The Frontier

Theo argues Google is not a serious company, pointing to a year-plus period of no notable frontier releases from its AI labs since Gemini 1.5 Flash, which he describes as a disaster.

Theo says Google's 'Omni' model concept of anything-in, anything-out has one real use case: video-to-video generation for tasks like adding fire to a background, which he finds not broadly useful.

Theo argues most LLM-generated text is never read by humans but is still useful for reasoning and tool calls, while unviewed images and videos have no inherent value, questioning heavy investment in those modalities.

Theo cites a survey showing Midjourney's user share dropping from 45% to 8%, illustrating the rapid churn in image generation tools as capabilities become commoditized within broader platforms.

Theo states OpenAI's GPT-4o image generation is now useful for creating UI mockups and dashboards as business assets, not just artistic consumption, marking a shift in the modality's value.

Theo criticizes Google Cloud's reliability, citing a four-year history of issues and a recent incident where Google's algorithm mistakenly deleted Railway's entire account without human oversight.

Ben reveals a private software engineering benchmark showing GPT-4o and Claude 3.5 Opus leading, with a steep drop to Sonnet 3.6 and Gemini 1.5 Flash, and a final cliff to Gemini 1.0 Pro at 10% performance.

Theo ranks the AI lab hierarchy as OpenAI and Anthropic far ahead, with XAI and Cursor as potential contenders, followed by Chinese labs, and Google in last place due to stagnant trend lines.

Ben discusses Anthropic's monthly compute spend on SpaceX servers, revealed in the SpaceX IPO filing, as $1.25 billion, which constitutes a majority of Anthropic's estimated $1.5-2 billion monthly revenue.

Theo describes the Manis-Meta acquisition fallout, where Beijing used a policy to undo the completed $2B deal after employee onboarding, forcing Manis to try raising $1B to buy itself back from Meta.

Theo argues the Manis case and China's move to close-weight models like Qwen signify a deliberate decoupling from Western AI development, ending the era of Chinese open-weight models feeding the global ecosystem.

Ben details Cursor's Composer 2.5 training techniques, including reverting implemented features to generate synthetic chat logs for RLHF and using a teacher-student method to correct tool-calling errors without explicit context.

Theo contends Google's core failure is bureaucratic fragmentation, contrasting it with OpenAI's model of individual experts moving between teams and Vercel's company-wide 'unblock me' Slack channel that treats internal blocks as P0 issues.

Ben asserts Google's models fail at reasoning, citing their tendency to get stuck in loops or berate themselves in traces, and posits that adding reasoning was the moment Gemini fell apart competitively.

Ben introduces Lakebed, his integrated cloud framework built in four days with GPT-4o, designed to compile a full-stack app from code with three commands, eliminating the glue work between databases, auth, and hosting.

Ben argues the pain of deployment has become disproportionate now that AI can build apps in 40 minutes, making the traditional 3-5 hours of cloud configuration feel like an unacceptable bottleneck.

Ben states Lakebed automatically syncs environment variables from a local .env file to production on deploy, adding, updating, or deleting them as needed, which he calls the right approach for 90% of apps.

Theo describes a novel prompt injection attack vector called 'font hacking', where a PDF uses custom glyphs to show one city name to a human but a different name to an LLM reading the underlying text encoding.