20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

AI's Leverage Will Force Businesses to Solve More Problems and Create 'Load-Bearing' 100x Engineers

Companies gaining AI leverage face a strategic choice: either tackle more problems with increased ambition or solve existing problems more efficiently with smaller teams. This decision isn't obvious, as it depends on market dynamics and strategic goals, but it inherently means re-evaluating the scope of their work.

The idea of a "100x engineer" isn't about the quantity of code produced, as AI can generate vast amounts of code, which might be low quality. Instead, it refers to "load-bearing individuals" within an organization. These are people whose removal would cause critical systems or projects to collapse, signifying their indispensable impact.

AI tools will amplify the capabilities of these already impactful individuals. Those who effectively utilize AI's leverage will achieve disproportionately greater results, widening the gap between them and those who don't. This effectively makes the most leveraged individuals even more valuable to businesses.

It's kind of like if you remove this person, things fall.

C-suites are shifting resource allocation to focus on core business competencies and output metrics.

Leaders must now contend with the complex resource allocation problem involving not just dollars and people, but also AI tokens. This challenge is expected to dominate C-suite discussions over the next two years, necessitating a strategic reevaluation of how resources are deployed within an organization.

The recommended approach is to identify the business's fundamental core competency—what truly matters for its operations—and then align all resource allocation accordingly. This means moving away from a focus on intermediate metrics, such as the number of features software engineers ship, which may not directly contribute to the core business objectives.

Instead, the emphasis should shift to tangible output metrics that genuinely impact the business, such as customer satisfaction, revenue, or market share. Resources, whether monetary or token-based, should be strategically allocated to dramatically move the needle on these critical business outcomes, rather than on internal, less impactful measures.

This shift represents a positive change, moving organizations away from the bloat that often resulted from an overemphasis on intermediate metrics. By tying every individual's work, from marketing to engineering, back to core business metrics, companies can achieve greater efficiency and effectiveness.

This resource allocation problem of token, it's not just tokens, it's like dollars, it's tokens, it's people. This is, I think, going to be the thing that over the next twenty-four months, every C-suite is going to be thinking about.

Kirkland's $500M AI Bet Raises Questions on Core Competency and Outsourcing

Matan Grinberg expressed surprise at Kirkland's $500 million investment over five years to build internal AI tools, like their own versions of Harvey or Logora. He questions this significant spend, noting that developing AI technology is not typically a core competency for a law firm.

Grinberg suggests that attempting to build complex AI solutions in-house might highlight the difficulty of such endeavors, ultimately leading firms to recognize the value of outsourcing to specialized experts. He uses an analogy of picking up lunch for a team: while one might know how to do it, it doesn't mean it's the most efficient use of their time.

The speaker argues that the traditional competitive advantage in software, based on proprietary building knowledge, is diminishing. Going forward, the ability to build almost any software will become more universal. The critical decision then shifts to efficient resource allocation—whether to dedicate internal time and energy to build non-core tools or to leverage external specialists.

My understanding is that building AI technology isn't a core competency of that firm, so I was surprised to see it.

The competitive landscape of value accrual in the AI industry

Businesses need to be ruthless about focusing on their core competencies. Tasks that are not central to a company's unique value proposition should be outsourced to ensure focus and efficiency. This principle applies to all areas, from internal operations to product development.

The AI industry is characterized by a "value war" where different segments, such as model companies, application developers, and infrastructure providers, are constantly trying to commoditize each other. Each segment claims to hold the primary value, much like a meme depicting different parts of a large corporation like Microsoft pointing guns at each other.

Value accrual in the AI ecosystem is a time-dependent phenomenon, not a fixed state. Pricing power and the capture of value shift between different players and segments over varying periods. There isn't a single entity that consistently holds all the value; it's an ever-evolving dynamic.

Companies operating in this landscape aim to be model-agnostic, seeking to provide customers with optimal pricing, performance, and speed for software development tasks. They leverage various AI providers like OpenAI, Anthropic, Google, and Microsoft, effectively treating the underlying models as commodities to deliver the best solutions.

The reality is, value accrual is a time dependent phenomenon. It's not like there is one person whose steady state gets all of the value.

Enterprises Experience AI Hangover and ROI Reckoning

The pace of AI model development is accelerating rapidly, with new models emerging every few days, particularly from open-source initiatives. This continuous churn creates fatigue among engineers within enterprises, who struggle to keep up with the latest advancements and determine the optimal model for specific tasks based on trade-offs between cost, quality, and speed.

Enterprises have gone through an initial phase of mandatory AI adoption, driven by executive pressure to implement an 'AI strategy.' This led to an 'AI at all costs' approach, characterized by 'token maxing,' where usage of AI, often frontier models, was encouraged and even measured, pushing for widespread adoption without stringent cost-benefit analysis.

This aggressive adoption has led to an 'AI hangover,' where companies are now reviewing their expenditures and realizing significant costs without clear return on investment. Many are finding they've spent hundreds of thousands of dollars monthly on advanced frontier models for basic, non-work-related queries that do not require such high levels of intelligence, prompting a re-evaluation of their AI spending.

Open-source models are emerging as a vital counterbalance, enabling enterprises to allocate resources more efficiently. By using intelligent routing, companies can select the appropriate level of intelligence for each task, avoiding the 'overkill' of using expensive frontier models when less powerful, more cost-effective options would suffice. This allows for better optimization across the cost, quality, and speed spectrum.

We've been spending hundreds of thousands of dollars per month on people asking Opus four point eight questions. We don't need the frontier of human intelligence to be doing this stuff for us.

The Growing Challenge of Managing AI Token Spend and Resource Allocation

Companies are experiencing rapidly increasing AI usage, often without clear initial spending strategies. This has led to scenarios where usage "goes crazy," prompting the need for user and token limits. Early approaches often involve blanket limits, which can initially cause a drop in usage.

A more effective approach involves a conscious decision-making process to identify which parts of a codebase should utilize AI tokens. This ensures resources are directed towards areas that generate the most business value, rather than being spent indiscriminately across all development activities.

As AI integration matures, organizations are moving towards nuanced resource allocation. This means different teams will receive varying token limits based on their strategic importance and specific project needs, moving beyond uniform restrictions.

Industry predictions suggest a future where AI token expenditure per developer could rival or even exceed their salaries. This shift necessitates careful financial planning and strategic resource management to avoid unsustainable costs, as seen in examples like large corporations spending hundreds of millions on AI models.

Open-Source Models Can Already Handle 80-90% of Frontier Model Tasks

The financial impact of AI tokens on individual roles will be highly variable, not a fixed percentage across the board. Some individuals may use zero tokens to deliver value, while others could leverage AI to achieve output equivalent to thousands of percent of their salary. This variance highlights that value derived depends on unique human skills rather than a standardized token expenditure.

Traditional job titles like 'developer' may become less distinct as nearly everyone in an organization, from sales to marketing, will likely interact with code through AI. The focus shifts to how individuals apply their skills, whether through token-intensive delegation to AI 'droids' or through human-centric activities like in-person meetings.

A significant portion of AI-driven work, specifically 80-90% of tasks currently relying on expensive frontier models, can already be accomplished using more cost-effective open-source models. The primary remaining need for high-end frontier models is for complex planning activities, suggesting a strategic trade-off where most operational tasks can be handled by efficient open-source solutions.

Organizations attempting to enforce a standard percentage of token use per engineer are likely oversimplifying, as the optimal leverage of AI will be highly individualized based on specific roles and skill sets.

What percent of tasks today be- using frontier models could be done with open-source models? Probably eighty to ninety percent. It's typically the planning that really needs the frontier models.

High-Value Tokens Drive Frontier Model Use, Requiring Polymath Engineers Focused on Business Outcomes

Most tasks (80-90%) can be handled by open-source models for implementation, but the critical 10-20%, termed 'high-value tokens,' necessitate more expensive frontier models. These tokens represent key decision-making points, similar to how leadership makes pivotal, often irreversible decisions that determine a company's fate, even if they don't consume the most hours. Organizations are increasingly allocating larger budgets to these specific, high-reasoning steps.

The traditional Silicon Valley hierarchy, which often prioritizes research and engineering over sales and marketing, is problematic. A company's 'product' is the entire end-to-end customer journey, encompassing not just software but also marketing, sales processes, and solution engineering. Treating any function as second-class, especially sales and marketing, will ultimately lead to competitive disadvantage when market conditions become tougher.

The role of a great engineer has shifted from focusing on coding speed, competition wins, or memorizing language nuances to owning full end-to-end business outcomes. This means engineers must understand customer behavior, drive changes that benefit users long-term, and enable sales teams. This 'general manager' approach emphasizes agency and entrepreneurial ownership over credentialism.

AI tools are bringing back the 'polymath' era, where individuals can quickly get up to speed on the frontier of multiple disciplines. Unlike previous decades where fields were too deep for multi-disciplinary mastery, AI allows engineers to understand systems and constraints across various domains, fostering a holistic approach to problem-solving and business impact.

The age of the polymath is back.

Mundane Engineering Tasks Are Prime for AI Automation

The discussion highlights the resurgence of the "polymath" era, where individuals can excel in multiple areas simultaneously. This shift is driven by AI taking over routine tasks, allowing professionals to innovate in diverse fields like developer marketing, token caching, or solution engineering. Companies are now seeking candidates who can demonstrate this multi-faceted capability.

The conversation touches on the emerging role of "agent operations," a function focused on creating and maintaining AI agents to streamline various organizational tasks, from social media content creation to marketing design visuals. However, the speaker emphasizes that individuals within specific departments should ideally be proactive in utilizing agents to enhance their own efficiency.

Looking ahead, some of the first jobs to be significantly impacted by AI automation are mundane engineering tasks. Specifically, the time-consuming process of writing release notes and extensive documentation is expected to be almost entirely automated. The example of Stripe's highly regarded documentation is cited to illustrate the significant effort previously dedicated to such tasks, which AI will soon handle.

For an engineering team, like writing release notes, that's crazy that people used to spend hours of time writing release notes or like writing documentation.

AI Agents Redefine Code Review with Enhanced Developer Experience

The initial wave of AI coding tools led to significant code generation, but often resulted in a flood of poorly formatted or non-standard pull requests, dubbed 'slop PRs'. This created a substantial review burden for staff engineers who had to ensure quality and adherence to organizational standards.

The evolution towards 'agent-native software development' highlights a critical shift: the return on investment (ROI) for developer experience (DevX) initiatives is dramatically amplified. Investing in elements that make agents more production-ready, such as up-to-date documentation, the ability for agents to run and iterate on code on remote machines, and robust CI/CD pipelines, becomes paramount.

Previously, investments in DevX primarily benefited human engineers linearly, making their work easier and onboarding smoother. However, when applied to AI agents, these same DevX improvements can multiply their impact by 10x or even 100x.

A superior DevX ensures that AI agents adhere more closely to coding standards, drastically reducing the time human engineers spend reviewing agent-generated code. This leads to significantly faster throughput in the software development lifecycle, transforming the efficiency of the entire process.

With agents though, the impact of that is now like ten x or a hundred x depending on how many agents you're using, because the better your DevX the better your agent ends up adhering to your standards, which means there's less time that that poor staff engineer has to go through reviewing your PR, which means you're kind of faster throughput in your software development.

Engineers will build software factories, not software

The future of software development involves a shift from human engineers directly writing code to designing and optimizing

software factories.

This means engineers will be responsible for creating the systems and processes that build software, much like how humans design and manage assembly lines in a car factory to optimize production.

The core idea is that as AI agents become more involved in code generation, human engineers will focus on the meta-level task of ensuring these agents produce efficient, clean, and maintainable code. They will set standards, manage integrations, and prevent

Why AI will ultimately lead to a net good for society

Matan Grinberg believes that while AI might cause short-term job displacement for engineers, he is not worried about its long-term impact. He argues that there are a vast number of problems in the world that could be solved or significantly helped by software, but currently only a small fraction of them are being addressed.

The increase in engineers due to AI will allow society to reallocate talent towards solving these unaddressed global issues. This shift will lead to a net positive outcome. For example, instead of many engineers working on optimizing ad platforms at large tech companies, they could be focused on critical areas like climate change or healthcare.

Matan highlights specific problems like pharmaceutical research and health issues, including dementia, that could be significantly advanced with better AI and software engineering. He criticizes those who advocate slowing down AI development, suggesting it delays solutions to severe human problems.

He contrasts the current allocation of engineering talent, sometimes on suboptimal tasks like "optimizing claw code," with the potential for them to work on more mission-critical areas like healthcare systems. This reallocation, though it might take time for the economy to properly incentivize, represents a significant societal benefit.

By saying you wanna slow down AI, that's saying like, these people who have relationships with loved ones who have dementia, you're like, no, no, no, sorry, you guys, you gotta maintain that relationship for a little bit longer, we're scared, we don't capacity.

Government Intervention, AI Infrastructure, and Human Adaptation Challenges

Government intervention in AI development can be useful for mission-critical areas like military or safety, and for societal problems where capitalism might not provide immediate feedback loops. While some level of incentivization can juice outcomes, there is general reluctance for broad government involvement. An example like climate change highlights the complexity, where rapid AI development might require short-term fossil fuel consumption, posing a dilemma between immediate and long-term benefits.

The idea of an AI infrastructure bubble is largely dismissed for the long term. While there might be short-term market blips or corrections, similar to past over-allocations by companies like Uber, these are not indicative of a prolonged bubble. The underlying demand for AI infrastructure is seen as robust and growing.

The most significant bottleneck to AI adoption in organizations is human behavior change. Experienced engineers, for example, often struggle to adapt to new patterns and workflows due to being set in their ways. Conversely, junior engineers are more eager to adopt new methods but may lack crucial skills in managing projects or delegating tasks, presenting an interesting balance challenge for companies implementing AI solutions.

I think the biggest bottleneck by far, working with all these organizations, is the human side of it. It's just like behavior change.

Matan Grinberg's unconventional journey from physics to startup founder

Matan Grinberg's early life was defined by an intense obsession with physics, fueled by a competitive drive. After being told he needed to retake geometry at age 12, he self-studied high school and college math textbooks, aiming to become a string theorist. This led him to Princeton to work with a famous physics professor, Juan Maldacena, and then to Berkeley for his PhD.

During his PhD at Berkeley, Grinberg realized his passion for physics was rooted in competitiveness rather than genuine interest. He explored alternative paths like quant finance but ultimately found his true calling in program synthesis, now known as code generation. He was "nerd-sniped" by the fundamental idea of code generating itself, a concept that appealed to his physicist's mind seeking universal solutions.

Recognizing that properly solving code generation challenges would require starting a company, Grinberg, with no business background, began self-educating. He read Peter Thiel's "Zero to One" and watched Y Combinator videos. This led him to discover a podcast featuring a theoretical physicist turned billionaire investor at Sequoia, whose paper he had cited at Princeton.

Grinberg reached out to the Sequoia partner, leading to an impromptu three-hour walk where they discovered shared reasons for entering and leaving physics. The partner encouraged him to drop out of his PhD and consider joining Twitter or his existing venture, Factory, which Matan had already known about but kept secret during their initial, non-transactional conversation.

There was something so fundamental about this idea of like code generating itself, and it just got me obsessed.

Securing Sequoia's First Check Based on Belief and Trust

Matan recalls meeting his co-founder, Eno, describing it as "intellectual love at first sight." They quickly collaborated, with Eno, a much stronger engineer, transforming Matan's initial demo into a polished version within 72 hours.

Following this, an investor challenged Matan to drop out of his PhD program to commit full-time to the project. This was a pivotal decision, as his parents, immigrants from the Soviet Union, held his PhD in high regard.

Matan and Eno then pitched Sequoia with a self-described "shitty deck," confidently presenting their vision for fully autonomous software development agents in April 2023, a concept that was significantly ahead of industry adoption at the time.

Despite the informal pitch, Sequoia offered a million-dollar seed check. Matan highlights that this investment was rooted in belief and trust, acknowledging that few others would have taken a chance on him given his lack of prior job experience, valuing this trust above financial terms.

No one else would have believed in me, except him.

The Story Behind Ivanka Trump's Investment in Factory

Identifying strong investors involves more than just their excitement when a company is doing well; their behavior during tough times is a critical indicator of their reliability and loyalty. This discerning approach ensures that investor relationships are built on a foundation of trust that can withstand market fluctuations and challenges.

One of Factory's best hires was Francesca, who Matan met at a random conference, seated next to her and Alex Paul of The Chainsmokers. Coincidentally, they also grew up in the same hometown. Francesca's persistence and clear aptitude were evident when she sought a larger investment allocation for her firm.

Her relentless pursuit of a greater stake in Factory led Matan to jokingly suggest she just join the company if she wanted more ownership. This offhand remark evolved into a serious proposition, demonstrating her commitment and the value she could bring beyond just an investment.

Francesca, being very close with the firm Affinity from her investing days, subsequently facilitated an introduction that led to Ivanka Trump's investment in Factory. This connection highlights the power of a strong network and the unexpected pathways to securing high-profile investors.

Look, Francesca, like, if you want more ownership of the factory, you could just join us.

Model Agnosticism Prevents Vendor Lock-in in the AI Market

The AI market, particularly for large language models, faces a critical challenge regarding vendor lock-in. When model providers also offer the applications built on top of their models, incentives become misaligned. The model provider, operating on an API business model, benefits from increased token usage and therefore has less motivation to improve token efficiency or offer competitive pricing, potentially leading to higher costs and a worse experience for the end consumer.

To avoid this, it is crucial for models and applications to be separate. An independent application layer allows enterprises to choose between various model providers on a task-by-task basis. This competition forces model providers to offer better prices and develop more efficient or specialized models, such as those excelling in specific programming languages, ultimately benefiting the consumer.

The AI industry can learn from the past mistakes of the cloud computing market. Cloud providers often enticed customers with discounts on long-term deals, only to significantly increase prices once customers were standardized on their platform. By demanding model agnosticism, businesses can prevent similar price hikes and avoid being trapped with a single provider, ensuring flexibility and competitive innovation.

Implementing model agnosticism does not mean forcing engineers to use multiple tools. Instead, platforms that can abstract away the underlying model, allowing for selection based on cost-efficiency or specific task performance, enable organizations to leverage the best available model without sacrificing developer experience.

In the world where they're, you're like vendor locked in, then you can slowly get like laziness and slower shipping and, and as a, you know, consumer, you end up getting a worse experience.

AI-Generated Code Security Risks, Geopolitical Concerns, and Europe's AI Lag

The market for low-code and no-code AI tools is currently in flux, with new competitive products entering the space. Developers using these tools are likely to pivot towards enterprise functions for non-technical teams, such as sales and marketing, to create customized applications. However, critical enterprise code generation will likely remain under the strict purview of engineers using specialized, secure platforms.

The exponential growth of AI-generated code is creating significant security risks, as security efforts are not keeping pace. This lag is expected to lead to major security incidents in the next few years. Beyond accidental vulnerabilities, the potential for adversarial actors to intentionally use AI tools for malicious purposes also represents a growing threat.

Concerns exist regarding US startups using Chinese open-source AI models, primarily centered on data exfiltration. While hosting an open model in the US is generally acceptable, sending sensitive data externally to countries like China poses a valid national security risk. There is a strong sentiment for the United States to develop its own frontier open models to ensure technological superiority.

Europe faces considerable challenges in catching up in frontier AI model development. This is largely attributed to slower democratic processes, which require extensive support and legislation for infrastructure projects like data centers. The continent also struggles with energy infrastructure build-out, with decisions like Germany's stance on nuclear power potentially hindering its ability to meet the high energy demands of advanced AI.

Code generated is growing exponentially, the security efforts aren't growing in kind, and there's kind of a lag there.

Core Competencies and Product Strength Define Success in AI Infrastructure and Enterprise Sales

Matan Grinberg analyzes the competitive dynamics in the AI infrastructure market, specifically between Nebius and CoreWeave. He suggests that companies should focus on their core competencies, implying that overambitious full-stack strategies might lead to unnecessary competition if they extend beyond a company's fundamental strengths.

The conversation also addresses customer concentration, pondering if having a significant portion of revenue from a few large clients is inherently problematic. While acknowledging the increased risk for investors, it's suggested that businesses can still achieve a stable operating state even with high customer concentration.

Matan offers strong opinions on selling to enterprises without a Field Development Engineer (FDE) model. He argues that a genuinely good product should enable sales on its own, with FDEs serving solely to accelerate customer adoption and consumption. If FDEs are required to make the product work or to close deals, it indicates a fundamental flaw in the product itself, distinguishing product companies from services providers.

If you're putting in FDES 'cause that's the only way you'll get a deal done, I'm sorry my friend, you have a shit product.

Elite Companies Will Optimize Employee Performance Like Professional Athletes

The concept of "grind slop" arises when companies focus on intermediate metrics like hours worked rather than actual output. This approach often misses the mark, as true performance, much like in sports, should be measured by results. Top companies prioritize getting the best players and then enabling them to score, regardless of how much visible effort they exert.

For high-stakes work requiring deep thought and every ounce of brain power, companies must invest in their elite employees. Treating them like Seal Team Six or NBA All-Stars means providing resources and conditions that optimize their performance. This includes ensuring they get adequate sleep and are in peak mental condition to make critical decisions, rather than mandating crazy hours or expecting work on two hours of sleep.

The best companies will increasingly treat their teams like professional athletes, but in the intellectual domain. This goes beyond superficial perks; it involves rigorous support for well-being, such as monitoring diet, ensuring mental recovery, and providing tools to enhance cognitive function. The goal is to maximize individual intellectual output to meet ambitious goals.

I think generally, we will see the best companies treat teams more and more like, whatever, SEAL Team Six or NBA, like professional athletes.

OpenAI and Anthropic are Business Equivalents, but the 'AI Takes Jobs' Message is Harmful

When evaluating investment opportunities in OpenAI versus Anthropic on IPO day, the speaker considers them largely equivalent from a business perspective. Both companies are deemed well-suited and well-positioned in the market. The primary factor differentiating their enterprise value, it is argued, is company volatility, with OpenAI having experienced more "random, chaotic, turbulent events" historically.

However, the speaker strongly criticizes Dario Amodei's narrative that "AI takes jobs." This message is considered disingenuous, factually incorrect, and psychologically damaging to developers and the general public. It's suggested that such a narrative does a disservice to AI by emphasizing job displacement rather than its potential to solve societal problems.

The motivation behind this "AI takes jobs" rhetoric is attributed to a fundraising strategy. By implying that "all of capitalism is gone" and only their company will survive, it provides a powerful incentive for investors to commit unprecedented amounts of capital. This approach contrasts with the potential shift in messaging that might occur when these companies eventually go public and seek investment from the very human population they suggested would be replaced.

I think that has been not only like disingenuine and wrong, but it's like really hurt the psychology of a lot of developers, like just people in the world.

Matan Grinberg Changes His Prediction on Frontier AI Company Dominance

Matan Grinberg shares a significant shift in his outlook regarding the future landscape of frontier AI. He initially believed that only one or two companies would emerge as the undisputed leaders, outcompeting others in advanced AI development.

However, his current perspective is that at least four companies, and potentially more, will achieve approximately equal and highly advanced AI capabilities. This suggests a more diversified and competitive top tier rather than a concentrated duopoly or monopoly.

Matan views this development as a positive outcome, explicitly calling it a "win for humanity." He believes that a greater number of equally powerful AI entities is more beneficial than having just one or two dominant players.

He acknowledges this updated prediction might be considered a "hot take," noting that many people currently remain fixated on the idea of only one or two companies eventually leading the AI race.

What seems pretty clear to me is it's probably gonna be at least four that are gonna probably be approximately as good, and that is a win.