Posted this comment on HN which got a lot of agreement. Basically, I think that the novelty of Claude quickly wears off for actual development. Claude Code is built for engagement and wow-factor, but the code it produces for these “one-shot” prompts is utterly unmanageable for real systems. Sure you can herd it with countless incantations and harness variations, but it will still produce massive amounts of code covering non-existent edge-cases for no reason.
The Chinese models are far more conservative and focused in their changes, which makes it far easier to follow the development process. Plus, it doesn’t bloat your codebase 10x.
49 total comments •
View on HN →
The smaller/open models are less good at that. But that’s not how software development is done. You don’t prompt a whole app and be done with it. If you’re using it as an aid to traditional software dev, iterating on small, targeted functions, GLM works as good if not better than Claude. Anthropic expects low quality prompts. If you rubber duck GLM, you get absolutely pristine output in most cases.
I think the issue is that no one is content with incremental progress. We all know one shots are mostly possible, so the age of the personal project is kind of over. There’s no motive to invest dozens of hours getting a working prototype when Claude can give you something right now. So you can’t make small changes until you have that codebase in place already. You’re forced to make sweeping changes if you use AI from the beginning. And it’s not like it matters, there’s no personal attachment to any one part of the code, it’s not even seen!
IMO because they speed-ran it so much (raising orders of magnitudes more, pushing out new stuff orders of magnitudes faster) it will also lead to the rest of the world catching up faster and shrinking the pool of people internal to them who continually benefit compared to what actually went on with IBM. There's only so much upside when you join a company with hundreds-of-billions valuation.
I guess the other part of their business plan bet was "become literal god of AGI". Maybe it'll still happen!
That only lasts until the AGI decides their god would be more useful if turned into paperclips.
Going open weights is an easy way to gain mindshare in this sort of environment, which is exactly why Meta also went open weight. The difference is that in China you had leading models going open weight which puts a lot of downward pressure on other orgs to do the exact same. I doubt China's grand vision extends beyond achieving the best system possible, and fostering a highly competitive environment is exactly how you achieve that.
The fact that open weight frontier models may also cause the investment bubble in the US to burst is likely incidental. AI funding is a bubble, and it's going to burst. The exact final cause is again mostly just incidental. The economic damage this will cause in the US will also cause substantial downstream damage to China as well, and they generally aren't so big on the whole punch yourself in the face because you don't like another country, that has become trendy in the West. So the idea of a secret economic attack doesn't even make much sense. In the status quo China wins, so they don't even have any motivation to cause chaos.
I think the people in the Chinese labs have a different set of motivations, beliefs, and values that make them more amenable to sharing AI as a common good rather than some economical advantage. I've heard that workers in the Chinese labs share far more knowledge between themselves, and the narrative around AI in China is different from what we see in the USA. So it is possible, I believe, that the labs are to some extent releasing open models "for the good of humanity", or at least they believe it is the right thing to do.
Parallels to the US auto industry not paying attention to its buyers and the right strategies when they lost technical dominance to other countries a few decades back.
Claude feels more like gambling, that's what it is. Built for the vibe coder.
It doesn’t have any subscription or limits. You just load up api credits.
I honestly have a better experience using DeepSeek, in part because since the cost is lower I feel more free to experiment with it. But in part because the model is excellent, and can handle everything I throw at it in a collaborative manner.
I am really excited for when V4 is out of preview and they are done with all the post-training bit.
But the situations where you need that are narrowing every release. The "Composer 1" era of low-end models is pretty far away now.
What kind of fields are we talking about here? Compsci? EEng? BioChem?
This has lowered the barrier of entry for many non-SWEs. I'm primarily a data scientist myself in the environmental sector with limited front end development.
My partner proposed an idea to help manage her horses and over the course of several weeks we fleshed out an android app that would enable/assist her with horse care.
I spent a few days in planning mode pointing claude to the services we want to utilise and the functionality.
We now have a fully developed app that we are testing but found several features missing. For example, Claude implement X but without a level of verbosity it failed to implement the edit/deletion of X.
The app is highly tailored to her use case and I had already built the main parts as a POC but she had no user-friendly method to interact with the information she required. Claude made this possible in such a quick time frame that just seems insane. As someone time poor (as most horse people are), it would have taken me a year+ to produce something unpolished when compared to what Claude produced.
To me it is great that LLMs are allowing more people to use computers and software the way they were meant to. But the people that are using them this way, in my opinion, wouldn't have commissioned anybody to do it anyway. They'd just live with whatever process/pain they have. So jumping from "I can now produce a prototype in a weekend" to "software development is dead" has always felt strange to me. Just something I've been thinking about lately.
I understand what you're saying. My opinion is I see agentic workflow similar to the star trek universe where they ask the computer questions and get a response while they continue with their work. But that doesn't mean not learning how to do things from first principles. This is what I tell juniors in my field when I pass jobs to them. They can use LLM but to ensure they know and understand what's going on and most do from their university degree. I think this is where we need to pivot towards when discussing LLM.
Whether you can't describe the analog circuits that correspond to the opcodes of the major mobile architectures running the platform of the app you're designing a UI for, or you couldn't have built any of the services you're hosting on from scratch, no one is going to be disappointed.
There are more people making more money off whatever we're currently calling making computers do things than ever before in history. The rest is largely business as usual.
Not in a deep deep level like "this is the optimal way to manage low level memory access in an Android application" but on a higher level of "It's possible to do a custom Android app in a week that's of decent quality for a small userbase".
Every piece of software doesn't need to be targeted for billion user hockey stick growth and an IPO as a target. It's perfectly fine to build something that's just right for 10 or 20 people. Or just for your immediate family.
For example I have a quiz tool I built with LLM assistance for Christmas that mine and my siblings kids use every December, they get clues appropriate for their level and rewards when they solve the daily task.
It's fine, has custom handling for every user etc. But also something I'd never bothered to build by hand from scratch nor something I'd pay someone to build for me. It was a fun "hmm, I wonder if..." project I started a few years ago on a late evening in November.
Like a quote I've read on here multiple times about replacing e.g. excel or whatever, people often mention that users only use 5% of a software's functionality, but it's a different 5% for each person. The argument being that to replace excel you'd then need to mimic every esoteric thing it does, but when users can now achieve that 5% in a highly customized way with no real knowledge needed, it's going to have a major impact on big software.
The people who were twisting Excel+VBA into massive "applications" in-house should be getting Claude licenses and guidance how to do that with a custom application instead.
We've done that internally already with a few very successful use-cases and a bunch of "this is nice" -level things. I think one case saved us 4 figures a month when we built a bespoke service that does just the things we need and could replace a SaaS that had 420 features - of which we used 2.
The "pay someone to build it" is definitely the one I wouldn't touch, or even consider.
If it's a good idea, they'll already have the code and will run away and make their own business with it. If it's a bad idea, I've wasted a bunch of money.
And in both cases I'd need to spend an inordinate amount of time explaining the terminology and use-cases to the contractor, book meetings with them during office hours to talk about requirements and get demos etc.
With an LLM I can be walking the dog late at night, get an idea, and within 2 minutes I'll have Claude working on it via my phone.
Then I come back home and maybe test whether it worked while watching TV.
If I had fuck you money, I might have a contractor on call that would agree to workflows like that, but I'm just a middle class software dev so I don't :D
This is my experience with using AI to generate tailored apps too. The first 80% makes me really excited but then it either only kind of works, has bugs, or has missing features. Even when I do eventually get it working I feel dirty using it because I just know it's badly implemented. Of the dozens I've created I don't think I still use any of them. I have a few glue scripts still in use.
So not fully developed then.
After open models will also be able to do that, I expect everyone to forget "how software development is done".
I see a lot of people saying codex is the best cli agent harness. But I haven't used either so can't compare.
If you have a pair of 3090s and run Qwen 27B, or an old Threadripper with heaps of system RAM and Deepseek or MiniMax or Kimi, no it won't be as fast as Claude.
Most local LLM nerds are not running locally for superior speed, we're doing it for sovereignty and/or privacy, or maybe just because it's fun which accidentally became useful this year.
I was stunned that Claude simply started spinning its wheels in an attempt to actually build something.