AI-Assisted Coding Start-Up Kite Is Saying Farewell and Open-Sourcing Its Code 32
Kite, a start-up that has been developing artificial intelligence technology to help developers write code for nearly a decade, is saying farewell and open-sourcing its code. Silicon Republic reports: Based in San Francisco, Kite was founded in 2014 as an early pioneer in the emerging field of AI that assists software developers in writing code -- an 'autocomplete' for programming of sorts. But now, after eight years of pursuing its vision to be a leader in AI-assisted programming, founder Adam Smith announced on the company website that the business is now wrapping up. According to him, even state-of-the-art machine learning models today don't understand the structure of code -- and too few developers are willing to pay for available services. "We failed to deliver our vision of AI-assisted programming because we were 10-plus years too early to market, ie, the tech is not ready yet," Smith explained. "You can see this in GitHub Copilot, which is built by GitHub in collaboration with OpenAI. As of late 2022, Copilot shows a lot of promise but still has a long way to go."
Copilot was first revealed in June 2021 as an AI assistant for programmers that essentially does for coding what predictive text does for writing emails. Developed in collaboration with OpenAI, GitHub had kept Copilot in technical preview until this summer, during which time it had been used by more than 1.2m developers. The AI was made available to all developers in June, at a cost of $10 a month or $100 a year. However, Smith said that the inadequacy of machine learning models in understanding the structure of code, such as non-local context, has been an insurmountable challenge for the Kite team. "We made some progress towards better models for code, but the problem is very engineering intensive. It may cost over $100m to build a production-quality tool capable of synthesizing code reliably, and nobody has tried that quite yet."
While the business could have still been successful without necessarily increasing developer productivity by 10 times using AI, Smith said he thinks that Kite's delay and unsuccessful attempt at monetizing the service prevented the start-up from taking flight. "We sequenced building our business in the following order: First we built our team, then the product, then distribution and then monetization," he explained, adding that Kite did not reach product-market fit until 2019, five years after starting the company. Despite the time taken to get to the market, Smith said Kite was able to capture 500,000 monthly active developers using its AI with "almost zero marketing spend." But the product failed to generate revenue because the developers refused to pay for it. Smith says most of their code has been open sourced on GitHub, including their "data-driven Python type inference engine, Python public-package analyzer, desktop software, editor integrations, GitHub crawler and analyzer, and more more."
Copilot was first revealed in June 2021 as an AI assistant for programmers that essentially does for coding what predictive text does for writing emails. Developed in collaboration with OpenAI, GitHub had kept Copilot in technical preview until this summer, during which time it had been used by more than 1.2m developers. The AI was made available to all developers in June, at a cost of $10 a month or $100 a year. However, Smith said that the inadequacy of machine learning models in understanding the structure of code, such as non-local context, has been an insurmountable challenge for the Kite team. "We made some progress towards better models for code, but the problem is very engineering intensive. It may cost over $100m to build a production-quality tool capable of synthesizing code reliably, and nobody has tried that quite yet."
While the business could have still been successful without necessarily increasing developer productivity by 10 times using AI, Smith said he thinks that Kite's delay and unsuccessful attempt at monetizing the service prevented the start-up from taking flight. "We sequenced building our business in the following order: First we built our team, then the product, then distribution and then monetization," he explained, adding that Kite did not reach product-market fit until 2019, five years after starting the company. Despite the time taken to get to the market, Smith said Kite was able to capture 500,000 monthly active developers using its AI with "almost zero marketing spend." But the product failed to generate revenue because the developers refused to pay for it. Smith says most of their code has been open sourced on GitHub, including their "data-driven Python type inference engine, Python public-package analyzer, desktop software, editor integrations, GitHub crawler and analyzer, and more more."
It is basically bullshit (Score:4, Insightful)
Even if you code the most simplistic business logic, there are just far too many options for continuation unless you understand the problem being solved.
Re: (Score:2)
Because we have a dilapidated framework due to budget issues, there's a lot of repetitive code I need to write that AI should have a relatively easy time with, at least for making a draft.
Re: It is basically bullshit (Score:2)
What would be more helpful might be for the AI to suggest a refractor to get rid of the repetitive code in the first place. Although no ai required... Intellij already shows duplicated Code fragments..
Imho, if an AI can figure out what you are doing, you are probably doing it wrong and writing redundant code to begin with...
Re: (Score:1)
> What would be more helpful might be for the AI to suggest a refractor to get rid of the repetitive code in the first place.
I requested to refactor but it doesn't fly with the team for messy reasons per TLDR. At this point I'd be happy to automate the repetition. Plus, it's probably a simpler goal for an AI engine.
Re: (Score:2)
At this point I'd be happy to automate the repetition.
What's wrong with just copy/paste? If you're stuck with the repetition, I don't see any reason not to just make a "template" for you or your team to use.
Re: (Score:2)
Indeed. Or a macro or a function. Die you know you can even work on code with, say, Perl scripts, if stuff gets really mechanical?
Re: (Score:2)
The trouble is that we know absolutely nothing about his problem other than it involves the production of repetitive code. Tablizer is more competent and thoughtful than most. I doubt he'd overlook something like a function, the single most common way to deal with repetitive code. For whatever reason, that doesn't seem to be a possibility in his case.
I don't know anything about his opinions on code generation, but his post here suggests that he was resistant to it for some reason until recently. Personall
Re: (Score:1)
> What's wrong with just copy/paste?
It doesn't propagate change well. It's more code to rework if the original "template" turns out wrong or needs changing. One can't just re-copy-and-paste because copies are typically adjusted for spot needs. That's typically the purpose of templates: paste something that's 80% there and change 20% for fit. If you re-paste a changed template, you wipe out that 20% at each spot.
Re: (Score:2)
I didn't mean in general, I meant in your specific case!
Re: (Score:3)
Imho, if an AI can figure out what you are doing, you are probably doing it wrong and writing redundant code to begin with...
And that is the core of the issue. "AI" cannot help you unless what you do requires no thinking.
Re: (Score:2)
For repetitive code pieces there is a far better tool. It is called a "function". Some languages even offer a syntactic version called a "macro". And, as a last resort, there is something called "copy & paste". As soon as that does not cut it anymore due to variations, Artificial Ignorance will not cut it either.
Re: (Score:3)
Me: printf("Hello,
AI: World!");
Thanks, AI!
Re: (Score:2)
I guess you could automate the generation of CRUD templates and other common sw-design-patterns, but copy-paste and find/replace would work almost as well.
Re: It is basically bullshit (Score:1)
If an AI can figure out the code while developing, why write that code in the first place instead of letting the AI do the actual work..
Re: (Score:1)
Many frameworks and/or shops like "consistent bloat" over refactored stuff. I find it frustrating, but they do. I can understand as my fav refactoring technique may be very different from another's. Team-friendly code tends not to be parsimonious, I hate to say, because convention & consistency overrides parsimony in most team environments.
Re: (Score:2)
convention & consistency overrides parsimony
Those things are not as different as you might think.
Re: (Score:1)
Maybe if both are done really well, but in practice they usually are not.
Re: (Score:3)
If some of your coders can't make and/or comprehend and deal with abstraction and factoring, fire them immediately.
Simpletons on a
Re: (Score:2)
If some of your coders can't make and/or comprehend and deal with abstraction and factoring, fire them immediately.
Simpletons on a codebase are toxic, infectious, and fatal.
They are also one of the core reasons why so much bad code is out there. But, if the PHBs do not want pay for competent coders, then incompetent coders producing crap it is. It seems the "60% of software cost is maintenance" average is still not known to many "decision makers" and hence they mess it up and invest far too little at the start which is ending up being very expensive.
Re: (Score:1)
The real world is Dilbert-esque, unfortunately. I can't refactor American Business practices nor humans, and doubt you can either. We're stuck with (mostly) hairless apes who happen to talk. We have to live with humans as-is for now.
Factoring indigestion (Score:2)
Multiple team leaders from diverse orgs said they've witnessed too much factoring confusing later coders/maintainers when I suggest factorings. After that many samples, I mostly have to believe them. They see the real world, not just theory.
Re: (Score:2)
Exactly.
Re: (Score:3)
Copilot is really hit and miss I've found.
Sometimes its downright spooky how good it is, and dumps out half a page of really good code that passes the unit tests, matches the docstring description perfectly, validates for everything, and has no obvious side effects in half the space I had envisioned it being.
Other times its pasting a javascript cookbook database into a rust file handler function.
And I can never quite predict which one I get. But the amount of times it does actually suggest something bang on
Re: (Score:3)
What about copyright? Remember the respective legal proceedings of the last two decades or so, where sometimes a single line was evaluated at billions of damages and similar stupidity.
Ten years translated by XKCD (Score:5, Insightful)
"We haven't finished inventing it yet, but when we do it'll be awesome" [xkcd.com]. I'm still waiting for the expert system to replace my Doctor as is sure to happen in 1985 .
I'm forming a startup to (Score:5, Funny)
...use AI to assist in the bankruptcy process of AI and crypto co's.
Tentative name: Autochoke. Alternatives welcome ("Zucked Up" taken).
from the 1980s (Score:3)
Before ML....
https://groups.csail.mit.edu/i... [mit.edu]
This guy needs to write a very detailed paper (Score:2)
The lessons this guy learned need to be published as a detailed paper on what he learned in the process of failing to accomplish his goals. Iâ(TM)m pretty interested in the failures of his approach, which do not get shared often in any sort of detail.
Failing with grace though it seems (Score:4, Interesting)
Re: (Score:2)
It's a shame more companies can't/won't do this. In a lot of cases, the investors hold out hope that they may be able to sell the IP for something, so don't allow it to be open sourced. A year or two later when its obvious no buyer will ever materialise, no one's around to do anything about it, so the knowledge is lost.
Tech companies no longer do tech (Score:1)
"We failed to deliver our vision of AI-assisted programming because we were 10-plus years too early to market, ie, the tech is not ready yet,"
What was their actual plan, wait until someone else invented it? Do Silicon Valley companies actually innovate any more?
Who could have known? (Score:3)
From TFS: "It may cost over $100m to build a production-quality tool capable of synthesizing code reliably, and nobody has tried that quite yet.""
Really? Coding is hard? Who could have known? And $100M seems kind of low for something that synthesizes code *reliably*.