If AI Can Write Code, What Should Software Engineers Do? | Bolt Picks

"Why is Cursor so popular?"

"Are AI dev tools actually useful?"

"If AI can write code for software engineers, what should engineers even do?"

These have been hot topics ever since AI coding tools emerged and entered the industry's field of vision.

Recently, the podcast 20VC interviewed Guy Podjarny, former CTO at cloud computing company Akamai (following its acquisition of his startup Blaze), founder of AI coding startup Tessl, and active angel investor. In this episode, Podjarny and the host started with an analysis of NVIDIA's market trajectory, moved on to the popularity of AI coding tools, and focused on AI's role in software development — what it means for developers and how it affects their career paths. We've selected and translated portions of the interview; you can listen to the original episode via the "Read More" link.

Image | Podcast Shownotes

📝 Summary

1. NVIDIA's Market Outlook: NVIDIA's position in the AI semiconductor market is expected to keep growing. The company will continue to dominate, with its technical advantages and strategic positioning keeping it ahead.

2. AI's Trough: Despite continuous progress in AI technology, many AI tools have yet to deliver expected returns on investment, suggesting the market may experience a downturn.

3. Cost vs. Benefit of Achieving AGI: Artificial general intelligence requires massive investment, but its potential economic impact is enormous. Societal acceptance of AI will take time; both technical and social factors must be considered in investment decisions.

4. The Future of AI and Specialized Models: While more efficient and accurate specialized models may emerge in the near term, capital dynamics will drive the market toward more general-purpose models over the long run.

5. Challenges and Opportunities in AI Dev Tools: AI development tools deliver value by reducing repetitive work and assisting with code completion, but their instability limits their utility.

6. Concerns About Closed vs. Open Development Platforms: Closed development platforms and ecosystems could create problems for future software development, limiting developers' control and capacity for innovation.

7. The Future Role of Software Developers: As AI assistants and code completion platforms proliferate, software developers may shift toward roles more akin to architects or product managers, adapting to increased software output and evolving consumer expectations.

Part.01

NVIDIA's Market Outlook

1) Harry Stebbings: Masayoshi Son says NVIDIA is currently undervalued. Do you agree?

Guy Podjarny: This breaks down into three parts. First, will NVIDIA's market continue to grow? I think yes — the AI semiconductor market will definitely keep expanding. More players will enter, and overall demand will keep increasing.

Second, what market share will NVIDIA capture? I think they'll continue to dominate. I'm not sure about the exact share, but they'll maintain a substantial lead. NVIDIA is doing some very smart things. Their current technical advantage is massive, and they're strategically leveraging it. Cloud providers have distribution advantages, so NVIDIA is building its own cloud infrastructure and using its semiconductor strength to stay ahead. So I believe they'll maintain market dominance for the long term.

Third, there's the current valuation. This isn't about whether NVIDIA will grow — it's whether that growth can support the current stock price, and whether it will outperform other investments. NVIDIA's growth is almost certain, but whether it's worth investing at today's price is a more complex question.

Part.02

Will We See an AI Trough?

2) Harry Stebbings: I have a market question. Many people say we'll experience a trough — companies will find that AI returns aren't proven on the first batch of AI tools. There could effectively be an AI "bust" that leads to reduced demand for NVIDIA chips next year. What do you think?

Guy Podjarny: First, NVIDIA already has a certain number of orders on the books, so revenue is largely guaranteed. I also think their core technology keeps improving and evolving, getting better and better. So they'll likely be the company that can produce most cheaply, or save costs through minimal compute or other means. It's not just about handling the biggest, most complex models — though that's part of it. So manufacturing, IP, the processes to meet these challenges, CUDA, the development environment — all of these are NVIDIA's durable advantages.

3) Harry Stebbings: Do you think we'll experience a trough? Will enterprises conclude these tools haven't delivered the promised value?

Guy Podjarny: I think there will be a phase like this, but not because AI is less promising — it's because the sheer number of AI tools right now is somewhat insane.

4) Harry Stebbings: Is it too many tools, or a timing issue?

Guy Podjarny: I think a lot of AI budgets currently lack resilience. People are spending money, experimenting with various solutions, and holding them to very high expectations. In the long run, those expectations are correct, but it's hard to imagine people adapting quickly to the process changes AI brings, or enterprises adapting fast enough to see returns within a certain timeframe. So there will definitely be some exceptional winners, but I think a lot of things will end up marginalized. The biggest problem is these small startups — sometimes they're all doing the same thing.

5) Harry Stebbings: Masayoshi Son also said achieving AGI will require $9 trillion in cumulative capex, but the benefit will be $9 trillion in GDP growth annually. Your take?

Guy Podjarny: The investment will definitely be massive, but I think people underestimate the non-technical aspects of AI adoption. I believe AI is a transformative, disruptive technology that requires responsible change-making, so society will need considerable time to absorb it.

For example, resistance to AI lawyers won't come from technology — it'll come from legal or insurance concerns. If that lawyer gives you wrong advice, who do you sue? The same applies to self-driving cars.

Technology is part of it, but the bigger issue is social. So I think this investment is necessary, but it will also take time. The overall amount is probably in the right ballpark, though it's hard to assess: is it $9 trillion, $15 trillion, or maybe $3 trillion would suffice? These are all enormous numbers. I think the end result is absolutely worth it; the only hard thing to assess is the timeline. The issue isn't technological progress.

Part.03

The Future of AI and Specialized Models

6) Harry Stebbings: Larry Ellison (Oracle co-founder) has said entering frontier model competition requires $100 billion. Do you agree? Or do you think costs to enter are decreasing?

Guy Podjarny: I don't think costs in the model space are decreasing. Generally, I agree: "To be competitive in foundation models, you need massive capital."

There are currently two different theories. One is that Scaling Laws will continue to hold — you need to invest more and more to make progress, and general-purpose models will keep expanding and eating competitors' lunch. The other view is that specialized models will actually benefit people. If some big companies trying to continue with ultra-large-scale models fail — recent rumors suggest GPT-5, Gemini, and Anthropic haven't achieved expected results on some tests — if that's true, and they've already invested so much money and time in ultra-large-scale models, then companies focused on building specific code or specific robotics models, training with less capital, might actually have a good opportunity. But I still think that opportunity is temporary.

Over a 5- or 10-year horizon, in the short term we'll see more efficient and accurate specialized models. But in the long term, capital still plays an enormous role, and the market will ultimately move toward more general-purpose models.

Part.04

Challenges and Opportunities in AI Dev Tools

7) Harry Stebbings: Why does everyone love Cursor?

Guy Podjarny: When you think about AI solutions, from a broader framework, what's easiest to adopt right now are solutions that don't require you to change how you work at all. They provide some "magic" — you don't need to fully trust the results, just that they work often enough. So coding assistants, and Cursor specifically, are genuinely very good.

Cursor goes a step further: it can make changes across many parts of code, multiple files, and makes those changes very easy to verify — you can tell at a glance whether they're correct. So you don't worry about incorrect output, and it integrates into your workflow; you just keep coding.

8) Harry Stebbings: Are AI dev tools actually useful?

Guy Podjarny: I think they do provide value in certain specific areas. Their main contribution is reducing repetitive work. By repetitive work, I mean the long stretches of adding comments, writing documentation, and writing tests. AI tools can work like email templates — you don't have to start completely from scratch.

I think code completion, specifically, helps because verification costs are low. Beyond that, they're not yet amazing, mainly because they remain unstable. Sometimes they work great; the next moment they completely fall apart, making them hard to fully rely on.

And this loss of control is hard to predict. Their errors fall into roughly two categories. One is errors humans might make — AI sometimes makes these, maybe never makes them, but they're within the realm of human acceptability. The other is errors AI makes that humans would never make, which is incredibly frustrating. For example, if there's a leaf in the road and the car thinks it's some creature and won't drive around it, people get upset and think it's stupid.

Part.05

Closed Platforms vs. Open Platforms

9) Harry Stebbings: We've talked about Magic, Cognition, and many others. These platforms operate in closed environments. How do you think about open vs. closed in future software development?

Guy Podjarny: I'm actually quite concerned about this. Closed environments, closed development platforms, closed ecosystems — like large platforms. When they add AI capabilities, it's relatively easy for things to be adopted and used. Eventually they create these "black boxes" you can't interact with or understand how they work. I find this worrying. If the future web becomes two or three companies controlling these very powerful compute models, where users just give instructions and get results, but have no ability to build tools that connect to these models or modify certain functions — that would be a very unfavorable situation.

10) Harry Stebbings: Why would this happen? Take OpenAI — they're clearly pursuing a platform strategy, not trying to build all tools around the platform themselves. They want to develop an ecosystem, attract multiple participants, and serve different customer segments.

Guy Podjarny: I think LLM foundation models themselves do want to be platforms, because they want everyone building new things on top of them. So I'm not sure how deep they'll go into the application layer. OpenAI is already entering the application layer now; the question is how deep they'll go.

What big companies should be more concerned about is companies like Cognition — they're not in the application layer yet, still early stage. But if this model works, where you just give instructions and it executes successfully, and they're building great things — what happens next? You're building these models, building these platforms, giving them instructions, letting them execute. How does the rest of the dev tools ecosystem plug in? If the core of software creation becomes dependent on this "magic" system that understands your code, your application, all your domains, and you're merely a customer of that system, I worry we'll enter a situation where only a small set of players have broad capabilities, and the remaining dev tools ecosystem becomes secondary. Many thriving dev tools companies that exist today will no longer exist.

11) Harry Stebbings: How likely are these platforms to become the only or few dominant players?

Guy Podjarny: I think the probability we're heading down this path is quite high. The ease of software development is already remarkable. You can see platforms like Vercel where you can very easily generate applications and complete everything in one place. Now GitHub is getting into application deployment too — from code editor to deployment, and they're also writing a lot of code. I think the probability of moving in this direction is very high. And the more you depend on this "agentic magic creation," the less control developers have.

---

The Future Role of Software Developers

12) Harry Stebbings: With assistants and code completion platforms emerging, how will the software developer's role change?

Guy Podjarny: I think the best software developers aren't best because they write the best code, but because they view development as a whole. They're systems thinkers who understand what matters in requirements and optimize around those things. They can foresee the trade-offs and downstream effects of different choices. In fact, many developers' favorite career path is the architect path. Architects don't write much code.

I think coding will still exist in 10 years, but as edge cases — only when you need to get close to底层 hardware or deal with some legacy technology. Most developers will either move toward architecture, or invest their energy in these trade-offs and systems thinking.

13) Harry Stebbings: For those who don't know, what does "moving up to the architect path" actually mean?

Guy Podjarny: It means they think about software and systems from a more macro perspective, making important decisions. Whenever you build a system, you always need to make trade-offs. For example, how do you balance scalability versus simplicity? The more scalable a system, the more complex it becomes. This is something you must consider — is it important? Is it necessary? You can optimize your system for a specific environment (like AWS), or choose to make it compatible across different cloud platforms. These are architectural decisions that affect how the system evolves. I think these decisions will continue to matter, because neither large language models nor humans can fully foresee all problems. What they can do is judge, based on current conditions, which changes are more likely to occur.

14) Harry Stebbings: Will all developers move toward architecture? Because maybe you don't want 20 developers all telling you their architectural opinions.

Guy Podjarny: I think so. Because software development speed is increasing dramatically, there will be more software, so more decisions need to be made, and more urgently.

Besides the architecture path, there's a more product-oriented path — greater emphasis on user empathy. Like Henry Ford said, "If I had asked people what they wanted, they would have said faster horses." This way of thinking is closer to the product manager role. So developers can choose one of these two paths. As software output increases, developer roles will become more diverse.

In the future we'll produce more software than ever before. Look at websites in 2000 — they were boxy, cluttered messes. Making a nice website, something we'd consider normal today, required enormous time, money, and technical expertise. Now, building a beautiful website is extremely easy. So consumer expectations have changed. We now have much higher requirements for website functionality and performance, and much lower tolerance for latency. System improvements let us create software more efficiently, which gives developers greater leverage to produce high-quality software — and consumer and enterprise expectations rise accordingly.

---

Linear Bolt is Linear Capital's dedicated investment program for early-stage, global-market-facing AI applications. It upholds Linear's investment philosophy, focusing on technology-driven transformative projects, and aims to help founders find the shortest path to their goals — whether in speed of action or investment approach, Bolt's commitment is lighter, faster, and more flexible. In the first half of 2024, Bolt invested in seven AI application projects including Final Round, Xinguang, Cathoven, Xbuddy, and Midreal.