Guy Podjarny: The Future of AI Software Development - What is Real & What is BS | E1232

Intro (0s)

• The idea that a SaaS business is only about the software it creates is not true, as these businesses are more than just their software, and having only this as a differentiation is not sustainable (10s).
• The expectation of when GP5 scale models will arrive is now thought to be further away, with some experts like Sam Altman changing their focus to reasoning and other areas as progress is made (15s).
• There are architectural problems that need to be addressed, which may cause a delay of a year or two before these models are ready to be implemented (27s).
• The conversation is expected to be informative and exciting, with the guest having shared interesting insights in the past (42s).
• The guest is happy to be back on the show and is looking forward to the conversation, which is expected to be engaging and possibly "spicy" (47s).

On NVIDIA’s Market Position (57s)

• Massa Sun stated that NVIDIA is undervalued today, which raises three questions: whether the market NVIDIA is in will continue to grow, how much of that market NVIDIA will capture, and whether the company's current 35x multiple on revenues is justified (57s).
• The market for semiconductors in AI is expected to continue growing, with multiple players entering the space, but NVIDIA is likely to dominate due to its substantial lead and strategic moves (1m11s).
• NVIDIA is taking advantage of its lead by building a cloud and leveraging its semiconductor advantage to maintain its market position, making it likely to remain the market leader by a significant margin for a long time (1m35s).
• The question of whether NVIDIA's 35x multiple on revenues is justified is more complex, as it depends on whether investors should put their money into other stocks that may grow faster (1m59s).

Will We See a Trough of Disillusionment in AI (2m12s)

• There is a possibility that the market will experience a trough of disillusionment with AI, where companies realize that the return on investment (ROI) for many AI tools has not been proven and this may lead to a reduction in demand for NVIDIA chips in the next year (2m22s).
• Despite this, NVIDIA's revenue is likely to be guaranteed due to existing commitments, and the company's core technology continues to evolve and improve, making it well-positioned to produce cost-effective solutions (2m43s).
• NVIDIA's advantages, including manufacturing, IP, processes, and the CUDA development environment, are durable and will help the company maintain its position (3m13s).
• A trough of disillusionment is expected in the enterprise sector, where companies may feel that AI has not delivered the expected value, but this is not because AI is less promising, but rather because the numbers and timing are unrealistic (3m27s).
• Many AI budgets are currently non-resilient, with companies spending money and expecting quick returns, but it is unlikely that they will adapt quickly enough to see significant returns in the short term (3m46s).
• While there will be some successful companies, many will struggle, and the biggest issue is the large number of startups doing the same thing, leading to redundant investment and waste (4m16s).
• The market is saturated with companies building similar products, such as note-taking apps for specific industries, which will likely lead to a significant amount of waste and unsuccessful companies (4m28s).

Is AGI Worth the $9 Trillion Investment? (4m40s)

• The cumulative cost to achieve Artificial General Intelligence (AGI) is estimated to be $9 trillion in capital expenditures, but the benefit would be a shift in GDP to $9 trillion per year (4m46s).
• The investment required to achieve AGI would be substantial, but people often underestimate the non-technology parts of adopting AI, such as changes in accountability, insurance, and law (4m58s).
• AI is considered a transformative or disruptive technology, not a sustaining one, and its adoption requires significant societal changes, which can cause delays (5m2s).
• The main holdback for adopting AI in various fields, such as law and self-driving cars, is not the technology itself, but rather societal factors like law and insurance (5m26s).
• The time horizon for achieving AGI is difficult to assess, and estimates vary widely, with some predicting it will be achieved soon, while others believe it is much further away (5m44s).
• The definition of AGI is poorly defined, making it easy for people to claim that it is coming soon, and those who need to continuously raise money, like Sam Altman, tend to have a more optimistic timeline (6m10s).
• In contrast, those who do not need to raise money, like Zach and Demis, tend to have a more pessimistic timeline, highlighting the role of incentives in shaping people's views on the timeline for achieving AGI (6m32s).
• Incentives can sometimes lead to biased views of the world, as people tend to believe what they want to believe, and this can influence their predictions and estimates for achieving AGI (6m47s).

Is $100 Billion Necessary to Compete in Frontier AI? (6m51s)

• To compete in the frontier model race, vast amounts of capital are necessary, and it is unlikely that the cost to enter this race will be democratized (7m1s).
• There are two competing theories: one is that scaling laws will continue to show themselves, requiring more money to make progress, and the other is that specialized models will benefit from not trying to scale as much as the big generic models (7m15s).
• The big generic models, such as GPT 5 and equivalent models in Gemini and Anthropic, are rumored to be failing in some tests, which could give smaller companies building specialized models an opportunity to compete (7m44s).
• Specialized models, such as code-specific or robotic-specific models, may have a shot at being able to train with relatively small amounts of money, but this advantage is likely to be short-lived (7m58s).
• In the short term, specialized models may be more efficient and accurate, but in the long term, more generalized models are likely to dominate (8m18s).
• The big generic models are apparently running into some limitations, which could give specialized models an opportunity to prove themselves in the next three-year timeline (8m36s).
• Building one's own model is not a viable option, as it is better to tap into the innovation of the market and use the best technology available, rather than trying to compete with companies like Open AI and Anthropic (8m44s).
• Different models, such as Anthropic and Open AI, have different strengths, such as reasoning and code generation, and it is better to be able to use the best technology available rather than trying to pick one (8m49s).
• The goal is to be the best user of the available technology, rather than trying to compete with the companies that are developing it (9m12s).

Is Benioff Right to Criticize Copilot? (9m19s)

• There are numerous AI Dev tools available, including note-takers, and Marc Benioff has been critical of GitHub Copilot, stating it doesn't work and lacks accuracy, while Gartner notes it spills data everywhere, causing customers to clean up the mess (9m32s).
• Despite criticism, GitHub Copilot and other coding assistants provide value, with many developers relying on them for software development and not wanting to go back to not using them (9m53s).
• However, these tools can produce software of questionable quality, as developers are only reviewing generated code rather than writing it themselves, resulting in less thought and consideration being put into the code (10m5s).
• Large Language Models (LLMs) tend to average everything, generating average code unless given very specific instructions, which can be a concern as it duplicates a lot of average code (10m27s).
• Coding assistants like GitHub Copilot and Kite's Cursor are popular because they don't require significant changes to a developer's workflow, providing a "magic" solution that works often enough and doesn't require trusting the result (10m48s).
• Cursor, in particular, has made it easy for developers to use AI assistance by making changes in multiple places in the code and allowing for easy review and verification of the changes (11m3s).
• The ease of use and integration of these tools into a developer's workflow makes them appealing, as they can be used while coding without worrying about incorrect results (11m20s).

Are AI Dev Tools Actually Any Good? (11m23s)

• AI Dev tools have proliferated, but their effectiveness is difficult to gauge, and they provide value in very specific areas, such as reducing toil, which is the notion of repeated work, like creating documentation and tests (11m24s).
• These tools help by providing templates, making it easier to start tasks, such as code completion, which has been helpful due to the low cost of verification (11m34s).
• However, AI Dev tools have not been dramatically helpful beyond these areas, mainly because they are still unreliable and unpredictable, often making mistakes that humans would not make (12m12s).
• There are two types of mistakes that might happen with AI Dev tools: those that humans might make, which are more forgivable, and those that AI will make that humans will never make, which can be frustrating (12m37s).
• The unreliability of AI Dev tools makes it difficult to rely on them, and they are not yet capable of replicating complex tasks or businesses, such as SAS companies (12m30s).
• SAS businesses are more than just the software they create, involving data, distribution, switching costs, and customer relationships, making it challenging to replicate them using AI Dev tools alone (13m28s).
• The idea that AI Dev tools can replicate a SAS business by simply generating code is considered unrealistic and overlooks the complexities of building a successful business (13m58s).

What Is Agentic Development vs. AI Dev Tools? (14m5s)

• Agentic development is a type of AI software development where the machine is given a high-level task and is trusted to complete it, including product research, exploration, and application building, with minimal human intervention (14m6s).
• This approach differs from most AI Dev tools today, which provide more control to the software developer and assume their involvement in the development process (15m14s).
• Companies like Cognition and Magic Dev are trying to build agentic systems that can handle the entire development process, but they are still secretive about their approaches (14m58s).
• There is a gap between the current AI Dev tools and agentic development, with the former providing more control and the latter relying on the machine to make decisions (15m28s).
• To bridge this gap, a new software development methodology may be needed, which would involve working together with the machine to define what is being created (15m45s).
• Agentic development can be seen as embracing the customer perspective and chaos, but it has also faced skepticism, particularly towards Cognition's abilities (16m3s).
• The skepticism may be due to Cognition's initial video, which overplayed their hand and created unrealistic expectations (16m15s).
• Despite the skepticism, many people who have tried Dev think it's cool, but it doesn't work reliably, highlighting the gap between demos and working products (16m42s).
• Companies like Magic, Cognition, and Tassel have raised significant amounts of money, which is partly used for training models and GPUs, but also raises questions about peak enthusiasm from venture crowds (16m51s).
• The cost of training models and running evaluations is high, but having good reserves is necessary for "go big or go home" type propositions (17m34s).
• Raising a lot of money can reduce optionality, but it's not an issue for everyone, and some companies are not looking for a quick exit (17m59s).

Open vs. Closed: The Future of Software Development (18m8s)

• The future of software development may be dominated by closed environments, closed development platforms, and closed ecosystems, which could lead to a lack of control and customization for developers, with the web potentially becoming controlled by a few large companies with powerful compute models (18m24s).
• Open AI, for example, is taking a platform approach, allowing multiple participants to build on top of their foundation models, and it is unclear how deep they will go into the application layer (19m6s).
• The next layer of big companies, such as Cognition, are still in the beginning stages of building their platforms, but if their model succeeds, it could lead to a concentration of power in the hands of a few players (19m43s).
• The core of software creation may become dependent on a single magical understanding of code, applications, and domains, making the rest of the tooling ecosystem minor and delegated (20m8s).
• The concentration of platforms being the sole providers or few providers to dominate the space is considered a high probability, with ease of use of software development already being amazing and platforms like Vercel making it easy to generate applications (20m50s).
• The more developers lean into the agentic, magical creation of software, the less control they will have, and it is likely that the rest of the dev tooling ecosystem will become minor and delegated (21m23s).

When Will Enterprises Move Beyond Experimental Budgets? (21m29s)

• The adoption of AI software development is expected to be incremental, with some assistants already providing value and moving out of the experimental budget phase into real budget, while others, such as resolution and SDR ones, are still in specific fields and mostly experimental (21m30s).
• The limiting factor for the adoption of AI software development is not the Large Language Models (LLMs) themselves, but rather the processes and systems around them to make the LLMs more reliable (22m21s).
• AI models are expected to go all the way in terms of entrance into the application layer as long as they are dealing with the same data and just a different way to interpret it, with search being a good example of this (22m46s).
• AI models are likely to replace web search because it's the same type of data they need, and they don't need a different type of expertise, and may also fall under the same bucket as user experience of code generation or image generation (22m52s).
• However, AI models may not be suitable for domains that require different data or very dedicated elaborate workflows, such as processing and whole system around how to process those (23m17s).
• When choosing between CH, Open AI, Anthropics, or Xod AI, the decision would be to go with Anthropics due to the quality of the team, stability of the team, and their opportunity to grow, as well as the value arbitrage (23m58s).
• The safety and alignment team moving out of Open AI may be a concern, as it could be suffering substantially from the churn in leadership, which could rock the boat (24m15s).
• In the search game, Perplexity may not go as far as people think, and Open AI may have an advantage, but it's not the main reason, and the outcome is uncertain (24m46s).
• Google's search product is deeply ingrained and has a strong distribution, making it difficult for other companies to compete, despite Google being generally terrible at product development on many fronts (24m58s).
• Perplexity is considered a better product than Google's search, with some users, including the speaker's mother, switching to Perplexity (25m40s).
• People tend to change their habits slowly, and it's underestimated how hard it is to change people's habits, especially when it comes to something as ingrained as Google's search product (25m5s).
• Google has the technical capabilities to compete with Perplexity, but lacks the product chops, and would likely need to copy Perplexity's features rather than inventing them (26m19s).
• Google's core technology is still good, and the company has the advantage of incredible distribution and access to users, making it a significant competitor in the search market (26m47s).
• The Trump Administration may have an impact on IP and M&A, with top companies affiliated with liberal or anti-Trump views potentially facing more challenges, while companies just below them may find it easier to acquire other companies (26m57s).

Why Would Companies Like Snyk Choose to Go Public? (27m15s)

• Companies like Snyk may choose to go public despite the availability of private capital and private markets, as going public is often necessary for building a long-term sustainable company (27m16s).
• Going public provides benefits such as brand recognition, constant liquidity for employees, and assurance for enterprise customers through financial transparency (27m31s).
• The decision of when to go public involves weighing the benefits against the heavy toll of being a public company (27m50s).
• The ideal time for Snyk to go public is a topic of ongoing conversation, with the goal of eventually becoming a public company (27m57s).
• Handing over the CEO role can be challenging, and finding one's place in the company years later can be difficult, as experienced by the individual who stepped down as CEO of Snyk (28m11s).
• Peter, the current CEO of Snyk, is considered incredible, and the transition has had positive implications for the company (28m17s).

How Will the Role of Software Developers Evolve? (28m28s)

• The role of software developers will evolve as the structure of teams and product creation changes, with a focus on systems thinking and understanding the requirements and tradeoffs of development (28m34s).
• The best software developers are not necessarily the best coders, but rather those who think about development as a whole, understand the important bits, and can anticipate tradeoffs and future implications (28m48s).
• The coding piece of software developers' work will diminish substantially in the next 10 years, becoming an edge case for tasks that require being close to the bare metal or working with older technology (29m20s).
• Most developers will progress up the architect path, investing in tradeoffs and system thinking, and making decisions about software and systems in a broader context (29m38s).
• Architects will think about software and systems in bigger paths, making decisions about tradeoffs such as extensibility vs. simplicity, and optimizing for specific environments or portability (29m49s).
• Architectural decisions have implications downstream, and will continue to be important as they require human assessment and understanding of what is likely to change in the future (30m16s).

How Many Devs Should Move to Architectural Thinking? (30m29s)

• The challenge lies in transitioning developers to more architectural thinking and strategic decision-making processes, which seems to be a progression towards leadership-style thinking, but there are only so many big decisions or leadership positions available (30m29s).
• Not all developers need to progress up the architect path, but rather, some can take a more product-focused route, understanding user empathy and product management, similar to the role of a product manager (31m9s).
• With the increasing ease of building software, consumer expectations have changed, and websites are now expected to have high levels of functionality and performance, leading to increased demands on developers and product managers (31m56s).
• The role of product managers (PMS) will change, becoming more autonomous, with some blurriness between the product manager and software developer roles, especially for technical PMS (32m34s).
• Some PMS roles will merge with software development, while others will remain focused on understanding users and making decisions, which will continue to require human accountability (32m58s).
• In the future, the structure of a technology company may not change dramatically, but the roles and scope of different positions will evolve, with a continued need for decision-making power, strategy, and system management (33m51s).
• The development teams will be able to produce more software, which will be more adaptable, personalized, and secure, with a greater emphasis on management and teamwork (34m25s).

Why Is Security Becoming Harder with AI Dev Tools? (34m54s)

• Security is becoming a more significant problem with the increasing use of AI Dev tools, particularly with Large Language Models (LLMs), as they make it harder to control and review the produced code, leading to potential security risks (35m7s).
• The lack of control and review in AI-generated code results in a significant amount of unreviewed or poorly reviewed code being deployed, which poses a real security risk (35m28s).
• Another issue is that people produce software using AI coding systems, but they often do not maintain the code, leading to outdated and vulnerable code that can be exploited by attackers (35m38s).
• The problem of unmaintained code is exacerbated by the fact that technology never dies, and old code can remain an entry point for attackers, making systems more prone to attacks over time (35m44s).
• The main challenge is the "messiness" of AI-generated code, which requires figuring out layers of control to ensure that software is created, maintained, and owned responsibly, with guardrails in place to evaluate its quality (36m0s).
• To address these issues, it is essential to think about maintenance, ownership, and control alongside the creation of software, ensuring that the produced code is reliable and secure (36m8s).

Spicy Questions (36m21s)

• A new round of questions is introduced, with the option to either answer the question or donate $1,000 to a charity of choice per question (36m23s).
• The first question is about how much liquid was taken out from Sneak, and the response is that about a third of ownership was sold, resulting in a nine-figure amount (37m0s).
• Acquisition offers for Sneak were discussed, with early offers being in the $200 million range, but later talks involved multiple billions, although no concrete numbers were reached (37m48s).
• The worst investor meeting experience was with Ser Investment Firm in the UK, where two partners were on their phones and seemed uninterested, which was insulting given the high demand for investment at the time (38m30s).

Quick-Fire Round (39m21s)

• The expectation of when GP5 scale models would arrive has been pushed back, with some architectural problems potentially causing a year or two of stagnation in their development (39m36s).
• There is a lack of focus on long-term, high-impact AI projects, with most people instead focusing on short-term gains and optimizing existing workflows (40m1s).
• Founders should look for ideas that are not immediately obvious or widely accepted, as these can provide a unique value proposition and reduce competition (40m26s).
• Competitive markets can be challenging, and it's essential to consider the number of players in the market, with a market of five or 10 companies being more desirable than a market of 50 or 100 (41m18s).
• When investing in a crowded market, it's crucial to prioritize operational excellence and find the best company builders, product teams, and go-to-market (GTM) builders (41m54s).
• With high interest rates, storing cash can be a strategic decision, and in the case of Tessle, the company is taking a conservative approach and earning high interest rates without taking on excessive risk (42m17s).
• The interest earned on a large cash reserve, such as Tessle's $125 million, can be substantial, potentially exceeding $10 million (42m31s).
• Raising a large amount of money allows a company to build for the long run, invest when things are working, and have good partners, providing insulation from market craziness and potential busts of evaluations or realizations (42m36s).
• Urgency in a company comes from market timing, not from runway, and having more runway gives flexibility, but losing urgency can cause a company to lose the opportunity (43m16s).
• Two primary mistakes of raising too much money too early are spending it too quickly and not spending it fast enough (43m40s).
• Spending money too quickly can lead to fake product-market fit or making it harder to turn the company around when needed (43m51s).
• Not spending money fast enough can cause a company to lose the pressure of the runway and the forcing function to execute on an idea (44m19s).
• Market timing is critical, and having more runway gives flexibility to execute on an idea within a window of time (44m31s).
• Having a large amount of money can cause a company to lose urgency and think they have more time to execute, leading to losing the opportunity (44m55s).
• Cloudinary was a successful angel investment, going from effectively zero to about $4 million worth (45m6s).
• Security Scorecard was another good investment, made at a $6 million valuation (45m15s).
• Money does not make one happy, but rather provides a mindset and freedom to focus on donating and passing it on (45m25s).
• Donating large amounts of money requires effort and investment (45m46s).
• The key to a happy marriage is communication, where both partners should discuss their issues and not let things stew, even if it means talking about problems before going to sleep (46m3s).
• A recent company product strategy that has been impressive is Intercom's bet on autonomous agents, which was a bold and correct move that required conviction and good execution (46m43s).
• Another impressive company is Lightdash, which is building the future of BI tools by allowing users to embed reporting platforms into their products, a solution that is needed by many SaaS companies but is not their core focus (47m4s).
• Lightdash's open-source BI solution has been successful, with its dashboards being sold like hotcakes, and its founder Hamza has been praised for identifying and tapping into this opportunity (47m42s).