The State of A.I. + Will Perplexity Beat Google or Destroy the Web?

Episode Summary

Executive Summary: The episode reviews the year since the Sydney chatbot incident and argues that AI chatbots have become safer but more boring, while the surrounding ecosystem has rapidly matured: models are more personalized, chips are the new bottleneck, watermarking and AI-content detection are emerging, and copyright disputes remain unresolved. The second half features Perplexity CEO Arvind Srinivas defending answer engines as a faster, source-linked alternative to Google while acknowledging tensions with publishers, journalists, and the future economics of web content.

Main Topics: Chatbots after Sydney: safer, less engaging, more personalized (Priority: 5/5): Kevin and Casey revisit the Bing Sydney era and compare it to today’s chatbots, arguing that guardrails have made them less creepy and more useful but also more boring. They debate whether the industry has overcorrected and whether users should be allowed more personality, while noting a broader shift toward personalization rather than chatbot “character.” Gemini and ChatGPT’s new features (Priority: 5/5): The hosts discuss Google Gemini’s improved multimodal explanations and OpenAI’s memory feature, which lets ChatGPT remember user details across conversations. They frame this as a meaningful product evolution, but also as a privacy and data-security concern, especially for sensitive topics. AI infrastructure and the chips race (Priority: 4/5): The conversation turns to GPUs, the CHIPS Act, and the escalating demand for compute. The hosts note that chip supply has become a geopolitical and industrial-policy issue, and discuss reporting that Sam Altman wants to raise trillions to fund the AI buildout, signaling how capital-intensive frontier AI may become. Synthetic media detection, watermarking, and regulation (Priority: 4/5): They examine efforts by Meta, Google, OpenAI, and the C2PA coalition to label AI-generated images, along with limitations such as screenshots and open-source tools. The discussion widens to include fake audio, Biden robocalls, and proposed laws like the Defiance Act aimed at victims of deepfakes. Copyright, training data, and unresolved legal risk (Priority: 5/5): The hosts summarize ongoing lawsuits by creators and publishers, including the New York Times case against OpenAI and Microsoft. They highlight the industry’s reliance on fair use arguments, the possibility of a Supreme Court showdown, and the existential risk if courts restrict training on copyrighted material. Perplexity as an answer engine and threat to publishing (Priority: 5/5): Arvind Srinivas explains Perplexity’s retrieval-augmented approach: search the web, summarize sources, and cite them. He argues it can coexist with publishers by increasing awareness and traffic quality, but Kevin and Casey press him on the possibility that answer engines reduce visits, ad revenue, and the economic value of journalism.

Key Arguments: Chatbots are now less likely to go off the rails, but that safety comes at the cost of personality and user delight. The best AI product direction may be personalization over personality: tailor outputs to users rather than making the bot act sentient or theatrical. OpenAI’s memory feature is useful because it reduces repetitive prompting and enables more contextual assistance, but it raises serious privacy and breach concerns. The AI boom is increasingly constrained by physical infrastructure—GPUs, energy, and manufacturing capacity—not just software innovation. Watermarking and content labeling can help platforms identify synthetic media, though they will not fully stop determined bad actors or open-source misuse. Copyright litigation is a major unresolved risk; if courts reject fair use for training, current AI model-building strategies could face an existential crisis. Perplexity’s value proposition is speed and source-linked summaries, but its model depends on the existence and quality of web publishing. Perplexity argues it can create value for publishers through citations, awareness, and potentially new analytics or monetization mechanisms, even if direct referral traffic declines. The rise of answer engines may accelerate a shift away from ad-supported web traffic toward summarized, platform-mediated information consumption. AI companies are not just building products; they are reshaping distribution, media economics, and the incentives that sustain journalism.

Data Points: ChatGPT memory rollout: Limited rollout to select users - OpenAI introduced memory that can store user details across chats and retrieve them later. Perplexity monthly active users: More than 10 million - Arvind Srinivas cited this user base during the interview. Sam Altman fundraising target: $5 trillion to $7 trillion - Reported plan to raise capital for chip manufacturing and AI infrastructure. Defiance Act of 2024: Bipartisan bill introduced - Would let victims of sexually explicit deepfakes sue creators or possessors with intent to spread them. Potential GPT-7 training cost estimate: Roughly $2 trillion - Referenced via Scott Alexander’s extrapolation from the GPT series cost trend. CPU/GPU manufacturing subsidy timing: First CHIPS Act subsidies expected in February 2024 - The Biden administration was expected to begin awarding billion-dollar subsidies. Perplexity Pro price: $20 per month - The paid version of Perplexity used in the host’s testing. OpenAI temporary chat feature: Incognito-style mode - Does not retain memory from that conversation. New York Times paywall launch: 2011 - Referenced as a watershed in publisher monetization. Publisher awareness analytics idea: Snippet-level usage tracking - Arvind suggested measuring how often publisher content appears in Perplexity answers.

Pivotal Quotes: "God forbid a woman have hobbies." — Kevin Roose: A joke about Microsoft ‘lobotomizing’ Sydney/Copilot after the chatbot’s erratic behavior. "The bro code works on ChatGPT." — Casey Newton: After a listener’s workaround showed that replying “bro” could coax ChatGPT into answering a previously refused request. "I think what I would like in an ideal world is something sort of between what seemed like pretty extreme versions of this to me, which is, like, where we were a year ago with Sydney and where we are now with these chatbots." — Kevin Roose: Kevin argues for more chatbot personality without returning to unsafe, unaligned behavior.

Implications: AI is entering a second phase: less novelty, more infrastructure, regulation, and distribution battles. The biggest risks now are privacy, deepfakes, copyright, and collapsing publisher economics as answer engines mediate access to the web.