> But I also think the demand for "fast/cheap/good-enough" models is just about to take off.
There's a sort of "revelation" I had in ~early '24 when I used a 7B local model with a library called Guidance (initially out of MS, then the team moved) to create a flow where the model would receive pseudocode for tests, first write the tests, and once I approved then started writing code until the tests passed. This was before "thinking" models, and yet using that library I was able to "guide" the model in the required "prompt / instruct" context such that it was working towards completion, and I saw the first things like we see now in the thinking traces "oh, test x doesn't pass because blah, I need to..." and so on.
Anyway, the revelation was "even if the models never improve, I'll have years of fun finding out all the ways I can use these things". And, obviously, the models improved a lot since then. But I think that revelation can still be applied, as a sort of "truism". We have, right now, access to things that 10-20 years ago would be considered magic. We are still finding ways of cobbling together systems with glue, duct tape and prayers and find new things they can do.
I think the "good-enough" stage has come not just for API models (cheap, fast, etc) but for local as well. Even if slower, even if clunkier, but they are good enough for a set of ever increasing tasks, and what's more it's incredibly fun to work with them.
The infancy phase of this technology is represented by the pursuit of making wildly grand, wildly expensive, all-purpose models that somehow discern a user's full accurate intent from a lazy, underdeveloped, vague idea that they ambiguously and poorly express in a couple dozen words.
The adolescence will arrive as those outsized and ill-considered ambitions collapse and we instead see a cambrian explosion of restrained but efficient model+harness-tuples that have been distilled, finetuned, and rigged to deliver on narrowly scoped but idiosyncratically-shaped tasks with incredible efficiency and erogonomics.
This idea has failed to pan out time and time again - people have an instinct that hand-crafted finely-tuned specialized AI systems must be optimal, but throwing more scale and compute to something more generally smart always wins out. It's especially palpable just looking at the last few years of LLM's: a frontier model with all the world knowledge you can stuff in it and every tool at its disposal has always performed the best at all tasks. Suggesting otherwise has become an extraordinary claim requiring extraordinary evidence.
The Bitter Lesson is very popular right now. It seems true right now. It’s having its moment right now. That doesn’t actually mean it’s axiomatically true.
Commenter below gets it absolutely correct: stockfish, which runs on your 5 year old phone, is dramatically better at chess than Fable. Like, so much better that it’s not even remotely comparable. The theory of the Bitter Lesson, and it’s only a theory, is that LLMs could eventually outperform stockfish. It’s not true today and it remains to be seen whether it will ever be true. For now, specialized models are absolutely better at specialized tasks.
This idea has not failed to pan out at all. I work for a startup that is exactly what GP described, and am set for life because of how wildly successful it is. Notably, we are successful, in a genuine sense of the word: we bootstrapped from running tiny models to larger and larger models on our own slowly improving fleet of GPUs, and now have millions in revenue without a single dime of outside investment. Conversely, you cannot call taking on ~1 trillion in debt and purchase commitments to scale "success". OpenAI and Anthropic are underwater financially. To be precise, they're in the Mariana Trench.
Bitter lesson #1: don't waste time optimizing code when a faster processor is around the corner.
What countered it: Moore's law stopped working.
Bitter lesson #2 similarly relies on scaling laws that might have diminishing returns wrt model runtime vs intelligence. Runtime matters for turnaround on the problem you're solving.
This is a misunderstanding of either the bitter lesson or what was being claimed, on multiple accounts. Firstly, the bitter lesson is merely about human expertise-tuned algorithms vs. throwing raw compute at a domain. But, notably, it is still domain-specific. No matter how much compute you throw at training an LLM, it is never going to beat a Chess engine at Chess. If you give a Chess engine 1,000,000 compute units and a general-purpose LLM 1,000,000 compute units, the Chess engine is obviously superior at Chess; ergo, there is value in throwing compute units into training models for specific tasks. This is true for within several orders of magnitude of compute, in fact. It's also true that if you give the Chess engine 1000 compute units it'll still beat the all-purpose model with 1,000,000 units, so actually there's a lot of value in training for specific tasks.
Secondly, the bitter lesson is predicated on compute being cheap. There was a period where a hand-tuned algorithm informed by human expertise would outperform a raw alpha-beta search at Chess. Then compute got cheaper, and DeepBlue ascended to the top. Compute is now expensive again relative to the tasks being performed. We are absolutely still in a period where human expertise in training LLMs will outperform a naive approach with more raw compute.
I don't know much about chess engines; do they still use hand-tuned algorithms, or are they more like AlphaZero, where they learn through self-play to beat any/all possible human contenders? I don't believe DeepBlue was automated to that extent, but it may have been.
In the latter case, the chess example would tend to support the Bitter Lesson, rather than refute it.
I would also be VERY slow to claim that general-purpose models will never be competitive at chess. It wasn't so long ago that transformers couldn't add two-digit numbers reliably without resorting to tool use. They are now as good at "mental arithmetic" as any human savant. It wouldn't surprise me at all to see someone come up with a model that just happens to be really, really good at absorbing the portions of its general training data having to do with chess.
In fact you could argue that AGI demands such a model, if we are to assume that LLMs are a guidepost in that direction.
The issue is that GP is misusing the bitter lesson. Yes, search + learn tends to be more effective than human rules based strategies, but that's not what's being considered here. The original claim is effectively that AGI isn't needed for most tasks and more value can be created by using search + learn to solve specific problems instead of applying general models to every problem. Then GP commented a non sequitur
Suggesting otherwise has become an extraordinary claim requiring extraordinary evidence.
VibeThinker 3B constitutes extraordinary evidence, IMO. The first such evidence I've seen myself. Very small model, very low literacy, almost no world knowledge, but it is as good at math and logical reasoning as models a hundred times larger.
The Bitter Lesson is a valid and trenchant observation about how about we got here, but I think it's a mistake to assume it tells us very much about where we're going. Too much has changed recently and is still doing so.
There's A LOT low hanging fruit still out there for sure. And with antigenic systems being able to do the boring repetitive work of looking for that low hanging fruit I think we will see interesting things indeed. Also I think heuristics is where its at for such things. Once you describe some good heutistical structures for the research models to always follow related to "creativity" and such things, thats where we will see biggest difference. The agentic systems know the scientific method well and can follow it they just need the ability to be "creative" so their sampling becomes less rigid.
Yep, I've been having excellent experiences with the models even from the 2023 era. They required a lot of "holding it right" (mostly: being very precise in what went into the context) but their raw coding capabilities were astonishingly good even then.
However, back then I was getting the AI to write individual functions or classes or a test suite. I was decomposing the larger task into smaller tasks, delegating some of them to the AI, reviewing the results and composing the codebase from those. I was also essentially the harness.
Today the models can write and test and deploy an entire project. In terms of the code quality, I actually don't think today's frontier models would have written it much better than the 2023 models did. So in terms of raw coding capabilities i.e. converting a high-level specification into working code, I think we hit the peak way back in 2024 itself.
What has changed is the AI has learned how to do the task I was doing (besides being the "harness"!), which was the mid-to-higher level "engineering" aspects like decomposing a task, specifying it to a reasonable level, reviewing the outputs, and course correcting as needed.
I'm not sure if that is something the AI labs explicitly focused on during training (which may be why Meta is having its highly paid engineers do annotation work), or an emergent property of "better reasoning" (which I believe Dario implied in a podcast), or some mix of both.
But the fact remains that even the weaker models are more capable than we realize, and many being open weights, are here to stay.
To me, most local models work just fine for anything you can be patient for. If I want something quicker, I will go to a SOTA model via API, but with multiple 3090s, I have never really needed a hosted model for a lot of my experiments.
For code, they are great, but for creativity for NPC controllers, they leave something to be desired, but work well enough for testing, so I don't burn tokens until I'm actually playing my games.
But nothing one-shots a prototype better than Fable 5. I can have a prototype built in 30 minutes, hooked up to my local LLMs and Claude Code is very good at testing the interactions and even tuning the prompts of the NPCs for better experiences.
Having multiple 6 year old cards doesn't seem like it's that big of burden for local LLMs.
I get that a lot of people don't have them. And a single one can be VERY performant. And the smaller models like a 7B can run on much smaller hardware like a mid-range [3|4|5]060.
My entire AI Dev Box cost $4500 in parts. 128GB RAM, i7-10700, 1TB and 2TB SSD, and 2x 3090s. Today's prices and inflation have definitely made that price tag seem a lot better than it was, but it was an investment in all things GPU that were happening in 2020 (crypto, blender, image gen), then LLMs exploded.
It seems roughly similar to the pricing level of personal computers in the early eighties (i.e. IBM PC and Apple Macintosh). I’d expect prices to come down significantly over the next few years. Not so much in the next year or two, but after that.
Just like for warships, the complexity and cost of building cutting edge hardware has grown exponentially up to a point where a significant chunk of the world's computing is dependent on 2 companies: ASML, TSMC. We shouldn't extrapolate linearly from examples from the 80s.
I don't think there is tunnel vision. I'm just saying that I have a couple 3090s I invested in a handful of years ago, and they are still going strong today as multiple GPU-needing technologies emerged.
I'm not saying everyone has to run local LLMs, because the APIs are in a race to the bottom, and my $10 of OpenRouter credits I bought months ago is down to $8.94 because most models give you MILLIONS of tokens for a US Quarter.
I mean, I'm not rushing out to buy that kind of hardware myself, but it is a matter of perspective. People commonly spend an order of magnitude more on a car, and that's just the sticker price.
I got a great deal on ~72 TB of NVMe right before storage prices shot up, doesn't make it any less ridiculous that I have it or any more relevant to people talking about building a NAS now. 99% of people, even in tech, do not have the stupid amounts of hardware people like us hobby on.
Most people in the US have a car, and the average new car is $40,000. Hell where I live a middle class consumer will spend double that on a Boat or an RV and think nothing of it. These aren’t elite tech workers.
It’s not unfathomable that if a personal, generally intelligent local AI provides enough utility and doesn’t require you to tweak CLI flags millions of Americans would want one.
They cost more to run than hosted anyway. But that isn't the point of having them. They are a playground, a backup when the internet is down, or claude is down. They can render Blender scenes pretty well. They play any game I want.
You can do each of those at various hosts and own nothing. Or own a couple "over priced" cards and do it all at home on battery power for a few hours while the power is out.
I keep seeing this comment. This is _hacker news_ where, back in the day, people just hacked on things, because it was a hobby. They weren't "moneymaxxing" or desperately trying to be as insanely efficient as possible. They hacked on stuff with a can of surge at 3am because it was fun.
Your comment is like a meta comment of "LLMs are generating everything, after a while the ouroboros will eat itself. (Which I agree with)" If people aren't hacking on this shit just because, you have completely conceded control of software to a handful of sociopaths, and open source software is dead.
I have trouble getting simple extraction to work sometimes. I have a block of text describing people and their roles at a company and their ages, and i asked for structured results of an array of these things with the text span that it appears in and all i can say is: nope.
While they are improving rapidly, or as you say even if they don't. The next stage is for hardware companies ( cough Apple cough ) to ship these Local Model ready hardware in their products.
It will be interesting to track the improvements of these 7B model over time.
There will be a turning point in the next few years where it attract enough consumer attention to create yet another Smartphone and PC super cycle.
I see it in a slightly opposite way: even the good models are relatively cheap, and so I worry what we might miss by spending too much time playing with the Sonnets of the world when the Opuses are still objectively a bargain for the power they bring.
I know companies that are using github, even using public repo, and request their teams to not use SOTA models, but are ok with local models.
Just stupid policy.
I think there's something subtle about language and ambiguity that means they aren't designed to become superintelligent autonomous machines. They're value is as information repositories that actual intelligent autonomous machines (us) mine and string together.
Yes LLMs are a beautiful way to compact knowledge. It would be such a cool technology to develop and worked with if it wasn’t linked to such a toxic industry
I think you're just observing ppl in one of these rare instances where enough of them come together because they are motivated. 'Toxic' is the clamoring sound of a crowded room where what gets through to your ears are just the most annoying snippets of incomplete conversations. I dare you to hang out with any actual people here, understand their viewpoint and listen to what they actually have to say in person, within the context of watching them do it.
Same. Mistral 7b has been more than I ever needed for text for years now.
Unless you must 1-shot with no harness it’s the same amount of power, maybe more because the big “good” models make too many assumptions and tend to become rigid.
Mistral 7b can do anything, and it’s basically instant even on an M3
What kind of work are you doing? For example, if I have some code in the hot path and I want to do all the usual tricks to help the compiler vectorize it, such a small model is not able to do much.
RAG is your friend (or any vector db). No model can vectorize an entire codebase in context.
Even a big mainstream product (like Gemini) cannot handle more than ~1k lines without missing details and making mistakes. And about every 1k lines, it seems to forget the previous 1k, doesn’t it? So you can never hold more than a file or 2 (or 3) in context at a time without losing details.
What you find is that the big models like Gemini are doing vector storage and retrieval too, and breaking prompts down into chunks for various models to handle to assemble a thorough response.
If you want that kind of control in your outputs, and be able to hold a lot in your inputs, I don’t see any other way regardless of which model you use.
BTW structured/constrained generation has so many places to trivially enable jailbreaking/alignment/safety problems that closed source models heavily limit the full expresivity of grammars and capabilities, particular of on-the-fly dynamic grammar construction/reconstruction.
ppl keep talking about the supposed unexplored and untapped "model overhang" but very few things in the world are where you can write elaborate test criteria to before using ai.
A sales person sending a prospect email doesnt have a way to write a test harness for it. Yet these tasks dominate what humans do compared to writing a crud app . otherwise anthropic wouldnt have trillions dollar valuation
I've amassed access to 4 different GPU rigs with 128GB to 72GB; I didn't this before I event touched an agentic engineering harness. It was sometime in February/March when I set them to first tackle small problems, and now with deer-flow, they're scaffolding full project/scope implementation and I'm finishing off the fine details around the problematic edges.
A dream of mine is to be able to host a LLM-powered video game that I can host on a home server running a decent mid-range GPU like the RTX 5060, and the LLM is fast and intelligent enough to make for a fun game experience for a few dozen concurrent players. People can ask for features and they just get made and added to the game on the fly for the lobby to enjoy. The hosting costs would be manageable enough that I don't have to charge anything for the game.
I think with one more year or so of small model progress, that might just be possible to accomplish.
It makes sense that we’ll see “room at the bottom” strategies. Currently, large parameter counts seem to be slush funds of world knowledge, language skills (because language’s nuances and open vocabulary make it high-dimensional), and reasoning primitives, the general belief being that the latter takes up the least space in the model.
There are many applications where world knowledge is unnecessary or even a negative, and in which only a small amount of language skill is necessary, and there we can expect small models more intelligently used to beat large ones naively used.
Perhaps we'll get to a point where believing any un-sourced information from an LLM will feel crazy. I don't want my model to know more than it needs to perform logic and use tools. Once it is capable of using tools I would much rather it looked up information or sourced it from existing context rather than just divine it from it's weights.
Probably. You can solve it with either some grounding context, or spending hundreds or thousands a month extra on a model that has more knowledge baked in. With modern harnesses, the choices is obvious.
Everyone wants this to be it but over and over we discover that the bigger a model is the better it is at all tasks, even ones far outside the domain it was optimized for. IE claude fable is better at writing both code and prose than smaller code- and prose-specific models.
The way vision and language models converge into the same geometric space should be extremely alarming for the "you don't need global knowledge for local tasks" type dreams.
And to be clear I'm not saying that smaller models don't or can't work well, or that we shouldn't be heading in this direction. And it's not quite the case that broad knowledge is strictly necessary. But it never seems to be negative! And so far it is the best way we've found to do... everything. Small models are good to the extent they are like big models, not to the extent that they are small.
On narrow domains, it is very common for small models to match or outperform larger ones at a fraction of the parameter count.
For example in language, this is called the “curse of multilinguality”. Small models that handle a single translation direction can easily outperform big ones that try to handle them all.
In any case, for most tasks the question is not “how many tasks can this model kind of do well” but “given time/cost constraints, what is the maximum level of quality we can achieve”. And for that, small models are usually very competitive.
Yes small models are and will be useful for lots of stuff for several reasons.
But the idea they’d be better than a bigger model is cope, you’re pretty much always better off running the biggest one you can bring to bear within your constraints.
I find it quite funny all these folks who are addicted to chasing frontier models, only just noticing that small models became "good enough" for most tasks. Those of us without fable-sized expense accounts noticed this quite a while back
Exactly! Composer 2/2.5 were amazing, cheap, and fast. Everyone else was Gaga about GPT 5.5 and such, while we were over here doing the work with less cost and more speed
I’ve been playing around with Luna, Terra and Sol and for the type of work I’ve been doing lately I actually think Sol is just a likely to trip up as Luna. Examples were Sol over assuming, persisting in the wrong direction, over engineering a little script to do some exploration of api. They can all be fixed but it’s a waste of tokens, I rather have Luna do it because course correction on small pieces of work is cheaper.
I've found the distinction to be in how much I care about how the final product looks. If I want high-quality code I typically find a smaller model with a well-designed spec to do better, if I want it to just run and produce something close to my vague description typically Sol does better. For most actual business use-cases I think the first is likely better but the experimentation speed up with the frontier is very nice.
The word “most” is doing a lot of work here. On a percentage basis perhaps most tasks a typical SWE needs to do when they aren’t in meetings or writing docs are just glorified autocomplete. But that’s boring and that’s why people don’t usually talk about it.
People are addicted to chasing frontier models because they all have memories of spending a week on a deeply challenging algorithm problem or even have crazy complicated algorithms they cannot implement themselves and want to have the models achieve this technical breakthrough. It’s the kind of productivity boost from spending one week on a problem to spending one hour. In contrast the productivity boost from spending ten minutes to spending one minute just doesn’t occupy people’s mind.
> crazy complicated algorithms they cannot implement themselves
I'm not sure I know very many engineers who would fall in this bucket. Or do you mean the business types who suddenly think AI can replace all the engineers?
It probably depends on the background and the company. For example if one works at a startup that happens to use technology, it’s unlikely to happen because SWEs just translate business rules to code. But if one works at the place where the technology itself is the focus, then yes most people will fall in that bucket.
In fact I noticed that this is the one place where people discussing AI on HN tend to talk past each other. On the one hand people are talking about supreme intelligence like designing new algorithms (on the same vein as finding counter examples for the Jacobian conjecture) and on the other hand people are just satisfied using AI to automate a few quotidian tasks that hitherto couldn’t be automated.
> There's obviously a lot we can optimize here, but if you're charging what the WSJ or The Economist charges, you'd better be delivering similar value.
Gosh, watching paint dry has been a better value than reading The Economist in the last 5 years or so.
That aside, I had good results with Luna. I'd be interested in hearing about a comparison that takes into consideration response time (not TPS), cost and performance of the popular models at different settings. That chart has some of that. For instance, is Luna Max a better value than Terra Medium?
I’m kind of cautiously excited for the next five to ten years, with these AI chips becoming incredibly fast and RAM capacities ramping up its in the cards that we’ll have chips like today’s ATMEL microprocessors that fit on a single board computer and can run small models locally, then all our gizmos can have local AI and I can have a truly intelligent home. Of course there will be a huge push to put all of it in the cloud but maybe we have a chance to take this technology home for good as it’s hard to imagine people will submit to this kind of surveillance required for AI home automation
24/7 (then again I might be wrong). Exciting times.
> the "token spewer" work. being ultra responsive, pushing the ball forward across dozens of different fronts... ~95% of the work he does falls into bucket 2. It's hopping on calls. Nudging people. Blocking and tackling.
I recently had some relevant experience: for a couple of months now I've been experimenting with on device models to summarize feeds in a Bluesky client I am developing. The feature extracts topic areas, categorizes posts, and creates a summary under each topic.
At first the results were hot garbage, and progress was slow. I hooked up the settings to download models from Hugging Face conveniently, so I could run experiments faster, and I massaged the prompts a bit. Last week this feature made a qualitative jump from science experiment to something I'd actually use.
The fact that all runs on the device means I've got no variable costs associated with adding this to what will be, at best, a pretty low revenue product. I've tested it on trailing edge devices like an M1 Mac and a Pixel 8, and performance is very tolerable.
The key is I'm not asking for open ended answers to open ended problems. When it proves to be useful it's not going to get less useful or more expensive.
There are vast domains of uses for LLM models with similar characteristics and likely similar results.
The tokens per second speed measurement is highly inflated nowadays because most of the tokens went into thinking. I wonder if there is a more realistic measurement for "effective speed", which accounts for thinking efficiency.
IMO big models are not a product in and of themselves. Inference is just a new type of compute. I'm confident that in two or three years, every product will have inference capabilities integrated into the experience, and models will become less and less distinctive from one another.
What most products need from a model is a pretty short list: the ability to make tool calls well, accurate recall, and the ability to follow directions without wavering (whether or not those directions are baked into the weights or provided in a system prompt). That covers 95% of inference utility in products.
We're nearly there, and I believe these capabilities will fit on small models.
Because of this though, I predict hardware demand will stay high despite demand for "hosted" inference dropping. Unless there's some regulatory shenanigans that step in to say otherwise.
100% agreed. Small, cheap, and hosted models. Luna (and open weight models and others) is ridiculously cheap @ $0.2/$1.2, easily accessible, and more than good enough for basic use cases (e.g. summarization, simple tool calling, etc.).
Small is relative. I'm looking for models that I can with run around 100 MiB mark (RAM just for the weights) to demo what you can do with this little memory.
I know of SmolLM 2 which in Q4 is borderline regarding the size and rather dated. There is also TinyStories, which is also old and also focussed on children's stories.
Is there anything newer in this category? Or should I try to distill something down to this size?
Maybe I'm being super reductive here, but operating small models at the core of your business kind of moves the needle from making external API calls (against frontier models) to running internal API calls (against your locally-run models). It seems like if we want local models to take off, it will need to become easier to run local models for cheap. I'm thinking like reducing the barrier of entry for running "local models" in the cloud providers like DigitalOcean, AWS, etc.
I only run local models and I don't give them access to much externally. I don't do anything serious with it, but it comes in handy and I know that they can do so much more. I'm on a meager RTX 3060 12GB and a GTX 1660 Ti with 6GB for some extra vram space. When I first started playing with local models, I was really impressed with what I was able to achieve locally.
That's great, but the thing that worries me is that many companies have billions invested in the AI bubble. It's around 1.5 trillion last time I looked. It's all circular spending between the companies building out the infrastructure, and the models. None of it is profitable. They will want to recoup that 1.5 trillion from consumers, which means using online-only pay-as-you-go cloud models. They will inevitably see that people using capable local AI are "lost customers" and they will try to kill the ability to locally host AI or somehow enshitify it enough to make paying a subscription more palatable.
I'm not saying I believe that will happen, I'm just worried that it will. Is anyone else worried about that as well?
I am very excited that more makers will come up with fast memory for consumers rather than enterprise. Companies can only pre order so much RAM.
At some point there will be a surplus of fast memory and even in a crash the current generation of SLMs are bounced to be plenty to build a lot of intelligence at home.
I forked my Big Serious Harness™ that models construction projects into a harness for building a vibe coded family assistant. I couldn't figure out how to make the toy operate at toy prices until Luna. Now you can vibe code all the little apps you might want for your fam for like $5 and operate it day to day for a few cents.
> Peter runs multiple companies. Beyond Segment, he's raised $100m+ for Charm Industrial, and just recently closed a Series A for Revoy. He's incredibly organized and efficient with his time.
You can do this before an exit? Build and fundraise for multiple (3?) companies at the same time?
I have trouble seeing the points of using less capable models.
I just want the smartest, best, and most capable models. It feels smaller models for speed and cost are just transitions towards better hardware allowing the very best model.
And that is why i always carry my groceries with an Antonov An-225 Mriya. Is it really needed? No, but i refuse to compromise on what is(was/will be) the best.
My experience has been that responsiveness is value. For tasks where you need steering, responsiveness allows for better steering. For tasks which you want unattended, better models are just better.
There are still tasks that even Fable is bad at doing. And many are just mundane things. Because of the fact that you have to steer it on those tasks, you might as well steer an 80% model that is 5x faster. And those do exist.
Naturally there’s a bit of a gap because the faster models need steering on tasks the slower models don’t so there’s no smooth transition but I find it worth it. Especially if you want to stay in flow.
Ironically this sometimes means starting a plan with a great model, planning with a worse model, iterating, then submitting it to a better model for review, and then having the better model do the implementation.
If you're hacking a US-based entity, using a high-performance Chinese model through a VPN is probably safe enough. I doubt a local model is going to be sufficiently smart to hack any major company.
It depends on what you're trying to do. For non-coding tasks luna is quite often enough. Flash models are more than enough for summarizing a text, for example, or whipping up a small script to save me fifteen minutes. If you're on a 200/month plan, I see your point. If you're on a dollar limit - or worse, paying per token out of your pocket - you look to be more efficient.
There's a difference between want and need. I want a 650hp V8 supercar. I need a 150hp I4 toyota corolla. Why choose a less capable car? Because I don't want to spend 10x as much money to get groceries.
Some reasons:
- Smaller models will always be cheaper
- Smaller models will always use less energy, therefore better for the environment
It's a bit like saying you always want the fastest and best car; Sure, you can have it if you keep paying for it. But a small car will also get you from A to B, will use less gas and will be much cheaper.
> But I also think the demand for "fast/cheap/good-enough" models is just about to take off.
There's a sort of "revelation" I had in ~early '24 when I used a 7B local model with a library called Guidance (initially out of MS, then the team moved) to create a flow where the model would receive pseudocode for tests, first write the tests, and once I approved then started writing code until the tests passed. This was before "thinking" models, and yet using that library I was able to "guide" the model in the required "prompt / instruct" context such that it was working towards completion, and I saw the first things like we see now in the thinking traces "oh, test x doesn't pass because blah, I need to..." and so on.
Anyway, the revelation was "even if the models never improve, I'll have years of fun finding out all the ways I can use these things". And, obviously, the models improved a lot since then. But I think that revelation can still be applied, as a sort of "truism". We have, right now, access to things that 10-20 years ago would be considered magic. We are still finding ways of cobbling together systems with glue, duct tape and prayers and find new things they can do.
I think the "good-enough" stage has come not just for API models (cheap, fast, etc) but for local as well. Even if slower, even if clunkier, but they are good enough for a set of ever increasing tasks, and what's more it's incredibly fun to work with them.
Yes.
The infancy phase of this technology is represented by the pursuit of making wildly grand, wildly expensive, all-purpose models that somehow discern a user's full accurate intent from a lazy, underdeveloped, vague idea that they ambiguously and poorly express in a couple dozen words.
The adolescence will arrive as those outsized and ill-considered ambitions collapse and we instead see a cambrian explosion of restrained but efficient model+harness-tuples that have been distilled, finetuned, and rigged to deliver on narrowly scoped but idiosyncratically-shaped tasks with incredible efficiency and erogonomics.
This idea has failed to pan out time and time again - people have an instinct that hand-crafted finely-tuned specialized AI systems must be optimal, but throwing more scale and compute to something more generally smart always wins out. It's especially palpable just looking at the last few years of LLM's: a frontier model with all the world knowledge you can stuff in it and every tool at its disposal has always performed the best at all tasks. Suggesting otherwise has become an extraordinary claim requiring extraordinary evidence.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Recent comment touching on this in relation to LLM's in more depth: https://news.ycombinator.com/item?id=49322695#49323341
The Bitter Lesson is very popular right now. It seems true right now. It’s having its moment right now. That doesn’t actually mean it’s axiomatically true.
Commenter below gets it absolutely correct: stockfish, which runs on your 5 year old phone, is dramatically better at chess than Fable. Like, so much better that it’s not even remotely comparable. The theory of the Bitter Lesson, and it’s only a theory, is that LLMs could eventually outperform stockfish. It’s not true today and it remains to be seen whether it will ever be true. For now, specialized models are absolutely better at specialized tasks.
This idea has not failed to pan out at all. I work for a startup that is exactly what GP described, and am set for life because of how wildly successful it is. Notably, we are successful, in a genuine sense of the word: we bootstrapped from running tiny models to larger and larger models on our own slowly improving fleet of GPUs, and now have millions in revenue without a single dime of outside investment. Conversely, you cannot call taking on ~1 trillion in debt and purchase commitments to scale "success". OpenAI and Anthropic are underwater financially. To be precise, they're in the Mariana Trench.
Wait, you actually found a viable counter to The Bitter Lesson? Please say more
Perhaps an analogy to Moore's law?
Bitter lesson #1: don't waste time optimizing code when a faster processor is around the corner.
What countered it: Moore's law stopped working.
Bitter lesson #2 similarly relies on scaling laws that might have diminishing returns wrt model runtime vs intelligence. Runtime matters for turnaround on the problem you're solving.
Moore's Law has nothing to do with processors getting faster. Dennard scaling stopped working but Moore just slowed somewhat, not stopped.
This is a misunderstanding of either the bitter lesson or what was being claimed, on multiple accounts. Firstly, the bitter lesson is merely about human expertise-tuned algorithms vs. throwing raw compute at a domain. But, notably, it is still domain-specific. No matter how much compute you throw at training an LLM, it is never going to beat a Chess engine at Chess. If you give a Chess engine 1,000,000 compute units and a general-purpose LLM 1,000,000 compute units, the Chess engine is obviously superior at Chess; ergo, there is value in throwing compute units into training models for specific tasks. This is true for within several orders of magnitude of compute, in fact. It's also true that if you give the Chess engine 1000 compute units it'll still beat the all-purpose model with 1,000,000 units, so actually there's a lot of value in training for specific tasks.
Secondly, the bitter lesson is predicated on compute being cheap. There was a period where a hand-tuned algorithm informed by human expertise would outperform a raw alpha-beta search at Chess. Then compute got cheaper, and DeepBlue ascended to the top. Compute is now expensive again relative to the tasks being performed. We are absolutely still in a period where human expertise in training LLMs will outperform a naive approach with more raw compute.
I don't know much about chess engines; do they still use hand-tuned algorithms, or are they more like AlphaZero, where they learn through self-play to beat any/all possible human contenders? I don't believe DeepBlue was automated to that extent, but it may have been.
In the latter case, the chess example would tend to support the Bitter Lesson, rather than refute it.
I would also be VERY slow to claim that general-purpose models will never be competitive at chess. It wasn't so long ago that transformers couldn't add two-digit numbers reliably without resorting to tool use. They are now as good at "mental arithmetic" as any human savant. It wouldn't surprise me at all to see someone come up with a model that just happens to be really, really good at absorbing the portions of its general training data having to do with chess.
In fact you could argue that AGI demands such a model, if we are to assume that LLMs are a guidepost in that direction.
The issue is that GP is misusing the bitter lesson. Yes, search + learn tends to be more effective than human rules based strategies, but that's not what's being considered here. The original claim is effectively that AGI isn't needed for most tasks and more value can be created by using search + learn to solve specific problems instead of applying general models to every problem. Then GP commented a non sequitur
Do you have a website?
Suggesting otherwise has become an extraordinary claim requiring extraordinary evidence.
VibeThinker 3B constitutes extraordinary evidence, IMO. The first such evidence I've seen myself. Very small model, very low literacy, almost no world knowledge, but it is as good at math and logical reasoning as models a hundred times larger.
The Bitter Lesson is a valid and trenchant observation about how about we got here, but I think it's a mistake to assume it tells us very much about where we're going. Too much has changed recently and is still doing so.
Someone will eventually figure out how to package it all into a single, cheap chip
That you can then write text to program and make applications with.
There's A LOT low hanging fruit still out there for sure. And with antigenic systems being able to do the boring repetitive work of looking for that low hanging fruit I think we will see interesting things indeed. Also I think heuristics is where its at for such things. Once you describe some good heutistical structures for the research models to always follow related to "creativity" and such things, thats where we will see biggest difference. The agentic systems know the scientific method well and can follow it they just need the ability to be "creative" so their sampling becomes less rigid.
Yep, I've been having excellent experiences with the models even from the 2023 era. They required a lot of "holding it right" (mostly: being very precise in what went into the context) but their raw coding capabilities were astonishingly good even then.
However, back then I was getting the AI to write individual functions or classes or a test suite. I was decomposing the larger task into smaller tasks, delegating some of them to the AI, reviewing the results and composing the codebase from those. I was also essentially the harness.
Today the models can write and test and deploy an entire project. In terms of the code quality, I actually don't think today's frontier models would have written it much better than the 2023 models did. So in terms of raw coding capabilities i.e. converting a high-level specification into working code, I think we hit the peak way back in 2024 itself.
What has changed is the AI has learned how to do the task I was doing (besides being the "harness"!), which was the mid-to-higher level "engineering" aspects like decomposing a task, specifying it to a reasonable level, reviewing the outputs, and course correcting as needed.
I'm not sure if that is something the AI labs explicitly focused on during training (which may be why Meta is having its highly paid engineers do annotation work), or an emergent property of "better reasoning" (which I believe Dario implied in a podcast), or some mix of both.
But the fact remains that even the weaker models are more capable than we realize, and many being open weights, are here to stay.
To me, most local models work just fine for anything you can be patient for. If I want something quicker, I will go to a SOTA model via API, but with multiple 3090s, I have never really needed a hosted model for a lot of my experiments.
For code, they are great, but for creativity for NPC controllers, they leave something to be desired, but work well enough for testing, so I don't burn tokens until I'm actually playing my games.
But nothing one-shots a prototype better than Fable 5. I can have a prototype built in 30 minutes, hooked up to my local LLMs and Claude Code is very good at testing the interactions and even tuning the prompts of the NPCs for better experiences.
"with multiple 3090s" is quite a bit of burying the lede for "most local models work just fine", don't you think?
Having multiple 6 year old cards doesn't seem like it's that big of burden for local LLMs.
I get that a lot of people don't have them. And a single one can be VERY performant. And the smaller models like a 7B can run on much smaller hardware like a mid-range [3|4|5]060.
My entire AI Dev Box cost $4500 in parts. 128GB RAM, i7-10700, 1TB and 2TB SSD, and 2x 3090s. Today's prices and inflation have definitely made that price tag seem a lot better than it was, but it was an investment in all things GPU that were happening in 2020 (crypto, blender, image gen), then LLMs exploded.
A single, used 3090 costs more than I have ever spent on a computer.
Yes, the tunnel vision around local models on this site is crazy. The percentage of people in the world who can afford the hardware is extremely low.
It seems roughly similar to the pricing level of personal computers in the early eighties (i.e. IBM PC and Apple Macintosh). I’d expect prices to come down significantly over the next few years. Not so much in the next year or two, but after that.
Just like for warships, the complexity and cost of building cutting edge hardware has grown exponentially up to a point where a significant chunk of the world's computing is dependent on 2 companies: ASML, TSMC. We shouldn't extrapolate linearly from examples from the 80s.
It's a decreasing pool as well I'd say, the dev machine I built last summer would make less financial sense to me now for example.
I don't think there is tunnel vision. I'm just saying that I have a couple 3090s I invested in a handful of years ago, and they are still going strong today as multiple GPU-needing technologies emerged.
I'm not saying everyone has to run local LLMs, because the APIs are in a race to the bottom, and my $10 of OpenRouter credits I bought months ago is down to $8.94 because most models give you MILLIONS of tokens for a US Quarter.
I mean, I'm not rushing out to buy that kind of hardware myself, but it is a matter of perspective. People commonly spend an order of magnitude more on a car, and that's just the sticker price.
My single 3080 runs so hot I don't need to warm my room in winter, and have to play games in my underwear in summer.
I got a great deal on ~72 TB of NVMe right before storage prices shot up, doesn't make it any less ridiculous that I have it or any more relevant to people talking about building a NAS now. 99% of people, even in tech, do not have the stupid amounts of hardware people like us hobby on.
Most people in the US have a car, and the average new car is $40,000. Hell where I live a middle class consumer will spend double that on a Boat or an RV and think nothing of it. These aren’t elite tech workers.
It’s not unfathomable that if a personal, generally intelligent local AI provides enough utility and doesn’t require you to tweak CLI flags millions of Americans would want one.
where? i would love that.
> 2x 3090s
You could sell those and have enough money to pay for hosted inference for years.
They cost more to run than hosted anyway. But that isn't the point of having them. They are a playground, a backup when the internet is down, or claude is down. They can render Blender scenes pretty well. They play any game I want.
You can do each of those at various hosts and own nothing. Or own a couple "over priced" cards and do it all at home on battery power for a few hours while the power is out.
For me it’s more that you can show them your financial and medical data without BigCo looking over your shoulder.
But after all those years you’d still have 2 3090s, which are now about 6 years old and still holding value.
I keep seeing this comment. This is _hacker news_ where, back in the day, people just hacked on things, because it was a hobby. They weren't "moneymaxxing" or desperately trying to be as insanely efficient as possible. They hacked on stuff with a can of surge at 3am because it was fun.
Your comment is like a meta comment of "LLMs are generating everything, after a while the ouroboros will eat itself. (Which I agree with)" If people aren't hacking on this shit just because, you have completely conceded control of software to a handful of sociopaths, and open source software is dead.
I have trouble getting simple extraction to work sometimes. I have a block of text describing people and their roles at a company and their ages, and i asked for structured results of an array of these things with the text span that it appears in and all i can say is: nope.
While they are improving rapidly, or as you say even if they don't. The next stage is for hardware companies ( cough Apple cough ) to ship these Local Model ready hardware in their products.
It will be interesting to track the improvements of these 7B model over time.
There will be a turning point in the next few years where it attract enough consumer attention to create yet another Smartphone and PC super cycle.
I see it in a slightly opposite way: even the good models are relatively cheap, and so I worry what we might miss by spending too much time playing with the Sonnets of the world when the Opuses are still objectively a bargain for the power they bring.
> when the Opuses are still objectively a bargain for the power they bring.
The cost isn't just what you're billed. There are security, privacy etc. concerns.
I know companies that are using github, even using public repo, and request their teams to not use SOTA models, but are ok with local models. Just stupid policy.
I think there's something subtle about language and ambiguity that means they aren't designed to become superintelligent autonomous machines. They're value is as information repositories that actual intelligent autonomous machines (us) mine and string together.
Yes LLMs are a beautiful way to compact knowledge. It would be such a cool technology to develop and worked with if it wasn’t linked to such a toxic industry
I think you're just observing ppl in one of these rare instances where enough of them come together because they are motivated. 'Toxic' is the clamoring sound of a crowded room where what gets through to your ears are just the most annoying snippets of incomplete conversations. I dare you to hang out with any actual people here, understand their viewpoint and listen to what they actually have to say in person, within the context of watching them do it.
If someone has an old GPU laying around, say a GTX 1080 with 8 GB of memory, would that be enough to get a (small?) local model running?
Same. Mistral 7b has been more than I ever needed for text for years now.
Unless you must 1-shot with no harness it’s the same amount of power, maybe more because the big “good” models make too many assumptions and tend to become rigid.
Mistral 7b can do anything, and it’s basically instant even on an M3
Sounds interesting. Can you give more details on your workflow and what tasks you use it for?
Code, creative writing, email summaries, automated email replies, and I prefill my invoice notes and daily updates for work.
Actually built a full invoicing product for that, using it too.
I use Mistral 7b and LlamaIndexTS on Node, I run it on a MacBook M3 and on a Linux server with only 8GB VRAM (old gaming PC).
Basically flawless, runs very fast and I don’t even know what paying for “tokens” is :)
What kind of work are you doing? For example, if I have some code in the hot path and I want to do all the usual tricks to help the compiler vectorize it, such a small model is not able to do much.
RAG is your friend (or any vector db). No model can vectorize an entire codebase in context.
Even a big mainstream product (like Gemini) cannot handle more than ~1k lines without missing details and making mistakes. And about every 1k lines, it seems to forget the previous 1k, doesn’t it? So you can never hold more than a file or 2 (or 3) in context at a time without losing details.
What you find is that the big models like Gemini are doing vector storage and retrieval too, and breaking prompts down into chunks for various models to handle to assemble a thorough response.
If you want that kind of control in your outputs, and be able to hold a lot in your inputs, I don’t see any other way regardless of which model you use.
What are some limitations you have found with using a smaller model like that?
BTW structured/constrained generation has so many places to trivially enable jailbreaking/alignment/safety problems that closed source models heavily limit the full expresivity of grammars and capabilities, particular of on-the-fly dynamic grammar construction/reconstruction.
ppl keep talking about the supposed unexplored and untapped "model overhang" but very few things in the world are where you can write elaborate test criteria to before using ai.
A sales person sending a prospect email doesnt have a way to write a test harness for it. Yet these tasks dominate what humans do compared to writing a crud app . otherwise anthropic wouldnt have trillions dollar valuation
I've amassed access to 4 different GPU rigs with 128GB to 72GB; I didn't this before I event touched an agentic engineering harness. It was sometime in February/March when I set them to first tackle small problems, and now with deer-flow, they're scaffolding full project/scope implementation and I'm finishing off the fine details around the problematic edges.
A dream of mine is to be able to host a LLM-powered video game that I can host on a home server running a decent mid-range GPU like the RTX 5060, and the LLM is fast and intelligent enough to make for a fun game experience for a few dozen concurrent players. People can ask for features and they just get made and added to the game on the fly for the lobby to enjoy. The hosting costs would be manageable enough that I don't have to charge anything for the game.
I think with one more year or so of small model progress, that might just be possible to accomplish.
> Across his various startups, Peter has seen two kinds of work:
> 1. the "IQ 180" work. some mad scientist genius type comes up with some crazy solution you've never thought of.
> 2. the "token spewer" work. being ultra responsive, pushing the ball forward across dozens of different fronts.
Interesting comp to pg's Maker's Schedule, Manager's Schedule https://www.paulgraham.com/makersschedule.html
I'm curious about not only which of these roles models will fill, but also how they will empower us to be in the mode we prefer.
It makes sense that we’ll see “room at the bottom” strategies. Currently, large parameter counts seem to be slush funds of world knowledge, language skills (because language’s nuances and open vocabulary make it high-dimensional), and reasoning primitives, the general belief being that the latter takes up the least space in the model.
There are many applications where world knowledge is unnecessary or even a negative, and in which only a small amount of language skill is necessary, and there we can expect small models more intelligently used to beat large ones naively used.
Small amounts of world knowledge seems like it would inherently be tied to more hallucinations.
Perhaps we'll get to a point where believing any un-sourced information from an LLM will feel crazy. I don't want my model to know more than it needs to perform logic and use tools. Once it is capable of using tools I would much rather it looked up information or sourced it from existing context rather than just divine it from it's weights.
Only if we require the knowledge to be built into the weights. Give it access to a search engine and a big library of ebooks, and it might do better.
Probably. You can solve it with either some grounding context, or spending hundreds or thousands a month extra on a model that has more knowledge baked in. With modern harnesses, the choices is obvious.
Everyone wants this to be it but over and over we discover that the bigger a model is the better it is at all tasks, even ones far outside the domain it was optimized for. IE claude fable is better at writing both code and prose than smaller code- and prose-specific models.
The way vision and language models converge into the same geometric space should be extremely alarming for the "you don't need global knowledge for local tasks" type dreams.
And to be clear I'm not saying that smaller models don't or can't work well, or that we shouldn't be heading in this direction. And it's not quite the case that broad knowledge is strictly necessary. But it never seems to be negative! And so far it is the best way we've found to do... everything. Small models are good to the extent they are like big models, not to the extent that they are small.
On narrow domains, it is very common for small models to match or outperform larger ones at a fraction of the parameter count.
For example in language, this is called the “curse of multilinguality”. Small models that handle a single translation direction can easily outperform big ones that try to handle them all.
https://arxiv.org/pdf/2311.09205
In any case, for most tasks the question is not “how many tasks can this model kind of do well” but “given time/cost constraints, what is the maximum level of quality we can achieve”. And for that, small models are usually very competitive.
I think the context here is that small models run locally, not rented from a cloud.
Yes small models are and will be useful for lots of stuff for several reasons.
But the idea they’d be better than a bigger model is cope, you’re pretty much always better off running the biggest one you can bring to bear within your constraints.
yeah, I think they will get smaller so they can be run everywhere, and really just be an interface to various non AI systems.
I find it quite funny all these folks who are addicted to chasing frontier models, only just noticing that small models became "good enough" for most tasks. Those of us without fable-sized expense accounts noticed this quite a while back
Exactly! Composer 2/2.5 were amazing, cheap, and fast. Everyone else was Gaga about GPT 5.5 and such, while we were over here doing the work with less cost and more speed
I’ve been playing around with Luna, Terra and Sol and for the type of work I’ve been doing lately I actually think Sol is just a likely to trip up as Luna. Examples were Sol over assuming, persisting in the wrong direction, over engineering a little script to do some exploration of api. They can all be fixed but it’s a waste of tokens, I rather have Luna do it because course correction on small pieces of work is cheaper.
I've found the distinction to be in how much I care about how the final product looks. If I want high-quality code I typically find a smaller model with a well-designed spec to do better, if I want it to just run and produce something close to my vague description typically Sol does better. For most actual business use-cases I think the first is likely better but the experimentation speed up with the frontier is very nice.
> for most tasks
The word “most” is doing a lot of work here. On a percentage basis perhaps most tasks a typical SWE needs to do when they aren’t in meetings or writing docs are just glorified autocomplete. But that’s boring and that’s why people don’t usually talk about it.
People are addicted to chasing frontier models because they all have memories of spending a week on a deeply challenging algorithm problem or even have crazy complicated algorithms they cannot implement themselves and want to have the models achieve this technical breakthrough. It’s the kind of productivity boost from spending one week on a problem to spending one hour. In contrast the productivity boost from spending ten minutes to spending one minute just doesn’t occupy people’s mind.
> crazy complicated algorithms they cannot implement themselves
I'm not sure I know very many engineers who would fall in this bucket. Or do you mean the business types who suddenly think AI can replace all the engineers?
It probably depends on the background and the company. For example if one works at a startup that happens to use technology, it’s unlikely to happen because SWEs just translate business rules to code. But if one works at the place where the technology itself is the focus, then yes most people will fall in that bucket.
In fact I noticed that this is the one place where people discussing AI on HN tend to talk past each other. On the one hand people are talking about supreme intelligence like designing new algorithms (on the same vein as finding counter examples for the Jacobian conjecture) and on the other hand people are just satisfied using AI to automate a few quotidian tasks that hitherto couldn’t be automated.
Exactly. Even 32b parameter models you can run locally on consumer hardware are "good enough" at this point for some workflows!
no they are not good enough for "most" tasks
> There's obviously a lot we can optimize here, but if you're charging what the WSJ or The Economist charges, you'd better be delivering similar value.
Gosh, watching paint dry has been a better value than reading The Economist in the last 5 years or so.
That aside, I had good results with Luna. I'd be interested in hearing about a comparison that takes into consideration response time (not TPS), cost and performance of the popular models at different settings. That chart has some of that. For instance, is Luna Max a better value than Terra Medium?
"Small models" nowadays work like someone who has IQ 100+ while SOTA ones are like 150, "relatively".
Given sheer number of turns I can make with small models, I can do a lotta stufff
- cheaper, and faster
Harness makes differences: There have been many HN posts about how one made tiny models work better at certain tasks using harnesses.
These "small" models with right context, and guidance, they work wonders.
---
I've been saying Luna has been my go-to AI in previous comments and why Luna is still more compelling than GLM-5.3-flash.
- https://news.ycombinator.com/item?id=49450353#49452248
I’m kind of cautiously excited for the next five to ten years, with these AI chips becoming incredibly fast and RAM capacities ramping up its in the cards that we’ll have chips like today’s ATMEL microprocessors that fit on a single board computer and can run small models locally, then all our gizmos can have local AI and I can have a truly intelligent home. Of course there will be a huge push to put all of it in the cloud but maybe we have a chance to take this technology home for good as it’s hard to imagine people will submit to this kind of surveillance required for AI home automation 24/7 (then again I might be wrong). Exciting times.
> the "token spewer" work. being ultra responsive, pushing the ball forward across dozens of different fronts... ~95% of the work he does falls into bucket 2. It's hopping on calls. Nudging people. Blocking and tackling.
This is a good insight broadly!
I recently had some relevant experience: for a couple of months now I've been experimenting with on device models to summarize feeds in a Bluesky client I am developing. The feature extracts topic areas, categorizes posts, and creates a summary under each topic.
At first the results were hot garbage, and progress was slow. I hooked up the settings to download models from Hugging Face conveniently, so I could run experiments faster, and I massaged the prompts a bit. Last week this feature made a qualitative jump from science experiment to something I'd actually use.
The fact that all runs on the device means I've got no variable costs associated with adding this to what will be, at best, a pretty low revenue product. I've tested it on trailing edge devices like an M1 Mac and a Pixel 8, and performance is very tolerable.
The key is I'm not asking for open ended answers to open ended problems. When it proves to be useful it's not going to get less useful or more expensive.
There are vast domains of uses for LLM models with similar characteristics and likely similar results.
Running small models locally beats wrestling with API latencies and rate limits. The compute trade-off is 100% worth the privacy and DX gains.
The tokens per second speed measurement is highly inflated nowadays because most of the tokens went into thinking. I wonder if there is a more realistic measurement for "effective speed", which accounts for thinking efficiency.
Many benchmarks now measure the total cost or energy usage per completed task.
IMO big models are not a product in and of themselves. Inference is just a new type of compute. I'm confident that in two or three years, every product will have inference capabilities integrated into the experience, and models will become less and less distinctive from one another.
What most products need from a model is a pretty short list: the ability to make tool calls well, accurate recall, and the ability to follow directions without wavering (whether or not those directions are baked into the weights or provided in a system prompt). That covers 95% of inference utility in products.
We're nearly there, and I believe these capabilities will fit on small models.
Because of this though, I predict hardware demand will stay high despite demand for "hosted" inference dropping. Unless there's some regulatory shenanigans that step in to say otherwise.
100% agreed. Small, cheap, and hosted models. Luna (and open weight models and others) is ridiculously cheap @ $0.2/$1.2, easily accessible, and more than good enough for basic use cases (e.g. summarization, simple tool calling, etc.).
I wouldn't call $0.2/$1.2 "ridiculously cheap"
Small is relative. I'm looking for models that I can with run around 100 MiB mark (RAM just for the weights) to demo what you can do with this little memory.
I know of SmolLM 2 which in Q4 is borderline regarding the size and rather dated. There is also TinyStories, which is also old and also focussed on children's stories.
Is there anything newer in this category? Or should I try to distill something down to this size?
Maybe I'm being super reductive here, but operating small models at the core of your business kind of moves the needle from making external API calls (against frontier models) to running internal API calls (against your locally-run models). It seems like if we want local models to take off, it will need to become easier to run local models for cheap. I'm thinking like reducing the barrier of entry for running "local models" in the cloud providers like DigitalOcean, AWS, etc.
You should be glad to know digital ocean already offers this
I only run local models and I don't give them access to much externally. I don't do anything serious with it, but it comes in handy and I know that they can do so much more. I'm on a meager RTX 3060 12GB and a GTX 1660 Ti with 6GB for some extra vram space. When I first started playing with local models, I was really impressed with what I was able to achieve locally.
That's great, but the thing that worries me is that many companies have billions invested in the AI bubble. It's around 1.5 trillion last time I looked. It's all circular spending between the companies building out the infrastructure, and the models. None of it is profitable. They will want to recoup that 1.5 trillion from consumers, which means using online-only pay-as-you-go cloud models. They will inevitably see that people using capable local AI are "lost customers" and they will try to kill the ability to locally host AI or somehow enshitify it enough to make paying a subscription more palatable.
I'm not saying I believe that will happen, I'm just worried that it will. Is anyone else worried about that as well?
I am very excited that more makers will come up with fast memory for consumers rather than enterprise. Companies can only pre order so much RAM.
At some point there will be a surplus of fast memory and even in a crash the current generation of SLMs are bounced to be plenty to build a lot of intelligence at home.
I love how "Small Models" apparently is "Model of unknown size but probably smaller than another model that we also don't know the size of".
The demand for fast, cheap, good enough models has always been borderline infinite, it’s the supply that’s going to take off.
I forked my Big Serious Harness™ that models construction projects into a harness for building a vibe coded family assistant. I couldn't figure out how to make the toy operate at toy prices until Luna. Now you can vibe code all the little apps you might want for your fam for like $5 and operate it day to day for a few cents.
> Peter runs multiple companies. Beyond Segment, he's raised $100m+ for Charm Industrial, and just recently closed a Series A for Revoy. He's incredibly organized and efficient with his time.
You can do this before an exit? Build and fundraise for multiple (3?) companies at the same time?
Segment had a $3B+ exit to Twilio back in 2020.
An absolutely nothing post at #2 on the frontpage
I think we'll see more of this soon
replit is already leading the way with free luna usage
small models + a good application layer - are more than enough, good for routine business tasks.
the application Layer i.e having a good graph RAG & connecting it up together is the missing piece for most.
Can you elaborate on this?
The model doesn't have to be smart if all it's doing is pushing a few different buttons.
I don't have to be an automotive engineer to start my car and put it in drive.
I have trouble seeing the points of using less capable models.
I just want the smartest, best, and most capable models. It feels smaller models for speed and cost are just transitions towards better hardware allowing the very best model.
And that is why i always carry my groceries with an Antonov An-225 Mriya. Is it really needed? No, but i refuse to compromise on what is(was/will be) the best.
My experience has been that responsiveness is value. For tasks where you need steering, responsiveness allows for better steering. For tasks which you want unattended, better models are just better.
There are still tasks that even Fable is bad at doing. And many are just mundane things. Because of the fact that you have to steer it on those tasks, you might as well steer an 80% model that is 5x faster. And those do exist.
Naturally there’s a bit of a gap because the faster models need steering on tasks the slower models don’t so there’s no smooth transition but I find it worth it. Especially if you want to stay in flow.
Ironically this sometimes means starting a plan with a great model, planning with a worse model, iterating, then submitting it to a better model for review, and then having the better model do the implementation.
First, smaller models are fun for hackers: you can run them locally, or run them faster.
Second, when cloud models become unavailable or otherwise deteriorate, these will be all you have. May as well prepare.
If you're hacking a US-based entity, using a high-performance Chinese model through a VPN is probably safe enough. I doubt a local model is going to be sufficiently smart to hack any major company.
You misunderstood what I was referring to by “hacker” there.
It depends on what you're trying to do. For non-coding tasks luna is quite often enough. Flash models are more than enough for summarizing a text, for example, or whipping up a small script to save me fifteen minutes. If you're on a 200/month plan, I see your point. If you're on a dollar limit - or worse, paying per token out of your pocket - you look to be more efficient.
Can you define your use of the word 'smartest' here just in case some of us don't quite know what you mean?
There's a difference between want and need. I want a 650hp V8 supercar. I need a 150hp I4 toyota corolla. Why choose a less capable car? Because I don't want to spend 10x as much money to get groceries.
Some reasons: - Smaller models will always be cheaper - Smaller models will always use less energy, therefore better for the environment
It's a bit like saying you always want the fastest and best car; Sure, you can have it if you keep paying for it. But a small car will also get you from A to B, will use less gas and will be much cheaper.
Cost is the point
I like the analogy on ways to make small model useful.
Well said