DeepMind's Ed Chi: Future of Personalized General-Purpose Assistants

Summary

• Ed Chi says general-purpose assistants are still at a very early stage, and expects their relationship and interface with AI Mode Search to become clearer in 6–12 months. Permissions and reliability are the sticking points: people worry that giving an assistant their credit card will lead to a spree of unauthorized purchases. There is still no real-world data showing whether assistants actually book flights correctly or get it wrong. His benchmark for maturity is the Grandma Test: “When your grandma points at your robot and yells, ‘How can you be so stupid?’（你怎么这么笨）, you’ll know the era of general-purpose assistants has really arrived.”
• He argues that economics ultimately comes down to capability over cost, and that cost control could become the most important moat if model capabilities become widespread. Google’s advantage is its full stack—from models and applications to hyperscaling, token factories, chips, and energy—and its ability to deliver tokens at the lowest cost. Ed also notes that many people are not focused on cost yet.
• He says Google cannot walk away from model development because users expect it to have its own model technology. The smaller models released in recent months are deliberate: general assistants and personalized applications need small models, while larger models are also in development. He says some AGI-pilled tech bros in the Bay Area believe “software took over the world” and are drawn to ultra-size models; personally, he does not want every mobile app running a huge model, given the energy costs of computation.
• He says it is too early to call either the commoditization of capabilities or the trajectory of the gap between open- and closed-weight models. Open models have caught up in part by using methods from the Chain-of-Thought paper he co-authored to study and replicate reasoning. Large players are trying to hide their reasoning chains to make copying harder. It is also unclear whether capabilities such as multimodality will become commoditized. The market is not one big bubble, though there may be bubbles in specific areas; Physical Intelligence may face unrealistic expectations, and many models are poor at visual reasoning.
• If he were starting a company today, he would expect more value creation to flow to applications and token factories as gaps in model capability narrow. On the application side, the key factors are user lock-in, stickiness, and experience. For hyperscalers, the open questions include energy use, chip design, the relationship between memory and computation, and how much compute will happen on-device.
• Memory is an underappreciated bottleneck, and it involves ranking as well as storage and retrieval. Transformer context buffers of 1–2 million tokens are nowhere near enough, in his view; from a cognitive perspective, memory is also a kind of program. On shortages of DRAM, NAND, and HDDs, he says memory and computation can substitute for each other, while long-term investment planning can reduce concern about temporary shortages.
• He expects the next wave of commercialization to come from consumer products: much of Anthropic’s and OpenAI’s revenue is driven by enterprise subscriptions, while consumer services are one reason he stayed at Google. Reasoning has advanced from solving grade-school math problems with 33% accuracy 3–4 years ago to tackling mathematical problems such as Navier–Stokes, but remains concentrated in specific domains. More people played Go after AlphaGo, and Google has seen search usage rise, not fall, in the large-language-model era. Solving software problems, he argues, does not mean every problem is solved.

Deep dive

1. Reasoning reaches a peak in 3–4 years: from grade-school math at 33% to Navier–Stokes, but still concentrated in specific domains

• Ed Chi’s starting point: for the past 30 years, and especially the first 20, much of the focus was on recommendation and search. Only over the past 10 years, with neural networks, next-token prediction, and Chain-of-Thought, have researchers begun to explore machines’ basic cognitive and reasoning abilities. The next academic questions include more complex reasoning, memory, and RSI (recursive self-improvement).
• The leap he describes is striking: when he wrote the CoT paper, models could solve grade-school math problems with just 33% accuracy. Today, the largest models can solve mathematical problems from the Millennium Prize Problems, such as the Navier–Stokes equations. In his view, reasoning reached a peak in just 3–4 years.
• But he immediately qualifies that those reasoning abilities remain concentrated in specific domains. He uses his own interest in golf as an example: how well do models understand and reason about golf’s principles, clubs, balls, and equipment? “These are all different from math,” he says; before building applications, it is essential to understand how well models reason across different fields.

2. General-purpose assistants are at a very early stage: the Grandma Test and where the interface is headed in 6–12 months

• Host 曹卿云 raised the personal-assistant discussion prompted by Instinct and the newly released Muse Spark. Ed’s framing: Google’s AI Mode Search is a consumer service, and search has in fact become one of the largest categories of large-language-model applications, evolving into experiences like AI Mode that combine search and answers. Muse and Instinct aim to move beyond answering questions to booking flights and hotels and buying things.
• Permissions and reliability are the sticking points: people worry that giving an assistant their credit card will lead to a spree of unauthorized purchases. In early 2026, experiments like OpenClaw prompted companies large and small to start testing the waters. But there is still no real-world data showing whether assistants actually book flights correctly or get it wrong. He expects the relationship and interface between general-purpose assistants and AI Mode Search to become clearer over the next 6 months to 1 year.
• On product moats, he emphasizes user experience and tolerance for error: one serious mistake can change a consumer’s view of a product. At another podcast in Taiwan, he introduced the Grandma Test: one day, when your grandma points at the robot and yells, “How can you be so stupid?”（你怎么这么笨）, you’ll know the era of general-purpose assistants has really arrived. Grandmas have high expectations of these assistants, so a serious mistake can make the product seem broken; the industry is still exploring that boundary.
• Asked which company might produce the next breakout product, he said the source may not matter as much as how the product reaches users and what the right human-machine interface looks like. A startup could build a product and then be acquired by a large company. He hopes the industry can sustain a healthy ecosystem where everyone has room to experiment.

3. Google will not abandon model development; over the long term, the measure is capability over cost

• Asked about DeepMind personnel changes and the pace of model releases, Ed said Google has not stopped developing models. The smaller models released in recent months are designed for general assistants and personalized applications; larger models are also in development. He described the team as rich in talent, but said many people want to build large models and resources are never enough.
• He said Google users expect the company to own the technology rather than rely on someone else’s model or an open-weight model. From that perspective, Google cannot walk away from model development.
• There is no single right answer to how resources should be divided between small and large models. He linked the ultra-size-model boom to the beliefs of some AGI-pilled tech bros in the Bay Area: they believe “software took over the world,” and that getting coding and software right is enough to take over the world. Personally, he does not want every mobile app running a huge model, given the energy costs of computation.
• In response to the host’s question about capex and cash-flow pressure, Ed distinguished between short- and long-term investment horizons. In the short term, companies focus on quality and accuracy; investing with a 5- or 10-year horizon means deciding which metric to optimize—not just capability, but capability over cost. Google is pursuing multiple paths, innovating across models, applications, chips, and the full stack, and can deliver tokens at the lowest cost. If model capabilities become widespread, cost control could become the most important moat.

4. Memory involves ranking as well as storage and retrieval; a context buffer of 1–2 million tokens is nowhere near enough

• On potential shortages of DRAM, NAND, and HDDs, Ed cited research showing that memory and computation can substitute for each other. With a long-term investment plan, he said, there is less reason to worry about temporary memory shortages because computation can be used instead. Memory research also needs to account for how much memory and bandwidth different methods require, and whether neural-network designs should reduce memory requirements. He cited Timi’s Delta Attention Model, which uses computation and memory differently from full attention, O(n²) attention, or sparse attention.
• That is only the memory challenge for sequential Transformer models; personalized applications, KV cache, and everyday notions of memory raise their own questions. From the perspective of human cognition, memory is also a program: it is not inert, but constantly changing, guiding what happens, and helping program AGI machinery. Research therefore also examines the underlying use of memory and how information is organized in the context buffer.
• He identified 2 bottlenecks. Transformer context buffers currently reach around 1 million or 2 million tokens, which he considers nowhere near enough. More fundamentally, people often treat memory as a retrieval mechanism for storage and recall, rather than as part of human cognition and thinking.
• His correction: “Memory is really a ranking problem.” When someone mentions Paris, for example, people recall the most important memories associated with a trip there—not everything related to Paris all at once.
• Asked whether startups building memory layers can find a place in the market, he said there are new ideas in hardware beyond the large companies, as well as work underway at those companies. The relationship between memory and computation could see new developments over the next few years.

5. Narrowing capability gaps shift value toward applications and token factories; open models’ progress owes something to CoT methods

• His view on where to start a company: model capabilities are showing some signs of commoditization, but are not fully commoditized. If the gap between the best and worst models narrows, more value creation will flow to applications and token factories. On the application side, the focus is user lock-in, stickiness, and experience; on the hyperscaler side, the questions include energy use, chip design, the relationship between memory and computation, and how much compute happens on-device. These remain open questions.
• On open models catching up with closed models, he said one reason is the Chain-of-Thought paper he co-authored: researchers use CoT methods to surface models’ reasoning chains, study how they think, and replicate their reasoning. Large players are now trying to hide those chains to make copying harder; capabilities such as multimodality may also be difficult to replicate. He said it is too early to treat capability commoditization as settled: some capabilities may become more commoditized through distillation, while others show no clear signs of commoditizing yet. There is still plenty of debate.
• Asked whether the AI market is in a bubble, he said he does not think the entire market is a bubble, though different areas may have their own bubbles. Physical Intelligence, for example, may face unrealistic expectations. He also noted that many models have very poor visual reasoning capabilities.
• He said reality and belief do not always align. If you believe that solving software means solving every problem, then much of what he discussed earlier does not apply. But if you think solving software still leaves plenty of other problems to tackle, “there’s still plenty left to play for.”

6. The next wave is consumer: hardware and human-machine interfaces remain open territory

• On opportunities beyond coding, Ed said much of Anthropic’s and OpenAI’s revenue is driven by enterprise subscriptions, and he expects consumer to rise next. One reason he stayed at Google is that consumer services have always been one of the company’s strengths, though he does not rule out new categories emerging.
• Looking across the industry, he said future products could take the form of a super app or involve a wholesale change to the phone. Glasses, earrings, headphones, and even clothing could all change. Consumer hardware still offers plenty of room to explore, including how on-device and off-device models interact; Apple has introduced the concept of hybrid intelligence.
• He recalled that after 4 years of smart speakers in the home, it suddenly seemed as if no one was building them anymore. People once talked about smart fridges; from the perspective of a personal general-purpose assistant, putting a screen on the fridge and connecting it to the internet is now enough to make it smart—much simpler than 7 or 8 years ago. In the era of general-purpose assistants, people should not still have to download apps for their robots all day; that was a concern from 15 years ago.
• On AI and education, he drew an analogy to AlphaGo. Nearly 10 years ago, when Google developed AlphaGo, many people said that after it played a Go champion, humans would stop playing Go. Instead, more people took up the game and wanted to study it. The same could happen with math: AI solving Navier–Stokes does not necessarily make people less interested in learning math; it could draw more people in. The motivation for research may shift from pushing the boundaries of knowledge toward understanding, entertainment, and enjoyment, with more people researching out of curiosity.
• He also noted that Go apps lowered the barrier to learning and encouraged more children to play. Since large language models emerged, Google has seen search usage increase, not decrease. AI will change existing structures and may create new opportunities; in both enterprise and consumer markets, “creativity seems to be everywhere.”