The Future of Google | EP 126
Summary
Google’s AI Mode is an attempt to defend a search business that generated $54 billion last quarter by making Google behave more like ChatGPT. The experiment, initially hidden behind Google One AI Premium and Labs, replaces ten blue links with synthesized answers and was announced alongside an upgrade of AI Overviews’ underlying model to Gemini 2.0; if it becomes search’s default front door, Kevin Roose argues, “that’s the ball game.”
The link economy faces a potentially severe traffic shock even if the most alarming estimates prove exaggerated. A TollBit study found AI-search interfaces delivered 91% less click-through traffic than conventional Google searches and pure chatbots 96% less, though Google disputed the methodology; Casey Newton notes that even a halved estimate—roughly 45% less traffic—would still be “a cataclysmic event” for much of media.
Chegg is the episode’s warning for businesses whose paid information can be synthesized directly inside search. The homework-answer service offered more than 100 million answers for about $15 a month and was valued at $12 billion in 2021, but Google’s AI Overviews helped eliminate the need to visit it; now its stock is near penny-stock territory and it is exploring “strategic alternatives.”
The newly signed US crypto structure is narrower than President Trump’s earlier rhetoric: seized Bitcoin already held by the government forms a strategic reserve, while forfeited non-Bitcoin assets enter a separate stockpile. Treasury and Commerce may develop “budget-neutral” ways to acquire additional Bitcoin, but the order says the government will not purchase other crypto assets—materially different from the earlier suggestion that XRP, Solana and Cardano might receive direct federal buying.
Crypto’s political victory is exposing a conflict between price appreciation, ideological independence and insider enrichment. XRP and the other named assets rose after Trump mentioned them; Ripple CEO Brad Garlinghouse had cultivated Trump-world ties, Solana hosts the Trump and Melania meme coins, and World Liberty Financial holds a large token portfolio. Even Coinbase’s Brian Armstrong and Tyler Winklevoss favored Bitcoin alone, while Nic Carter warned against turning Bitcoin into “the plaything of the government.”
The larger systemic risk lies in crypto’s growing connection to conventional savings, not merely in the reserve’s composition. A stablecoin framework that appears capable of passing with bipartisan support, efforts to shift authority from the SEC to the less aggressive CFTC, and ETFs accessible through ordinary brokerage and retirement accounts mean “the walls are coming down”; another crash might therefore reach people who never consciously joined the crypto trade.
Vibe coding is simultaneously expanding software creation and potentially undercutting narrow-purpose paid software. Non-coders built a 400-word-per-minute accessibility reader, an anesthesiology simulator in six weeks, a flavor-pairing tool and reusable photo-processing scripts; the catch is brittle maintenance, weak interoperability and Kevin’s admission that when generated software breaks, “you have no idea what you’re doing.” The durable insight is broader than efficiency: “It is fun to learn and it is fun to make things.”
Deep dive
1. Google is testing the interface that could replace ten blue links
AI Mode is Google’s early answer to ChatGPT and Perplexity: a chatbot-like search experience introduced alongside an upgrade of AI Overviews’ underlying model to Gemini 2.0, initially available only to Google One AI Premium subscribers who also opt in through Labs. Google calls it an experiment, but Kevin sees “a once-in-a-generation chance to reinvent the search experience” through a fundamentally different presentation of information.
Instead of returning a conventional results page, AI Mode writes a bespoke answer and embeds source links through small chain icons and horizontally scrolling website carousels. In Google’s example—“Explain how déjà vu works and how it relates to memory”—the response reads like a short essay, while the four underlying sites sit behind a compact control rather than commanding the page.
That design preserves the formal existence of links without necessarily encouraging clicks. Casey jokes that many users may interpret the chain-link citation as “a fancy period,” making AI Mode less a navigation tool than Google “doing the Googling for you.”
2. Search’s $54 billion quarter makes reinvention an incumbent’s dilemma
Google still handles around 90% of searches and remains internet infrastructure for billions, but it must protect a core business that generated $54 billion last quarter while responding to a product that often answers questions without ads or link-hunting. The risk is not simply losing queries; it is adopting the rival interface while preserving the economics the old interface created.
The competitive movement is small but meaningful. Analysis cited by Casey found Google’s share slipped below 90% during the final three months of 2024 for the first time since 2015, while some analysts expected ChatGPT to reach 1% of search by year-end. Bing, after decades in market, held roughly 4%—making ChatGPT’s path from zero to one notable.
Kevin already uses chatbots for at least half the questions he once took to Google, including product recommendations, appliance repairs and explanatory queries. Google remains his choice for navigational needs such as a train schedule or restaurant menu; Casey likewise favors Google or Google Maps for local businesses but increasingly uses chatbots for trivia and how-to answers.
A Verge survey found 42% of respondents thought search was becoming less useful. The hosts acknowledge several causes—heavy advertising, deteriorating websites and AI-generated “slop”—but argue that direct, one-sentence answers can simply be a superior interface for many everyday questions.
3. Reliability and politics still constrain Google’s rollout
AI Mode’s hidden, paid and opt-in launch reflects the scars from AI Overviews that advised users to eat rocks and put glue on pizza. Casey doubts inference cost is the primary brake because Google says serving costs keep falling; the more immediate concern is discovering “obvious horrible problems” before exposing the product to the full user base.
Google’s warning that answers might accidentally adopt “a persona or reflect a particular opinion” strikes the hosts as evidence of political anxiety. Kevin says Gemini is unusually likely to refuse political questions after earlier image-generation controversies, but calls total neutrality unrealistic: people increasingly seek judgments from these systems, and “on some level you want to have a product whose opinions you stand behind.”
User habits nevertheless change slowly—AOL, Yahoo and perhaps Hotmail retain large populations, Casey notes—so Kevin pushes back against assuming immediate disruption. His answer is the power of defaults: once AI Mode stops being a hidden tab and becomes Google’s front door, users will receive the new behavior without changing habits. That is why today’s experiment may preview Google in “two, three, five years.”
4. AI answers could break the traffic-for-content bargain
Kevin divides queries into two categories. Most users want an answer that is “good enough,” where being 5% wrong in either direction scarcely matters; specialists need the original source because their health, money or job depends on it. As a journalist, he opens The New York Times, The Wall Street Journal or Bloomberg and verifies the paragraph himself—but concedes that this is “a minority use case.”
TollBit estimated that AI-search interfaces produced 91% less click-through traffic than standard Google and pure chatbots 96% less. Kevin questions its calculation, which compared traceable referrals with scraper visits, and relays Google’s objection that a crawler hit is not equivalent to a search. Yet the direction feels intuitive because both hosts rarely open citations when using chatbots casually.
Casey’s stress test is more important than the headline figure: even if TollBit were twice as pessimistic as reality, a 45% decline would devastate much of publishing. Kevin relays Google’s claim that AI Overviews induce more searches overall, but the hosts have seen no indication that the resulting traffic to individual publishers is positive.
The affected set extends beyond journalism. Chegg’s roughly 100 million homework answers once supported a $15 monthly subscription and a $12 billion valuation in 2021; after Google began answering those questions directly, Chegg fell toward penny-stock status and sued, arguing that the traffic rightfully belonged to it and Google destroyed it with AI Overviews. Kevin calls it an unsympathetic but important bellwether.
5. The official Bitcoin reserve stopped short of open-market altcoin purchases
The hosts added an update after the prerecorded conversation: Trump signed an executive order establishing a strategic Bitcoin reserve funded with Bitcoin already held by the government, including assets seized or forfeited through criminal and civil proceedings. Treasury and Commerce may devise “budget-neutral strategies” for acquiring more Bitcoin without additional taxpayer cost.
A separate U.S. Digital Asset Stockpile will retain non-Bitcoin crypto and other digital assets obtained through criminal or civil forfeiture, but the order says the government will not purchase those assets. That distinction resolved a central uncertainty in the prerecorded conversation, which had reacted to Trump’s suggestion that Solana, XRP and Cardano could join Bitcoin and Ethereum in a broader reserve.
David Yaffe-Bellany explains the industry’s informal vocabulary: a “stockpile” means keeping crypto the government already owns, while a “reserve” implies buying more. Retaining seized Bitcoin requires only a policy change; major market purchases would raise funding and congressional-appropriation questions, notwithstanding speculative legal theories that Trump might act alone.
6. The reserve’s strategic rationale remains underspecified
David gives two arguments. The reserve-like case imagines Bitcoin or crypto eventually displacing the dollar as the world’s operating currency, making a large U.S. holding strategically useful. Kevin presses on what the government would practically gain by owning it—whether it could borrow against or spend it—and David concedes, “It is not totally clear how this would work.” Crypto cannot power heavy machinery or feed people in the way some traditional reserve assets can.
The investment case resembles a sovereign wealth fund: buy Bitcoin at about $90,000 and benefit if its value grows dramatically over ten years. That is less an emergency reserve than a directional government wager, sitting awkwardly beside an administration simultaneously promising to cut spending and shrink the state.
Large institutional purchases would ordinarily happen over the counter, directly with counterparties, to limit market impact—not through a Treasury official clicking “buy” on Coinbase. But before the order, nobody had explained how broader purchases would be financed or reconciled with promised savings.
7. Altcoin inclusion made conflicts of interest impossible to ignore
Trump first named XRP, Solana and Cardano, then added Bitcoin and Ethereum almost as an afterthought. XRP is associated with Ripple; Solana competes with Ethereum and is the platform on which the Trump and Melania meme coins were built; Cardano came from a figure involved in Ethereum’s early movement. David’s distinction: these are established altcoins, “not some random coin that got invented a day ago,” but establishment does not itself create strategic national value.
David says the market consequence arrived before any policy: the named assets rose after Trump’s post, though some gains later reversed. He compares the symbolic effect of a U.S. government seal on Cardano to Trump announcing a strategic Facebook stock reserve after spending time with Mark Zuckerberg.
Ripple CEO Brad Garlinghouse spent time at Mar-a-Lago and was a major supporter of Fairshake, the crypto PAC. Separately, the Trump family’s World Liberty Financial accumulated a large portfolio of tokens while its promised DeFi application had yet to materialize, giving the family potential exposure to policy-driven appreciation.
David’s description of the industry’s reaction is deliberately unsparing: participants “love enriching themselves” but are also paranoid about schemes that might favor someone else. Some ignore conflicts when SEC cases disappear, then discover the principle when a rival coin enters the proposed reserve.
8. Bitcoin loyalists resisted becoming Washington’s favored trade
Coinbase CEO Brian Armstrong said a Bitcoin-only reserve “would probably be the best option,” and Tyler Winklevoss also opposed including other coins. Part of the objection is quality—Bitcoin has the longest record and greatest value—but part is ideological resistance to turning a renegade monetary project into a government-managed basket.
Crypto advocate Nic Carter opposed even a Bitcoin reserve, warning that it “would transform Bitcoin from an apolitical asset into the plaything of the government, subject to Washington’s political cycles.” David notes the contradiction in libertarian holders asking the state to accumulate Bitcoin so their asset appreciates: most of the industry may no longer prioritize those principles, “but some still do.”
The internal dispute matters because it prevents a simple pro-crypto versus anti-crypto reading. Federal support creates immediate gains, yet it also subjects assets to political rotation, lobbying and suspicions that well-connected insiders—not technological merit—determine which ticker receives privileged treatment.
9. Quiet legislation may connect crypto more deeply to ordinary savings
One major legislative priority is a stablecoin bill that appears capable of passing with bipartisan support. It would establish rules for stablecoin companies operating in the United States and effectively grant that branch of crypto a government seal of approval. Critics described it as a potential “gateway drug,” not because the framework itself is necessarily extreme, but because it could ease broader crypto integration.
The second priority is a market-structure bill that would transfer authority from the SEC to the CFTC, which David characterizes as weaker and less aggressive. The personnel context has changed too: Gary Gensler had taught a crypto course at MIT and was running the SEC, but he is now out and “the crypto people have the run of the town.”
Skeptical expertise has also faded through exhaustion. Critics warned during the previous boom, watched the market crash and Sam Bankman-Fried go to prison, and felt vindicated—only to see crypto return politically stronger. Kevin worries that the remaining policy conversation is dominated by people whose investments gave them reason to master L2s, DEXs and stablecoins.
Kevin’s broader warning is that “the walls are coming down.” A Bitcoin ETF already exists, and XRP or Solana ETFs could follow, creating more avenues for people to put traditional savings into crypto. Kevin compares the information asymmetry to mortgage-backed securities and credit-default swaps before the financial crisis; David adds that ETFs let people buy crypto from ordinary brokerage accounts and factor it into retirement investments.
10. Vibe coding turns personal frustrations into software—with maintenance debt attached
More than 60 listeners responded after the hosts’ earlier vibe-coding segment. Casey built a working platform game about a bathrobe-wearing frog; Casey reported that Kevin’s Hot Tub Time Machine began emailing daily and prescribing quarterly maintenance after one week. Kevin’s core limitation: when generated software fails, “you have no idea what you’re doing” beyond asking the model to fix it.
Mike’s accessibility reader flashes one word at a time at about 400 words per minute, versus the roughly 250 he cited for “normies,” helping dyslexic ADHD friends avoid jumbled text and lost focus. Though not a coder, his product-management experience transferred directly: “The better I defined the project up front, the better the results.” A loose specification produced laborious iterations, much like managing a human developer.
An anesthesiologist, Lauren, began building Consult Craft with Claude in November 2024. Six weeks of spare-time work produced a functional simulator for anesthesiology case discussions; about three further months added cases, feedback and a transcript database for research. Intended for residents preparing for oral-board exams, it also became a refresher for experienced colleagues who had not handled a particular case recently.
Other projects exposed both leverage and friction: Flavor Finder combines The Flavor Bible’s data with a color-palette-style interface to generate five-ingredient flavor pairings, while photographer Zach used ChatGPT, Claude and DeepSeek to rename and organize nearly 1,000 images and export layered Photoshop files as JPEGs and TIFFs. Debugging took perhaps twice as long as manual work, but left a reusable script—and reinforced Casey’s conclusion that “it is fun to learn and it is fun to make things.”
Ashley’s proposed “bit of a wife” would ingest invitations and children’s social logistics from email, WhatsApp, texts, Evites and Messenger, then summarize tasks and update a shared calendar for a working mother with children aged nine, six, four and two. Kevin pledged to attempt it, while acknowledging the hardest constraint: generated code cannot easily connect services that lack compatible interfaces.