The case for the return of fine-tuning
welovesota.com
welovesota.com
- many, if not most MLEs that got started after LLMs do not generally know anything about machine learning. For lack of clearer industry titles, they are really AI developers or AI devops
- machine learning as a trade is moving toward the same fate as data engineering and analytics. Big companies only want people using platform tools. Some ai products, even in cloud platforms like azure, don’t even give you the evaluation metrics that would be required to properly build ml solutions. Few people seem to have an issue with it.
- fine tuning, especially RL, is packed with nuance and details… lots to monitor, a lot of training signals that need interpretation and data refinement. It’s a much bigger gap than training simpler ML models, which people are also not doing/learning very often.
- The limited number of good use cases means people are not learning those skills from more senior engineers.
- companies have gotten stingy with sme-time and labeling
What confidence do companies have in supporting these solutions in the future? How long will you be around and who will take up the mantle after you leave?
AutoML never really panned out, so I’m less confident that platforming RL will go any better. The unfortunate reality is that companies are almost always willing to pay more for inferior products because it scales. Industry “skills” are mostly experience with proprietary platform products. Sure they might list “pytorch” as a required skill, but 99% of the time, there isn’t hardly anyone at the company that has spent any meaningful time with it. Worse, you can’t use it, because it would be too hard to support.
I have avoided fine tuning because the models are currently improving at a rate that exceeds big corporate product development velocity.
More than once I've just done labeling "on my own time" - I don't know the subject as well but I have some idea what makes the neurons happy, and it saves a lot of waiting around.
I've found tuning large models to be consistently difficult to justify. The last few years it seems like you're better off waiting six months for a better foundation model. However, we have a lot of cases where big models are just too expensive and there it can definitely be worthwhile to purpose-train something small.
In true hacker spirit, I don't think trying to train a model on a wonky GPU is something that needs an ROI for the individual engineer. It's something they do because they yearn to acquire knowledge.
I ask a version of this every six months or so, and usually the results are quite disappointing.
This time I had more credible replies than I have had in the past.
Here's my thread with highlights: https://twitter.com/simonw/status/1979254349235925084
And in a thread viewer for people who aren't signed into Twitter: https://twitter-thread.com/t/1979254349235925084
Some of the most impressive:
Datadog got <500ms latency for their language natural querying feature, https://twitter.com/_brimtown/status/1979669362232463704 and https://docs.datadoghq.com/logs/explorer/search/
Vercel run custom fine-tuned models on v0 for Next.js generation: https://vercel.com/blog/v0-composite-model-family
Shopify have a fine-tuned vision LLM for analyzing product photos: https://shopify.engineering/leveraging-multimodal-llms
I'm including things like RL metrics as data here, for lack of a better umbrella term, though the number of proposed projects that I've seen that decided that ongoing evaluation of actual effectiveness was a distraction from the more important task of having expensive engineers make expensive servers into expensive heatsinks is maddening.
On the "applying X" problem - this almost feels to me like another argument against fine tuning? Because it seems like Applying can be a surprisingly broad skill, and frontier lab AIs are getting good at Applying in a broad fashion.
Even worse, even if you DO get an improvement you are likely to find that it was a waste of time in a month or two when then next upgraded version of the underlying models are released.
The places it makes sense from what I can tell are mainly when you are running so many prompts that the cost saving by running a smaller, cheaper model can outweigh the labor and infrastructure costs involved in getting it to work. If your token spend isn't tens (probably hundreds) of thousands of dollars you're unlikely to save money like this.
If it's not about cost saving, the other reasons are latency and being able to achieve something that the root model just couldn't do.
Datadog reported a latency improvement, because fine-tuning let them run a much smaller (and hence faster) model. That's a credible reason if you are building high value features that a human being is waiting on, like live-typing features.
The most likely cases I've heard of for getting the model to do something it just couldn't do before mainly involve vision LLMs, which makes sense to me - training a model to be able to classify images that weren't in the training set might make more sense than stuffing more example images into the prompt (though models like Gemini will accept dozens of not hundreds of comparable images in the prompt, which can then benefit from prompt caching).
The last category is actually teaching it a new skill. The best example here are low-resource programming languages - Jane Street and OCaml or Morgan Stanley and Q for example.
Jane Street OCaml: https://www.youtube.com/watch?v=0ML7ZLMdcl4
Morgan Stanley Q: https://huggingface.co/morganstanley/qqWen-1.5B-SFT
Depending on use case you could either insert the LoRA weights as their own layers at runtime (no time to create, but extra layer to compute each token), merge them with existing layers (initial delay to merge layers, but no runtime penalty after), or have pre-merged models for common cases (no perf penalty but have to reserve more storage).
My current mental model of LoRA is that this would be unlikely to Work, but I've never used them so I don't really know what I'm talking about. Would be a very interesting experiment!
- PaddleOCR, a 0.9B model that reaches SOTA accuracy across text, tables, formulas, charts & handwriting. [0]
- A 3B and 8B model which performs HTML to json extraction at GPT-5 level accuracy at 40-80x less cost, and faster inference. [1]
I think it makes sense to fine tune when you're optimizing for a specific task.
[0] https://huggingface.co/papers/2510.14528
[1] https://www.reddit.com/r/LocalLLaMA/comments/1o8m0ti/we_buil...
I've played around with doc recognition quite a bit, and as far as I can tell those two are best-in-class.
Our thesis was that fine tuning would be easier than deep learning for users to adopt because it was starting from a very capable base LLM rather than starting from scratch
However, our main finding with over 20 deployments was that LLM fine tuning is no easier to use than deep learning
The current market situation is that ML engineers who are good enough at deep learning to master fine tuning can found their own AI startup or join Anthropic/OpenAI. They are underpaid building LLM solutions. Expert teams building Claude, GPT, and Qwen will out compete most users who try fine tuning on their own.
RAG, prompt engineering, inference time compute, agents, memory, and SLMs are much easier to use and go very far for most new solutions
Otherwise, you should just use gpt5
Preparing a few thousands training examples and pressing fine tune can improve the base LLM in a few situations, but it also can make the LLM worse at other tasks in hard to understand ways that only show up in production because you didn’t build evals that are good enough to catch them. It also has all of the failure modes of deep learning. There is a reason why deep learning training never took off like LLMs did despite many attempts at building startups around it.
Andrej karpathy has a rant about it that captures some of the failure modes of fine tuning - https://karpathy.github.io/2019/04/25/recipe/
Then try the same thing using fine tuning. See which one wins. In ML class we have labeled datasets with breeds of dogs hand labeled by experts like Andrej, in real life users don’t have specific, clearly defined, and high quality labeled data like that.
I’d be interested to be proven wrong
I think it is easy for strong ML teams to fall into this trap because they themselves can get fine tuning to work well. Trying to scale it to a broader market is where it fell apart for us.
This is not to say that no one can do it. There were users who produced good models. The problem we had was where to consistently find these users who were willing to pay for infrastructure.
I’m glad we tried it, but I personally think it is beating a dead horse/llama to try it today
Why would they join you rather than founding their own company?
Pick the right customers.
> Why would they join you rather than founding their own company?
The network effects of having enough resources in one place. For having other teams deal with the training data, infrastructure, deployment, etc.
That’s fair, one market segment of this is sometimes called sovereign compute.
Another common model that I have seen is to become the deepmind for one very large and important customer.
I think this works.
Once you (as in you the person) have the expertise, what you need all the people for exactly? To fine-tuning you need to figure out the architecture, how to train, how to infer, pick together the dataset and then run the training (optionally setup a pipeline so the customer can run the "add more data -> train" process themselves). What in this process you need to hire so many people for?
> Why would they join you rather than founding their own company?
Same as always, in any industry, not everyone wants to lead and not everyone wants to follow.
Read Andrej’s blog that I linked earlier in the thread if you want to understand why.
Debugging requires knowing some small detail about your data distribution or how you did gradient clipping which take time and painstakingly detailed experiments to uncover.
Right, but why does that mean you need more employees? You need to figure out how to surface failures, rather than just adding more meat to the problem.
Depends on what you want to achieve, of course, but I see fine-tuning at the current point in time primarily as a cost-saving measure: Transfer GPT5-levels of skill onto a smaller model, where inference is then faster/cheaper to run. This of course slows down your innovation cycle, which is why generally this is imo not advisable.
But a recent trend that cut into the cost savings is that foundation model companies have started releasing small models. So you can build a use case with qwen 235B, then shrink down to 30B, or even all the way down to 0.6B if you really want to.
The smaller models lose some accuracy, but some use cases are solvable even by these smaller and much more efficient models.
The problem is easily avoided by not using it for other tasks.
This is a reason why general purpose models shine. You don’t have to carefully characterize a task and put guard rails around it.
However, I personally think that this intuition applies to products and interfaces, not to AI.
Intelligence and learning is general. Intelligence without generalization is memorization, which seems to be less useful in practice.
They'll pay for anyone that can personalize models to be meaningfully diverse.
Training an LLM from scratch is trivial - training a good one is difficult. Fine tuning is trivial - doing a good job is difficult. Hitting a golf ball is trivial - hitting a 300 yard drive down the middle of the fairway is difficult.
We have a lot of more capable open source models now. And my guess is that if you designed models specifically for being fine tuned, they could escape many of the last generation pitfalls.
Companies would love to own their own models instead of renting from a company that seeks to replace them.
One annoying part was switching to new and better models that came out literally every week.
I don’t think it substantially changes anything. If anything I think the release of more advanced models like qwen-next makes things like fp4, moe, and reasoning tokens an even higher barrier of entry.
Naturally I reached for CLIP+ViT which got me a ~60% success rate out of the box. Then based on that, I created a tiny training script that read `dataset/{slide,no_slide}` and trained a new head based on that. After adding ~100 samples of each, the success rate landed at 95% which was good enough to call it done, and circle back to iterate once I have more data.
I ended up with a 2.2K large "head_weights.safetensors" that increased the accuracy by ~35% which felt really nice.
Yes, 100s of housands of them
From fine-tuning for coding assist, to medical applications, customer support, legal and financial use cases, various classification tasks, for government work, statistics, language learning, music, education, even for role playing game character AI...
I'd rather have a fine-tuned model specialized to any of those tasks and countless others when I'm doing one of those tasks, than a jack of all trades...
I discuss a large-scale empirical study of fine-tuning 7B models to outperform GPT-4 called "LoRA Land", and give some arguments in the discussion section making the case for the return of fine-tuning, i.e. what has changed in the past 6 months
For training data, I was thinking you could just put all the stuff into context, then give it some prompts, and see how the responses differ over the baseline context. You could feed that into the fine tuner either as raw prompt and the output from the full-context model, or as like input="refactor {output from base model}", output="{output from full-context model}".
My understanding is that LoRA are composable, so in theory MCPs could be deployed as LoRA adapters. Then toggling on and off would not require any context changes. You just enable or disable the LoRA adapter in the model itself. Seems like this would help with context poisoning too.
There is growing emphasis on efficiency as more companies adopt and scale with LLMs in their products.
Developers might be fine paying GPT-5-Super-AGI-Thinking-Max prices to use the very best models in Cursors, but (despite what some may think about Silicon Valley), businesses do care about efficiency.
And if you can fine-tune an 8b-parameter Llama model on GPT-5 data in < 48 hours and save $100k/mo, you're going to take that opportunity.
Together with speed and const, this is from my point of view this is the only "case" for the return of fine-tuning here. And this can be managed by context management.
With growing context sizes, first RAG replaced fine-tuning and later even RAG was replaced by just a good-enough prompt preparation for more and more usage pattern.
Sure, speed and costs are important drivers. But like with FPGAs vs. CPUs or GPUs, the development costs and delivery time for high-performance solutions, eliminate the benefit most the time.
It requires no local gpus, just creating a json and posting to OpenAI
Curious to hear others’ thoughts on this
I don't think anyone thought fine tuning was dead.
The main claim was that new models were much better than anything you could get your hands on to fine tune.
IMO, intuitively that never made sense. But I never tested it either.
In any case, platforms like tinker.ai support both SFT and RL.