53. Why Wasn’t Google the One to Build Muse? — Lewis | Microsoft / ex-Gemini
Summary
Muse’s lead is first and foremost a productization win, not a breakthrough in underlying technology. Tools such as Claude Code could already connect to Gmail, but Muse turned “check subscriptions, chase hospital emails, handle refunds” into clickable ideas, giving ordinary users their first concrete sense of what an Agent can do. Lewis’s summary: “It’s not technology-led; it’s product-led.” Asked why Google, which controls the 2 critical data gateways of Gmail and Calendar, did not get there first, he answered directly: “Google is too slow,” arguing that Google has long been reactive, responding to the market with countermeasures.
Early in the foundation-model war, Google treated OpenAI as its primary rival, concentrating resources on consumer products, model benchmarks and “fancier demos”—and leaving Anthropic an opening through core engineering and enterprise. Lewis describes Anthropic’s bet this way: “We’re not going to do all the fancy stuff you’re doing. We’re going to bet on core engineering and bet on enterprise.” Once that strategy was validated, OpenAI and Google shifted toward coding and enterprise products, showing how product choices can change the competitive order.
The real challenge for enterprise Agents is not the chat window but the combination of memory, knowledge, permissions and a context graph. Microsoft’s advantage is the enormous volume of PPTs, documents and internal data accumulated in SharePoint, but employees, 2-person conversations, teams and companies all have different permissions; whether a successful trace can become procedural memory—and whether only safe content can be promoted from the individual layer to the team layer—will determine whether Agents evolve from personal toys into enterprise infrastructure.
For Personal Agents, the key data gateways are Gmail, Calendar and long-term behavioral data, making privacy and trust a core barrier. Lewis declined to try Instinct because its terms allowed it to use the data it accessed, but was willing to use Muse; he also cautions that distrusting Zuckerberg does not mean one should automatically trust a newly founded startup. “Basically, who you are and what you do is all in there”—the source of personalization value as well as the central risk of a Personal Agent.
Open APIs in the US gave Muse a window, while China’s walled gardens make the Agent opportunity more dependent on super-apps. Google’s principle is closer to “the data belongs to the user,” allowing third parties to call Gmail; platforms such as WeChat and Xiaohongshu keep data siloed, leaving users facing “10 people standing there” and forcing them to ask each one separately. WeChat has the best shot at a general-purpose Personal Agent, while other players could build vertical Agents around search, consumption and interest data—but China’s fragmented data makes a unified assistant across life, work and preferences harder to build.
For To C, a $200 entry fee or a one-off refund cannot create durable LTV; the more rational early model is freemium plus token-based pricing, while To B can price against measurable productivity. Lewis’s test is: “Are you charging $200 because you solved the problem, or are you simply making someone prepay that many tokens?” Cancellations, price matches and refunds can be useful once, but may not recur; enterprise workflows run repeatedly, making ROI directly calculable from hours saved and employees’ hourly pay.
Privacy, security, inference costs and the chip roadmap are all pushing Personal Agents toward the edge: a high-end Mac Studio costing about $12K is advertised by Apple as capable of running trillion-parameter models, and online examples already exist; Qwen 27B is sufficient for a large share of day-to-day tasks. Lewis believes “the era of Edge Compute has arrived,” citing a chip-industry view that in 6 or 7 years, many models that run in the cloud today may run on phones. The unresolved question is not whether the hardware can carry the load, but who can turn installation, permissions and task execution into a one-click software experience.
Lewis still believes Gemini will catch up, but release cadence has become a structural burden: a new model needs at least a week of SOTA expectations, while competitors are resetting the benchmark every few days. Google’s research base, compliance discipline and refusal to misuse user data are long-term assets that may also be short-term speed disadvantages; DeepSeek’s arrival showed that open source can still produce real innovation, but the industry’s competitive window has compressed from “6 months, 3 months” to potentially “1 week after a paper is published.”
Deep dive
1. Lewis’s path from the early cloud era to the front line of foundation-model productization
Lewis studied computer science at university and began as a software engineer before joining the team working on Oracle’s public cloud. He remembers that around 2011–2012, AWS was still centered mainly on EC2 and databases, and the industry was only beginning to accept that “everything was going to the cloud.”
After more than 7 years in infrastructure, he moved to Google and switched into product management because he was “probably more interested in what to build than how to build it.” Before the foundation-model wave, he helped take 3 or 4 AI products from zero to one, then became one of the first people at Google to productize foundation models.
In August or September 2022, Lewis joined the PaLM 2 effort. The team worked with Labs to fine-tune a research model that had pre-training but almost no post-training for specific use cases, then built API serving, pricing and a developer community around it. The initial core team had only 8 or 9 people, and the project started about 3 months before ChatGPT launched.
2. The first bottleneck foundation models addressed was Document AI’s generalization problem
Lewis uses Document AI to explain the original product motivation: every person and company has different documents, and “training one model for every document” to reach 99.9% precision would be highly inefficient. The real question was whether one model could be trained to generalize across all documents.
Google had therefore been looking internally for “more general intelligence” well before ChatGPT. ChatGPT’s breakthrough was not only technical; it brought what had largely been a To B capability to consumers, allowing ordinary users to ask all kinds of questions and feel for the first time that “the AI era had arrived.”
3. ChatGPT exposed the gap between Google research and product release
Google was not unprepared when ChatGPT launched. The company already had chatbots such as Mina and had worked on them for a long time, but legal and responsible-AI concerns kept them from going public. Another camp also acknowledged that OpenAI’s models in many situations felt “more like a conversation with a person,” delivering a better experience than Google’s internal products at the time.
The crisis had 2 layers: whether the models were good enough, and whether Google could “turn a research project into an API and let users actually use it.” Google Brain, DeepMind and Google Research were still operating with dispersed resources, which were later concentrated to catch OpenAI.
Sergio’s involvement escalated gradually, from occasional quarterly meetings to monthly meetings, weekly meetings and daily stand-ups. He also pressed teams for technical details and forecasts in chat groups. Lewis’s description is blunt: when the founder tags you in a group, that is already a powerful forcing function—“the intensity can’t really go any higher.”
4. Gemini 2.5 Pro’s peak became the point at which Lewis left
Lewis left Google in the summer of 2025, after Google released Gemini 2.5 Pro. One reason was that the main path for model productization had already been laid: the API, pricing, use cases and developer community were relatively mature, leaving the next phase focused more on competing over models and benchmarks.
Raymond compares the market to a water utility: intelligence is the water, and the product is the faucet. Once the faucet’s shape, size and color are defined, competition shifts to water quality, price and model architecture. Lewis agreed that the contest had moved closer to research, while he wanted to work on a problem whose shape was still undefined.
Another factor was organizational bloat. Once Gemini became a company-wide priority, the number of participants and layers of management grew. Middle layers would sometimes “withhold information, because otherwise there would be no reason for them to exist,” turning collaboration into a “game of telephone.”
Lewis did not start a company immediately. He chose Microsoft because of its unique position with enterprise users, internal data and the developer ecosystem. Rather than continuing to “compete over the faucet,” he wanted to solve how Agents could use enterprise knowledge safely.
5. Microsoft’s Agent opportunity is built on enterprise knowledge, not the operating system
The internal scale estimate Lewis heard was that the data in SharePoint amounted to “many, many times the entire internet.” Large companies have accumulated huge volumes of PPTs, documents and internal records, but usually lack a Google-scale search engine, leaving much of that data present but practically undiscoverable.
Lewis works on 2 areas at Microsoft: Agent memory and the knowledge supplied to Agents. The latter extends into the context graph, with the goal of helping Agents access vast enterprise datasets more efficiently.
These capabilities are aimed at developers and are not tied to Windows. Whether the final product runs on Mac, Windows or the Web, it can call the same backend and store personal preferences, historical memory and searchable knowledge. Microsoft is providing backend capabilities, not a single end-user application.
6. Memory is both a personalization layer and a way for Agents to improve themselves
The first type of memory tells an Agent “who you are, what you’ve done and what you talked about before.” It can remember Lewis’s last conversation or proactively remind him about an untracked refund. That is especially direct in customer support and long-term assistant use cases.
The second type is procedural memory. When an Agent executes a task, it may “try one thing, fail, then try another,” wasting large numbers of tokens. If the system can read historical traces and identify past failures and successful paths, it can choose a more effective method next time.
Lewis sees little distinction between procedural memory and a skill: the former is experience in solving a problem that needs to be compressed, while a skill is one way to package it. The same experience can be stored as procedural memory or as a skill. Agent self-improvement requires turning execution traces into reusable assets first.
7. Enterprise memory’s core problem is layered permissions and knowledge promotion
Consumer memory startups usually distinguish only between User A and User B. Enterprises are entirely different. A CEO’s memory, the context a developer sees while debugging a bug and an ordinary employee’s conversations all require strict enforcement; a developer cannot gain access to a CEO’s private context simply because they are investigating a bug.
Memory exists simultaneously at the individual, 2-person chat, team and company levels. Lewis gives the example of 2 employees discussing work: the resulting knowledge should not belong to only those 2 people forever, but deciding what can be consolidated into a team asset and what must remain at the original level creates a large number of corner cases.
Raymond uses Feishu as an example: when an employee leaves, a manager can take over the employee’s documents but cannot truly inherit the higher-value implicit context embedded in chats. Even with 3 months and millions of characters of chat history, no one could read it all. An Agent’s value is in extracting the signal, but extraction does not automatically grant permission to propagate it.
Lewis’s current principle is to have the model identify candidate information first, then ask the participants whether it can be “promoted to the next level.” Because privacy boundaries are difficult to infer automatically, the people in the conversation should decide what enters team or company memory.
8. Fully transparent collaboration reduces permission complexity—but changes employee behavior
Raymond cites Shopify’s approach: employees interact with the same Claude Agent, and the relevant windows are open to the organization. The benefit is that knowledge is shared by default, eliminating the need to solve complex access control; the cost is that employees know from the start that others may see what they write.
Lewis’s warning is that if every chat feels like “someone is watching you,” employees will stop using it for private questions or casual conversation. The experience approaches “your boss standing behind you and watching you work the whole time.” That sense of surveillance pushes genuine communication into a second window such as WhatsApp or WeChat.
For most companies, he believes private chats should still not be readable without cause. Access would typically require a formal investigation into harassment, an unfair workplace or a similar issue. An AI-native company with 30 employees can operate through shared group chats; a company with 100,000 employees cannot simply copy that model.
9. Where the model is deployed determines whether privacy promises are enforceable
Raymond asks whether Microsoft, which is more respectful of employee privacy, could lose ground if model vendors such as Anthropic inspect user data line by line. Lewis’s response is that the first question is enterprise policy: if a company explicitly requires access to employee data, employees can at least choose whether to participate, but ordinary companies will have a hard time legally reading private chats arbitrarily.
For hyperscalers, an OpenAI or Anthropic model can be deployed inside the enterprise’s own cloud wall. Lewis says that in this setup, the model will not be logged or recorded. The host jokingly calls it “privacy bought back for $1T.”
Lewis still leaves room for uncertainty. Every company promises on paper not to view or record user data, but whether those promises are strictly followed is something outsiders cannot know “unless they’re on the inside.” That is why Personal Agent trust cannot be solved by terms and conditions alone.
10. Muse wins by turning existing capabilities into visible tasks
Lewis’s journey began with OpenClaw. After the 2.0 upgrade repeatedly broke and migration also failed, he asked Hermes, running on the same virtual machine, to retrieve all the old data and then went all in on Hermes. Hermes was powerful, but still required users to maintain the environment themselves.
Grok Bot targeted pro users but sealed the VM too tightly for security reasons. After code was generated, users could not directly access websites or the contents of the virtual machine; they could only view or download artifacts in the built-in browser. Lewis’s analogy: “You thought you bought a computer, but it turns out you only bought a screen”—without a mouse or keyboard to operate it.
Muse feels like an ordinary app, with a complete presentation and, in particular, an idea tab. It not only lists what the Agent can do but also recommends tasks based on the user’s context. For ordinary users staring at a blank chat window and unsure what to type, that discovery layer is itself a product advantage.
Raymond’s own example: as soon as Muse received Gmail access, it identified that he needed to follow up with a hospital and organized a long email thread involving more than 10 people. Claude Code could technically connect to Gmail as well, but users previously had to find a YouTube tutorial and copy a prompt. Muse turned the workflow into a clear “click once to unsubscribe” option.
11. Gmail and Calendar are both data gateways and trust gates
Raymond notes that much of daily life in the US runs through email: hospital appointments, children’s registrations, insurance claims, credit-card chargebacks and subscription cancellations all leave an email trail. Once Gmail and Calendar are connected, an Agent can largely know “who you are and what you do.”
Lewis refused to try Instinct on privacy grounds. Based on the terms he read, the platform had the right to use the data it accessed, potentially including emails to train models or improve the product. Muse has its own model, while Instinct mainly provides Agents plus memory, but that distinction does not change the authorization risk for Lewis.
At a conference hosted by The Information, the moderator asked how many people trusted Zuckerberg, and the room showed intense distrust. Lewis’s counterpoint was sharper: if users do not want to give Meta their email, why would they automatically trust a newly formed startup whose team they barely know?
12. Google did not build Muse first because it was “too slow”
Asked why Google did not build the product when the data was already inside Google, Lewis’s immediate answer was: “Google is too slow.” He believes the Google seen from outside is often reactive—the market produces a product, and the company then builds a countermeasure.
During the Gemini period, Google treated OpenAI as its biggest enemy. OpenAI was strong with consumers, so Google devoted substantial energy to better Web applications, fancier demos and model behavior that was more appealing to users. That focus caused both companies to overlook another product path.
Lewis sees Anthropic’s strategy as a successful product bet: “We’re not going to do all the fancy stuff you’re doing,” but instead bet on core engineering and enterprise. Once that direction was validated, OpenAI cut back on some To C experiments and refocused on coding, while Google also began to catch up.
A ubiquitous assistant had already appeared in Google I/O and in the product definition around Gemini and Bard 2 years ago, but iteration speed was “not good enough.” Lewis had heard that Google already had a similar product internally and might launch it soon. The issue was never that nobody had thought of it; it was that Google had not iterated and released it quickly enough.
13. Big tech can launch later, but open APIs are not easy to close
Lewis believes xAI’s Grok Bot fired the first shot, but he quickly gave up on it because of the $200 subscription and the lack of sufficiently practical utility. Meta spent a long time building and refining its product internally, then entered more effectively; its downloads have already surpassed OpenAI’s, making it the App Store’s No. 1 app.
Raymond worries that Google could lock down Gmail and Calendar to protect its own Agent. Lewis thinks that is “not very likely”: Google’s culture is generally more open, many of its products are built on its APIs, and the company’s basic definition of data ownership is that “the data belongs to the user.”
That creates a dilemma for large companies. Raymond believes Google faces compliance, bad-advice and privacy-liability risks if it launches its own product, while smaller companies can move more aggressively. But if big tech kills later entrants by closing its APIs, it would also undermine the ecosystem position it has built over the long term.
14. China’s walled gardens turn the general-purpose Agent into platform power
When using Agents such as Muse and Hermes, Lewis most wants access to WeChat data, but WeChat is almost completely inaccessible. The Xiaohongshu website he logged into sometimes still allowed the Agent to read content and account stats; Raymond says he is often forced to log out, while WeChat has shut down its Web entry point altogether.
Raymond summarizes the legacy of China’s mobile internet as a walled garden: each platform offers a complete internal experience, but the platforms do not interoperate. The result is that the top 10-plus internet companies each have their own knowledge-base Agent, leaving users facing “10 people standing there.” They can ask each one separately but cannot dispatch one general Agent to call across them.
Lewis believes a Personal Agent needs 2 categories of critical data: the user’s context, and preferences expressed through behavior such as search and browsing. Xiaohongshu therefore has an opening in interest and consumption scenarios such as what to eat, where to go and how to spend leisure time. But China’s fragmented data separates life, work and preferences, so startups may be better off starting with a vertical domain.
This also explains why the market still believes in Tencent. Even if its model progress is slower, WeChat’s social graph and mini-program ecosystem cannot be replaced in the short term. Tencent can even wait for US Personal Agents to complete a first competitive cycle and for the product form to converge before copying it, because the core gateway will still be there.
15. A Personal Agent business model cannot rely on one-off savings
Muse offers free usage with a weekly limit, then charges after the limit is exceeded. Lewis sees that as more rational than Grok Bot’s $200 entry fee because users do not yet know what an Agent can do, and charging a high price upfront is likely to produce extremely low retention.
He pushes the question further: “Are you charging $200 because you solved the problem, or are you simply making someone prepay that many tokens?” If the product is mainly OpenClaw-inspired and competitors can copy it, collecting tokens upfront does not amount to building a moat.
Refunds, cancellations, price matches and finding cheaper products can all be monetized through outcome-based sharing, but they are often one-off tasks. After an Agent successfully cancels a subscription once, the user may not have a similar task the following month. Looking useful “on a single task” does not mean the product can support long-term LTV.
If Muse is free because it earns shopping commissions, platform conflicts appear immediately. Raymond says Amazon has already blocked Agents such as Muse and Instinct. Merchants want real people to enter their own pages and transact; they do not want a third-party Agent controlling traffic, selection and commissions.
16. To C should charge by usage; To B can price against ROI
Lewis separates the 2 markets: early To C products are better suited to freemium and token-based pricing, with heavier use costing more. That is easy for users to understand. Raymond adds that once advertising matures, a product could also rely on ads rather than depending entirely on freemium.
To B workflows are more repetitive and easier to measure. Enterprises can divide employee compensation by hours worked, calculate how many hours the Agent saves, and price against the productivity improvement. That measurable ROI is more capable of supporting a stable budget than “helping an individual worry a little less.”
Over the long term, the price per unit of intelligence from frontier models will continue to fall, and intelligence may become commoditized like water or electricity. Much day-to-day reasoning can happen locally, with only frontier tasks such as cancer research or complex coding sent to the cloud.
17. Local models are already viable; the missing layer is one-click software
Raymond connects edge models with recent Agent safety incidents: people still do not know what Agents are doing, so they want better control and privacy through local execution. Lewis believes “the era of Edge Compute has arrived.” Apple’s official marketing for the highest-end Mac Studio says it can run a trillion-parameter model, and people have already done so online; the price is about $12K. NVIDIA DGX Spark also offers a local compute device.
For everyday tasks, his favorite example is Qwen 27B. As long as the task is not extremely complex coding, it is already sufficient for a large share of day-to-day needs. Bills, invoices and personal errands do not require trillion-parameter models, nor do they require sending private data to the cloud every time.
Lewis’s view from the chip industry is that model capabilities are hard to forecast 6 months out, but the hardware roadmap is visible 6 years out. In 6 or 7 years, many models that run in the cloud today may run on phones. The progression could resemble the shift from phones that could barely display a choppy GIF to 4K and 6K video, and from Snake to Genshin Impact.
Apple and Google could both choose to launch later and still overtake, but the opportunity remains in the software experience. After a user buys a Mac Studio, does the model run automatically, or must the user install an environment, configure an API and manage permissions? The fact that OpenClaw installation services can sell for RMB500 on Xianyu shows that “runs locally” remains far from “usable by ordinary people.”
18. Gemini has the foundation to catch up, but is boxed in by the SOTA clock and secrecy
Lewis’s conclusion on Gemini remains that it “will catch up,” because Google’s research foundation and talent pool are still extremely strong. The short-term pressure is that Google said in May that it would release Pro, yet still had not done so by October; any new version must at least match the leading model available at that point.
Model releases have become a race against the clock. When one company releases a formal Flash or Lite version, 10 companies may release products at the same time, and another new version may appear every 2 days. Raymond says every leading model is expected to remain SOTA for “at least 1 week”; otherwise, the company may choose to keep training, pushing the release date further out.
When DeepSeek appeared, Lewis says the first thing the team did was read the paper and validate the techniques. His view was that the consensus was “there’s real substance there,” and that it was not entirely based on distillation; it included techniques that the US side had not adopted at the time. Open-source papers allow the industry to “learn and grow from one another.”
Lewis can speak only for his time at Google: his team did not log user data or train on other companies’ model data. He cannot say whether other companies operate differently. But for Google, the need to find new data internally—and the possibility that excessive adherence to rules becomes a speed disadvantage—remains real.
As the technology lead time compresses from 6 months to 3 months and shorter, leading companies are publishing fewer papers and blog posts and allowing employees to appear on fewer unofficial podcasts. Raymond believes a team working on a bottleneck area may have only about an 18-month critical window; once the secret is public, “what could have lasted 6 months may last only 1 week after a paper is published.”