AI Startups Revolutionize Coding Industry, Leading To Sky-High Valuations 39
Code generation startups are attracting extraordinary investor interest two years after ChatGPT's launch, with companies like Cursor raising $900 million at a $10 billion valuation despite operating with negative gross margins. OpenAI is reportedly in talks to acquire Windsurf, maker of the Codeium coding tool, for $3 billion, while the startup generates $50 million in annualized revenue from a product launched just seven months ago.
These "vibe coding" platforms allow users to write software using plain English commands, attempting to fundamentally change how code gets written. Cursor went from zero to $100 million in recurring revenue in under two years with just 60 employees, though both major startups spend more money than they generate, Reuters reports, citing investor sources familiar with their operations.
The surge comes as major technology giants report significant portions of their code now being AI-generated -- Google claims over 30% while Microsoft reports 20-30%. Meanwhile, entry-level programming positions have declined 24% as companies increasingly rely on AI tools to handle basic coding tasks previously assigned to junior developers.
These "vibe coding" platforms allow users to write software using plain English commands, attempting to fundamentally change how code gets written. Cursor went from zero to $100 million in recurring revenue in under two years with just 60 employees, though both major startups spend more money than they generate, Reuters reports, citing investor sources familiar with their operations.
The surge comes as major technology giants report significant portions of their code now being AI-generated -- Google claims over 30% while Microsoft reports 20-30%. Meanwhile, entry-level programming positions have declined 24% as companies increasingly rely on AI tools to handle basic coding tasks previously assigned to junior developers.
Re: (Score:2)
The days of 6-figure dev salaries are rapidly closing. This is nothing short of a watershed moment in the democratization of software creation and 'on demand' app creation for the masses is only a few days away. In other words, rather than searching for an app that does what you want it to do, examining the license, possibly paying money to obtain it, working around shortcomings or bugs, etc ... you will just tell the AI what you are trying to do and it will create an app for you 'on the fly.' Owned by you, doing exactly what you want it to do without reporting PII back to some mothership. This is the end game here and I for one cannot wait.
I'm not sure what in the hell fantasy you're playing out here, but exactly who do you think is going to simply give away that kind of compute power?
Yeah. You might be able to ask AI to create $your_app 'on the fly', but that sure as shit ain't gonna be "owned" by you (code, license, or otherwise), and you'll likely be paying money for both the app-generating query and the compute effort no matter how short it may be.
You meant to say the days of free to 99-cent apps, are over. No PII? That'll be $139.99.
Wrong Problem (Score:3)
Re: (Score:3)
>> they will still need to describe the program with complete precision
It isn't as hard as you might think. You have to get a good grasp of what the AI is appropriately capable of doing and prompt it in incremental steps. This comes with a little practice. Code generation happens in a minute or two so the penalty for failure is very small, if something didn't quite work out you can reject it and try a different approach.
You do need to have a reasonably clear idea of what the outcomes should be as you
Re: (Score:2)
You have to get a good grasp of what the AI is appropriately capable of doing and prompt it in incremental steps.
It depends on what you are doing and how specific you need your code to be. The more precise the harder, and harder it is to get the AI to do what you are asking in many cases. For example, when I asked ChatGPT what the statistical signfiicance of a specific scientific result in it said "very significant", when I told it to be more precise in its response it responded that it was a "high statistical significance". I then asked it to specify the number of sigma significance, it responded "more than 5 sigma"
Re: (Score:2)
It sounds like you tried to push the AI beyond the edge of its limits. If you "get a good grasp of what the AI is appropriately capable of doing" you are more likely to get satisfactory results.
Re: (Score:2)
Re: (Score:2)
>> the AI can't write the code we need
From the account of your experience above you "asked ChatGPT what the statistical signfiicance of a specific scientific result". That doesn't sound like coding to me. Also ChatGPT is not the only option out there. One of the others might be more capable for your purposes.
I'm hearing that a large proportion of software developers are using an AI model for assistance these days, clearly they find it useful. I certainly do, and I've seen an amazing improvement in cap
Re: (Score:2)
It may be that those limits have expanded in the past few months although that example is only a month old. However, I'd be surprised if my original point does still not
Re: (Score:2)
I do data science work with AI assistance on a regular basis, and I can tell you that it is well versed in all of the relevant packages. More so than any human is likely to be. I'm curious to know what kind of scientific analysis you are doing where AI isn't able to provide a boost for you.
Re: (Score:2)
>> who do you think is going to simply give away that kind of compute power?
They don't give it away, but its real cheap for the end user. Windsurf is currently $15/month for 500 compute units and gives you access to all the modern AI variants. The one I use does a fine job at a cost of .25 units per prompt. An 'enterprise' subscription gives you a whole bunch more for $50. Do you own the resulting code? Damn straight you do.
Yes they are losing money, but this is the paradigm these days. You burn thro
Re: (Score:2)
AI can write code... but here is the rub... the code needs debugged, unit tested, smoke tested, refactored, made to company standards, commented, audited, and so on. That all takes time. Will the AI written stuff even work? Who knows. I've had code that tried to use commands that were not even existing. It might just be that it takes more time to debug and fix as opposed to writing from scratch.
Best thing is to do small pieces... a function or some easily debugged chunk, by chunk. Slow, but sure.
Yes,
Re:Overpriced dev divas in shambles (Score:5, Insightful)
It's a giant string generator, copying from other people's strings. It's a good giant string generator, but that's what it is - another tool in the box. Most of programming is not just the syntax, it's the ideas. "Doing exactly what you want it to do" - hah, most people absolutely cannot specify exactly what they want a thing to do.
Re: (Score:2)
hah, most people absolutely cannot specify exactly what they want a thing to do.
Exactly. I'd also add that doing proper validations is also hard for "generic business person" to understand, much less ask for.
Re: Overpriced dev divas in shambles (Score:2)
Makes me think of all of the times I've debated the correct behavior with my colleagues in various edge cases.
A lot of new developers think they're being clever by quietly disregarding errors that they think are just minor and can be safely ignored, not realizing that they're quietly introducing bugs that they just made harder to debug.
And then you get older developers like angle o sphere who are stuck in their ways, and you try to explain to him why he should have used a deque instead of a linkedlist, and
Re: (Score:2)
Reminds me of people who wrote a program I am working on, which is Angular -> GraphQL -> C# -> SQL Scripts which do the main body of the work. If everything you have is a hammer etc....
Re: (Score:3)
...hah, most people absolutely cannot specify exactly what they want a thing to do.
This, in spades. I often write specification documents for large projects that my clients want to build - automated people mover projects, two-way public-safety-critical radio systems, light-rail passenger systems, etc. It is really hard to write a tight, complete spec that doesn't have ambiguity that must be dealt with later when the contractor actually attempts to go out and build it.
Dross in, dross out.
Re: (Score:2)
Heard this so, so many times over the last 35 years. 3GL, 4GL, graphical-style (Powerbuilder etc.), object orientiation...so, so many times.
Note that a lot of those technologies actually worked. Because of open source shared libraries, I can now build tremendously complex products with very little effort. Now both of these are relatively easy [xkcd.com].
And yet despite that, programmer salaries have kept going up.
Re: (Score:2)
Owned by you, doing exactly what you want it to do without reporting PII back to some mothership. This is the end game here and I for one cannot wait.
I doubt the above, I believe it is far more likely than you will be granted a licence to use the app, but the ownership of it will stay with the startup.
Re: (Score:2)
What democratization are you referring to? Every platform where an AI can generate an application for you already has tools you can use for free to do that.
Popcorn (Score:3)
Re:Popcorn (Score:4, Insightful)
If you thought users were frustrated with their computers and apps now, just wait until they have to spend hours and hours asking an AI to "fix" their app, only to have it create increasingly bizarre problems.
I can't wait for the days of troubleshooting turning into, "Which crap AI did you use to create your app suite? That'll give us a starting point to figuring out exactly how fucked your system is and how lost and irretrievable your data is."
Lessons from feeding the amateurs (Score:2)
Whenever amateurs and tinkerers use "low code" or Rapid App tools* to create apps, almost 90% turn into maintenance headaches in my experience. They "mostly work" up front, but either the original author leaves, or the app grows into too big a mess for the original author to fix.
The user group calls in the formal IT team to fix it, and then get angry when the formal IT team says, "they didn't use standard practices, made poor documentation, and we are not familiar with their tool". I expect AI "solutions"
AI coding (Score:3)
A fool and his money (Score:3, Insightful)
Re:A fool and his money (Score:5, Insightful)
Re: (Score:2)
Re: (Score:2)
Worldcom was not a .com startup, it started as a small LD reseller in the 1980s. By the time of its bankruptcy two decades later, it was the second largest telecommunications provider in the US.
They should donate their valuations to open source (Score:3)
Tainted by garbage. (Score:2)
we don't want open source to be tainted by garbage.
plain English commands (Score:2)
Chatgpt is amazing for simple coding (Score:2)
Re: (Score:2)
Last time if I didn't know better I'd think it was fucking with me... 'Generate code following the rules I gave earlier where your output was correct.' 'No, you didn't a or b.' 'Ok you did a and b,
Be careful (Score:2)
You mean like builder.ai [techspot.com]? Because the sky-high valuations may not turn out entirely reflect the actual value.
Re: (Score:1)