For AI chip companies, the promise of AI-based coding agents is compelling: generate the low-level kernels needed to get a model running on new hardware, optimize them for performance, and do it without the software engineering resources traditionally required.
But according to Chris Lattner, CEO of AI software startup Modular (recently acquired by Qualcomm, where he now serves as executive VP of software engineering), chip companies should be wary of the limits of AI code generation. Making a kernel run fast is a very different problem from building software infrastructure we can depend on, he told EE Times in an exclusive interview.
“Agents are good, but they’re like anything else; it’s a tool, and you have to use it the right way,” Lattner said. “What I’ve seen with agents is they make tons of throwaway code. So, if you’re building a product, you don’t want to agent-generate everything.”
Agentic code generation makes sense for projects that are well-defined and limited in scope (e.g., “make this kernel go fast”), which can be quantified, measured, and validated, Lattner said, but bigger projects require product design and software engineering and can’t be delegated to agents.
“If you’re a startup and you only care about getting one model working, or getting it to go fast, [then go ahead,] use AI tools to prove it’s possible,” he said. “But that’s not going to be great for when you want to sell your chip to someone else, and they want to run a model you don’t have access to.”
Lattner said that despite some high-profile companies’ performance claims for their AI-generated low-level software, there are still many unanswered questions, including the dollar cost to generate large amounts of code and its applicability to large, open-ended problems.
“I have heard of hardware companies that say they don’t believe in software anymore; they just want to use agents to generate everything,” Lattner said. “I think that’s interesting because they’re definitely using LLVM and still investing in compilers, but saying they’re not. So there’s a divergence between the party line and what’s actually happening.”
Writing performant low-level code for new and novel hardware has long been a difficult task for chip companies, and some are now relying on agents for this.
“Various people use [agentic code generation] as a talking point, particularly when their software is known to be problematic,” Lattner said. “So, I don’t really know what to believe in all these cases… great claims need great proof.”
Porting code
While AI is useful in some use cases, it can’t be used for everything, Lattner said. AI might help get a demo up and running, but this isn’t the same as writing a usable software stack with determinism and adaptability.
“The hard part about building a company or building a product or building something meaningful in the world is understanding the product requirements,” he said. “What is the marketing story? What is the product-market fit that you have to get for your product? I think that remains quite hard.”
At Modular, teams use agentic code generation for tasks such as porting PyTorch code into the company’s Max AI framework and for porting CUDA code to its programming language Mojo. Lattner said that porting code is a great use case for AI, along with specializing existing software (for example, specializing a matmul kernel for a specific size of matrix), since the agent can do this quickly and efficiently.
People will always need to know how to code, he said, since prompting in English isn’t precise enough to get a good result.
“It’s true things are changing very rapidly—I use AI to write code, and I wouldn’t go back, personally—but that doesn’t mean that writing code was the job,” he said. “If you go back five years, more code was read than written.”
Open source under pressure
The downsides to AI code generation are still making themselves apparent in the world of software. Lattner identified open-source projects, the backbone of much of modern software, as being particularly under threat.
“Open-source projects like LLVM are under assault,” he said.
Before generative AI, people who wanted to contribute to open-source projects were motivated by the satisfaction of identifying and fixing problems, most often because they were using the software themselves and wanted to contribute. Making changes meant they had to invest a certain amount of time. Maintainers, people who had been working on the project for a long time, reviewed contributions for quality and worked with the community to teach them the nuances of the codebase. Lattner said the investment in teaching contributors improved the code.
“Today, that model is broken, because people are just vibe coding tons of patches and sending them to tons of projects, and the maintainers are getting completely overwhelmed with AI crap,” he said.
The deluge of AI-generated code is increasing the workload for maintainers, who have to review and iterate code before accepting it, Lattner said.
“And then the submitter doesn’t actually care, so the investment back into the human doesn’t lead to a future maintainer; it leads to someone who walks away, or just sets up their agents to spam a bunch of different projects to try to get GitHub badges,” he said. “This is a huge issue, so a lot of projects have stopped accepting contributions.”
The response from open-source projects such as LLVM has been to consider working only with established contributors, a move Lattner described as “terrifying.”
“The value of open-source projects is the humans, it’s the relationships, it’s the tribal knowledge,” he said. “If you destroy that, there’s going to be a huge problem… Running an open-source project and making it successful has always been super hard, so this is another pretty big threat.”
This is a huge challenge for all open-source projects today.
Modular’s programming language, Mojo, was recently open-sourced, but contributions aren’t yet being accepted. Lattner said the company intends to open to submissions gradually.
“It’s not just about the lines of code; it’s about the social side of things—if you have an employee that’s investing their effort into someone else, traditionally the thought would be, we’re investing in you, and if you’re good we’ll hire you, or [something might come out of it],” he said.
Things will stabilize, Lattner said. AI code review tools will improve, easing some of the burden, but ultimately, code review is difficult to automate completely because it requires judgement.
“What you’re building doesn’t just need to pass a test; it needs to be able to support the architecture we need for tomorrow and the future,” he said. “Models tomorrow will be better and faster, but I think these are inherent structural problems.”
Max alliance
For startups without the resources to build a state-of-the-art software stack, Lattner is planning on opening up Max, Modular’s AI framework, by starting an alliance for hardware makers. The alliance is planned to kick off around the end of the year, he said.
“This will be a big deal,” Lattner said. “We’re saying to everyone who has chips: You now have software. A lot of startups have been asking for help, but we always had to say no; we couldn’t scale to support them. Now we’ll be able to say yes, you can do it for yourself.”
Coding agents can solve for a specific kernel for a specific chip, but today’s complex, disaggregated, heterogeneous data center needs orchestration, scaling, and management, all of which must be handled by software, Lattner said.
“We’ve been building this for four and a half years—now you’ll be able to use it and scale it for yourself,” he said.
This is part of Qualcomm’s plan for Modular, Lattner said.
“[Qualcomm] wants it to be an open ecosystem so [its data center chip] can compete on its own merits,” he said. “Qualcomm is modest enough to understand it can’t [overtake Nvidia]. There will be a lot of different players with great contributions in different parts of the value chain, but it needs to be an open ecosystem.”
The Max alliance will enable edge chip companies and markets such as China, which Lattner is looking forward to.
“This is going to be very, very fun and exciting,” he said.
Also read:
How Much Impact Will AI Have on IoT Software Engineering?
Can Agentic AI Solve the Embedded Software Problem?
Qualcomm Bets on Modular to Break Nvidia Moat
Leave a Reply
You must Register or Login to post a comment.