Ep9. GPT-4o, Astra, Multi Modal, China Tariffs, Tech Earnings | BG2 with Bill Gurley & Brad Gerstner

Episode Summary

Executive Summary: The episode centers on AI’s consumer breakthrough—especially voice, multimodality, tutoring, and translation—while debating who wins economically as models improve and prices collapse. The hosts argue consumer AI will reshape apps, pressure startups, and favor platforms with distribution and data. They also cover Washington’s tentative AI posture, U.S.-China trade and tariffs, and a cautious market outlook amid slowing growth and uncertain earnings.

Main Topics: AI’s consumer breakthrough and the 'Her' moment (Priority: 5/5): The discussion frames recent OpenAI and Google demos as a consumer-inflection point: faster voice interaction, more natural interruption, and a more human assistant experience reminiscent of the film Her. Multimodal AI as a single-model leap (Priority: 5/5): The hosts emphasize that the major advance is not benchmark gains but a single model reasoning across voice, text, and vision, which may unlock new interfaces and use cases. Voice as the new UI and on-device assistants (Priority: 4/5): They explore voice as a replacement or supplement for typing and mouse input, and discuss the challenge of getting memory and actions onto phones and edge devices. Competitive dynamics: platforms vs startups (Priority: 5/5): The conversation weighs who benefits from AI consumerization, suggesting large platforms with users, apps, and data may win while many AI-native startups face compression or displacement. Policy, regulation, and federal spending on AI (Priority: 4/5): They examine Washington’s approach to AI regulation, favoring output-based oversight over input restrictions, and criticize a proposed large federal AI spending program. Trade, tariffs, and deglobalization (Priority: 4/5): The hosts argue tariffs and industrial policy are inflationary, reduce competition, and conflict with free-trade principles, especially in the context of U.S.-China competition. Markets, rates, and earnings caution (Priority: 4/5): The episode ends with a market check: inflation is easing but growth is slowing, software guidance is weak, and the hosts are turning more defensive despite AI-favored winners remaining attractive.

Key Arguments: AI’s most immediate value is as a patient, always-available tutor and assistant, especially for education and language learning. OpenAI and Google’s latest demos mark a consumer-facing shift, with voice interaction becoming faster, more natural, and more central. The real technical milestone is multimodal reasoning in one model across voice, text, and vision, not just benchmark improvement. Voice-based AI could become the new graphical interface, but only some apps are naturally voice-in/voice-out; others will remain screen-based or hybrid. Apple, Google, Meta, OpenAI, and Anthropic may all compete for the same consumer assistant prize, creating a battle royale among the largest tech firms. Startups focused on tutoring, translation, language learning, meeting assistants, and some BPO functions may see multiple compression as AI eats into their future cash flows. Regulating AI outputs makes more sense than regulating inputs, because existing laws already cover fraud or harmful acts regardless of the tool used. The proposed $32 billion annual AI spending initiative in Washington looks excessive given current private investment and the U.S. debt burden. Tariffs and deglobalization will raise prices, reduce competitiveness, and weaken U.S. firms by insulating them from global competition. Given slowing growth, rising uncertainty, and record highs in equities, the prudent stance is to own clear AI winners but reduce broader risk exposure.

Data Points: Time from science fiction to product iteration: 11 years - The hosts noted the span from the movie Her to contemporary AI demos. Typical science-fiction-to-product timeline: 15 to 40 years - A ChatGPT query estimate cited during the discussion of the Her analogy. OpenAI model benchmarks: Only modest improvement on human eval, with major inference pricing gains - They referenced a chart showing GPT-4 Omni improved pricing more than benchmark quality. Voice recognition improvement: Better than Siri / near-human cadence - Qualitative assessment of OpenAI’s voice demo and the new interface potential. Model modalities: 3 modalities in one model - Karpathy’s point that one model processes voice, text, and vision together. Google / OpenAI context window references: 2 million tokens - Mentioned as a benchmark-style stat that the public barely understands but vendors highlight. Estimated OpenAI revenue impact on pricing: Price cliff of 20x differential - Backing off one release generation reportedly saves 90% to 95% of costs. Duolingo stock move: Down 4% to 5% - Cited as a market reaction to AI tutoring/language-learning disruption. Consumer AI company revenue reference: About $20 million in revenues - Mentioned for Perplexity as an example of a consumer AI startup with early traction. Proposed federal AI spending: $32 billion per year - Schumer-related AI policy discussion of a federal funding initiative. NIH budget comparison: A little over $40 billion - Used to show the scale of proposed AI spending in Washington. National Science Foundation budget comparison: $9 billion - Used to compare the AI spending proposal to existing science funding. U.S. debt level: $38 trillion - Raised as a concern about the fiscal context for new spending. Interest payments: Nearly $1 trillion - Cited as a warning sign for fiscal constraints. Tariffs announced by Biden: 25% steel/aluminum, 50% semiconductors, 100% EVs, 50% solar panels - Used in a critique of U.S. tariff policy. CPI market reaction: 10-year Treasury around 4.35% - Referenced after a softer inflation print. AI-related market multiples: 20x to 30x earnings - Described as reasonable for Microsoft/Nvidia-like AI beneficiaries. Software earnings guidance: 54% disappointed vs consensus - The hosts said more than half of software companies guided below expectations.

Pivotal Quotes: "This is classic freshman economics. If someone's better at doing something than us, we should let them do it, and we should buy it from them, and we should send them what we're good at." — Speaker 1: Argument against tariffs and deglobalization, during the trade-policy discussion. "The winner, if you will, of the week was OpenAI's voice recognition." — Speaker 2: Assessment of the key product advance from the week’s AI demos. "LLM stands for large language models. So things that involve language are what this thing was designed for and what it's really great at." — Speaker 2: Explanation of why language and code are especially strong early-use cases for AI.

Implications: AI is moving from novelty to consumer infrastructure, threatening many software, education, and service businesses while rewarding platforms and clear AI leaders. Investors and operators should expect multiple compression, faster product cycles, and a widening disruption surface across language, voice, and automation.