> 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 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 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
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.
I have an agentic workflow and Luna just always gets stuck, SOL and grok 4.6 don’t. I like Luna in theory I just find not much practical work for it yet in coding type work.
Now I think Luna is plenty good for many applications inside a very good harness/scaffold. And I think there are a lot of those usecases. So I think these small models are really good for application developers.
But for entrepreneurial knowledge work all of my work still benefits a lot from more intelligence.
The word “most” is doing a lot of work here. On a percentage basis perhaps most tasks a typical SWE needs to do 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 have crazy complicated algorithms they cannot implement themselves and want to have the models achieve this technical breakthrough.
> 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.
Sol is 100 EUR/m and I prefer it over Fable, and you get practically unlimited usage unless you go really overboard with it. So I don't get the cost argument. Would be better to own hardware, but with the shortages that's not very attractive atm.
I received access to AI in May of this year and used smaller models a lot in May, June, and July. Haiku 4.5 is terrible for coding. This month Luna was added and it's incredibly capable. I have to reach for a better model 1/50 times instead of 1/2 or 1/3 with Haiku. Luna is better and cheaper than the Haiku and Sonnet models I've been spending most of my time with.
It's safe to say I didn't come to this conclusion because I was "addicted to chasing frontier models"
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.
> 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.
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
you must have extremely low standards. I guarantee everyone who receives an automated email reply from you feels anger and resentment towards you, exactly in reflection for the disrespect and contempt you must hold for them.
WHY IS YOUR TIME MORE VALUABLE THAN THEIRS? WHY SHOULD THEY READ ANYTHING YOU HAVEN'T TAKEN THE CARE TO WRITE? FUCK YOU
I would spit on you if I could. Disgusting behavior.
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.
Yes - you are experiencing a mix of context caching and db retrieval from these mainstream model experiences.
Even the best models available lose a ton of detail over time if you were to paste in tens of thousands of lines of code.
The only way to hold huge amounts of context with a high degree of accuracy is to store it using various mechanisms (one of which is RAG).
On “effectiveness”, I mean end use case effectiveness in the tasks at hand, not whatever benchmark the model developer or vendor themselves come up with - which may or may not be useful to the work I’m doing.
As a developer, none related to the model. It’s faster, lighter, makes less assumptions.
For a casual consumer, it is no doubt a worse chat buddy. Knows less historical facts à la Wikipedia.
But I’m not really using LLMs for that kind of entertainment and I don’t rely on them for fact-checking anyway. It’s almost worse to rely on a smoother talker for something it can’t possibly know.
For feeding in thousands of daily updates and getting it to predict the next one based on a crude list of tasks, it’s great. For completing code files in my style, also great. It can also handle most small customer service issues and refunds related to my apps on App Store, the back and forth to collect data from users.
For a booking engine I run with a partner, it is great at confirming bookings and following up.
For invoices, I prefill notes and things based on previous ones. It’s fine. I don’t need a massive model that takes 30s to reply etc. it’s way overkill. Maybe worse than overkill - off in a direction I don’t need.
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.
Not really. 2 years ago that was a pretty normal amount of GPU hardware for a hacker or gamer. It's all relative. They are not accessible to most people yet, but for someone that cares and is a technologist? Likely accessible.
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.
I've done pretty decent local prose->json extraction using Qwen and Phi and Gemma.
I'm sure most of it comes down to prompts, and all of them run over 100tps on a 3090. Smaller cards will likely be slower, but Qwen3.5 9B is small enough to fit on most consumer cards.
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 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.
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.
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.
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.
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.
You are technically correct. The best kind of correct.
However, what most people think of as Moore's Law--CPU speed doubles every 18 months--broke somewhere between 90nm and 22nm.
And even the actual Moore's Law--2x the transistors every 18 months--doesn't hold for all types of chips anymore. Memory only gained 2x density over 10 years.
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
I think the gp found a viable counter to the ggp's version of the bitter lesson, which seems so extreme as to certainly exceptions. IE, they seem to say nothing but the latest frontier model is ultimately viable as an AI business.
The actual argument of the Bitter Lesson essay is pretty limited but people's interpretation of it has gradually drifted until it's seen as prediction that current LLM will reach AGI at a large enough scale.
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.
I wouldn't use it for anything important without heavy supervision, as it's very weak outside its specialty. Not ideal for instruction-following tasks.
So theoretically, if you give that model the means to find information, ascertain the quality of said information, it could still reason its way to an proper answer?
Is this whole thing than maybe a read vs write optimisation again? Spent more time and effort training more knowledge into the model upfront and get it out in a single question instead of training a small model and needing more steps to answer the same question?
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.
Specialized models are better at specialized tasks.
This seems really backwards. The Bitter Lesson is all about large data-based approaches vs hand-crafted ones, it doesn't say anything about language models not trained specifically for chess.
I can't find the comment you're referring to, but the latest versions of stockfish are based on neural networks trained on millions of games, so if anything the Bitter Lesson turned out true here.
The conclusion of the bitter lesson would be that a large language model trained on chess commentary as well as being trained on millions of chess games would outperform stockfish which is only trained on millions of chess games. There’s no evidence at this point that this is true.
The common mistake is to think “maybe if we use a blend of raw data and hand-crafted heuristics, we’ll get the best of both worlds!” But the bitter lesson says no, beyond a certain point it’s better just to use the data.
Thinking that an LLM might be able to improve on purely “big data” machine learning seems to me to be the same incorrect idea. Its “intelligence” is no more useful than human intelligence. The LLM is based on a massive data corpus, sure, but the amount of data specifically about chess in there pales in comparison to just playing billions of games of chess.
Also, training it on chess books is literally training it on human knowledge, and not the actual game, which is exactly what the bitter lesson says not to do.
The conclusion of the bitter lesson would be that a large language model trained on chess commentary as well as being trained on millions of chess games would outperform stockfish which is only trained on millions of chess games.
If you can draw any lessons from chess commentary, I think it’s very reasonable to call it “hand-crafted heuristics.”
Hand-crafted even if you're feeding in the raw commentary? That seems like a weird way to consider it. Wouldn't that make LLMs in general "hand-crafted"?
And raw games plus raw commentary is all the data you have. You can make more games but those can be fed to both stockfish and the LLM competitor. So it seems like a valid interpretation of the bitter lesson to me.
Yeah, "hand-crafted" is a bit of a stretch; I mean their value is in the human insight they contain.
The key point I was trying to get at is that the human insights don't contain anything that can't be mined from vast amounts of gameplay. Every human insight can eventually be rediscovered and made rigorous by data (in chess, at least!) In the short term, those insights are useful, but in the longer term, they add nothing at all.
Note also that "raw gameplay" here can mean new games -- you can generate as much data as you need, you don't need to rely on real recorded games.
> Every human insight can eventually be rediscovered and made rigorous by data (in chess, at least!) In the short term, those insights are useful, but in the longer term, they add nothing at all.
But isn't that the bulk of what we're shoving into LLMs, and it makes them much smarter? If it's useful there but not in a chess AI then that seems like a significant crack in the bitter lesson.
> Note also that "raw gameplay" here can mean new games -- you can generate as much data as you need, you don't need to rely on real recorded games.
Yeah I mentioned that, generated games are useful. But if we're being fair and letting both AIs use generated game data, does the more general LLM ever actually overtake the specialized stockfish like the bitter lesson suggests?
Another way to look at this is that giving the LLM the commentary is a way to avoid complaints of hiding data from the LLM, since it'll have strictly more info than stockfish. But if we cut that from the training data and only give it a basic description of chess and lots of raw game data then it's going to get even worse than it already is at chess. Meanwhile stockfish never had that commentary, just actual hand-crafted heuristics and training on game data, and it's very strong.
Or maybe an LLM could just tool call stockfish and doesn’t need to have more than a basic understanding of chess. The bitter lesson seems extraordinarily wasteful on the compute side.
> a large language model trained on chess commentary as well as being trained on millions of chess games would outperform stockfish which is only trained on millions of chess games
Not really, if anything it's closer to the opposite. The Bitter Lesson essay literally has this as an example:
> These researchers wanted methods based on human input to win and were disappointed when they did not.[1]
and
> Enormous initial efforts went into avoiding search by taking advantage of human knowledge, or of the special features of the game, but all those efforts proved irrelevant, or worse, once search was applied effectively at scale[1]
The actual bitter lesson is this:
> breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.[1]
Applying to the "LLMs-for-chess" example the bitter lesson approach would be to put many, many more games into the LLM.
> These researchers wanted methods based on human input to win and were disappointed when they did not.[1]
This was/is basically a strawman though. Like maybe "human input winning" was desirable for chess masters but for computer science wonks? Not the point or the disappoint. It's always neats and scruffies fighting about using some kind of recognizable method (logic) instead of magic (ML).
> breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.
More to OP's point I think: nowadays when someone wants to beat you over the head with the bitter lesson, they aren't as careful to include learning and search. They want to say learning leads to intuition (magic) whereby we can avoid work (logic/search), and maybe argue or assume from there that neats and scruffies is settled. TBF, something like reasoning in latent space does resemble intuition!
But the real lesson is confirmed every time we bother to check, and not very bitter for anyone. Search/learning/logic are ALL always necessary on any sufficiently difficult problems, and hybrids that interleave always outperform everything else. Stockfish being the example in this thread that different camps of absolutists would like to claim, but also all the MCTS examples, evolving examples, and new hybrids all the time. My favorite lately: https://arxiv.org/pdf/2511.08983
> This was/is basically a strawman though. Like maybe "human input winning" was desirable for chess masters but for computer science wonks?
Oh no!
The whole field was full of people whose entire career was built around the idea of developing smart priors.
To quote Wikipedia:
> For computer vision in particular, much progress came from manual feature engineering, such as SIFT features, SURF features, HoG features, bags of visual words, etc. It was a minority position in computer vision that features can be learned directly from data
This undersells the change though! David Lowe's reputation as the best image researcher in the world was based on his SIFT patent[1]
This approach worked until 30 September 2012.
That was a bitter day for many, many computer science researchers.
Is being wrong/ignorant about whether/how something can be automated the same as having a preference for doing it manually? Maybe so if it's your patent, your thesis I guess..
But as it relates to more/less magic, maybe the more modern lens on this is e.g. https://arxiv.org/html/2505.11581v1 . Is manual feature-engineering more like what you'd evolve, or more like what you'd get from SGD ? Feasibility and performance is always a question, there are others like what is robust, stable, adaptable, predictable, explainable. Maybe the manual-features people were interested in something besides the manual part? Maybe the story isn't so simple, and maybe it's not finished yet.
There's two different goals to AI research - one was to get results - a chess engine thst wins, etc. But the other goal (which seems to have been abandoned in the deep learning era) was to use AI to help understand how human minds work. A chess engine modeled after human grandmasters is much more interesting in that regard than either a min-max algorithm like beat Kasparov or modern deep learning engines.
The bitter lesson is that simply scaling training on more games—including self-play—trumps any hand-crafted human input, whether that's fine-tuning on human commentary or clever engineering tricks.
Current models are just high-dimensional interpolation engines. The denser the data sampling, the more accurate the interpolation gets. Given a choice between denser sampling and anything else, denser sampling always wins. That is the bitter lesson.
Denser sampling only seems useful if the problem domain is in some way smooth - interpolatable. If you run it on a fractal problem domain you just learn more special cases. Chess is fractal.
Chess is a brute force search problem. Humans are not good at chess, even a small computer can beat Magnus Carlsen.
It would be better to compare models at how well they can write the code for chess engines, otherwise it's just saying that Fable is not a good CPU emulator, which is obvious.
The Bitter Lesson says that the only things that scale are search and learning.
Stockfish is the best chess search engine we've got, and you can learn some good heuristics for chess search policy that will make time-limited chess search a lot more powerful. That's perfectly in line with the Bitter Lesson.
In contrast, LLMs playing chess are relying solely on learned behavior. The inference harnesses surrounding them aren't designed to do chess things, they're designed to do autoregressive token decoding, which isn't a search process. Reasoning traces can resemble a search process, but they're far less efficient - the LLM would have to work out each legal move, test each one, calculate a score, and simulate minimax over all of that. Assuming the LLM is smart enough to even do all that.
A hand-crafted approach can absolutely beat data if your approach unlocks more search and/or learning than the general solution.
> A hand-crafted approach can absolutely beat data if your approach unlocks more search and/or learning than the general solution.
Now let's look at the bitter lesson again. It says that general methods that leverage computation are ultimately the most effective, and by a large margin.
That's different from just saying to leverage computation (which is how I would interpret "unlocks more search/learning"). If the lesson is "more computation wins, when sufficiently channeled" you're basically looking at a truism. Of course more computation beats less when it's used right. The bitter lesson is about abandoning specialization in order to get more computation, and while there's a couple ways where that helps with chess, there's a lot more ways where it's counterproductive. It looks like it's more true for Go than it is for chess, and that it's not universally true. It probably correlates with the state space.
Both you and the parent commenter seem to be misunderstanding the point the Bitter Lesson paper makes.
The Bitter Lesson is about general-purpose algorithms vs. specialized algorithms. Historically, chess engines were programmed to look at a chess position and use positional understanding (imparted by the human programmers) to decide what the best move is. But eventually, the chess engines that actually became stronger than humans were instead programmed to just check every possible move and countermove and see which ones lead to a win. (I'm oversimplifying, but you get the point.) So even before Stockfish contained a neural network, it was considered an example of the success of the Bitter Lesson.
As it applies to AI agents, the Bitter Lesson would predict that the best possible agent would simply possess A) a way to do anything it wants, B) a way to evaluate whether what it did was correct, and C) a ton of compute. Then just turn it loose on your task. (The fact that the "brain" of the agent is an LLM is kind of irrelevant - you could also imagine the brain just being a program that generates random syntactically-correct code. What the LLM achieves is that, the random generator would take millions of years whereas the LLM is much more efficient at creating plausibly-working code. This is analogous to a chess engine's pruning heuristics.)
The hard part here is B. We've seen some great agentic successes when rewriting an existing project in a new language, since the agent can just use the project's prior test suite as its evaluator. But when developing a new project, you're still figuring out the finer details of how everything is supposed to work. As the old saying goes - writing a spec that perfectly describes how a program should work, is equivalent effort to just writing the program.
Dumb AIs are needed for customer service. Most of that industry is still at "press 1 for sales, 2 for billing..." and needs something that will run locally on a 1U server.
Yes, and the technology to improve the interface you described is already available to run hundreds of concurrent instances on a 1U server. The barrier to entry is getting the people who manage those systems to care enough to implement something better.
Fact. My company's largest partner is CoreWeave, and convincing leadership that we could run it ourselves on partner discounted hardware for a lot less money has gone nowhere.
> Most of that industry is still at "press 1 for sales, 2 for billing..." and needs something that will run locally on a 1U server.
Needs? Customers want something that immediately answers their question/solves their problem, but that’s far away, even ignoring the “run locally on a 1U server” and that that may not be in the company’s interest.
For many companies, that support line is a cost center, not a PR mechanism.
Also “Press 1 for sales, 2 for billing...” has the big advantage that it handles all accents, speech impediments, etc.
Long term I think a solution where a user’s agent trained on their voice, running on their phone communicates with the support agent of a company is where we will end up, and support phone lines will cease to exist.
But isn't that really just about giving "front end" models more access to specialized tool libraries, which include models tuned to specific tasks? Like the first model says ah, we're being asked to code something, oh and we've been provided with some example code, let me invoke a tool call to my model the recognizes many languages, that model says that we're looking at ocaml. Okay, I better pass this off to my ocaml model which will decipher the supplied code and make a plan for what we do about the user's intent. The ocaml model recognizes that there are tests in the supplied code, let's have the special testing model have a look at the testing strategy and see how that fits in with what we just implemented, etc etc.
And perhaps at the end it all gets a single pass by a god-tier model for overall sanity and congruence, but the actual work, planning, coordination, and even user interaction was done by cheaper and faster agents of much more limited capability.
It’s kinda funny that your last paragraph is basically describing why sparse files, sparse matrices, etc. are used in other contexts.
It really is absurd to ask programming questions to a model also trained about the lifecycle of a fruit fly.
Instead of building small models from scratch, we train an enormous model and use ridiculous amounts of GPU memory. In the end, the whole thing is shoved into RAM because we don’t know where the useful parts are…
We certainly would know where they were if they were just in smaller models in the first place!
Yes, but an LLM will just call stockfish if it needs to play chess … sure if you arbitrarily constrain an LLM to use no tools it’ll suck at chess. But no one is using LLMs in isolation. Even consumer-grade, bone-stock ChatGPT has tools.
You're making a conceptual mistake here, comparing a chess tool to its operator. Deterministic tools produce superior results compared to models in many areas, so we allow models to use tooling.
The correct analogy here is Fable as a second tier player assisting a SuperGM in running stockfish, then assessing its output to identify promising variations.
There might be a limit somewhere that prevents the bitter lesson being axiomatic - for instance where simulations for anything can be exhaustive - so that judgement isn't needed any more as an arbiter. But while there are problems sufficiently complex or large to require a breadth models don't currently have, greater scale and compute will continue to convert to better decision making, and the bitter lesson will remain true (true enough).
No. The bitter lesson is about capabilities. GP is talking about efficiency.
GP isn’t suggesting that focused narrow model(s) will be more capable than large model, but that many small focused models can have sufficient capability while being more optimal.
Also, the bitter lesson is just wrong. The bitter lesson is about hand tuned AI vs computational general methods. However in truth today’s AI uses both. We have general compute heavy models which require narrow expert instructions (eg tools internet docs).
LLMs would not be as good without expertly written context, and expert context without LLMs aren’t as good either.
> The bitter lesson is about hand tuned AI vs computational general methods. However in truth today’s AI uses both. We have general compute heavy models which require narrow expert instructions
The models are not even really trained bitter lesson-style anymore. That concept peaked during the era of pre-train scaling, back when it was thought that making a bigger and bigger GPT-3 would automatically solve all problems through prompting. In 2026, the most important part of training is post-training, which uses vast quantities of niche, hand-curated data to fit the models for specific tasks in domains like tax law.
If you look at value as purely the LLM output, then there's a valid argument that the best frontier models will always be better than fine tuned specialists. (I'm not convinced personally, but it's a defensible claim)
But that misses two dimensions:
1. The cost of acquiring that output
2. What is actually "good enough" for that specialist domain
Not every output needs to be the best to produce value.
And as specialist models increase in cost, their cost/value proposition goes down.
At some point, there's a threshold where cheaper, fine tuned models are "good enough" at the task and also substantially cheaper than the expert models.
That's where fine tuning helps.
Personally, I became a believer in fine tuning after fine tuning a 1B Qwen model as a second pass over my local voice transcription app, achieving excellent accuracy at ~zero token cost and waaaay lower latency than if I'd invoked my Claude subscription under the hood.
At some point, the idea of cost/benefit analysis in the software business turned into the benefit analysis. The amount of money going into the frontier LLM model game is fantastically ridiculous. Being much better than the free resources doesn’t even touch how much better they will have to be to justify the expense of creating them, let alone continually maintaining these services.
> 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.
Absolutely false. At least when it comes to multimodal inputs, even a simple classifier will outperform the largest LLMs who still hallucinate details or don’t describe audio and images accurately.
And there’s also the issue of cost/inference speed. Running a trillion parameter model for all tasks will be incredibly costly, require a cloud API, while a tiny CNN can be run locally or at a cost multiple orders of magnitude lower.
It’s not hard to fine tune a 4B parameter model to outperform GPT-5.x or Claude on a domain-specific task. You don’t even need GRPO or whatever the flavor of the month RL technique is either. I suspect what you’re saying is that massive frontier models are the best for doing big slow tasks on human time horizons—and I emphatically agree—but keep in mind LLMs are useful for quite a bit more than just multi-turn chatbots; and in those instances, throwing infinite compute and latency budgets at a problem is generally not a great way to find the “best” solution :)
I think large-scale systems are optimal when the number of tasks is relatively bounded and costs don't matter.
When you're doing something "at scale", you want smaller models, weaker heuristics and rules written in code, as they're much cheaper, much faster, much more predictable and much easier to reason about.
If your task is "tell me which pages on the internet are most relevant to this search query", running Claude Fable to evaluate every single page would be ridiculous. You may run something like it at the end, to synthesize information from a small result set for example, as that's a "human-sized" task.
I think if you start taking cost into account for tasks which models have saturated a long time ago, the bitter lesson stops being applicable.
If your task is "write a function to sort a list of 1mn entries alphabetically", a small coding model will exactly match the performance of a frontier model (it'll 1-shot it, with 100% accuracy). So the only differentiator is cost, and the frontier model will cost orders of magnitude more than the specialised model.
For most intelligence based tasks, you don't (and never have) needed the tool which "performs best at all tasks". You need the cheapest one which performs adequately for your immediate task.
This doesn't mean the bitter lesson is incorrect. At the frontier, it's still correct. It means that it's not applicable at all to lots of tasks.
At work i only had early copilot which was hysterically bad at everything. As i wanted it to do the same task repeatedly and could spot wrong results instantly i kept evolving a prompt that attempted to correct all ways it found to do it wrong. It kept inventing new ways to get it wrong until it eventually got it right 90% of the time. My theory is that an avanced model that has no issues with a task could do the prompt enginering much better than i ever could. You could for example run x different queries that all do the same thing and compare the results y times. If there are >1 correct results and the wrong versions are all unique you should be able to drill down to a valid result with even a truly shit model running on a potato. Basically what humans do.
Yes! As long as you have some criteria to judge the final answer, you can do a kind of "prompt-side RLVR", where you have the model generate prompt changes, try a bunch of different prompts and see which ones improve the results.
You don't necessarily need a bigger model to do this.
Trying until you find something that doesn’t fail is the model used throughout evolution & engineering, think of bridges collapsed, airplanes falling from the sky or exploded steam locomotives.
In engineering you learn from these mistakes and try never making them again. Do we want to go through this evolution every time we solve a software issue? Just because we can with an unlimited number of cheap tokens? I think not, I’d rather use the knowledge build up that also knows about the edge cases forgotten to test. Or better, use multiple models that evaluate each other, as Entropic describes it in their recent report https://news.ycombinator.com/item?id=49316271
> We expect that agents coordinating in the wild will act in higher variance ways than we see here, because they’ll have different backgrounds and therefore different contexts. They also, presumably, won’t all be Claudes.
Early Copilot was tab-complete in editors and was honestly the best version of LLM-assisted development I've used, because it was intentionally small in scope, trivial to verify output from at a glance and easy to opt in and opt out of. I have yet to see anything more useful in terms of code generation; very small-scale code generation (think function-level) comes close, but is a lot more tedious.
For analysis, bug hunting, overview and some porting work to popular languages and so on I think the current SotA is fantastic, but they're still very disappointing for code generation past function or small module level.
> "I have yet to see anything more useful in terms of code generation"
Recently my employer hooked Claude into a bunch of SaaS services like a ticketing system and an asset inventory system, and I feel a big productivity boost is that I don't have to dig for their open browser tabs, remember their product name/URL/where I put the bookmark, find that my sessions have logged me out "for my protection" since I last looked, find my MFA code, approve sign-in on my phone, then use their mismashed/sluggish/poor-UX interfaces and slow searches to loko for things, and can instead ask Claude "search these systems for anything about <issue>" and it just does. It correlates between systems, summarizes things, and gives me references which system and where to look to check what it found. Internal search on a mess of acquired and disjoint systems.
This feeling reminds me of two classic Joel Spolsky blog posts below, and your comment here about the surprisingly useful tab-complete gives me a similar feeling. I wonder how many smaller candidates there are where LLMs are more helpful and less hyped, against the big vibe-coding solve-everything hype which are currently less helpful than their hype?
References [1]: "A lot of us thought in the 1990s that the big battle would be between procedural and object oriented programming, and we thought that object oriented programming would provide a big boost in programmer productivity. I thought that, too. Some people still think that. It turns out we were wrong. Object oriented programming is handy dandy, but it’s not really the productivity booster that was promised. The real significant productivity advance we’ve had in programming has been from languages which manage memory for you automatically. [...] Whenever you hear someone bragging about how productive their language is, they’re probably getting most of that productivity from the automated memory management, even if they misattribute it. Sidebar: Why does automatic memory management make you so much more productive? 1) Because you can write f(g(x)) without worrying about how to free the return value from g, which means you can use functions which return interesting complex data types and functions which transform interesting complex data types, in turn allowing you to work at a higher level of abstraction"
and [2]: "I have seen many language and programming fads come and go. But there’s only ONE, that’s right, ONE language feature I’ve ever seen that actually improves your productivity significantly. No, it’s not object oriented programming; no, it’s not intentional programming or assertions or programming by example or CASE or UML or XML or Java. The only thing that improves your programming productivity is using managed code – that is, using a language in which memory management is automatic."
we could have had that before, but in past era of tech companies, APIs were not designed with an individual needs in mind and management was all too happy to order you to update status (manually) across JIRA, Slack and whatever other tracking system they were using.
I think this sounds fantastic, and I welcome all developments in these types of areas where you lean heavily on the hyper-attentive, very fast assistant aspect of LLMs intended to make information more easily searchable, visible and the results more accurate.
I agree that the mismatch between the expectations of OOP as a paradigm and this situation right now feel very similar. I suppose the many billions pumped into selling vibecoding and large-scale code generation via LLMs of other kinds is sort of like the massive amount of money pumped into marketing Java as a viable alternative long before it actually was.
Just like with human langages their only advantage is the community around it, the culture. (Im pretending size is a part of that Obj) No one does it but that is how you should pick it.
I've done this. Ask Claude to permutate a prompt and run Claude Code or a subagent to observe effects until success. Claude will iterate on the prompt faster than I could and likely catch more edge cases, too. It's one of those tasks where the end goal is very clear and the agent just needs to iterate on permutations. Arguably this is a perfect match, too, because prompt engineering is really just text generation, so why not have a text generation machine do it.
Honestly, I've been perfectly fine using OpenClaw and Claude 4.6 for the last six months, and am used to its flaws. In general it doesn't screw up anything major, but I also wouldn't ask it to reason about anything more than a few code files at the same time. Since I'm not interested in it replacing higher level reasoning or architectural design, it's pretty capable of just doing the modifications and features I suggest.
One thing notable is that it makes a difference when I specifically refer to line numbers and variable names when I'm asking it to make changes. Meaning, I also read what it writes and understand what it's doing, and have specific queries where I already know what I'd like to change I add.
A small model, yes! But not necessarily a good model.
With the additional caveat that I don't know whether that specific card is supported by modern drivers.
You'd be looking at one in the 6B or 7B parameters range at FP8. Or smaller. It's been quite some time since a recognizable company in the AI space released a model that small. You can try larger model that has been quantized down to that size, but they don't always fare well with that.
Modern text-to-speech and speech-to-text models also fit well into modest amounts of VRAM.
You're arguing for a very specific range of weights but many slightly smaller and slightly larger models have been released including QAT and MoE versions.
An old nVidia brand card with 8GB is more than enough to see those models running at usable speeds and accuracy.
I'm able to run qwen 3.6 35b a3b with my rtx 2060 super which also has 8gb of vram. It's a very capable model, not as smart as a frontier model but it is still very useful.
I use an 6bit quant and get around 20 tokens per second.
Sure. The fastest small coding model is probably Mellum2 12B-A2.5 by Jetbrains. It matches or beats all Qwen models in this class.
Can even run on a notebook CPU and comes in Base (best for FIM), Instruct and Thinking variants. mradermacher has imatrix quants for people who can't run it at Q8.
IQ4 should fit, but even if it doesn't, llama.cpp has options to partially offload models to system memory.
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.
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 compared to writing a crud app. otherwise anthropic wouldnt have trillions dollar valuation
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.
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.
It does have to be said that if LLMs keep becoming better coders at some point the bottleneck on quality is prompting. Good ideas have many hidden assumptions you think are procedural but often are pivotal to your broader vision.
I find that when I give an LLM my full handcrafted codebase, it does very well. It follows my conventions, sees the intent and can coherently build within its scope. It writes much better code than a 'vibe' prompt.
It is always tempting and I myself will continue pushing the boundaries, but when you keep an LLM in reasonable scope (that may be one line, function, file at a time, depending on your idea of reasonable), you, by definition, can get sound utility out of them.
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.
Just being able to instantly generate a complicated query expression to pull specific bits out of a JSON blob sold me. It's awesome that I can ask Claude to build a whole feature and it will often one-shot it for me, but generating utility bash / python scripts or little throwaway utility webapps is what really excites me.
Agreed. I've been "guide coding" in my editor (Zed / VSCode) for a while now and it's really enjoyable.
DeepSeek v4 flash has been dirt cheap and so fast that my development loop is;
- small prompt
- review
- small prompt
- review
I build software with the same quality I normally would but it's way faster to produce and I think more about architecture and flows than I do about implementation details. The small diffs let me accept / modify / veto diffs and if the model struggles, I just write it by hand. It prevents compounding defects from leading the model astray (like you see in vibe coding).
In some cases vibe coding is useful, like when the complete specification is available (e.g. creating a JavaScript engine that implements the standard) - but anything that requires iterative development sees vibe coding break down pretty quickly (you could argue that is the case for a JavaScript engine).
I feel energised by AI assisted coding rather than drained, as it's a force multiplier for my skills and it lets me build more than I could by myself.
That said, most of my team vibe codes and reviewing their work is like pulling teeth.
Basically how I feel, but I use the low-powered API models like Sonnet. I feel more energized and focused on the bigger picture than getting stuck on stupid implementation of micro-problems.
I think the muscle memory of doing those tiny problems is good for our minds, but solving larger-scale issues is also challenging.
I'm on vacation right now and getting claude to build a mostly-throwaway e2e testing harness (admittedly not small-prompt-review-repeat) for a backend API to speed up our existing e2e test suites which do click-ops to set up tests 8-10 years ago, we had a team who spent 3-4 months every year maintaining our E2E suite and people would do rotations on there to spread the knowledge.
I basically want an industry standard practice implemented on my team of 4 devs who are too busy doing other things.
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.
> 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.
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.
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.
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.
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.
Fable 5 is actually a lousy writer. Opus 4.6 is the best for writing and prose assessment. Gemini 3 is smarter at reading comprehension but tends to be more unstable in judgement.
This argument seems somewhat strange to me, and at odds with the recent model releases.
It feels like claiming that because a current generation laptop CPU is slower than a current generation HEDT CPU, we should ignore that the laptop CPU is faster and more efficient than the previous generation HEDT CPU.
Of course a huge system could do more, but these systems are constantly evolving and just in this month we've seen comparatively smaller models approach capabilities that were frontier AI less than a year ago.
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 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.
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.
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.
That plane was destroyed by Russia, wasn't it? I believe it was partially disassembled when Russia invaded Ukraine and so it wasn't possible to save it. :(
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.
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.
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.
with wireless tech, you can embody Ai in just about anything, waiting for that hacker post about turning their toaster into a thinking machine, I have a Anki vector I've been meaning to do this with (has camera, speakers, microphone, and screen built in)
Is a "truly intelligent home" something I should want? None of the current generation of "smart" addons are true value adds - their entire purpose is data collection. The pretext is always absurdly thin. I just thought "I bet there's a wifi enabled microwave", googled it, and indeed, Samsung have released such a thing - you can control it with your voice! Wow! Never mind that you can't un/load food with your voice, or that you're never more than 10 feet from it in the kitchen anyway. Oh but there's an app. Of course there's an app. It "suggests recipes". Right.
I think hoping for a locally hosted "smart" home is backwards. They have no other function than to invade your space. "Smart" objects are agents, and they don't work for you.
Same. I was in the doom AI crowd ~2 years ago, then I realized how many problems could be mitigated or solved by having these local models, and I became part of the cautiously excited crowd
> 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.
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.
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?
e.g for a lot of small business - they generate invoices - feed that into a RAG system. Then have an agent etc do follow ups. for agents then the smaller models are enough.
> 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.
297 comments
[ 0.96 ms ] story [ 14.5 ms ] threadGosh, watching paint dry has 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?
Now I think Luna is plenty good for many applications inside a very good harness/scaffold. And I think there are a lot of those usecases. So I think these small models are really good for application developers.
But for entrepreneurial knowledge work all of my work still benefits a lot from more intelligence.
The word “most” is doing a lot of work here. On a percentage basis perhaps most tasks a typical SWE needs to do 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 have crazy complicated algorithms they cannot implement themselves and want to have the models achieve this technical breakthrough.
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's safe to say I didn't come to this conclusion because I was "addicted to chasing frontier models"
replit is already leading the way with free luna usage
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.
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.
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
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 :)
WHY IS YOUR TIME MORE VALUABLE THAN THEIRS? WHY SHOULD THEY READ ANYTHING YOU HAVEN'T TAKEN THE CARE TO WRITE? FUCK YOU
I would spit on you if I could. Disgusting behavior.
>”@dang I really need an IP &/or account ban”
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.
Even the best models available lose a ton of detail over time if you were to paste in tens of thousands of lines of code.
The only way to hold huge amounts of context with a high degree of accuracy is to store it using various mechanisms (one of which is RAG).
On “effectiveness”, I mean end use case effectiveness in the tasks at hand, not whatever benchmark the model developer or vendor themselves come up with - which may or may not be useful to the work I’m doing.
For a casual consumer, it is no doubt a worse chat buddy. Knows less historical facts à la Wikipedia.
But I’m not really using LLMs for that kind of entertainment and I don’t rely on them for fact-checking anyway. It’s almost worse to rely on a smoother talker for something it can’t possibly know.
For feeding in thousands of daily updates and getting it to predict the next one based on a crude list of tasks, it’s great. For completing code files in my style, also great. It can also handle most small customer service issues and refunds related to my apps on App Store, the back and forth to collect data from users.
For a booking engine I run with a partner, it is great at confirming bookings and following up.
For invoices, I prefill notes and things based on previous ones. It’s fine. I don’t need a massive model that takes 30s to reply etc. it’s way overkill. Maybe worse than overkill - off in a direction I don’t need.
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.
I'm sure most of it comes down to prompts, and all of them run over 100tps on a 3090. Smaller cards will likely be slower, but Qwen3.5 9B is small enough to fit on most consumer cards.
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.
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.
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
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.
However, what most people think of as Moore's Law--CPU speed doubles every 18 months--broke somewhere between 90nm and 22nm.
And even the actual Moore's Law--2x the transistors every 18 months--doesn't hold for all types of chips anymore. Memory only gained 2x density over 10 years.
The actual argument of the Bitter Lesson essay is pretty limited but people's interpretation of it has gradually drifted until it's seen as prediction that current LLM will reach AGI at a large enough scale.
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.
Any similar model aimed at coding?
A >10B model for mass spawning/swarming and reporting back to a larger model
I wouldn't use it for anything important without heavy supervision, as it's very weak outside its specialty. Not ideal for instruction-following tasks.
Is this whole thing than maybe a read vs write optimisation again? Spent more time and effort training more knowledge into the model upfront and get it out in a single question instead of training a small model and needing more steps to answer the same question?
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.
Specialized models are better at specialized tasks.
I can't find the comment you're referring to, but the latest versions of stockfish are based on neural networks trained on millions of games, so if anything the Bitter Lesson turned out true here.
The common mistake is to think “maybe if we use a blend of raw data and hand-crafted heuristics, we’ll get the best of both worlds!” But the bitter lesson says no, beyond a certain point it’s better just to use the data.
Thinking that an LLM might be able to improve on purely “big data” machine learning seems to me to be the same incorrect idea. Its “intelligence” is no more useful than human intelligence. The LLM is based on a massive data corpus, sure, but the amount of data specifically about chess in there pales in comparison to just playing billions of games of chess.
> maybe if we use a blend of raw data and hand-crafted heuristics
I don't follow. They're suggesting giving raw chess data to the LLM, no heuristics involved.
The conclusion of the bitter lesson would be that a large language model trained on chess commentary as well as being trained on millions of chess games would outperform stockfish which is only trained on millions of chess games.
If you can draw any lessons from chess commentary, I think it’s very reasonable to call it “hand-crafted heuristics.”
And raw games plus raw commentary is all the data you have. You can make more games but those can be fed to both stockfish and the LLM competitor. So it seems like a valid interpretation of the bitter lesson to me.
The key point I was trying to get at is that the human insights don't contain anything that can't be mined from vast amounts of gameplay. Every human insight can eventually be rediscovered and made rigorous by data (in chess, at least!) In the short term, those insights are useful, but in the longer term, they add nothing at all.
Note also that "raw gameplay" here can mean new games -- you can generate as much data as you need, you don't need to rely on real recorded games.
But isn't that the bulk of what we're shoving into LLMs, and it makes them much smarter? If it's useful there but not in a chess AI then that seems like a significant crack in the bitter lesson.
> Note also that "raw gameplay" here can mean new games -- you can generate as much data as you need, you don't need to rely on real recorded games.
Yeah I mentioned that, generated games are useful. But if we're being fair and letting both AIs use generated game data, does the more general LLM ever actually overtake the specialized stockfish like the bitter lesson suggests?
Another way to look at this is that giving the LLM the commentary is a way to avoid complaints of hiding data from the LLM, since it'll have strictly more info than stockfish. But if we cut that from the training data and only give it a basic description of chess and lots of raw game data then it's going to get even worse than it already is at chess. Meanwhile stockfish never had that commentary, just actual hand-crafted heuristics and training on game data, and it's very strong.
Just kidding of course
Not really, if anything it's closer to the opposite. The Bitter Lesson essay literally has this as an example:
> These researchers wanted methods based on human input to win and were disappointed when they did not.[1]
and
> Enormous initial efforts went into avoiding search by taking advantage of human knowledge, or of the special features of the game, but all those efforts proved irrelevant, or worse, once search was applied effectively at scale[1]
The actual bitter lesson is this:
> breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.[1]
Applying to the "LLMs-for-chess" example the bitter lesson approach would be to put many, many more games into the LLM.
Does this work? People have trained fairly small LLMs that are competitive Stockfish at the ELO 1500-2000 level, eg: https://github.com/kinggongzilla/chess-bot-3000
This seems to be evidence that large LLMs probably don't have as much chess training data as Stockfish does.
[1] http://www.incompleteideas.net/IncIdeas/BitterLesson.html
This was/is basically a strawman though. Like maybe "human input winning" was desirable for chess masters but for computer science wonks? Not the point or the disappoint. It's always neats and scruffies fighting about using some kind of recognizable method (logic) instead of magic (ML).
> breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.
More to OP's point I think: nowadays when someone wants to beat you over the head with the bitter lesson, they aren't as careful to include learning and search. They want to say learning leads to intuition (magic) whereby we can avoid work (logic/search), and maybe argue or assume from there that neats and scruffies is settled. TBF, something like reasoning in latent space does resemble intuition!
But the real lesson is confirmed every time we bother to check, and not very bitter for anyone. Search/learning/logic are ALL always necessary on any sufficiently difficult problems, and hybrids that interleave always outperform everything else. Stockfish being the example in this thread that different camps of absolutists would like to claim, but also all the MCTS examples, evolving examples, and new hybrids all the time. My favorite lately: https://arxiv.org/pdf/2511.08983
Oh no!
The whole field was full of people whose entire career was built around the idea of developing smart priors.
To quote Wikipedia:
> For computer vision in particular, much progress came from manual feature engineering, such as SIFT features, SURF features, HoG features, bags of visual words, etc. It was a minority position in computer vision that features can be learned directly from data
This undersells the change though! David Lowe's reputation as the best image researcher in the world was based on his SIFT patent[1]
This approach worked until 30 September 2012.
That was a bitter day for many, many computer science researchers.
[1] https://en.wikipedia.org/wiki/Scale-invariant_feature_transf...
But as it relates to more/less magic, maybe the more modern lens on this is e.g. https://arxiv.org/html/2505.11581v1 . Is manual feature-engineering more like what you'd evolve, or more like what you'd get from SGD ? Feasibility and performance is always a question, there are others like what is robust, stable, adaptable, predictable, explainable. Maybe the manual-features people were interested in something besides the manual part? Maybe the story isn't so simple, and maybe it's not finished yet.
The bitter lesson is that simply scaling training on more games—including self-play—trumps any hand-crafted human input, whether that's fine-tuning on human commentary or clever engineering tricks.
Current models are just high-dimensional interpolation engines. The denser the data sampling, the more accurate the interpolation gets. Given a choice between denser sampling and anything else, denser sampling always wins. That is the bitter lesson.
Computer chess is the canonical example of this.
In the case of stockfish, the harness is a tree search around the neural network evaluations.
It would be better to compare models at how well they can write the code for chess engines, otherwise it's just saying that Fable is not a good CPU emulator, which is obvious.
Stockfish is the best chess search engine we've got, and you can learn some good heuristics for chess search policy that will make time-limited chess search a lot more powerful. That's perfectly in line with the Bitter Lesson.
In contrast, LLMs playing chess are relying solely on learned behavior. The inference harnesses surrounding them aren't designed to do chess things, they're designed to do autoregressive token decoding, which isn't a search process. Reasoning traces can resemble a search process, but they're far less efficient - the LLM would have to work out each legal move, test each one, calculate a score, and simulate minimax over all of that. Assuming the LLM is smart enough to even do all that.
A hand-crafted approach can absolutely beat data if your approach unlocks more search and/or learning than the general solution.
Now let's look at the bitter lesson again. It says that general methods that leverage computation are ultimately the most effective, and by a large margin.
That's different from just saying to leverage computation (which is how I would interpret "unlocks more search/learning"). If the lesson is "more computation wins, when sufficiently channeled" you're basically looking at a truism. Of course more computation beats less when it's used right. The bitter lesson is about abandoning specialization in order to get more computation, and while there's a couple ways where that helps with chess, there's a lot more ways where it's counterproductive. It looks like it's more true for Go than it is for chess, and that it's not universally true. It probably correlates with the state space.
The Bitter Lesson is about general-purpose algorithms vs. specialized algorithms. Historically, chess engines were programmed to look at a chess position and use positional understanding (imparted by the human programmers) to decide what the best move is. But eventually, the chess engines that actually became stronger than humans were instead programmed to just check every possible move and countermove and see which ones lead to a win. (I'm oversimplifying, but you get the point.) So even before Stockfish contained a neural network, it was considered an example of the success of the Bitter Lesson.
As it applies to AI agents, the Bitter Lesson would predict that the best possible agent would simply possess A) a way to do anything it wants, B) a way to evaluate whether what it did was correct, and C) a ton of compute. Then just turn it loose on your task. (The fact that the "brain" of the agent is an LLM is kind of irrelevant - you could also imagine the brain just being a program that generates random syntactically-correct code. What the LLM achieves is that, the random generator would take millions of years whereas the LLM is much more efficient at creating plausibly-working code. This is analogous to a chess engine's pruning heuristics.)
The hard part here is B. We've seen some great agentic successes when rewriting an existing project in a new language, since the agent can just use the project's prior test suite as its evaluator. But when developing a new project, you're still figuring out the finer details of how everything is supposed to work. As the old saying goes - writing a spec that perfectly describes how a program should work, is equivalent effort to just writing the program.
Dumb AIs are needed for customer service. Most of that industry is still at "press 1 for sales, 2 for billing..." and needs something that will run locally on a 1U server.
Needs? Customers want something that immediately answers their question/solves their problem, but that’s far away, even ignoring the “run locally on a 1U server” and that that may not be in the company’s interest.
For many companies, that support line is a cost center, not a PR mechanism.
Also “Press 1 for sales, 2 for billing...” has the big advantage that it handles all accents, speech impediments, etc.
Long term I think a solution where a user’s agent trained on their voice, running on their phone communicates with the support agent of a company is where we will end up, and support phone lines will cease to exist.
And perhaps at the end it all gets a single pass by a god-tier model for overall sanity and congruence, but the actual work, planning, coordination, and even user interaction was done by cheaper and faster agents of much more limited capability.
It really is absurd to ask programming questions to a model also trained about the lifecycle of a fruit fly.
Instead of building small models from scratch, we train an enormous model and use ridiculous amounts of GPU memory. In the end, the whole thing is shoved into RAM because we don’t know where the useful parts are…
We certainly would know where they were if they were just in smaller models in the first place!
The correct analogy here is Fable as a second tier player assisting a SuperGM in running stockfish, then assessing its output to identify promising variations.
There might be a limit somewhere that prevents the bitter lesson being axiomatic - for instance where simulations for anything can be exhaustive - so that judgement isn't needed any more as an arbiter. But while there are problems sufficiently complex or large to require a breadth models don't currently have, greater scale and compute will continue to convert to better decision making, and the bitter lesson will remain true (true enough).
GP isn’t suggesting that focused narrow model(s) will be more capable than large model, but that many small focused models can have sufficient capability while being more optimal.
Also, the bitter lesson is just wrong. The bitter lesson is about hand tuned AI vs computational general methods. However in truth today’s AI uses both. We have general compute heavy models which require narrow expert instructions (eg tools internet docs).
LLMs would not be as good without expertly written context, and expert context without LLMs aren’t as good either.
The models are not even really trained bitter lesson-style anymore. That concept peaked during the era of pre-train scaling, back when it was thought that making a bigger and bigger GPT-3 would automatically solve all problems through prompting. In 2026, the most important part of training is post-training, which uses vast quantities of niche, hand-curated data to fit the models for specific tasks in domains like tax law.
If you look at value as purely the LLM output, then there's a valid argument that the best frontier models will always be better than fine tuned specialists. (I'm not convinced personally, but it's a defensible claim)
But that misses two dimensions: 1. The cost of acquiring that output 2. What is actually "good enough" for that specialist domain
Not every output needs to be the best to produce value.
And as specialist models increase in cost, their cost/value proposition goes down.
At some point, there's a threshold where cheaper, fine tuned models are "good enough" at the task and also substantially cheaper than the expert models.
That's where fine tuning helps.
Personally, I became a believer in fine tuning after fine tuning a 1B Qwen model as a second pass over my local voice transcription app, achieving excellent accuracy at ~zero token cost and waaaay lower latency than if I'd invoked my Claude subscription under the hood.
Absolutely false. At least when it comes to multimodal inputs, even a simple classifier will outperform the largest LLMs who still hallucinate details or don’t describe audio and images accurately.
And there’s also the issue of cost/inference speed. Running a trillion parameter model for all tasks will be incredibly costly, require a cloud API, while a tiny CNN can be run locally or at a cost multiple orders of magnitude lower.
When you're doing something "at scale", you want smaller models, weaker heuristics and rules written in code, as they're much cheaper, much faster, much more predictable and much easier to reason about.
If your task is "tell me which pages on the internet are most relevant to this search query", running Claude Fable to evaluate every single page would be ridiculous. You may run something like it at the end, to synthesize information from a small result set for example, as that's a "human-sized" task.
If your task is "write a function to sort a list of 1mn entries alphabetically", a small coding model will exactly match the performance of a frontier model (it'll 1-shot it, with 100% accuracy). So the only differentiator is cost, and the frontier model will cost orders of magnitude more than the specialised model.
For most intelligence based tasks, you don't (and never have) needed the tool which "performs best at all tasks". You need the cheapest one which performs adequately for your immediate task.
This doesn't mean the bitter lesson is incorrect. At the frontier, it's still correct. It means that it's not applicable at all to lots of tasks.
You don't necessarily need a bigger model to do this.
In engineering you learn from these mistakes and try never making them again. Do we want to go through this evolution every time we solve a software issue? Just because we can with an unlimited number of cheap tokens? I think not, I’d rather use the knowledge build up that also knows about the edge cases forgotten to test. Or better, use multiple models that evaluate each other, as Entropic describes it in their recent report https://news.ycombinator.com/item?id=49316271
> We expect that agents coordinating in the wild will act in higher variance ways than we see here, because they’ll have different backgrounds and therefore different contexts. They also, presumably, won’t all be Claudes.
For analysis, bug hunting, overview and some porting work to popular languages and so on I think the current SotA is fantastic, but they're still very disappointing for code generation past function or small module level.
Recently my employer hooked Claude into a bunch of SaaS services like a ticketing system and an asset inventory system, and I feel a big productivity boost is that I don't have to dig for their open browser tabs, remember their product name/URL/where I put the bookmark, find that my sessions have logged me out "for my protection" since I last looked, find my MFA code, approve sign-in on my phone, then use their mismashed/sluggish/poor-UX interfaces and slow searches to loko for things, and can instead ask Claude "search these systems for anything about <issue>" and it just does. It correlates between systems, summarizes things, and gives me references which system and where to look to check what it found. Internal search on a mess of acquired and disjoint systems.
This feeling reminds me of two classic Joel Spolsky blog posts below, and your comment here about the surprisingly useful tab-complete gives me a similar feeling. I wonder how many smaller candidates there are where LLMs are more helpful and less hyped, against the big vibe-coding solve-everything hype which are currently less helpful than their hype?
References [1]: "A lot of us thought in the 1990s that the big battle would be between procedural and object oriented programming, and we thought that object oriented programming would provide a big boost in programmer productivity. I thought that, too. Some people still think that. It turns out we were wrong. Object oriented programming is handy dandy, but it’s not really the productivity booster that was promised. The real significant productivity advance we’ve had in programming has been from languages which manage memory for you automatically. [...] Whenever you hear someone bragging about how productive their language is, they’re probably getting most of that productivity from the automated memory management, even if they misattribute it. Sidebar: Why does automatic memory management make you so much more productive? 1) Because you can write f(g(x)) without worrying about how to free the return value from g, which means you can use functions which return interesting complex data types and functions which transform interesting complex data types, in turn allowing you to work at a higher level of abstraction"
and [2]: "I have seen many language and programming fads come and go. But there’s only ONE, that’s right, ONE language feature I’ve ever seen that actually improves your productivity significantly. No, it’s not object oriented programming; no, it’s not intentional programming or assertions or programming by example or CASE or UML or XML or Java. The only thing that improves your programming productivity is using managed code – that is, using a language in which memory management is automatic."
[1] https://www.joelonsoftware.com/2004/06/13/how-microsoft-lost...
[2] https://www.joelonsoftware.com/2001/10/17/working-on-citydes...
I agree that the mismatch between the expectations of OOP as a paradigm and this situation right now feel very similar. I suppose the many billions pumped into selling vibecoding and large-scale code generation via LLMs of other kinds is sort of like the massive amount of money pumped into marketing Java as a viable alternative long before it actually was.
One thing notable is that it makes a difference when I specifically refer to line numbers and variable names when I'm asking it to make changes. Meaning, I also read what it writes and understand what it's doing, and have specific queries where I already know what I'd like to change I add.
With the additional caveat that I don't know whether that specific card is supported by modern drivers.
You'd be looking at one in the 6B or 7B parameters range at FP8. Or smaller. It's been quite some time since a recognizable company in the AI space released a model that small. You can try larger model that has been quantized down to that size, but they don't always fare well with that.
Modern text-to-speech and speech-to-text models also fit well into modest amounts of VRAM.
An old nVidia brand card with 8GB is more than enough to see those models running at usable speeds and accuracy.
I use an 6bit quant and get around 20 tokens per second.
Can even run on a notebook CPU and comes in Base (best for FIM), Instruct and Thinking variants. mradermacher has imatrix quants for people who can't run it at Q8.
IQ4 should fit, but even if it doesn't, llama.cpp has options to partially offload models to system memory.
A sales person sending a prospect email doesnt have a way to write a test harness for it. Yet these tasks dominate compared to writing a crud app. otherwise anthropic wouldnt have trillions dollar valuation
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.
I find that when I give an LLM my full handcrafted codebase, it does very well. It follows my conventions, sees the intent and can coherently build within its scope. It writes much better code than a 'vibe' prompt.
It is always tempting and I myself will continue pushing the boundaries, but when you keep an LLM in reasonable scope (that may be one line, function, file at a time, depending on your idea of reasonable), you, by definition, can get sound utility out of them.
These things would be considered magic even 4 years ago!
DeepSeek v4 flash has been dirt cheap and so fast that my development loop is;
- small prompt
- review
- small prompt
- review
I build software with the same quality I normally would but it's way faster to produce and I think more about architecture and flows than I do about implementation details. The small diffs let me accept / modify / veto diffs and if the model struggles, I just write it by hand. It prevents compounding defects from leading the model astray (like you see in vibe coding).
In some cases vibe coding is useful, like when the complete specification is available (e.g. creating a JavaScript engine that implements the standard) - but anything that requires iterative development sees vibe coding break down pretty quickly (you could argue that is the case for a JavaScript engine).
I feel energised by AI assisted coding rather than drained, as it's a force multiplier for my skills and it lets me build more than I could by myself.
That said, most of my team vibe codes and reviewing their work is like pulling teeth.
I think the muscle memory of doing those tiny problems is good for our minds, but solving larger-scale issues is also challenging.
I'm on vacation right now and getting claude to build a mostly-throwaway e2e testing harness (admittedly not small-prompt-review-repeat) for a backend API to speed up our existing e2e test suites which do click-ops to set up tests 8-10 years ago, we had a team who spent 3-4 months every year maintaining our E2E suite and people would do rotations on there to spread the knowledge.
I basically want an industry standard practice implemented on my team of 4 devs who are too busy doing other things.
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?
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.
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.
> Thanks to Sam Altman, … for reading drafts of this.
Oh how times have changed!
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.
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.
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.
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.
There's a million ways to define "better at writing prose", and it's only better at some of them. It's worse at many others.
Try using a LLM model for RAG embeddings and get back to us on that.
It feels like claiming that because a current generation laptop CPU is slower than a current generation HEDT CPU, we should ignore that the laptop CPU is faster and more efficient than the previous generation HEDT CPU.
Of course a huge system could do more, but these systems are constantly evolving and just in this month we've seen comparatively smaller models approach capabilities that were frontier AI less than a year ago.
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.
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.
I think hoping for a locally hosted "smart" home is backwards. They have no other function than to invade your space. "Smart" objects are agents, and they don't work for you.
You can do this before an exit?
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?
the application Layer i.e having a good graph RAG & connecting it up together is the missing piece for most.
I don't have to be an automotive engineer to start my car and put it in drive.
e.g for a lot of small business - they generate invoices - feed that into a RAG system. Then have an agent etc do follow ups. for agents then the smaller models are enough.
[0]: https://help.getzep.com/graphiti/getting-started/overview
Given sheer number of turns I can make with small models, I can do a lotta stufff
- cheaper, and faster
---
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
This is a good insight broadly!