Ex-NVIDIA engineer who built Unsloth explained RL, kernels, reasoning, quantization, and agents in 2 hours 42 minutes - better than $5000 fine-tuning bootcamps. pick the base model -> write triton kernels for 2x faster fine-tune -> quantize to 4-bit -> run GRPO/DPO -> ship a reasoning model on your... single GPU. That loop is why Unsloth is the default way to fine-tune Llama, Qwen, Gemma, and Phi on hardware you already own. Unsloth + Triton kernels + 4-bit quantization + GRPO/DPO + single-GPU fine-tuning - that's the stack. Watch and save it, then fine-tune your first model tonight.show more

Thanks for the support! We did another 3 hour one at AIE last week! Look out for the recording on YT hopefully in weeks!

@danielhanchen YouTube link:

@grok can you find the video on youtube

Source:

YT link:

This is the kind of talk that’s worth watching all the way through instead of just reading the highlights

love unsloth team very cracked

I also wrote an article on Loop Engineering:

I tuned into a Gemini live stream that featured Daniel a couple months ago. It was that stream that convinced me that AI is the real deal. When obviously-smart MFs like him are pursuing this stuff, it's like wildfire.

Spent a month debugging browser automation edge cases before my agent stopped breaking in production. No bootcamp preps you for that.

been self-taught my whole life. the 2h42m from someone who actually shipped it > any bootcamp

This is a solid overview of the modern open-source fine-tuning stack. Understanding the concepts matters more than memorizing the tools

Crazy that this is becoming something you can just try at home.

@grok what video means?

the goat

Oh this is the last year's session. Have they does this year's one?

Solid overview of the single-GPU stack. Just finished fine-tuning a Qwen model with Unsloth last week-2x speed really holds up.

unsloth makes fine-tuning accessible without a massive gpu cluster

Most people are learning how to use AI. The real leverage comes from learning how to build and fine-tune it. That's the difference between being a user and becoming a builder.

@ycombinator TIL there are fine-tuning boot camps 😂

Game-changing approach! Who else is excited to ditch the $5000 bootcamps and optimize their workflow with Unsloth?

the single GPU line is what got me

The pipeline is real, but the sequence is backwards for most people. Quantize-then-GRPO only makes sense once you already know your reward signal is stable at full precision. Debug RL at 4-bit and you cannot tell if the policy collapsed or the quantization did.

It's refreshing to see someone explain the full pipeline instead of just sharing another fine-tuning tutorial. Understanding why each step exists makes a huge difference.

This is gold! Been waiting for a no-BS deep dive on GRPO + Unsloth stack. Watching tonight and shipping my first reasoning model on one GPU. Legend work @danielhanchen 🔥

do you think you will get another community notes saying the quoted article is not related to the speech unsloth founder gave?

Excellent resource. Practical insights into fine tuning, optimization, and AI agents are exactly what the community needs as open-source AI continues to evolve. Thanks for sharing. A follow back would be greatly appreciated. Looking forward to learning from your future posts!```

Great stuff

Also recommend this tool that helps you manage your loop

that kernel optimization part is the real differentiator because most people just import hf and pray it works

bootcamps are just tax

Fine tune for what? Been hearing it for ages. I want to see what people have created after fine tuning.

This is one of the most amazing introductions I've ever seen to a technical talk. Such unpretentious wonderful excitement and enthusiasm. Love it.

because the hard part was never writing kernels, it's optimizing performance without sacrificing reasoning