Felipe Mello

Currently @ TorchTitan RL/META
prev TorchTune/TorchForge

[Felipe Mello @fmello93]

@d1muncher69 @Im_IrushiK haiku 3.5 with low reasoning does it. It is not an intelligence issue. To have the model properly count characters, we would probably need tokens to be individual letters.

[Felipe Mello @fmello93]

@QGallouedec @SergioPaniego When the model is deployed, thinking will be stripped, but during training, past thinking is preserved. So the model is trained with a different distribution vs production. Do you know if anyone has analyzed if this matters?

[Felipe Mello @fmello93]

@Acon43158243 @arcprize 😂, thats fair, second example is also like that

[Felipe Mello @fmello93]

@arcprize on the 3rd example, why is the correct answer to fill the 2 spots at the top, and not to the bottom right?

[Felipe Mello @fmello93]

@vikhyatk loss fn peak?

[Felipe Mello @fmello93]

@maharshii Can you share the downsides? i.e. why isnt this the default?

[Felipe Mello @fmello93]

@vikhyatk i was considering doing some experiments with it soon. I am surprised that you couldn't use it with torch, since they mention it in their readme

[Felipe Mello @fmello93]

@drisspg I didn’t have a good experience with gpt-5 on cursor, but heard good things about codex from multiple people. I will have to give it a try

[Felipe Mello @fmello93]

@redtachyon @ShashwatGoel7 We want to get this right in TorchForge (let you worry about your crazy ideas, not infra). It is still early days, so there's a lot of room on the design. Let us know if you have any ideas/would like to contribute.

[Felipe Mello @fmello93]

@xidulu I see. If you need on the fly packing, I implemented it in tune but never merged: https://t.co/f7VQQY8vfv

You can also find it here by someone else that did it: https://t.co/FE9UUrfajY

[Felipe Mello @fmello93]

@redtachyon best of luck in the new journey!

[Felipe Mello @fmello93]

@_lewtun Glad you liked it! its WIP. I will make it a bit less crowded, give users some filtering options. Let us know if you have any feedback/ideas.

[Felipe Mello @fmello93]

@difficultyang @testttt1236 Thanks for your service, sir

[Felipe Mello @fmello93]

@fchollet Couldn’t the takeaway be that we can try to apply this curriculum to **any** thinking problem and have LLMs internalize thinking? Something like how Claude could achieve great results without a thinking mode.

[Felipe Mello @fmello93]

@giffmana Yeah, I feel unproductive and slower when I go pure vibe coding. Much better for me to take the lead and prompt the model every other step. I also find it helpful with planning, ie chugging thousands of lines of code and coming up with a proposal + pseudo code.

[Felipe Mello @fmello93]

@giffmana My Meta/Google interviews were pretty standard. But there was one google interview that caught me by surprise. I was asked to write unit tests and share the hardest bug I had ever solved. I wouldnt say it was "cool", but it carried more signal than leetcode.

[Felipe Mello @fmello93]

@_xjdr Gemini cli performs waaay worse than cursor. Have been trying Claude code. I like it more than cursor, it gets further and for better price, but it’s still stupid. I think that cursors advantage is that I can query N models to revise each others plan, something I miss with Claude

[Felipe Mello @fmello93]

@Daksh46559036 @coldhealing same idea: I think its a safe choice to do something technical, that allows you to explore multiple areas in that field, and gives you a way out if you find yourself happier at non-technical roles.