Summary

Brin sees AI as the fastest-compounding technological transition of his career, with progress that “dwarfs anything we’ve seen.” Unlike the early web, whose distribution accelerated while its underlying technology changed more gradually, AI systems change substantially month to month. That pace pulled him back from a retirement he began roughly a month before COVID, after an OpenAI contact called it the greatest transformative moment in computer science.
The near-term economic leap is not full AGI but machine intelligence operating at a volume no individual can match. Brin contrasted summarizing 10 search results, which he could do himself, with reading 1,000 results, launching follow-up searches and analyzing them deeply—“a week of work for me.” Post-training and thinking models have delivered another major step, and he said, “We don’t really know what the ceiling is.”
Software development and management already look like high-value automation markets, even inside Google’s bureaucracy. Brin overturned an internal rule putting Gemini on the prohibited list for coding, while Google now tests its own and external tools such as Cursor for productivity. He agreed that management is “the easiest thing to do with AI” and described an internal system that summarized team chats, assigned work and surfaced an overlooked engineer for promotion.
Brin expects model capabilities to converge into increasingly general systems, with specialization offering iteration and sometimes size, speed and cost advantages. Machine learning moved from separate vision, speech and text architectures toward transformers and “increasingly…just becoming one model”; discoveries from specialized models can usually be folded back into the general model. Open models remain an unresolved competitive variable: DeepSeek “closed the gap” in January or so, while Google’s smaller Gemma models run on one computer but remain less capable than Gemini.
Robotics may finally benefit from stronger software, but Brin is less bullish on humanoid form factors while refusing to discount them. Google acquired and later sold roughly five robotics companies because “the robots are all cool,” while the software was not ready to make them truly useful. He argues AI can learn through simulation and real-world experience without matching humans’ exact limbs, although “a lot of really smart people” pursuing humanoids keep him from dismissing the category.
AI has made education and career planning radically uncertain without giving Brin a confident prescription. With children in high school and middle school, he said AIs are already ahead in many areas, though they make mistakes humans would never make; they can win math and coding contests against some top humans. He wants his children to do what they like, choose something challenging and learn to overcome problems. On college, he said he lately did not think they should go, but then viewed his son’s interest in an SEC school as potentially valuable for social adjustment, handling failure and exploration.
The interface and pricing stack point toward voice, glasses, vast context and rolling access to prior-generation capability. Brin called unlimited context valuable, potentially spanning Google’s codebase, and said Google Glass was early because the technology was not ready, though current glasses are more sensible aside from battery-life issues. He attributed better voice experiences to smaller, faster models and stacked speech systems such as Whisper and ElevenLabs. Gemini 2.5 Pro offers limited free prompts and costs roughly $20 a month for heavier use; he expects top models to remain initially supply-constrained, while one step down in hardware cost could make the whole service free.

Deep dive

1. AI’s compounding pace pulled Brin out of retirement

Brin retired “in theory” about a month before COVID, planning to read physics in cafés. After running into Dan from OpenAI at a party, he heard: “This is the greatest transformative moment in computer science ever.” Brin had already started returning to the office and concluded Dan was right.
Comparing AI with Mosaic-era excitement, Brin recalled when the internet’s “What’s New” page could list only two or three new sites. The web spread rapidly but changed less technically from month to month or year to year; AI systems themselves change “quite a lot.”
His hands-on work began around pre-training—the compute-intensive phase most people simply call training—and shifted toward post-training as thinking models emerged. Those models delivered “another huge step up,” leaving the ceiling unknown.
Brin said current systems are not yet AGI or superhuman intelligence, but can surprise users. His test for an AI superpower is volume: processing 10 search results saves time, but processing 1,000, reading them deeply and generating follow-up searches replaces a week of human work.
Calacanis then described asking Gemini to estimate F1 deaths per mile rather than per decade. After being told to make its best attempt, the system proposed including 100 practice miles for every mile on the track and produced an estimate in minutes; Calacanis cross-referenced it and compared the result to an undergraduate term paper.

2. Human capital becomes harder to price before education adapts

Brin’s honest non-answer on parenting was, “I don’t really know how to think about it.” He has one child in high school and one in middle school, and said AI is already ahead in many areas. It can make mistakes a human would never make, but is “pretty damn good” at math and calculus and can win math and coding contests against some top humans.
Rather than optimize his children around a forecast he does not trust, Brin wants them to do what they like, face challenging problems and learn to overcome them. He explicitly said he does not know whether people can plan their lives around what AI will be in a year.
Calacanis noted that questions about college’s cost, vocational value and usefulness were already building before AI. Brin first said that, lately, he did not think his children should go to college; then, discussing his son’s desire to attend an SEC school for its culture, he decided that social adjustment, dealing psychologically with failure and having a few years of exploration could be “the best thing” to do.

3. Better software changes both robotics and programming

Google acquired and later sold roughly five robotics companies, including Boston Dynamics, and also built Everyday Robots internally before later transitioning it. Brin’s recurring conclusion was that “the robots are all cool,” but the software was not yet ready to make them truly useful.
His humanoid skepticism targets the core premise: copying the human body helps in a human-designed world and enables training on videos of people, but it may under-credit AI’s ability to learn through simulation and real-world experience. Wheels, arms and legs need not match humans exactly—though Brin explicitly would not discount the many smart teams pursuing humanoids.
Inside Google, Gemini had appeared on an internal list of tools employees were not allowed to use for coding. Brin fought the restriction for “a shocking period of time,” then got it fixed. Separately, when Calacanis argued that a junior person being able to push back on the founder was evidence of healthy culture, Brin agreed.
Google is now trying internal and external coding systems—including tools such as Cursor—to learn what raises productivity. Brin said the tools already make him more productive.
Brin also agreed that management is “the easiest thing to do with AI.” Using an internal chat-space tool, he had it summarize a full discussion and assign work, initially pasting the output back without telling colleagues it came from AI. When he asked who should be promoted, it identified a young woman engineer whose pull requests were excellent even though she was not especially vocal. Her manager agreed with the assessment, and Brin thought the promotion ultimately happened.

4. General models keep absorbing specialized breakthroughs

Brin’s base case is convergence: convolutional networks were used for vision and RNNs for text and speech, but much of that work shifted to transformers. The direction is increasingly toward “one model” handling languages, images, video and audio.
Specialized models still provide scientific leverage because teams can iterate against one target without solving every modality simultaneously. They can also be smaller, faster and cheaper, but successful capabilities are generally transferable back into a general model—so “the trends have not gone that way.”
Open versus closed remains unresolved. Brin credited DeepSeek’s January-or-so release as “a really surprisingly powerful model” that closed the gap to proprietary systems. Google’s open Gemma models are small and dense enough for one computer, perform well and remain explicitly less powerful than Gemini.

5. Context, compute and interfaces define the next bottlenecks

Brin endorsed effectively infinite context, including access to Google’s codebase, and imagined multiple persistent sessions running at once. He said there is “no limit” to the usefulness of larger context.
He added that for almost any compelling new AI idea, there are probably “five such things internally”; the real question is how well they work. Google is pushing the bounds of intelligence, context and speed.
Gemini mostly uses Google’s own TPUs, while Google also supports and purchases NVIDIA chips and offers them through Google Cloud. Hardware is not yet abstracted away: model-scale computation still depends materially on the exact chip, memory behavior and communication between components, and AI is not yet good enough to reason through all of that itself.
Brin said he got Google Glass’s timing wrong and was early because the technology was not ready. Current glasses are more sensible and a cool form factor, although battery life remains a problem.
Calacanis said voice mode had gone from unusably slow to fast enough for interruptions and follow-ups. Brin attributed the improvement to more capable smaller models and better inference, while noting that speech systems can be stacked: Whisper is strong at some tasks and ElevenLabs is an exceptional text-to-speech stack.
Brin recommended the dedicated Gemini app and said Gemini 2.5 Pro provides a few free queries before heavier use costs roughly $20 a month. He expects the top models will not be supplied infinitely at launch, but said a later generation can arrive after a few months; one step down in hardware cost, he argued, could make the whole service free.

The RAISE Act: Minimum Standards for Frontier AI Development, with NY Assembly Member Alex Bores Market Overview May 20, 2025