The state of modern answer engines, AI demos, and more! | E1924

Episode Summary

Executive Summary: The episode explores how AI is rapidly reshaping search, content discovery, and media creation. The hosts compare AI answer engines like Perplexity and U.com against Google, argue that falling inference costs will commoditize many AI experiences, and demo tools that generate video/podcast clips and emotionally aware voice interactions. The discussion concludes that voice- and context-aware AI could dramatically improve UX while raising serious ethical and manipulation risks.

Main Topics: AI answer engines vs. Google search (Priority: 5/5): The hosts compare U.com, Perplexity, and Google on a difficult, time-sensitive query about the Warriors’ playoff odds. They argue AI search interfaces are more natural and often more useful, but accuracy and freshness vary, especially when sources are dated or not explicitly labeled. Inference cost collapse and model economics (Priority: 5/5): They discuss how open-source model pricing has fallen sharply and how this trend could make AI search and summarization economically viable at scale. They frame this as an AI-era version of Moore’s Law, with costs dropping rapidly enough to enable free or low-cost products. Search behavior, SEO, and the future of web traffic (Priority: 5/5): The conversation suggests that answer engines may disrupt the SEO-driven publishing economy by giving users direct answers instead of sending them through ad-heavy intermediary pages. They debate whether summarization of web content is fair use and how publishers might demand compensation. Benchmarking chatbot quality and the rise of proprietary models (Priority: 4/5): They review the LM Sys Chatbot Arena leaderboard, noting Claude’s rise and the continued dominance of proprietary models at the top. The hosts infer that access to better data, RLHF, and human-in-the-loop training remains a key moat. AI-generated media and podcast/video synthesis (Priority: 4/5): A demo from Infinity AI shows a short, synthetic clip of an All-In-style Taylor Swift debate generated from a prompt. The hosts see this as a preview of AI systems that can generate increasingly coherent audio-video content, eventually reaching full episodes. Emotion-aware voice AI and ethical risk (Priority: 5/5): The Hume demo impresses the hosts with emotional detection from voice and facial cues, improving response timing and personalization. They also warn that the same capabilities could be used for manipulation, radicalization, or exploitative persuasion in consumer and enterprise contexts.

Key Arguments: Inference costs are falling so fast that AI search and summarization will become cheap enough to scale broadly, including for products that currently seem economically impossible. Google’s legacy search model is vulnerable because answer engines can give direct responses with citations, reducing the need for click-through behavior and ad-supported intermediary pages. The decisive product advantage is not just UI; it is correctness, freshness, and access to authoritative data sources. Proprietary models still dominate rankings because they likely benefit from better data, more human feedback, and more advanced training pipelines. AI-generated clips and audio/video content will rapidly improve, moving from seconds to minutes and eventually to full-length, coherent shows. Voice and emotion-aware AI can dramatically improve user experience by reducing latency and making systems more responsive to human frustration or urgency. The same emotional intelligence that makes AI helpful in support or companionship could also be weaponized for persuasion, radicalization, or behavioral manipulation.

Data Points: Grok Cloud devs: 70,000 - Reported as the number of developers using Grok Cloud as of that morning. Apps using Grok: 20,000 - Number of applications said to have Grok integrated. Mixtro API pricing: $27 per million input tokens; $28 per million output tokens - Used to illustrate how cheap open-source inference has become. Google revenue per query: $0.0161 - From a February 2023 analysis referenced during discussion of search economics. Google cost per query: $0.0106 - From the same analysis estimating Google’s search profitability per query. Google profit per query: $0.0055 - Calculated difference between estimated revenue and cost per query. ChatGPT-4 pricing: $10 per million input tokens; $30 per million output tokens - Compared against open-source pricing to show the gap narrowing over time. Perplexity active users: 50 million - Cited as evidence that AI answer engines can already operate at significant scale. Backblaze storage cost per gigabyte (2017): $0.32 - Referenced to show long-term storage deflation. Backblaze storage cost per gigabyte (2022): $0.15 - Shows storage price more than halved over five years. 10-terabyte hard drive price: $195-$200 - Used as a vivid example of falling physical storage costs. LM Sys leaderboard score for Claude 3 Opus: 1255 - Referenced as part of the chatbot arena ranking discussion. U.com Warriors playoff estimate: 48.6% - One cited result shown for Warriors reaching the playoffs. Ringer estimate cited by U.com: 92% play-in tournament; 26% playoffs - Shown as one of the sources surfaced in the U.com answer. DraftKings odds example: Pulled from a sports betting page - Used to show how answer engines and custom demos derive live answers from web results. Tesla stock price cited by Perplexity: $173.81 - Used to test whether the system can retrieve live market data. Tesla stock price cited by ChatGPT: $173.32 - Another live-data comparison point during the demo. Infinity AI output time: About 3-4 minutes - Estimated turnaround for generating the short All-In-style clip. Generated clip length: 9 seconds - The AI-produced Taylor Swift debate clip demonstrated by Infinity AI.

Pivotal Quotes: "If it can do nine seconds, it'll be able to do 90 seconds next year and nine minutes the year after, and then 90 minutes." — Jason / Sundeep discussion: Used to forecast the rapid improvement of AI-generated podcast/video content. "The question is: is that answer correct?" — Jason: Central concern in the debate over AI search engines and answer quality versus presentation. "I think the voice interface is going to be the winning one." — Jason: Summarizes the hosts’ belief that voice will become the dominant AI interaction mode.

Implications: AI search, voice, and generative media are converging quickly. Expect more direct-answer interfaces, less reliance on traditional search, faster content generation, and bigger debates over accuracy, licensing, and manipulation.