such benchmarks. Maybe we can find the evil ones from here haha

[Fatha @fathaa_s]

https://t.co/p25SbfuqYx

Cool! Can't wait for more things in this direction.

[Fatha @fathaa_s]

Still waiting on someone to train a vision decision model and make it play some games. The latency part is almost there - QwenVL models are sub 100ms per token.

[Fatha @fathaa_s]

@imjustnewatai Damm... Can you share the math problem scraper with us as well :)?

[Fatha @fathaa_s]

@quxiaoyin Valid point here, I can see a lot of LLM driven productivity gains coming out of rethinking the org chart. But in my domain(computer vision engineers), we are rethinking how gathering and managing data can be fundamentally changed. This is often offloaded to offshore conteactors.

[Fatha @fathaa_s]

Why does the frontier labs keep pushing bigger and better models when they plan on IPO? The bar is so high, I was on OAI last month and this month it's Anthropic again.

[Fatha @fathaa_s]

Just trained a vision-based JEV on Qwen3.5-9b clocking a lightning-fast 20ms per response. It handles almost anything you throw at it. But not ping pong :).

[Fatha @fathaa_s]

Honestly great software architecture makes the agents crazy good. Did something that would have otherwise taken weeks in few mins.

You can't vibe a good architecture yet!

[Fatha @fathaa_s]

@brancasi Prompting well is a big thing. And also depends on the harness 😅. I feel like the image, video generation models barely works with raw human prompts.

[Andrej Karpathy @karpathy]

Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.

[Fatha @fathaa_s]

Simply mind-blowing, an army fo AI agents solving humanity's most unsolved problems.

https://t.co/VIgrfAQujS