What Chinese firms are doing makes perfect sense from the commercial perspective actually because they understand how a classic commoditization spiral works. The reality is that models themselves are general commodities and there's just not enough difference between them. A company can get ahead of others by a few months, but then the rest quickly close the gap. It's a really low margin business because there's no way to differentiate yourself.
Chinese companies know that there's no profit in general purpose models in the long run, and they're treating models as shared infrastructure akin to Linux. They're amortizing the cost of research by keeping models open, and rapidly closing the gap and driving prices towards the marginal cost of inference. The money is going to be in customization niches. Companies will charge to tune models for specific use cases and charge support for that. There's also going to be money at the bottom for hardware vendors making chips and memory. But the middle tier of generic LLMs is seeing involution where there's relentless competition driving profits towards the bottom.
Network-based recovery refers to a rocket landing on a massive, ship-borne capture net, instead of landing on landing upright on its engines like a SpaceX Falcon 9. Hooks on the booster get snagged by a net of tensioned steel wires that ride robotic rails sliding into place to meet it. Catching instead of landing has the advantage of shedding the heavy legs and sparing the fuel a soft touchdown burns to hover.
The question is that of cost. Sure, there's no downside to having a smarter model, but if you have to pay orders of magnitude to use it, then it's just a waste of money. If you can get around on a bicycle then you're not gonna buy a monster truck.
There's definitely a saturation point depending on the complexity of the problem you're solving. For example, any model can write a small shell script to resize a video with ffmpeg for you right now, so it doesn't matter whether you're using a local Qwen model, GLM, or Fable. They'll all do a roughly comparable job and you'll end up with a working script that does what you need.
Then you have things like CRUD apps, where a model needs to write some SQL, make a service endpoint, serialize some JSON, etc. Here a local model might have a bit more trouble juggling all the pieces, but any hosted model will do just fine. If your day to day job involves working on CRUD apps, then it's basically a solved problem now.
The cases where frontier models matter are when you're solving genuinely complex problems, but that's not what most people are doing day to day. So, paying an order of magnitude for a model that has capabilities to solve problems outside the range of problems you actually work on becomes a waste of money.
There's going to be a market for these models from people who really do work on complex things on regular basis, but the question is how big that market is. Additionally, open models keep getting better, and GLM 6 or DeepSeek v5 could end up being another big jump in capability where they fully close the gap with Fable. At that point, even more of the market becomes covered by these models leaving truly complex cases on the frontier.
Another thing to consider is that most big problems can be broken down into smaller ones. That's the basis for how programming languages are structured. We have primitives which are arranged into functions, that get bundled into classes or namespaces, and so on. So, you don't need an infinitely capable model to solve big problems. You just need to be able to break large problems into smaller ones, and a model that's smart enough to decompose a problem to the point where it becomes tractable.
The anger doesn't come from the internet, which is just an outlet. What's driving the anger is the decline in living standards as more and more people are pushed to the brink economically. Majority of the population is barely making ends meet while being crushed by debt. The richest 20% of American consumers made up 50% of spending, that means most people are so strapped for cash they can only afford the essentials. That's the situation on the ground for you.
Except that capitalist competition is precisely what leads to centralization and concentration of capital. The way competition works is that winners grow making it harder for new players to enter the market. You need more up front capital to even try while the existing players get to leverage economies of scale, and if a new company becomes a threat they can just buy it out or run it out of business. That's precisely how current tech monopolies were born with smaller companies like WhatsApp, YouTube, Instagram, and so on, all being absorbed by a handful of giants.
You're just regurgitating slogans here without any actual understanding of the subject.
You're just doing sophistry here. There are patterns in the brain that form which allow it to model the world and make predictions about it. If we can understand the nature of these patterns and implement them in a different substrate, then we'd be able to do similar type of inference on a computer. That's what algorithms are.
Exactly, it'd be the same as regular chip designed evolving. You get a specific model version baked into the chip, if it does what you need then it's fine. If you need more capability in the future, you just buy a new chip.
I also think the dynamic would be really different if model inference can run at ridiculous speeds. You could make a genetic algorithm loop around it, so it can generate a population of proposals at each step, then have those tested and whittled down iteratively. If inference happens at thousands of tokens per second, then from user perspective it would still be really fast, and even a small model could solve complex problems.
Reverse engineering how algorithms in the brain work is a really promising path towards making genuine AI systems which would make the current crop of LLMs obsolete.
I was thinking how it would be interesting to make an environment where instead of LLM just crapping out a bunch of code really fast, it works more like a pair programming exercise going at human speed.
The LLM would explain what it's doing, then write a bit of code, then you have time to look at it and understand it, and go to the next step. At any point you can interject and discuss or change it.
I find the biggest problem is that once an LLM generates a bunch of code, it's really hard for a human to build up the context for what the code is doing and why. When you're coding normally or pairing, then you're gradually absorbing the context and what the code is doing throughout the process.
The reality is that writing code fast was never the bottleneck. It's understanding the code and making sure it's actually doing what's needed that's hard.
This whole situation is very reminiscent of how Microsoft was trying to get Linux and open source banned when NT started losing market share on the server.
Ah yes, voting is always effective. Thank goodness people in Germany kept voting in the early 1930s. Imagine what terrible things might have happened if they hadn't.
I look at what the model is doing in the loop and whether the harness is catching cases such as the model having to write scripts to balance parens, whether it's trying to do the same thing over and over again, and all the other cases I explained in detail in the blog post.
Even without having hard numbers, it's pretty easy to see from the log whether the model is getting stuck or not.
I haven't really seen anybody come up with a good test to show hard numbers on comparing agentic harnesses. It's a bit tricky to set up a definitive test given the whole non deterministic nature of LLMs. What I've been focusing on is watching the loop and seeing where model does things that it shouldn't have to. For example, I notice models doing stuff like writing python scripts to match parens for Clojure all the time using editors like Pi. So, having a mechanical way to repair parens, and when that fails, to give the model clear error regarding where syntax is broken removes that whole cycle.
As it stands, it's kind of subjective, you just have to try the harness and see if the model seems to be have better than with the other ones you've been using.
Thanks, glad to hear the harness is actually doing its job with smaller models on your end. There definitely seems to be a limit of how small a model can get before it can't do any practical work.
I find I tend to view agentic coding similarly to a genetic algorithm. The model is the mutator function, and the harness along with the tests acts as the selection function. Each round the model generates some plausible code, it gets tested against the constraints, the model gets feedback and iterates on it until it converges on something that's workable. So, the real trick is to make sure the environment is producing correct pressures to guide the model in the needed direction.
Can't wait to see what GLM 6.0 is gonna be like. So glad I got their subscription back when they had a discount at $250 a year. At the time, I really wasn't terribly impressed with 5.1, but now with 5.2 it's good enough for most of what I do.
[ my public key: https://keybase.io/yogthos; my proof: https://keybase.io/yogthos/sigs/JQklAIz-z2zRaANShfHgNDDmq_0mLbg24Mg2TzcYzw8 ]