The point that the major labs don’t seem to get is that the vast majority of use cases simply don’t need models that have 50 PhDs and can speak 12 languages. Most use cases are defined within constraints where costs matter a lot.
As open weight models and cheap fine tuning services become the norm the whole economic framework of these mega models the labs are in an arms race building just completely crumbles. As does the economic picture that justified the massive infrastructure building that’s now broadly funded by a complex network of debt.
This is what makes open weight models so threatening to them. The political and “it’s China” angle is mostly just a cover for the real reasons why they’re freaked out.
The fact that models are now a pure commodity is bad enough for the big labs. If small open weight models become the norm the big labs are toast.
The premise of your project seems compelling. Is it novel or building on existing work? Anything one could read or watch to get introduced to that area of research?
But then can't the case be made that narrower and stricter-defined use cases are better served by more conventional ML? If/Wherever efficiency is a concern, that is.
LLMs fit between well defined and poorly defined data; if you have well defined data, you can just feed it to ML models and they'll do what you want.
LLMs will take poorly defined data and "potentially" create well defined output.
So ideal systems will likely be constructed with a LLM on one end and a ML on the other and some how, if you can feedback poor data from the ML back into the LLM to clean up the data, you have a magic layer that doesn't care as much about the structure of the data.
That's a lot of supposition, but that's the difference I see between what we can do with LLMs and how ML models are structured. They might be classified as doing the same thing, but their data inputs are vastly different.
As someone who worked at multiple startups, I am pretty sure they get it but it doesn’t fit their goals.
They want to moat where they don’t need to compete with other companies because they have something that other companies can’t have.
In my opinion this is a short-sighted and greedy worldview. Haven’t seen it work personally. It is a different version of the month-to-month salary guy thinking he will be a billionaire and having that thrash “mindset”.
The reality is that practically none of those companies will amount to anything and they would be better off weighing the usefulness aspect of their output more. Instead they are imagining they will be Google.
Anthropic and openai ofc are the pinnacle of this greed culture and they correspond to FTX from the crypto trash hype so I don’t think they fit into the scale of sensibility.
Coming from this perspective, it is pretty easy to see what they are.
Yeah and these latest and greatest models universally suck hard for any use case outside of the few 'blessed' ones. Like I'm sure their ability to write prose and generally sound like a human being has regressed quite a bit, but even if not. Opus 5 is barely above GPT4 when it comes to stuff like home improvement advice.
Fine timing takes time and data, though. If smart enough models get cheap enough, then most people have lots of use cases that are cost insensitive enough that it's not worth the effort.
Of course, "smart enough" is a low enough threshold for most uses that this is still a problem for the frontier labs.
But at the same time, a truly smart enough closed model could also potentially command almost whatever they'd care to charge for it.
Whether they can actually get to that level remains to be seen, but I can definitely see a situation where most people are perfectly happy with cheap middle of the tree models while large corporations pay magnitudes more than current API pricing for access to models never even marketed as a mass market product and keep the labs afloat.
It's of course be a lot easier for them to find the path towards that of they didn't need to compete with open models in the meantime.
Yes. Fine tuning is an optimization. Like all optimizations, it’s downstream of figuring out your problem, implementing a naive solution, scaling the naive solution, and getting frustrated by SWAP-C. THEN you start poking at what optimizations are possible, choosing an approach, implementing the optimization, and redeploying.
This cycle happens when you have a well defined use case with high volume. It is the far opposite end of the spectrum from the general purpose intelligence on tap that frontier AI models purport to deliver.
I think focused fine tunes and big general models will coexist. Ideally with smart routing and caching to use small, specialized and local options when appropriate.
seems that, like software engineering and other areas before, they also have to rediscover that one single monolithic solution that handles everything is too inflexible and not maintainable, it's just a bad approach. people don't need the models they use to generate their codebase to also be able to translate Shakespeare into gen-z slang
Some specialized models may end solving themselves by helping fit the problem with the appropriate regular algorithms/formulas, which are a million times more efficient. So they are probably less attractive to dump money on. As of currently they still benefit from expressing lots of patterns that nobody bothered formalizing.
I would not say cheap per say, but in the context of "where is the business success going to be" that's one of the reasons why Mistral focuses on providing tuned model on premises to their customers.
> The point that the major labs don’t seem to get is that the vast majority of use cases simply don’t need models that have 50 PhDs and can speak 12 languages
I think they understand it but they think they can get away with it because they own the narrative. As long as they can make people believe they need such a model, it doesn't matter if it's true or not.
They are playing the cloud playbook, it didn't matter that most companies didn't need 99,999% uptime and instantaneous horizontal scaling, as long as people believed they did they are happily paying 10-100x the cost to AWS instead.
Every time I see this kind of story, two things bother me.
First, I have watched the free improvement of frontier models surpass the gains from retraining, many times now. Squeezing more out of the models that already exist, or simply doing nothing and waiting, is a real strategy and it often pays better. The fair comparison is not against today's frontier but against whatever ships while you are still maintaining your fine-tune.
Second, the $500 training bill is the cheapest line item in this story. The expensive parts are creating the data and maintaining the model afterwards. How many use cases can actually produce 177k scored episodes? Here they had to generate them synthetically from Amazon Berkeley Objects. To me, that dataset is the strongest evidence in the article of how hard fine-tuning is to apply: if the data existed naturally, nobody would need to manufacture it.
Regarding your first point, if the cost of not doing it while you wait exceeds the cost of training/maintaining the model, then this approach still makes sense. This really depends on how fast you expect cheap (and open-source) models to improve. So it might only be a temporary strategy, but still worthwhile. Also, I expect that specialized models will always be better (cheaper or better outcomes) than a general model, similar to how specialized HW like GPUs is still used even though CPUs have improved a lot too.
Maintaining in the way of ongoing training is relatively cheap, considering that the initial training run is only 500$ (in this example). Creating new examples to account for drift in the training data is more expensive, but these examples can then be reused when training a new model. And to some degree you need them anyways, to evaluate the models and prompt changes you make.
Overall, as you pointed out, this won't always make sense (either due to the cost of creating examples and training, or simply the lack of available data). But they point this out themselves in the diagram towards the bottom of the page: it only makes sense for frequent and verifiable tasks. This might result in this only being a sensible approach for very large companies (they talk about millions of decisions), but it might make sense for them.
Data drift. The data you are doing real world inference on can start changing, meaning your model performance starts degrading, so the model needs to be retrained on new data (that you have to collect and prepare).
Sure, but that apparently only costs $500. If you do that once a month then that's still basically free. Hell if you do it once a week it's still about as much as a single cheap employee.
you missed the parent comment's point that the $500 training run was merely the last step, generating the training data is likely to be significantly more expensive and time-consuming. might still pay for itself, but it's not a trivial "spend $500 every month" decision.
What I really started to notice is that SOTA models are really good at putting themselves out of the job.
We can see this already with GPT how luna can do 90% of what sol is used for. The only reason why china still bothers 'distilling' models is accurate training data generation, something that oai and anthropic had to spend years collecting while trying to dodge legal challenges.
The more intelligent models get, the more people will offramp to cheaper solutions that get the job done. There's no real benefit to using a sota model when the accuracy is already 99% and I think that is the biggest danger to US labs.
Sometimes that'll turn up real bugs, sometimes just overengineered designs, premature-optimization, and 1-in-a-million possibility "bugs".
And sometimes it's not about the model, it's just about refining the search space. E.g. I've had Opus write tests and GPT 5.5 write the implementation passing all the tests. Then ask about that specific implementation and find some real corner cases. Add those to tests, etc.
But the other fun trick that's been working better and better on the GPT-5.6 series is that even the lower-end models can find the things they didn't think of first when inspecting the already-written output.
I think there's still a bit of hard-to-quantify "creativity" to the bigger models - especially when trying to untangle (a) is this edge case that the model built a complicated way to avoid real/worth worrying about and also (b) even if it is real, is there not a better way to mitigate it? But it might be confirmation bias, in a way that definitely didn't use to be true about GPT-5.3 for planning and Composer 2 for implementation, say.
> even the lower-end models can find the things they didn't think of first when inspecting the already-written output
I use Deepseek V4 Pro for my hobby project (an OpenCode client). The economics will obviously be different at work.
It’s worse than the GPTs, but indeed, if you focus it on reviewing its own code (essentially spending more reasoning and changing perspective), it’s also quite capable at improving its own approaches.
The GPTs are better at general architecting, but I think a lot of the performance gains also came from my more careful prompting (“okay, this is a hard problem, let’s think this through …”) to make using the more expensive model worth it. I use them and they are worth it on subsidized rates, but not at API prices. Since my Codex sub ran out I sometimes miss the models, but it really hasn’t devastated me.
On my own time I still like the 'chat' format of me writing a short prompt and having the model turn it into a few dozen lines of code, or make a simple refactor (like turn a field into a method parameter.
I wasn't impressed with Chinese models (I tried Deepseek V4 Pro/Flash, and GLM5.2) the amount of questionable code and mistakes and misunderstandings meant the money saved and faster speed of models didn't translate to faster progress for me.
> They do tend to overengineer. The other day 5.6 Sol generated a while loop around a uuid4 call to make sure the generated ids were unique...
That’s not overengineering, it’s plain nonsensical, because presumably it doesn’t compare it to all IDs generated in the past. Which, if you wanted to do that, you’d use a database with a uniqueness constraint, in case you don’t already have that anyway.
Fixing this lack of reliable common-sense awareness seems to remain elusive for LLMs.
You can have sol write code and terra will find a plethora of bugs and rip the code apart. In reality this is just the nature of advisory prompting and why advisor from omp.sh is such a great feature. They get caught as they're being written.
Or even more poignant, have terra write code and it will find lots of bugs in its own code. But let it iterate code reviews/fixes/test cases a few times, and you’ll have something nice.
The only cases where weaker models fail entirely and the frontier models really come through is when you have a non-obvious bug in a larger codebase - one that requires tracing lot of calls through the ast (especially in multithreaded code) to understand what might be going wrong.
This is a general theme with technology and the 'S-curve'. Let's not even get into whether the improvement for AI reasoning ability has slowed - for practical purposes of writing a React frontend, it has.
But other tech is like that - I don't even remember when I bought my LCD TV - 2018 I think? I have no inclination of buying a new one.
Technology has a tendency to replace new technology, or intrude into vacant areas, but its very rare for technology to replace non-technology (like human interaction).
Most of the recreation humans do in front of screens is tending to (para)social relationships.
I can’t see the difference between terra and sol but I always use Sol. You guys can tell the difference. Even medium to xhigh is not that clear the difference.
Amazing! I've been in a similar autoresearch-y rabbit-hole lately with getting Apple's 3B Foundation Model to match Sonnet 4.6 on a very specific task.
The result was: 90% parity achieved with a weird combination of a fine-tuned adapter + 1 deterministic step.
I didn't read the Ramp article but this reads like a post hoc fallacy. Companies with 2x revenue have money to spend on AI. Companies with 1.15x revenue don't.
I like the 2x2 grid that describes when to fine-tune a model, when to use a frontier model, etc.
From the article it’s not clear how the scorer grades every episode - was it a frontier model that assigned the grade? How does that continue to work as the model that is being fine-tuned becomes better at the task than the frontier model?
I want small models to win and I am constantly experimenting with them. I have never tried fine-tuning and do not have that kind of budget. My approach is to remove some of the burden from models and bring into the agent.
Tool calling is an example - in some tasks RAG works really well, including coding agents where code, git log, Epics/Tasks, dependencies sources, etc. are all available in very structured manner. You can save many extra tool calls if you can run separate prompts and retrieve the source data needed for the actual work - rather its prompt.
And I really want to focus on search - this is the key technology if we want to use RAG instead of fine-tuning. If we can present really contextual sources in the prompts using a hybrid search approach - you can see how easily we get better results - either decisions or summaries from even small models.
This continuous cycle of fine-tuned open models beating frontier on (often vaguely labeled/defined) benchmarks doesn't provide an accurate comparison to the expanding generalized capabilities of the SoTA, which makes them effectively meaningless.
If we were to take these at face value, why is it that the frontier labs' models are making legitimate new discoveries (e.g. Erdős and Jacobian conjectures) and these models are not?
To me, a better signal of capability would be similarly performing novel work at the same or better level, which they presently are not. I say this as someone who very much looks forward to open models being more capable, but to deny the gap is misguided hopeful hype.
Why? Why is the premise that Fine-tuned models should be geared towards new discoveries?
The point of Fine-tuning small models is for specific downstream tasks, which SOTA models can do, but at higher costs. It is purely an economic play, not an attempt at pushing boundaries of SOTA.
I'm not suggesting that fine-tuned models don't have their place, all I'm saying is that the constant drumbeat of "cheap model X beats more expensive model Z" completely misses that the more expensive model is capable of doing more things at a higher level.
If the appropriate qualifiers were added to say "cheap model X does better at test Y than expensive model Z when we fine tune X to take Y test of existing knowledge" then it would be a more accurate statement, but naturally less impressive.
> I'm not suggesting that fine-tuned models don't have their place, all I'm saying is that the constant drumbeat of "cheap model X beats more expensive model Z" completely misses that the more expensive model is capable of doing more things at a higher level.
What if you have to do the task a billion times? Which model will you choose?
This is about saving money by using the right tool for the job. If you have a system that does a lot of mundane, repetitive work, you don't need a frontier model to do it.
It doesn't mean frontier models aren't good at harder tasks.
Huge SOTA models are like a floodlight. They elucidate a huge area at once.
A fine-tuned small model is like a laser pointer. It only illuminates a tiny specific spot. But it can illuminate it as brightly as the huge floodlight, for a tiny fraction of cost.
Depends on use case. That email classifier for legal emails: cheaper at scale as a small tuned model. Let alone better for the planet. Frontier model may have done that tuning!
Share of the maximum achievable score our GRPO-trained 9B open-source model reached on catalog review, vs 76.9% for the best frontier configuration: a 13.5% relative improvement over the frontier, and 36% over its own untrained base (64.2%). The five frontier models, even with optimized prompts, plateaued within a tenth of a point of each other; the trained specialist cleared that ceiling."
_______
This is hard for me to believe. I have a lot of skepticism that frontier models like GPT 5.5 that are likely 2T+ parameters in size only got about 12% more accurate than an untrained 9b parameter LLM.
Why? This is a very narrow task, it’d be surprising if the results were different actually; more interesting question would be how an even smaller model performs in the same finetune.
They built their own benchmark and then trained directly against its scoring function.. seems to be the rage, but nothing convincing from the article alone.
The point that the major labs don’t seem to get is that the vast majority of use cases simply don’t need models that have 50 PhDs and can speak 12 languages. Most use cases are defined within constraints where costs matter a lot.
As open weight models and cheap fine tuning services become the norm the whole economic framework of these mega models the labs are in an arms race building just completely crumbles. As does the economic picture that justified the massive infrastructure building that’s now broadly funded by a complex network of debt.
This is what makes open weight models so threatening to them. The political and “it’s China” angle is mostly just a cover for the real reasons why they’re freaked out.
The fact that models are now a pure commodity is bad enough for the big labs. If small open weight models become the norm the big labs are toast.
Very true. In fact, there's a lot more under the surface that will come to light once more efficient tiny models come out.
P.S. Building something that proves that you don't need those many params even.
https://github.com/guilt/tinytot
The premise of your project seems compelling. Is it novel or building on existing work? Anything one could read or watch to get introduced to that area of research?
But then can't the case be made that narrower and stricter-defined use cases are better served by more conventional ML? If/Wherever efficiency is a concern, that is.
That's assuming conventional ML and fine-tuned SLMs are interchangeable with acceptable behavioral change/degradation for any specific case
LLMs fit between well defined and poorly defined data; if you have well defined data, you can just feed it to ML models and they'll do what you want.
LLMs will take poorly defined data and "potentially" create well defined output.
So ideal systems will likely be constructed with a LLM on one end and a ML on the other and some how, if you can feedback poor data from the ML back into the LLM to clean up the data, you have a magic layer that doesn't care as much about the structure of the data.
That's a lot of supposition, but that's the difference I see between what we can do with LLMs and how ML models are structured. They might be classified as doing the same thing, but their data inputs are vastly different.
As someone who worked at multiple startups, I am pretty sure they get it but it doesn’t fit their goals.
They want to moat where they don’t need to compete with other companies because they have something that other companies can’t have.
In my opinion this is a short-sighted and greedy worldview. Haven’t seen it work personally. It is a different version of the month-to-month salary guy thinking he will be a billionaire and having that thrash “mindset”.
The reality is that practically none of those companies will amount to anything and they would be better off weighing the usefulness aspect of their output more. Instead they are imagining they will be Google.
Anthropic and openai ofc are the pinnacle of this greed culture and they correspond to FTX from the crypto trash hype so I don’t think they fit into the scale of sensibility.
Coming from this perspective, it is pretty easy to see what they are.
Yeah and these latest and greatest models universally suck hard for any use case outside of the few 'blessed' ones. Like I'm sure their ability to write prose and generally sound like a human being has regressed quite a bit, but even if not. Opus 5 is barely above GPT4 when it comes to stuff like home improvement advice.
Fine timing takes time and data, though. If smart enough models get cheap enough, then most people have lots of use cases that are cost insensitive enough that it's not worth the effort.
Of course, "smart enough" is a low enough threshold for most uses that this is still a problem for the frontier labs.
But at the same time, a truly smart enough closed model could also potentially command almost whatever they'd care to charge for it.
Whether they can actually get to that level remains to be seen, but I can definitely see a situation where most people are perfectly happy with cheap middle of the tree models while large corporations pay magnitudes more than current API pricing for access to models never even marketed as a mass market product and keep the labs afloat.
It's of course be a lot easier for them to find the path towards that of they didn't need to compete with open models in the meantime.
Yes. Fine tuning is an optimization. Like all optimizations, it’s downstream of figuring out your problem, implementing a naive solution, scaling the naive solution, and getting frustrated by SWAP-C. THEN you start poking at what optimizations are possible, choosing an approach, implementing the optimization, and redeploying.
This cycle happens when you have a well defined use case with high volume. It is the far opposite end of the spectrum from the general purpose intelligence on tap that frontier AI models purport to deliver.
I think focused fine tunes and big general models will coexist. Ideally with smart routing and caching to use small, specialized and local options when appropriate.
seems that, like software engineering and other areas before, they also have to rediscover that one single monolithic solution that handles everything is too inflexible and not maintainable, it's just a bad approach. people don't need the models they use to generate their codebase to also be able to translate Shakespeare into gen-z slang
Some specialized models may end solving themselves by helping fit the problem with the appropriate regular algorithms/formulas, which are a million times more efficient. So they are probably less attractive to dump money on. As of currently they still benefit from expressing lots of patterns that nobody bothered formalizing.
what are those cheap fine-tuning services these days?
I would not say cheap per say, but in the context of "where is the business success going to be" that's one of the reasons why Mistral focuses on providing tuned model on premises to their customers.
Tinker is pretty cheap. Prime Intellect if you want more flexibility.
> The point that the major labs don’t seem to get is that the vast majority of use cases simply don’t need models that have 50 PhDs and can speak 12 languages
I think they understand it but they think they can get away with it because they own the narrative. As long as they can make people believe they need such a model, it doesn't matter if it's true or not.
They are playing the cloud playbook, it didn't matter that most companies didn't need 99,999% uptime and instantaneous horizontal scaling, as long as people believed they did they are happily paying 10-100x the cost to AWS instead.
the major labs want to advance science. current business use cases are a happy accident
That party is over. They’re all on the clock now to show they can make money or the plug will be pulled.
Every time I see this kind of story, two things bother me.
First, I have watched the free improvement of frontier models surpass the gains from retraining, many times now. Squeezing more out of the models that already exist, or simply doing nothing and waiting, is a real strategy and it often pays better. The fair comparison is not against today's frontier but against whatever ships while you are still maintaining your fine-tune.
Second, the $500 training bill is the cheapest line item in this story. The expensive parts are creating the data and maintaining the model afterwards. How many use cases can actually produce 177k scored episodes? Here they had to generate them synthetically from Amazon Berkeley Objects. To me, that dataset is the strongest evidence in the article of how hard fine-tuning is to apply: if the data existed naturally, nobody would need to manufacture it.
Regarding your first point, if the cost of not doing it while you wait exceeds the cost of training/maintaining the model, then this approach still makes sense. This really depends on how fast you expect cheap (and open-source) models to improve. So it might only be a temporary strategy, but still worthwhile. Also, I expect that specialized models will always be better (cheaper or better outcomes) than a general model, similar to how specialized HW like GPUs is still used even though CPUs have improved a lot too.
Maintaining in the way of ongoing training is relatively cheap, considering that the initial training run is only 500$ (in this example). Creating new examples to account for drift in the training data is more expensive, but these examples can then be reused when training a new model. And to some degree you need them anyways, to evaluate the models and prompt changes you make.
Overall, as you pointed out, this won't always make sense (either due to the cost of creating examples and training, or simply the lack of available data). But they point this out themselves in the diagram towards the bottom of the page: it only makes sense for frequent and verifiable tasks. This might result in this only being a sensible approach for very large companies (they talk about millions of decisions), but it might make sense for them.
And the hidden, trial and error runs tweaking the hyperparams. guaranteed, it cost >$5000 in training runs alone.
What do you mean with maintaining the model? I am puzzled.
Data drift. The data you are doing real world inference on can start changing, meaning your model performance starts degrading, so the model needs to be retrained on new data (that you have to collect and prepare).
Sure, but that apparently only costs $500. If you do that once a month then that's still basically free. Hell if you do it once a week it's still about as much as a single cheap employee.
you missed the parent comment's point that the $500 training run was merely the last step, generating the training data is likely to be significantly more expensive and time-consuming. might still pay for itself, but it's not a trivial "spend $500 every month" decision.
What I really started to notice is that SOTA models are really good at putting themselves out of the job.
We can see this already with GPT how luna can do 90% of what sol is used for. The only reason why china still bothers 'distilling' models is accurate training data generation, something that oai and anthropic had to spend years collecting while trying to dodge legal challenges.
The more intelligent models get, the more people will offramp to cheaper solutions that get the job done. There's no real benefit to using a sota model when the accuracy is already 99% and I think that is the biggest danger to US labs.
Anything but Sol is not sufficient for complex (or even just large) enough code.
The upper end is all about coding. If I have terra on extra high write code, Sol will find a plethora of bugs and rip the code apart.
Anything else? Sure use a cheaper model.
Sometimes that'll turn up real bugs, sometimes just overengineered designs, premature-optimization, and 1-in-a-million possibility "bugs".
And sometimes it's not about the model, it's just about refining the search space. E.g. I've had Opus write tests and GPT 5.5 write the implementation passing all the tests. Then ask about that specific implementation and find some real corner cases. Add those to tests, etc.
But the other fun trick that's been working better and better on the GPT-5.6 series is that even the lower-end models can find the things they didn't think of first when inspecting the already-written output.
I think there's still a bit of hard-to-quantify "creativity" to the bigger models - especially when trying to untangle (a) is this edge case that the model built a complicated way to avoid real/worth worrying about and also (b) even if it is real, is there not a better way to mitigate it? But it might be confirmation bias, in a way that definitely didn't use to be true about GPT-5.3 for planning and Composer 2 for implementation, say.
> even the lower-end models can find the things they didn't think of first when inspecting the already-written output
I use Deepseek V4 Pro for my hobby project (an OpenCode client). The economics will obviously be different at work.
It’s worse than the GPTs, but indeed, if you focus it on reviewing its own code (essentially spending more reasoning and changing perspective), it’s also quite capable at improving its own approaches.
The GPTs are better at general architecting, but I think a lot of the performance gains also came from my more careful prompting (“okay, this is a hard problem, let’s think this through …”) to make using the more expensive model worth it. I use them and they are worth it on subsidized rates, but not at API prices. Since my Codex sub ran out I sometimes miss the models, but it really hasn’t devastated me.
On my own time I still like the 'chat' format of me writing a short prompt and having the model turn it into a few dozen lines of code, or make a simple refactor (like turn a field into a method parameter.
I wasn't impressed with Chinese models (I tried Deepseek V4 Pro/Flash, and GLM5.2) the amount of questionable code and mistakes and misunderstandings meant the money saved and faster speed of models didn't translate to faster progress for me.
They do tend to overengineer. The other day 5.6 Sol generated a while loop around a uuid4 call to make sure the generated ids were unique...
I wonder if that's also how I'd behave if I had a reinforcement learning harness around me that dived hard on and punished me for every small mistake.
> They do tend to overengineer. The other day 5.6 Sol generated a while loop around a uuid4 call to make sure the generated ids were unique...
That’s not overengineering, it’s plain nonsensical, because presumably it doesn’t compare it to all IDs generated in the past. Which, if you wanted to do that, you’d use a database with a uniqueness constraint, in case you don’t already have that anyway.
Fixing this lack of reliable common-sense awareness seems to remain elusive for LLMs.
You can have sol write code and terra will find a plethora of bugs and rip the code apart. In reality this is just the nature of advisory prompting and why advisor from omp.sh is such a great feature. They get caught as they're being written.
Or even more poignant, have terra write code and it will find lots of bugs in its own code. But let it iterate code reviews/fixes/test cases a few times, and you’ll have something nice.
The only cases where weaker models fail entirely and the frontier models really come through is when you have a non-obvious bug in a larger codebase - one that requires tracing lot of calls through the ast (especially in multithreaded code) to understand what might be going wrong.
This is a general theme with technology and the 'S-curve'. Let's not even get into whether the improvement for AI reasoning ability has slowed - for practical purposes of writing a React frontend, it has.
But other tech is like that - I don't even remember when I bought my LCD TV - 2018 I think? I have no inclination of buying a new one.
Technology has a tendency to replace new technology, or intrude into vacant areas, but its very rare for technology to replace non-technology (like human interaction).
Most of the recreation humans do in front of screens is tending to (para)social relationships.
> I don't even remember when I bought my LCD TV - 2018 I think? I have no inclination of buying a new one.
LCD backlights are usually rated for 5-10 years of normal use so your inclination might change soon.
Eh, I’m using an 15-year old LCD monitor and its still fine. One generally doesn’t use these at full brightness.
I can’t see the difference between terra and sol but I always use Sol. You guys can tell the difference. Even medium to xhigh is not that clear the difference.
A fine-tune will generally outperform just about any other method on a closed domain, non generative problem.
LLMs are great because they can handle open domain problems in part because they are generative.
Amazing! I've been in a similar autoresearch-y rabbit-hole lately with getting Apple's 3B Foundation Model to match Sonnet 4.6 on a very specific task.
The result was: 90% parity achieved with a weird combination of a fine-tuned adapter + 1 deterministic step.
I wrote the whole thing down here: https://alexisrondeau.me/tada/research/FMDiscovery/dashboard... which includes the question, the answer, the 96 experiments and their lineage etc. etc.
> I wrote the whole thing down here
I like the idea, but it looks really hard to read. Try reading it top to bottom without skipping.
Oh, thank you for noticing.
Is it a layout issue for you (it lays out better on desktop) or the language of the text or...? Let me know!
I didn't read the Ramp article but this reads like a post hoc fallacy. Companies with 2x revenue have money to spend on AI. Companies with 1.15x revenue don't.
I like the 2x2 grid that describes when to fine-tune a model, when to use a frontier model, etc.
From the article it’s not clear how the scorer grades every episode - was it a frontier model that assigned the grade? How does that continue to work as the model that is being fine-tuned becomes better at the task than the frontier model?
I want small models to win and I am constantly experimenting with them. I have never tried fine-tuning and do not have that kind of budget. My approach is to remove some of the burden from models and bring into the agent.
Tool calling is an example - in some tasks RAG works really well, including coding agents where code, git log, Epics/Tasks, dependencies sources, etc. are all available in very structured manner. You can save many extra tool calls if you can run separate prompts and retrieve the source data needed for the actual work - rather its prompt.
And I really want to focus on search - this is the key technology if we want to use RAG instead of fine-tuning. If we can present really contextual sources in the prompts using a hybrid search approach - you can see how easily we get better results - either decisions or summaries from even small models.
This continuous cycle of fine-tuned open models beating frontier on (often vaguely labeled/defined) benchmarks doesn't provide an accurate comparison to the expanding generalized capabilities of the SoTA, which makes them effectively meaningless.
If we were to take these at face value, why is it that the frontier labs' models are making legitimate new discoveries (e.g. Erdős and Jacobian conjectures) and these models are not?
To me, a better signal of capability would be similarly performing novel work at the same or better level, which they presently are not. I say this as someone who very much looks forward to open models being more capable, but to deny the gap is misguided hopeful hype.
Why? Why is the premise that Fine-tuned models should be geared towards new discoveries?
The point of Fine-tuning small models is for specific downstream tasks, which SOTA models can do, but at higher costs. It is purely an economic play, not an attempt at pushing boundaries of SOTA.
I'm not suggesting that fine-tuned models don't have their place, all I'm saying is that the constant drumbeat of "cheap model X beats more expensive model Z" completely misses that the more expensive model is capable of doing more things at a higher level.
If the appropriate qualifiers were added to say "cheap model X does better at test Y than expensive model Z when we fine tune X to take Y test of existing knowledge" then it would be a more accurate statement, but naturally less impressive.
Maybe because people who are target audience don’t need to have it spelled out like that?
People who are not really into it, don’t care.
> I'm not suggesting that fine-tuned models don't have their place, all I'm saying is that the constant drumbeat of "cheap model X beats more expensive model Z" completely misses that the more expensive model is capable of doing more things at a higher level.
What if you have to do the task a billion times? Which model will you choose?
Speed play also you can get much faster responses with 9b model.
This is about saving money by using the right tool for the job. If you have a system that does a lot of mundane, repetitive work, you don't need a frontier model to do it.
It doesn't mean frontier models aren't good at harder tasks.
That's fair and I agree with this framing.
That was always the framing. I don’t understand your pedantry.
Huge SOTA models are like a floodlight. They elucidate a huge area at once.
A fine-tuned small model is like a laser pointer. It only illuminates a tiny specific spot. But it can illuminate it as brightly as the huge floodlight, for a tiny fraction of cost.
Depends on use case. That email classifier for legal emails: cheaper at scale as a small tuned model. Let alone better for the planet. Frontier model may have done that tuning!
turns out the bottleneck was never model size, it was having someone who actually understood the problem define the reward function
I agree very much. I'm getting the same take-away even more now that I'm using autoresearch as my main strategy.
And I wonder: We've got all these amazing (programming) languages to define solutions; where are the languages to clearly define the problems?
"87.3%
Share of the maximum achievable score our GRPO-trained 9B open-source model reached on catalog review, vs 76.9% for the best frontier configuration: a 13.5% relative improvement over the frontier, and 36% over its own untrained base (64.2%). The five frontier models, even with optimized prompts, plateaued within a tenth of a point of each other; the trained specialist cleared that ceiling."
_______
This is hard for me to believe. I have a lot of skepticism that frontier models like GPT 5.5 that are likely 2T+ parameters in size only got about 12% more accurate than an untrained 9b parameter LLM.
Why? This is a very narrow task, it’d be surprising if the results were different actually; more interesting question would be how an even smaller model performs in the same finetune.
The problem i've had with finetuning models is that most of the time better prompting beats finetuning
Better prompting doesn't improve response time or price!
Yes, it can.
There seems to be no hold out data for test, so this is just overfitting?
RL is the gold standard and significantly beats self learning methods, but other than coding and computer refereed games it's cost prohibitive.
What benchmark is it? Is it super niche?
They built their own benchmark and then trained directly against its scoring function.. seems to be the rage, but nothing convincing from the article alone.
As AI agents become more autonomous, governance and auditability will become critical, not optional.
Comparing the revenue of the top quartile of AI-using companies to the average of all non-AI-using companies is beyond intellectually dishonest.
Tldr: we don't know what we are doing like the rest of 99% AI teams.