乐观、懒惰与「如果呢」：Google Docs 创始人的产品课

Schillace says he still owns the first-ever Google Doc, from around October 2005, and it still opens today. Writely started as a C# application on three rented Windows file servers in Texas, run by a sysadmin in the Philippines. After the acquisition the team ported everything to Java and moved the data into Bigtable, later followed by Spanner, with the front end rewritten twice. Nothing was lost along the way, which prompts the joke that it is a Ship of Theseus question.

The document is mostly a note to his co-founder Steve about whether the collaboration algorithm was colliding, plus a pasted picture of Edna from The Incredibles. The Computer History Museum has asked to curate it, which he says is awkward because the real artifact is a living file. He also recalls migrating Writely onto Google infrastructure over a read-only Sunday of eight hours, so smoothly that when Sergey Brin asked days later when they would move, the answer was that it had already happened.

Schillace says everything we use was a disruptive innovation once, and that anything new is threatening to some people because it forces a choice between "you're wrong" and "it's wrong". Humans are storytellers, so they readily invent reasons to reject it. He calls these why-not questions: the browser isn't ready, nobody will trust the cloud with files, what if there's no connectivity on an airplane. The more interesting questions are what-if questions, which ask how far a curve can be extended.

Engineers, he admits, are pessimists by trade, but he feels he has missed more by being pessimistic than by being optimistic too early. His mantra is that there is little prize for being pessimistic and right. He adds that when critics call a new thing "a toy", it often means it is real and threatening and they can't find a better criticism.

He also offers a litmus test for impact: a small group loves the product, sometimes more than the makers do, while others want it to "die in a fire". An indifferent bell curve means an incremental product, whereas ChatGPT split the world into amazed and appalled. In Docs' early days, with millions of users, Google executives still told him it was a stupid idea, and he says that is fine as long as you don't run out of people who love it.

Applying the what-if lens to generative AI, Schillace asks what happens if software can be built around models, or if a reasoning engine does meaningful work. He frames it as a second Industrial Revolution, with a surplus of cognitive energy in place of physical energy. He compares this with Tesla and SpaceX, where Elon Musk asked what if rockets could be reused or a car were basically a software product.

He says he is not an optimist by nature; it is a conscious, growth-mindset choice because he regrets what he missed through caution more than what he avoided. The second half is making it cheap to try things: his two Writely co-founders, experts in word processors, said the browser would never support it, and they were right and wrong at once. Because the tools were sharp, a couple of days was enough to see how it felt.

His strategy is to get your tools sharp and run lots of cheap experiments, and to stay receptive to results that fail in interesting ways. He gives a recent example from Microsoft: multiple chatbot agents with memory worked together poorly until the team gave them a shared whiteboard as working memory, which made them noticeably smarter, and the team still doesn't fully know why.

Asked how to separate real opportunities from hype, Schillace says he thought hard about crypto and could never find a user-value answer to "what if it works?". The best case seemed to be running security on his own finances, which he finds dystopian. For him the root principle is user value: nobody does anything unless it makes their life better, and marketing spend cannot substitute for solving a real problem.

He notes that whatever you think of Musk, he articulates clear end-user value, with Teslas promising real range, solved charging and cheaper fuel. Rachitsky draws the lesson as a simple exercise: ask whether you personally would want it.

Schillace says builders love their product for reasons that don't matter to users, who care only about their lives getting easier. He sees two consequences. Adoption speed tracks the amount of value added, as with ChatGPT's explosive growth versus modest AI features that lumber along. And a feature that looks better in isolation still requires the user to hear about it, remember it, learn it and form a habit.

Writely therefore did not ask for an email address at all at first; after about two minutes of typing it gently asked for one, with no password, so it could send a link to the document. The effort users spend must be outweighed by the ease gained, ideally by a large factor. In consumer products, he says, you have perhaps 15 to 30 seconds to show value.

He describes Docs as a classic innovator's dilemma story. In an era when Office was impregnable and software was physically installed, Writely offered zero install, easy use and collaboration without emailed attachments, while removing most features, perhaps a few too many, since early student users demanded a word count. Microsoft survived and kept the money, he says, while Google got the users. He adds that there is friction in being at Microsoft as the person who "messed them up a little bit".

Schillace says Docs felt like it was working almost immediately. He noticed that browsers offered contentEditable and JavaScript, and his team, who had built word processors before, tried building one in a browser. Collaboration came by accident because documents were simply objects on a server and nothing yet locked people out. The team loved working in the same document, then discovered edits collided.

Fixing it was brutal. They used three-way merge, because operational transform had not been invented yet, but browsers represented the same document differently in HTML, with Firefox and IE disagreeing even on how an empty paragraph looks. He believes that if they had met the difficulty before seeing the value, they might have given up. They got lucky in the order the idea was presented.

He recalls hostility inside Google, including an executive who refused him hardware in his first week and had to be threatened with legal action since Schillace had contracted earn-outs. Without that protection, he thinks the project would have died. Great product builders, he says, must be rigid about the mission and flexible about feedback, like a samurai sword that is hard on the back and soft on the edge.

He thinks documents will change radically. His team's chatbots double as documents, which he sums up as "bots or docs": the bot interviews you for an hour, producing a readable linear document while building semantic memories, so you can later ask for a diagram or a summary. He predicts that within a few years an application you can't talk to will feel as anachronistic as one not connected to the internet.

Rachitsky asks when it is fine to simply play with a technology before finding a problem. Schillace admits he thinks with his fingers, but has evolved toward picking what he calls north stars: interesting, useful goals that might be buildable. Right now his team is testing how much independent work multi-agent systems can do, using programming as the domain because an agent needs only a Python environment and a file system.

One goal is for the system to write an equivalent of an existing program, which wouldn't be valuable by itself but would show whether self-monitoring, self-correcting agents can build things at that scale. The effort has already produced insights. The team built a debugger agent that explains what went wrong, which they don't yet trust to fix things, and an agent that documented the code base fairly well.

He notes that even Writely's start was half play and half goal: could he build a word processor in a browser? He believes many good product ideas come from engineering, and that unusual combinations of tools can help, noting that he held code-readability approval in a backend, middle-tier and front-end language at Google.

Asked how people can get better at thinking about what's possible, Schillace says he notices today's students are more pessimistic and less engaged than the generation that dismissed old men calling the PC a toy. He suggests several causes: filtered social content that presents six-sigma successes as normal, a life lived out loud where failure hurts, highly curated paths into elite schools, and the end of unsupervised childhood play from the mid-1980s.

The result is less willingness to experiment to the point of failure. He sends goofy ideas to Satya Nadella and sometimes follows up a week later to say it was a bad one. His principle is that you can't dance if you're afraid to embarrass yourself or succeed if you're afraid to fail. His personal motto this year is "virtue from error", turning mistakes into something useful.

On whether failing hurts your career, he says tightly managed, surprise-free work yields a linear, boring return that takes thirty years, while extraordinary returns require bigger risks and more interesting shots. He admits this is not for everyone.

Schillace says the reason people get ahead is impact, which comes from picking something you're good at and doing it with intensity. We tend to undervalue what we are good at and assume work must be unpleasant. His advice is to find the thing you feel guilty about being paid for and do it as hard as you can. He notes this is privileged advice and that early in a career compromises are often necessary.

Rachitsky ties this to Buffett and Munger's advice to go big when you find an advantage, and later to Seth Godin's advice to do the best version of any job. Schillace agrees and adds that you should bring something uniquely yours to a role. He describes himself as binary, all in or all out, and unhappy when he cannot be intense.

He recalls that when he first became a manager he felt he wasn't producing anything, until his boss told him that everywhere he went things got better. That showed him he enjoyed talking to people all day. He suggests staying attentive to moments when someone lets you do something that surprisingly makes you happy, since others may see strengths in you that you don't, and being open to them allows pivots.

Schillace endorses his own lines that AI isn't a feature of your product, your product is a feature of AI, and that transformative value will come from apps that don't work without it. He says most of the industry focuses on building models, whereas he is an app and tool builder. After starting on GPT-4 at Microsoft, he saw it as a stochastic pure function that needs state, control flow and orchestration, which led to Semantic Kernel and then Infinite Chatbot.

He believes we are at the start of a gigantic disruption: as the internet made distributing information free, AI will make pixels free. Images already show it, since a picture of a cat riding a bike is now one prompt rather than Photoshop skills. If models plan better, handle multimodal UI, store state and personalize well, we will spend our time conveying intent to an agent, like a very rich person with a tireless staff, rather than poking at static apps.

He compares this moment to seeing the Palm or early iPhone, or to realizing in 2005 that the browser was a platform. It may take ten years, and he guesses a quarter to a half of what he said is wrong. He also dismisses the press take that Gemini being similar to GPT-4 meant AI was finished.

On how people should approach AI, Schillace says "just play with it" is weak advice. A better way to learn is to choose a goal, even a goofy one, and be stubborn about reaching it, ideally one that leads somewhere interesting, such as seeing how close you can get a GPT to do part of your job. He sees the exhaustion of keeping up as another sign, like the early dot-com era, that something big is happening.

He argues that going from zero to one is the hard part, and that once people understand something is there, the one-to-many elaboration happens fast and in parallel. We are in that elaboration phase. His advice is mundane: read a lot, think a lot, choose your battles, pick a goal in your own domain, and accept that you will be out of breath. Rachitsky sums it up as finding a problem someone has and asking how AI could provide that value.

Asked why Microsoft is so innovative, Schillace says he thinks very highly of Satya Nadella, who in his experience is as decent, empathetic and high-energy as he appears, and believes a leader's job is to raise the organization's energy. He also describes Microsoft's culture as humble, doing unglamorous work to make businesses succeed, and notes the sheer number of talented people. He credits Kevin Scott, who he works for, as probably the smartest person he has worked with directly.

He adds that luck matters too, and that Nadella and Scott's bet on OpenAI did not happen by accident. He says his views are his own as an engineer rather than Microsoft's. His closing message echoes the Steve Jobs line that the world was built by people no smarter than you: "You don't have to have permission, you just have to have energy."

Schillace recommends Invisible Cities by Italo Calvino, the disturbing novel The Wasp Factory, and Steven Johnson's Where Good Ideas Come From for its idea of the adjacent possible. His favorite interview question, which Google banned once it became known, asks how many zeros end 100 factorial. He uses it to see whether someone who is handed something seemingly impossible rolls up their sleeves or backs off.

He shares a story of selling both Writely and 200 pounds of blood sausage to Google, via a friend's artisan salumi company he had invested in. He writes a weekly Substack, Sunday Letters From Sam, and says he is most interested in people making progress on independent agent action and generated UI. He won't look at the 27th memo-writing chatbot: it has to be interesting and disruptive.