LLMs improve faster at coding due to its deterministic execution
Eric Simons · Inside Bolt: From near-death to one of the fastest-growing products in history | Eric Simons · Mar 13, 2025 · at 1:09:57
StackBlitz CEO Eric Simons compares software development to non-deterministic fields like law to explain why LLM reinforcement learning works especially well for code generation.
0:00 / 0:57exact quote · 57.3s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Software is deterministic. When you write code and you hit run, either runs or it doesn't. And that is why that's the key insight Anthropic really had. They just went deep. And this is what they're doing. It's just reinforcement learning. I'm basically permutating every type of app you could ever build. And just spinning up tens of thousands of courses whatever to do that, just building tons of training data and doing reinforcement learning and making their LMS the best in the world at building beautiful, reliable applications. I am, I'm extremely bullish, right? I mean, it makes technical sense. Why of anything LMs are going to get insanely better at writing code than probably most other types of applications for LMs, simply because it's something that, that can be extremely deterministic and permutated thousands and thousands and thousands of times per second.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Eric Simons
Simons: Bolt grew from zero to over $30M ARR in 4.5 months
“In the first two months, we went from zero to twenty million of ARR. And I think we're on like month four or four and a half or something like that at this point. And we've already crossed thirty million of ARR and we're on the verge of crossing 40.”
Simons: Cursor is the only code-gen startup ahead of Bolt in revenue
“Our company wasn't doing AI stuff the six months ago, we had no AI products and just out of nowhere, we you know, From almost death of the company to being, you know, the number one by traffic revenue, et cetera, like, you know, AI co-gen app that's, you know,…”
Client-side compute is the only web model scaling to a billion users
“If you look at the other productivity apps that have really worked on the web, they've all had this compute model, right? Figma, when you open a Figma document, there's not like some cloud VM that gets spun up for you to render the documents you're dragging th…”
Cloud VM-based AI coding tools cannot scale due to global capacity limits
“And that's how all these other AI you know, code text to apps sort of tools work. They're all using cloud VMs. And the problem is, you know, on a small scale, it can work, but as you scale it up, there, I mean, there, there's not even a hundred million VMs to …”
Simons: Prototyping in Bolt is faster than using Figma
“Now it's faster to do this with bolt in code and have a real working software products than Dragging around frames in Figma to actually make a static version of it.”
Simons: PMs will directly ship code via AI instead of Jira tickets
“How this is going to work, if you fast forward, you know, one, two, five years, whatever, like PMs are, they're going to be writing quote unquote, instead of just writing a Jira ticket and waiting for a developer to do it. The developers are going to be able t…”