May 16, 2025 · 43m · a16z
Who's Coding Now? - AI and the Future of Software Development
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this a16z panel discussion, Matt Bornstein, Yoko Li, and Guido Appenzeller examine how artificial intelligence and 'vibe coding' are transforming software development workflows, computer science education, and enterprise programming paradigms.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 0.0 Guest teaching 2.0 Guest disagreement 1.8 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 31:15 Rejecting Direct AI Transpilation of Legacy COBOL
Matt directly rejects the optimistic narrative that AI can automatically port legacy COBOL code, emphasizing that critical context and architectural intent have been irrevocably lost over decades.
Hardest push from the host ▶ 40:10 Pushing Back on Prompts as Standardized Protocol WaistlinesYoko explicitly refutes Guido's framing of prompts as a narrow waist analogy, arguing that unstandardized natural language prompts lack the formal constraints of historical protocol waistlines.
Biggest teaching moment ▶ 37:40 Reframing LLM Compliance Metrics around Human Error BaselinesGuido educates the panel on enterprise LLM deployment, explaining that financial institutions must reframe compliance metrics away from zero-tolerance determinism toward matching human error rates.
The host holds their own ▶ 21:50 Demonstrating Foundational Computer Science ExpertiseMatt demonstrates deep technical domain expertise by breaking down classical CS curriculum layers, citing hands-on experience with assembly, physical gate connections, Lisp, and OS file systems.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Why Coding Is AI's Biggest Homogenous Market | 0 | 1 | 1 | 0 | The panel collaboratively estimates market sizes for AI coding tools, with Guido noting the potential three trillion dollar value unlocked by doubling developer productivity. Host Steph Smith does not join the active transcript dialogue, keeping host scores at zero. | |
| How AI Has Already Changed Software Development Workflows | 0 | 2 | 1 | 0 | Guido explains how developer workflows shifted over nine months from simple copy-paste prompting to using models as sparring partners for drafting detailed technical specs. The panel maintains an agreeable, shared-experience discussion without host participation. | |
| Vibe Coding vs. Serious Engineering and Tooling Limits | 0 | 2 | 2 | 0 | Yoko and Guido explore the limits of AI coding assistants when applied to novel problems or distributed systems, noting how models confidently invent non-existent API functions. The conversation remains conversational and peer-driven. | |
| Vibe Coding and the Future of Computer Science Education | 0 | 2 | 2 | 0 | The guests debate whether computer science education will still require learning low-level abstractions like assembly and processor gates or shift toward higher-level problem definition. The interaction is a balanced exchange of opinions among peers. | |
| How Vibe Coding Affects Programming Languages and Legacy Systems | 0 | 3 | 3 | 0 | Matt challenges the assumption that AI can easily transpile legacy COBOL codebases due to lost institutional context, and Guido extends this by describing how enterprises create specifications first before rewriting legacy code. The technical debate among guests is constructive. | |
| Adjusting to Non-Deterministic Outputs in AI Systems | 0 | 3 | 2 | 0 | Guido explains how AI introduces chaotic non-determinism into software architecture and details how banks manage risk by benchmarking model compliance against human error rates rather than seeking absolute determinism. | |
| Are Prompts the Narrow Waist of AI Programming? | 0 | 2 | 2 | 0 | Yoko challenges Guido's framing of prompts as the narrow waist of AI, pointing out that natural language prompts lack formal specifications compared to classical network protocol standards. | |
| The Nexus of Vibe Coding and Enterprise Programming | 0 | 1 | 1 | 0 | Yoko and Guido define vibe coding in contrast to classical engineering, concluding that enterprise software developers will increasingly leverage vibe coding for high-level outcomes. |
Statements from this episode (16)
Appenzeller: LLMs as direct compiler inputs will transform software engineering
“If I look at a classic, say, compiler design or, you know, in, in, in, in programming languages, if I would have LLMs as a tool, I would probably think very differently about how I would build a compiler. And I don't think we've seen that work its way through …”
Bornstein: Coding is the second-largest AI market behind consumer chatbots
“We're pretty sure it's the second biggest AI market right now. Correct me, you guys, if I'm wrong, but, you know, consumer pure chatbot, I think, is number one, and I think coding is number two, just purely, purely looking at the numbers.”
Bornstein: Cursor is outperforming GitHub Copilot in AI coding adoption
“And I think companies like Cursor have just done a much better job with that now.”
Li: Coding excels in AI because software code is verifiable
“But I also think coding market is doing so well also because it's somewhat verifiable problem. Like you can verify like a coding function, like input and output are very clear compared to like user preference, you know and all the other problems.”
Appenzeller: Basic copilots boost enterprise developer productivity by 15 percent
“I think we, if I look at the data we've seen from some of the large financial institutions, they're estimating that the increase in developer productivity from just a vanilla copilot deployment is something like 15%.”
Li: AI coding agents still assume the current year is 2023
“When you ask the coding agent, when do you think today is? It's always like, 20, 23. And then all the specs that we'll give you are from, like, 20, 24 at best, depending on when the knowledge cutoff is.”
Bornstein: AI coding excels at front-end tasks but struggles with niche libraries
“And I've always found it works really well for high complexity, high kind of annoyance factor things like front end. Like if anybody on earth can remember All of the CSS classes that people use now for margins and padding, it's like, it's, you know, I don't th…”
Appenzeller: Novel coding lacking AI training data is only 0.01% of development
“I think the good news is that is 0.01% of all software development, right, for the, I don't know, you know, 100,000 ERP system implementation or so, right, that we have tons of training data, and I think these tools can be very, very powerful.”
Appenzeller: System architecture will supersede low-level coding as AI advances
“Explaining the problem statement, explain the algorithmic foundations, explaining architecture, and explaining data flows getting more important, and the nitty gritty coding, you know, what's the most clever way to unrule a for loop That's a very specialized, …”
Appenzeller: Formal programming languages will not be replaced by AI
“I think formal languages won't go away because ultimately they seem complicated, but I think effectively a formal language is often the simplest type representation you can find to specify intent, right?”
Bornstein: AI cannot solve legacy code migration due to lost context
“AI may be able to transpile, you know, COBOL to Java, but there's a huge amount of context in what went into creating that COBOL code that's been totally lost, right?”
Appenzeller: Creating specs first yields best results in AI legacy code migrations
“The most efficient way for them is to actually go first and try to create a spec, use the AI to create a spec from that code, right? And once they have the spec, then to re-implement the spec.”
Appenzeller: Guaranteeing LLMs never produce forbidden output is an unsolvable problem
“You're trying to have an LM that is very helpful and never even implicitly gives investment advice. That's sort of an unsolvable problem, right? You can get better and better and better, but you can never completely rule it out. And you can add a second LM tha…”
Appenzeller: Natural language prompts are the narrow waist abstraction of AI
“I think we have the waste, the narrow waste. I think it's the prompt.”
Appenzeller: AI output generation layers will delaminate from reasoning models
“Long-term, I actually wonder if the, like for reasoning model, where a lot of the thinking sort of happens internally, if the model is generating the user facing machine facing output, it's going to be a different model from the model doing the reasoning, if t…”