Summary

Google’s I/O pitch rests on efficiency, distribution, and product breadth: Gemini 3.5 Flash is four times faster and much cheaper than leading frontier models, while AI Mode accompanies what Google calls Search’s biggest change in 25 years. Kevin saw a strategy built to serve billions; Casey countered, “Fast and cheap is what you talk about when it isn’t the best,” with much of the promised utility still “coming this summer” or limited to trusted testers.
Sundar Pichai candidly conceded that Google is a bit behind in agentic coding, tool use, instruction following, and long-horizon work on complicated codebases. He argued that internal Antigravity usage is “doubling every week,” producing the data needed to close the gap, and that in today’s model race “30 to 60 days look like five years.”
Google plans to evolve Search methodically rather than abruptly retire its classic interface, preserving sources and links while expanding AI Mode. Pichai expressed confidence that greater user value will support some combination of advertising and subscriptions: “Adam Smith’s rules don’t change in this new world.”
Public acceptance is a strategic constraint Google must address beyond model benchmarks: a cited New York Times/CNN poll found 16% of people calling AI mostly good versus 35% mostly bad. Pichai called the anxiety natural because “humans aren’t evolved to process that much change,” acknowledged disruption and economic concerns, but rejected overly deterministic doom scenarios.
Google is deliberately constraining consumer agents until they earn trust, because an unexpected action could make users back off and the systems can be hacked. Pichai compared the rollout to persuading someone to enter the back seat of a self-driving car: users need control and transparency before Spark can progress from meeting preparation and calendar organization to consequential tasks.
Pichai described current agent systems as a continuum toward recursive self-improvement, not RSI itself. Antigravity can build a simple operating system in over 12 hours rather than the “multiple thousands of hours” a person might need, but “we aren’t quite there yet”; if that threshold approaches, he wants consultation beyond individual labs and no “race conditions” around AGI.
Selling TPUs to rivals is a platform-economics decision, not a concession that starves Google’s own models. Pichai said first-party AI and Google Cloud have separately planned capacity and cash flows, while outside researchers improve future hardware; he still could not choose between a three-to-five-year and five-to-10-year AGI horizon, but progress over the last one to two years made him feel it was “on the closer side than not.”
Meta’s response to falling behind is an unusually large organizational bet: roughly 8,000 layoffs, or 10% of workers, alongside 7,000 employees reassigned to AI initiatives. The hosts expect morale and surveillance concerns to intensify, but Casey’s thesis was stark: helping Mark Zuckerberg win the AGI race is now “the entire point of the company,” with “not a secondary objective.”

Deep dive

1. Google flooded I/O with AI, but much of the utility remains prospective

Casey treated the claimed “biggest change to Search in 25 years” cautiously: the demonstration initially looked like a box that grew with longer queries. Custom interfaces and other substantive changes were promised for summer, reinforcing his rule that productivity software cannot be judged “until it’s in your hands.”
Kevin’s standout was Google’s emphasis on cost and speed. Gemini 3.5 Flash was presented as four times faster and much cheaper than other leading frontier models, suggesting a strategy of efficient inference, massive distribution, and acceptable capability rather than waiting for undisputed benchmark leadership.
Casey’s pushback — worth keeping: Google surely wants the best model but “just hasn’t built it yet,” and “fast and cheap is what you talk about when it isn’t the best.” He also questioned which agent could reach a billion users; even Google’s version of OpenClau lacked an obvious mainstream task he wanted performed overnight.
Agentic Search was presented as a more powerful version of Google Alerts, such as notifying users about new Zillow listings or baseball scores. Casey’s own test—to alert him if Kevin wrote a good story—had produced nothing. The broader I/O approach still looked like “spray and pray”: Search agents, coding agents, video tools, and an omni model, capped by Demis Hassabis declaring that we were in “the foothills of the singularity.” The crowd was unusually receptive: Casey noted that I/O produced nary a boo at AI, unlike many recent gatherings.
Kevin saw more immediate value in Gemini Spark running agents inside a cloud virtual machine, removing the need to leave a laptop open.

2. Gemini 3.5 Flash wins on throughput, not power-user conversion

Kevin’s first Antigravity tests confirmed exceptional speed and fewer immediate token-limit problems, but not a dramatic capability jump. He noted early indications that Gemini 3.5 Flash might underperform Gemini 3.1 Pro on some coding evaluations, while stressing that the evidence was preliminary.
The pricing story was less clean than the keynote implied. Flash remained cheaper than other frontier models but was significantly more expensive than the preceding Flash version; for a company consuming billions of tokens daily, its throughput could matter, yet Kevin found no reason to make it his personal “daily driver.”
Casey prefers waiting and paying for the strongest answer, and said vocal AI power users appeared disappointed. His real test comes with 3.5 Pro next month; until then, I/O was a “base hit”—solid if everything ships as advertised, but without a product that created immediate “extreme physical distress” from being unavailable.

3. Pichai conceded Google’s coding gap and attributed some swings to the frontier’s cadence

Pichai described Google as frontier-level in text, multimodality, voice, audio, reasoning, and overall intelligence, but “a bit behind” in agentic coding, tool use, instruction following, and long-horizon tasks. He also claimed Google was the only large company actually at that frontier.
Coding matters because it is “very foundational” across Google. Pichai distinguished the company’s strength in single-shot web front ends from its weakness when serious developers work for long periods across complicated codebases: “There is a gap to the frontier where others are.”
His proposed catch-up mechanism is usage data. After getting Antigravity with 2.0 into Google’s internal workflows, token consumption began “doubling every week”; Pichai argued that those feedback loops supply the data flows Google previously lacked and help it “hill climb” on practical coding performance.
Asked about first-day pricing and quality complaints, Pichai focused on regressions, post-training fixes, and usage limits tightened to avoid outages. He promised quick improvement, while framing temporary swings in who is ahead as normal pretraining-cycle mismatches: “30 to 60 days look like five years.”

4. Search will become more agentic without abandoning the web or its economics

Pichai resisted declaring a date when classic Search or the 10 blue links disappear. Google intends to bring users along “methodically,” preserve speed and connections to the open web, and follow long-term usage metrics; AI Mode did not exist a year earlier but had already become much easier to enter.
Kevin relayed Casey’s claim that he had basically not performed a traditional Google search for a year, then asked whether that behavior threatened Google’s pretty good advertising business. Pichai’s answer was that agentic Search should accomplish far more for users than Search could 10 years earlier, increasing the total economic value available.
More competition and choice do not change his underlying monetization thesis. Pichai expects the right mixture of subscriptions and ads to emerge as user value rises, concluding, “Adam Smith’s rules don’t change in this new world.” Sources and links, he said, will remain part of the product.

5. Pichai treats the AI backlash as rational, not as a messaging defect

Kevin cited a New York Times/CNN poll in which roughly 16% called AI mostly good and 35% mostly bad. Pichai’s response began with scale: AI is “the most profound technology humanity will ever work on,” is advancing extraordinarily fast, and exceeds humans’ evolved capacity to absorb continuous change without anxiety.
The industry can demonstrate more concrete benefits and improve how it scales infrastructure, but Pichai said the deepest concern is economic. He considered public engagement healthy: in democracies, citizens need to understand the transition, express preferences, and cause society to act while the technology is still moving rapidly.
His optimistic case starts with changed baselines. Spreadsheets transformed who could perform financial analysis; AI coding tools should similarly let many more people program. Pichai expects higher productivity and more leisure, although he explicitly refused to minimize the disruption accompanying every major technology shift.
Healthcare supplied his nonlinear example: doctors trained to care for patients spend too little time with them, while AI could help them spend more time with patients. His scans contain roughly 10 times the information his father’s did, and he expects another 10-fold increase in 10 years—creating demand that AI will be needed to keep up with.

6. Agents must earn trust before Google lets them become consequential

Pichai has primarily used Spark professionally to prepare for meetings—including Hard Fork—and recently tested it on his personal calendar. Asked to categorize upcoming meetings, it proposed two color schemes, awaited his selection, then recolored personal, health, and work events: “It’s actually sci-fi.”
That modest task illustrates Google’s staged approach. Pichai compared agent adoption with getting someone into the back seat of a self-driving car: an unexpected action could make users retreat, so systems need visible control and transparency before permission expands.
Security is the harder boundary because agents “can be hacked.” Pichai said Google does not want to get ahead of the safety frontier in the wrong way, even as Spark moves toward general users later in the summer.
On prospective government oversight, he withheld judgment until the executive order’s full details appeared. Cross-industry coordination made sense for cyber exploits affecting agencies, but oversight should not unnecessarily slow national progress at the frontier; he allowed that the balance might need to change at more advanced capability levels. He also pointed to Google’s building and open-sourcing Synthy and forming a consortium as an example of industry coordination.

7. Recursive improvement is not here, but Google is planning for the threshold

Pichai said Antigravity can build a simple operating system from scratch in over 12 hours, work he said would genuinely take “multiple thousands of hours” for somebody. Agents, sub-agents, and orchestration already create a continuum of machine-assisted development, but this is not yet recursive self-improvement as commonly described.
If RSI arrives, it would represent “a next level of acceleration” with broad implications. His proposed protocol was deliberately non-corporate: it “shouldn’t be an internal conversation,” responsible labs should consult much more widely, and the industry must avoid race dynamics near advanced stages of AGI.
Google’s decision to sell TPUs externally does not, in Pichai’s framing, reduce first-party supply. Google DeepMind and internal services plan against company cash flows, while Cloud plans against its own revenue; without the external business, Google would not necessarily be planning those chips anyway.
External TPU users, including Anthropic researchers, help Google improve subsequent hardware, while economies of scale strengthen the platform. Google also uses NVIDIA chips, which Pichai called “incredible,” making the infrastructure strategy multi-platform rather than an attempt to hoard one accelerator architecture.

8. AGI moved closer as the surrounding industry absorbed legal and cultural shocks

Pichai called progress toward AGI “inevitable,” clarifying that even if AGI itself takes 10 years, systems three years from now will be powerful enough to require preparation today. He cannot confidently choose three-to-five versus five-to-10 years, but the last one to two years shifted him “on the closer side than not.”
He interpreted Hassabis’s “foothills of the singularity” as singularity meaning the advent of AGI, recalling, with the hedge “if I remember correctly,” a forecast of “by 2030 or so.” Pichai’s conclusion was not certainty but responsibility: frontier builders should articulate what they believe so society can internalize the possibility and prepare.
Elon Musk’s OpenAI claims were unanimously rejected after less than two hours because the three-year statute of limitations had expired, not because jurors resolved the merits. Musk’s lawyer said he would appeal, while Casey highlighted trial evidence that Musk had previously explored making OpenAI for-profit and absorbing it into Tesla—the conduct he later attacked.
Meta laid off roughly 8,000 people, or 10% of its workforce, while reassigning 7,000 to AI tools and apps. The hosts saw both a serious catch-up mobilization and a recipe for poor morale, particularly as Zuckerberg defended training computer-using and coding models on employees’ own work.
Publishing supplied the complementary trust crisis: readers accused “The Serpent in the Grove” of AI generation, though the hosts stressed “we don’t know” and the Commonwealth Foundation stood by the assurances given by the author. Separately, Steven Rosenbaum acknowledged fabricated or misattributed quotations in The Future of Truth after using ChatGPT and Claude; Kevin’s verification rule was blunt: “Most is not enough.”

Chip design from the bottom up – Reiner Pope Google I/O 2026, Karpathy Joins Anthropic, and Cerebras’ $95B IPO | EP #256