Robinhood CEO Says Majority of Company's New Code Written by AI (businessinsider.com) 66
Robinhood CEO Vlad Tenev has said that the majority of his company's new code is written by AI, with "close to 100%" of engineers using AI code editors. Speaking on the 20VC podcast, Tenev estimated around 50% of new code at the trading platform is AI-generated.
Tenev said the 50% figure is imprecise due to advanced "agentic" code editors that have made it difficult to distinguish human-written from AI-generated code. The company has progressed from GitHub Copilot to Cursor and now Windsurf, where "nearly all of the code is written by AI," he said. Tenev estimated only a "minority" of new code at Robinhood is written by humans.
Tenev said the 50% figure is imprecise due to advanced "agentic" code editors that have made it difficult to distinguish human-written from AI-generated code. The company has progressed from GitHub Copilot to Cursor and now Windsurf, where "nearly all of the code is written by AI," he said. Tenev estimated only a "minority" of new code at Robinhood is written by humans.
Hacker's paradise? (Score:5, Interesting)
Gotta wonder how many extra security vulnerabilities there will be in their trading infrastructure.
Re: (Score:3)
The Lady Marian backdoor.
She is not amused.
Re: Hacker's paradise? (Score:2)
Yep. Someone's going to fuck them hard from behind one day.
Re:Hacker's paradise? (Score:4, Interesting)
Likely a lot, because them doing it cheaper means they are not doing the code review that AI code needs. Also expect unreliability, outages, corrupted trades and other crap.
Re:Hacker's paradise? (Score:4, Interesting)
Re: (Score:2)
Indeed. And that has been known for a long time. These "AI coders" seem really intent on repeating history.
Re: (Score:2)
Oh I dont know. My workmates think I'm a goddamn horrible monster in code reviews.
That said, in 30 years in the job, you get a nose for that sort of things, cos god knows if theres something stupid you can do in code, chances are I've already done it and learned the hard way.
Re: (Score:2)
That said, in 30 years in the job, you get a nose for that sort of things
AI will mistakes that are different than the normal human mistakes, so you need to turn up your attention and look for things you're not expecting.
Re: (Score:2)
Sure, it can be done by somebody competent and experienced. But it does not save time to have AI write flawed code and then review it to fix it when you want the same quality level as something originally written by a competent coder. It takes more time.
Also, when you want high quality code you have two quality gates: (1) The coder and (2) the reviewer. With AI, you only have the reviewer as an instance that knows what they are doing. So coder+reviewer and AI+reviewer are not the same thing. The second is
Re: (Score:2)
That and a human, even in error, at least knows what he meant to do when writing code. When AI spits out code, it isn't a given that anyone understands what it does, much less what it is meant to do.
Re: Hacker's paradise? (Score:1)
There was the article before about AI automatically adding imports to stuff that is a Trojan horse. I think that will become more common over time until it starts to explode in the faces of various companies.
From what I see, even code that requires a lot of editing is retarded as AI code as well
Re: (Score:2)
None, because he just made this shit up.
Metrics drive behaviors (Score:2)
When the boss starts tracking your AI use, your behavior will change.
Expect major IT incidents in their future then (Score:5, Interesting)
Because that is an extreme risk. Having "AI" write code is only cheaper if you do not have competent people do careful review of that code. As soon as you have competent people do careful review, it becomes much more expensive, as reviewing code above a certain low complexity threshold is much harder than writing good code in the first place. AND AI code is harder to review on top of that because everything looks good, even when it is not.
Re: (Score:2)
You are hallucinating hard...
Re: (Score:2)
Re: Expect major IT incidents in their future then (Score:1)
I agree about AI generated code requiring competent review. I disagree that in the hands of a competent reviewer, AI generated code is slower. It's incredibly fast to build out a well-speced function.
And reviewing one function should be pretty easy for a skilled coder. Especially if you wrote the specs.
So I would never unleash an AI to build an entire application from top to bottom (although I've heard of some promising techniques that involve generating requirement documents, running them through multiple
Re: (Score:2)
Sorry, but code review is always slower than writing code on the same quality level above a very low complexity threshold. This is well established. It stems from (a) the code not being your code and hence the style making things more difficult for you and (b) review needing to use a top-down approach. There is nothing that can be done about these things.
Your statement "And reviewing one function should be pretty easy for a skilled coder." is bullshit and leads me to believe you have no real-world code revi
I call bullshit (Score:5, Interesting)
I recently used an agentic AI to push through an undocumented USB stack (I'm talking to you TinyUSB) while trying to implement a proprietary and poorly documented protocol as a custom driver. The AI was able to apply some reversed engineered logic I assume it learned from scraping open source projects.
The whole time I had to treat it like a hostile four-year-old brushing their teeth. "Brush up, now down, no you're not done yet, up again, then down, now move to the next set." And then I had to rewrite everything since it was 200% the size it needed to be. It did save me some time, but 50% sounds like an ass fact (a fact pulled from your ass). Oh wait, they probably left the bloat in, which exactly aligns with how my code would look if I didn't rewrite the whole thing to be maintainable.
Um, I meant I'm calling bullshit on anyone still claiming AI generated code will be full of security vulnerabilities. This is absolutely not the case if the AI is driven by someone competent. The old rules still apply - you get what you pay for. If the engineer is good, the code will be good no matter the tools.
Re: I call bullshit (Score:2)
"This is absolutely not the case if the AI is driven by someone competent."
That's the point, they are claiming to be saving both time and money. That means they're using incompetents, not reviewing the code properly, or both. Or just lying of course. No guesses which from me, any of these are plausible.
Re: (Score:2)
Re: (Score:2)
Someone competent today may not be very competent after a few months of doing only boring AI code reviews. That, or the reviews may degrade to rubber-stamping due to their tediousness.
Just get other AI to do the reviews...
Seriously, I've used AI to answer questions that I might have posted to a forum if something is working, or I want some ideas on how to do something. I have found the results to vary from helpful to useless. For me, its usefulness is getting ideas, asking it to explain them and then I write the code so I understand what it does, not just cut and paste. What find it most useful for is taking my code and adding comments explaining what it does so someone in the future
BULLSHIT indeed (Score:1)
Robin Hood's CEO knows nothing of what his coders are doing, and you can tell because he conflates "AI code" with "AI code editors."
In either case AI can't yet code at any level of trust worthy of a flower stand point of sale system let alone a commodities trading platform.
He knows nothing. He speaks as if he does. Those are the signs of an idiot pretending not to have impostor syndrome.
AI isn't coding ANY production code ANYWHERE today. If it were, we'd hear about it from REAL companies with REAL softwa
Re: (Score:2)
It'll be awesome when robinhood has its dictionary.com moment.
Re: (Score:2)
Not "by AI", it really "with AI assistance" (Score:5, Insightful)
The reality is 'with "close to 100%" of engineers using AI code editors'. So it's really "with AI assistance" not "by AI". Thank you AI editor for the code completion suggestions.
Maybe it offered a lint-like warning that some thingee was not deallocated. Thanks again.
Re: (Score:2)
Our executives pulled this trick recently. They created a survey that was engineered so it could be interpreted to say a large amount of code is AI generated. Basically they took the numbers of like "how many times a day do you use Copilot" and turned that into n% of code is AI generated. Never did they ask how much of the code we write is AI generated. We have a pretty good system with Copilot, and it can contextualize a good amount of code, but the only thing it replaces is Google search/StackOverflow. It
Re: (Score:2)
Tenev said the 50% figure is imprecise due to advanced "agentic" code editors that have made it difficult to distinguish human-written from AI-generated code. Tenev estimated only a "minority" of new code at Robinhood is written by humans.
You just got to trust him on all of this. A CEO would never lie.
He’s also working with the Trump White House...
Oh, can't trust a word out of his mouth.
Re: (Score:1)
Its really programmers using an AI tool and auto-complete in an editor is a tool as well.
So soon we will have to draw a line: when is AI writing code and when is it a tool? will it always be a tool? if one single prompt writes an entire bug free project, is that AI writing code 100%?
Re: (Score:2)
"if one single prompt writes an entire bug free project, is that AI writing code 100%?"
that's not the right question. A prompt may be broken down into a series of prompts, a prompt itself isn't really important. We can say that "AI" generates 100% of the output for its prompt, as long as we define what prompt and output are (and AI I suppose).
The answer is who knows, and who cares. Anyone claiming these things is doing so for non-technical reasons. What's important is (a) what's being done, (b) how it's
Re: (Score:1)
call me boring or irrelevant, that's fine.
On a good day ... AI can glue a few algos together (Score:2)
On a really good day, AI can notice that a particular problem can be solved by glueing together a few such well known and well documented algorithms. Not that it reasoned this or anything, its training probably included disc
Re: (Score:2)
They added an AI button to Gmail (we use web gmail for work) and let me tell you that our email communications are not 100% written 'with AI assistance'. But the button is there 100% of time, so our CEO would happily tell investors that they are.
Re: (Score:2)
They added an AI button to Gmail (we use web gmail for work) and let me tell you that our email communications are not 100% written 'with AI assistance'. But the button is there 100% of time, so our CEO would happily tell investors that they are.
I'd use it for spelling and grammar check. :-)
Re: (Score:2)
>> Thank you AI editor for the code completion suggestions.
I use it every day. Its way more than that, he isn't "misrepresenting reality".
You can work up some very good incremental development with nothing but prompts, the AI will write all the code for you. If you don't quite like what it did or want improvements you can just tell it. On new projects I generally start small and build new features as I get things working. In the past I had to write all the code, now I just describe what I woul
AI the ultimate cut and paste coder? (Score:2)
Re: (Score:2)
I'm thinking you haven't used these tools as a professional developer. They are surprisingly capable. The LLM's have been trained on all coding textbooks, all open source code in existence, all stack overflow interactions, etc so there's quite a bit of material to draw from.
And a lot of new code will be similar to what's been written previously elsewhere. Web servers and interactions with databases for example have a lot of boilerplate, and a LLM can put that together for you very quickly. Also generate uni
Re: (Score:2)
I'm thinking you haven't used these tools as a professional developer. They are surprisingly capable.
I've done precisely what is described at times. Iteratively adding to and refining generated code. Its pretty much doing as I described, stitching together an algorithm or three known to address the subgoal at hand.
The LLM's have been trained on all coding textbooks, all open source code in existence, all stack overflow interactions, etc so there's quite a bit of material to draw from.
How is that not what I said? An automation of what we already do?
Also generate unit tests, integration tests, and full documentation.
Again, an automation of what we were already doing. Note that testing and documentation is something different from the defensive coding I referred to.
Re: (Score:2)
If the generated code isn't 'defensive' enough you can always tell it to make some targeted enhancements, but I find that it is perfectly capable of catching exceptions, etc.
>> How is that not what I said?
"Thank you AI editor for the code completion suggestions."
It is way more than code completions at this point.
Re: (Score:2)
How is that not what I said?
"Thank you AI editor for the code completion suggestions." It is way more than code completions at this point.
Wrong post. The relevant post is "AI the ultimate cut and paste coder?" ..."
"Using AI is like automating looking up a reference implementation of an algorithm in a textbook. Or automating googling a problem and finding a solution on stack overflow. Between these two an AI might recognize when it can glue a few well known well document algorithms together to solve a particular problem
Re: (Score:2)
Whatever. I disagree with that statement also. I find that the AI can write very acceptable code even when the 'algorithms' are not well known. They may have been encountered in some obscure training documentation or code sets that hardly any human has seen.
Also I find that the better models can interpret your intentions from prompts quite well even if there are spelling errors, and often elaborate it better than you were able to state it. When companies say their new code is written mostly by AI I have no
Re: (Score:2)
Whatever. I disagree with that statement also. I find that the AI can write very acceptable code even when the 'algorithms' are not well known. They may have been encountered in some obscure training documentation or code sets that hardly any human has seen.
I'm not sure how you are disagreeing when you are saying pretty much the same thing. "So yes, if you carefully sequence questions to fit such automation, it can go find reference code for you and build it out step by step." An automated search leading to more possibilities to evaluate. It then "cuts and pastes",
Also I find that the better models can interpret your intentions from prompts quite well even if there are spelling errors, and often elaborate it better than you were able to state it. When companies say their new code is written mostly by AI I have no doubt they are truthful.
I think one of the new skills for humans to develop is how to phrase the questions to give the AI the best hints possible.
Re: (Score:2)
>> I think one of the new skills for humans to develop is how to phrase the questions to give the AI the best hints possible.
Totally agree with you about that.
Re: (Score:2)
Yep. The CEO's definition of "the code was written by AI" is apparently just "the author of the code used an IDE that has AI available."
Someone should ask the CEO whether their email client has AI available.
If yes, then "okay so AI writes all of your email now, I see, thanks."
Re: (Score:2)
So it's really "with AI assistance" not "by AI".
You don't know that though. You're just applying the thought of how you as an engineer would use AI, to someone else's work. Not everyone is the same. Some people take zero pride in their work and will shortcut it anyway possible.
Another "trust me, bro" benchmark? (Score:5, Interesting)
Meanwhile, I just had Claude 4.0 try to write some simple refactors earlier today and it randomly inserts semi-colons in the middle of Java method calls and it still can't match braces. I can't rely on it to do basic refactors and ChatGPT and Copilot seem a lot worse.
So does Robinhood.com have a new proprietary AI? Does he have infinitely better tools than are available to the rest of us?
I don't know what to believe anymore. It seems like every CEO is doing the AI song and dance and cheering it on...but I can't see an example of a good AI application or any major service built on AI...just various toolkits that promise glory someday...and various CEOs using it to justify layoffs.
I logically assume it has to do some things well....but are both Amazon and Apple pulling back their releases? Why are the vendors selling general AI platforms? If this can write the vast majority of Robinhood's code, why aren't the vendors offering more specialized services?...how about I pay you a few thousand a month and your magic AI guarantees my company's codebase is perpetually security patched?....or some service that will go through my legacy code and port it to a new language? How much money could you make porting COBOL to Java or Rust?...or VB to C#?...or it doesn't even have to be a legacy language...why not Python to RUST?...why not a JVM, CLR, node.js or Python runtime that converts your regular code to amazing assembly language or highly optimized code?...seems like a license to print money.
I want to believe...I just keep hearing promises of "someday"...and honestly just don't know what to think. This is all a scam seems too simplistic...but....why can't I see more concrete proof?
Re: (Score:2)
I think where we are is a weird culmination of sales-speak along with minor practical advances leaving us with a weird perception by the C-suites that AI is taking over, while in reality the real work is still being done by humans.
Our company has executives going to AI summits that preach AI in everything, all the time, for every reason. These people usually work for AI companies, and desperately want AI to have access to current codebases so that they can get more training data from stable, long-term codeb
Re: (Score:2)
how about I pay you a few thousand a month and your magic AI guarantees my company's codebase is perpetually security patched?
I think their ask would be more like a few hundred thousand, or a few million. The bros have been especially comparing the prices of their services to the prices of salaries. They need to box people into this paradigm of comparing AI to humans. Probably the obsession over the word "agentic" has something to do with this as well.
Framing the loss of talent as "increasing your AI headcount" serves a psychological purpose, making the C-suite less nervous about cutting off chunks of their company. Or at least pr
Old McDonald (Score:3)
Even if it's correct code (Score:1)
it's still just the 2025 version of a 1960s-era IBM, DEC, Honeywell, HP, DG, and Wang* saying majority of our new code is written by compilers where code = stuff the computer that will run the code understands.
Today's AI-generated source code is 60-years-ago's compiler output. Compilers have bugs, and so does AI.
* and Amdahl, NEC, and NCR [genius.com] too
Re: (Score:2)
are you saying that it's not different until AI emits machine code directly?
Why stop there, let's eliminate instruction sets and have AI schedule execution units on demand. It's still just a the same old Turing machine!
Previously 100% of code was written by IDEs (Score:2)
The C-Suite is filled with ignorant people who have made a career out taking credit for other people's work.
Now they want to pretend that the "AI is writing the code" just because a developer is making use of it.
All AI is doing is giving CEOs an excuse to called yet another group of workers, "unskilled."
CEO Marketing. Selling a dream. (Score:2)
What do CEOs do? They don't do any actual work. They are the face of a company and the ultimate marketing representative. Their job is to smile, talk and do anything it takes to excite potential investors and bring more money into the house. Usually, what they say have nothing to do with the actual state of the company because they are either too immersed in their own self delusion or just too disconnected from the ground floor.
It's no different this time. Robinhood CEO is just selling a dream to potential
AI - Actual Indian (Score:2)
I wonder how many of those AI agents are just outsourced contractors based in India like it was discovered to be the case with some big AI company a few weeks ago.
where did all the old code go? (Score:2)
You can't get to 100% without removing code that doesn't comply with a metric, in this case human-written code? Why was all the existing code removed? It wasn't, of course, it's just a lie.
Should investors be impressed if the story their being given is either (a) an obvious lie, or (b) the company clearly doesn't know how to responsibly manage its code?
Re: (Score:2)
Re: where did all the old code go? (Score:2)
written by AI (Score:2)
Public domain code (Score:2)
If nearly 100% of their code is written by AI that means it's not under copyright then. Interesting future when this becomes the common case across all companies.
job listings (Score:1)