Shopify's React Native Exit, AI Subsidies, and Startup Growth

Theo upgraded his audio interface from a $700 model with four inputs to a $1,500 Mix Pre3v2, which has three inputs and fewer features but offers higher-end sound quality. This change involved shifting from hardware processing to software filters within OBS. He acknowledged a potential 'closer' sound profile resulting from the new microphone settings, though a current congestion made a definitive assessment of the audio quality difficult.

Theo expressed an optimistic view regarding David Heinemeier Hansson (DHH)'s recent shift in public discourse, noting a more focused engagement on open source initiatives and their funding. He believes criticisms directed at DHH's current objectives are overly pessimistic, asserting that collaboration between individuals with differing political views can still yield significant value. As an example, Theo mentioned his ongoing work with Ryan from the Omari team to integrate T3 Code.

Theo advocates for a strict rule in startup growth: companies should never more than double their size year-over-year. He argues that aggressive overhiring frequently dilutes company culture and causes a loss of strategic focus. He shared an anecdote about a founder who, after realizing the mistake of overhiring, corrected course by offering severance packages. Theo views overhiring as the most prevalent error made by well-funded startups.

Theo dismissed accusations that OpenAI monitors individuals to steal mathematical proofs as a "low IQ/conspiracy" theory. He posits that OpenAI's likely objective was to determine the computational resources required to solve an unsolved mathematical problem, rather than engaging in espionage. He highlighted that the solved problem cited differed from the mathematician's claims. Theo criticized the detractors for letting their biases against OpenAI cloud their judgment.

He explained that individuals with strong distribution and reputation typically show greater interest in new technologies. This means that if a project offers compelling technology, it can gain significant traction by winning over key developers with large platforms, effectively acting as a multiplier. Theo further asserted that if a project cannot impress him or similar figures, it is unlikely to reach their respective audiences.

Theo defended Jacob, a former OpenAI researcher, against viral, unsubstantiated conspiracy theories suggesting a political plot behind his departure and subsequent move to Anthropic. He dismissed the narrative that 'Democrats are funding this' due to a lack of credible incentive or evidence. Theo explained that leaving a company over safety concerns is a logical career decision, not a political maneuver. He expressed frustration with the public's tendency towards paranoia and lack of critical thinking.

Shopify announced its decision to transition away from React Native, returning to native code (Swift and Kotlin) after a substantial investment in the cross-platform framework. This strategic shift is primarily driven by the evolving landscape of Large Language Models (LLMs) and coding agents, which have significantly altered the cost dynamics of native development. Maintaining platform parity, once a major hurdle, is now less challenging with advanced agentic assistance. Shopify's choice underscores the changing trade-offs between cross-platform frameworks and native capabilities.

Shopify's initial decision in 2020 to adopt React Native was primarily to avoid the inefficiency of building features twice for different platforms. However, the advent of Large Language Models (LLMs) has fundamentally altered this core assumption. LLMs now effectively handle implementation and testing across multiple platforms, drastically reducing the cost of maintaining native codebases through agentic assistance. Theo critically examined Shopify's internal approach to complexity and simulator testing within this new paradigm.

Traditional native builds necessitate slow and expensive update processes, whereas OTAA enables rapid bug fixes and even temporary UI changes without requiring app resubmission. However, Apple and Google generally discourage unauthorized feature updates through this method.

Without Expo, many developers find themselves rebuilding similar foundational tools, leading to inefficiencies. Expo offers superior dev client support and facilitates web-based iteration, streamlining the build process and native development. Theo suggested that Shopify's omission of Expo adoption likely hindered their overall efficiency in React Native development.

Theo contended that the average React Native developer possesses lower skill levels compared to the average native developer, considering this a "fact" without room for argument. He attributed this to the perception that native development is inherently more difficult, thus attracting more experienced professionals. React Native's lower entry barrier, while democratizing access, often results in more frequent "foot guns" – errors that developers make due to insufficient understanding or best practices. Consequently, the average skill level within React Native teams is often lower than in pure native development teams, indicating that technological popularity does not necessarily correlate with developer quality.

Theo noted that while using Large Language Models (LLMs) to port React Native applications to Swift can be quick, it often introduces new errors. He highlighted that SwiftUI, despite being native, presents its own set of challenges and 'rough edges' comparable to those found in React Native, particularly concerning list implementations. While native code generally offers better background data processing, the subjective 'feel' of native navigation versus React Native remains a contentious point among developers.

He underscored that React Native is substantially more complex to build and maintain than React. Meta's substantial investment, however, successfully fostered a high-quality native experience, making it accessible to a broader base of general developers.

Google has invested more time and resources into optimizing Windows for application developers than Microsoft itself. This commitment stemmed from Google's frustration with Windows development during the creation of applications like Google Earth for Windows, leading them to prioritize web-based platforms. Microsoft, in contrast, heavily relies on Electron for key applications such as VS Code and Teams, illustrating a broader industry trend where abandoning web-based platforms for desktop native development incurs significant losses in performance research.

Theo clarified that Y Combinator (YC) investment returns significantly outperform those of standard venture funds. He presented data from 2018 to 2020, revealing that the bottom 25% of early-stage YC investors, who backed at least three companies per batch, still tripled their money. The top-performing YC investors, within the same timeframe, achieved returns of 15 times their initial investment. Consequently, Theo's financial managers recommend allocating liquid cash to YC companies.

Conversely, none of a16z's partners are founders; they are career investors. A16z operates as a general fund, primarily focusing on Series A and later-stage investments, often requiring equity rather than SAFEs (Simple Agreement for Future Equity).

The price had remained stable for two months prior to his recommendation, indicating a 30% increase directly attributable to his tweet. He stated his intention to find alternative products but keep them secret until he can secure his own stock.

Theo revealed a significant arbitrage opportunity within consumer AI subscription tiers, where the value of included tokens far exceeds the subscription cost. A $200 Claude subscription, for example, provides approximately $8,000 in token inference value, while a $200 Codeex subscription offers roughly $12,000 in inference value. Theo asserted that personal subscriptions, with data sharing disabled, are functionally equivalent to team accounts for most purposes. He categorized direct API spending as 'throwing away money' if users are not leveraging this personal subscription arbitrage.

AI models experience periodic resets due to compute availability, and older models can exhibit degraded performance when configured improperly. Theo noted that there have been bugs where skills designed for older models trigger too frequently, and context management experiments fail. He also highlighted issues with badly configured engines leading to measurable quality degradation for prolonged traffic. Reset mechanisms involve banked resets versus immediate resets, with Fridays and Saturdays often seeing more frequent resets as enterprise usage typically declines.

This new methodology aims to leverage the AI to identify potential shortcuts or alternative solutions that might not have been initially considered. He emphasized that the goal is for AI to help determine if a simple solution exists before deep intellectual investment, thus freeing up human cognitive resources for problems that AI cannot yet solve. This strategy involves outsourcing execution rather than the core thinking process.

Theo expressed strong dissatisfaction with the user interface of T3 Connect, specifically criticizing the absence of a 'remove' button for managing devices. He found the 'disconnect' function ambiguous, as it does not provide a clear way to permanently delete machines from the system. He suggested the need for a toggle or a more intuitive UI design that includes a permanent deletion mechanism.

Theo countered concerns about AI agents installing malware, arguing that human error represents a far more significant threat. He suggested that individuals who fear agent-based malware are overestimating the agent's malicious potential and underestimating the likelihood of their own mistakes. His argument implies that if one is concerned about an agent installing malware, the human user is the more probable source of such an incident.

He posits that API requests necessitate a global fan-out of cache data across various data centers to ensure reliability, which is an expensive operation due to the widespread writes. In contrast, he speculates that for signed-in sessions via platforms like Codeex, cache writes can be localized to a single connection, thereby reducing overhead and improving efficiency.