Wookyoung Kim, Jisoo · July 14, 2025
Readwise, AI Training, Coatue, AI Crawlers, Early-Stage Startups, Windsurf
Knowing what is interesting to me, and why, matters. Weekly Commits collects the articles and information I read during the week in one place and introduces them through why they felt interesting.
(1) How Readwise turned user behavior into advertising Readwise, a productivity app for managing articles and notes, turns user behavior on x.com into advertising. When users tag @readwise on a tweet they want to save, the tweet is saved to Readwise. Users no longer need to open the Readwise app and add the link themselves. (A shorter UX path.)
Posts that tag @readwise also remain on x.com, so they serve as ads that attract other people's attention. Readwise's viral marketing is interesting because it shortens the UX flow while turning user behavior into advertising.
(2) Training AI on used books is not copyright infringement
Anthropic, the developer of Claude, bought millions of used printed books, scanned them into digital form, and then destroyed the original copies to collect training data for AI.
Anthropic went to trial over whether buying used printed books without permission from the authors and training AI on them constituted fair use. The judge ruled that it was fair use because the scanned books were transformed forms of the copyrighted works and were not shared outside the company. However, ebooks were deemed not to fall under fair use, and further proceedings are ongoing.
The key passage from the ruling:
Everyone reads texts, too, then writes new texts. They may need to pay for getting their hands on a text in the first instance. But to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable. For centuries, we have read and re-read books. We have admired, memorized, and internalized their sweeping themes, their substantive points, and their stylistic solutions to recurring writing problems.
One point worth noting is that when AI training uses physically published printed books, companies do not need to negotiate with authors. The AI industry will find loopholes one way or another.
(3) Coatue's 2025 Keynote
Coatue, a U.S. crossover investment firm that invests in both public and private companies, presents a keynote every year at EMW (East Meets West). I watch Coatue's Keynote Deck closely because it covers both public and private companies, unlike many other investment managers' keynotes.
The part that stood out most in Coatue's keynote was the section on the top 40 companies by market capitalization in 2030. The fact that Google and Apple were excluded from the top 40 is striking.
(4) Zapier CEO's job posting for an AI Automation Engineer
The CEO of the no-code workflow tool Zapier posted a job opening for an AI Automation Engineer. The role is to build employees' AI capabilities and automate the parts of their work that can be automated, helping them focus on what matters.
Seeing it made me think of a kind of HR role, and I wondered whether HR jobs that require development skills may become mainstream. (Recently, Samil PwC even created a separate recruiting track for candidates with digital capabilities.) It feels as if development skills are becoming essential across every function.
(5) AI revenue is hard to predict because of outcome-based pricing
In an interview with Navin Chaddha of Mayfield, a 55-year-old venture capital firm, I saw something interesting.
The point was that consulting firms like McKinsey charge based on time, while AI (LLM) services price per outcome, so at this stage they are targeting completely different markets. Since smaller companies have difficulty using consulting firms, I found it interesting that AI services, where you pay only when a problem occurs or an outcome is delivered, may find it easier to penetrate the early market.
But these AI services also have a problem. Time-based billing allows revenue to be forecast; outcome-based pricing makes revenue difficult to predict.
(6) Cloudflare introduces AI crawler blocking by default
Cloudflare, a web infrastructure company, introduced a default feature that blocks AI crawlers from crawling websites. Instead, it wants to build a Pay per Crawl model in which AI bots must pay each time they crawl a website.
According to Cloudflare, OpenAI's crawl-to-referral ratio is only 1,700:1, meaning content creators are losing significant traffic because of AI.
For AI startups, this is truly a bolt from the blue. It will become harder to collect training data, making it more difficult for latecomers to build new AI models, and it is interesting that Cloudflare and content creators appear to have found a new revenue model. That said, I suspect this revenue model for content creators may apply only to large companies such as news organizations.
To exaggerate a bit, future content creators may end up making content for AI bots.
(7) Fewer things matter than we think
This is a 2022 essay by ZVZO CEO Won Ji-hyun. It can be summarized as: 'Are we doing what really matters?'
About three years after founding the company, one day I was struck by the thought that we might not have been spending time on the 'important' things. From A to D, as the hoodie slogan said, we had worked while 'obsessing like hell,' but I began to suspect that perhaps A through D were not important, and that even scoring 100 on them might have almost no effect on success. As we looked at the data, that suspicion frighteningly turned into truth. I will never forget the helplessness and self-reproach I felt alone in the office at dawn that day. Since then I have always carried this question: Are we in a state where we can say, 'The most important things are A, B, and C, and we are focusing our efforts on conquering A, B, and C'? Are projects like A', A'', and A''' meaningful attempts to test the hypothesis that 'A matters'? In other words, are we making important attempts (A', A'', B', C', C''...) in the important areas (A, B, C)? If I cannot answer this question well, I cannot sleep well, just like that dawn.
What should we focus on most? I thought about it for a long time after reading this, but I still have not found the answer. One conclusion I reached is that 'the most important thing in business is not decided by me; it is decided by the market.'
(8) Thoughts on early-stage startup teams and products
This is a slide deck presented in 2013 by ZVZO CEO Won Ji-hyun.
Top-tier people who are called geniuses elsewhere hear at home, 'When are you graduating? Why are you starting a business?' while friends around them seem endlessly successful. -> Everyone may look fine on the outside, but inside they are always sensitive.
I felt that sentence cuts through startups. A startup can still have a 90% chance of failing even when it is full of A-player talent, and it made me feel that taking care of A-players' mental state is also a virtue of startup leaders. Perhaps because I had not thought deeply about team members' mental health, the line landed even harder.
The point that you should not take heavy users' words at face value was also interesting. It felt that way because it seems to conflict with the idea that helping people in extreme pain is a good way to find PMF.
I saw a similar message in another piece about product management: heavy users have the loudest voices, but their needs often do not match those of typical users.
Because I found it hard to accept the advice to focus on ordinary users instead of heavy users, this piece, which argued against my own thinking, left a strong impression.
(9) What should a new employee do?
Elena Verna, who leads Growth at Lovable, wrote this piece using her own joining Lovable as an example to explain what a new employee should do.
Brief summary:
Day 1: Understand why what works is working. Do not break what is already working. Days 2-30: Based on past experience, find and try something that has an 80% chance of succeeding. Focus on optimization and improvement. Day 30: Think about the bigger business, but do not decide what you must do yet. Think and ask other people. Days 30-90: Write down your perspective on the company. The deeper you understand the company, the harder it becomes to see the product like a customer, so this is a good time to record your viewpoint. Things not to do before day 90: 1) Try to prove yourself. 2) Move so fast that you fail to align with the team. 3) Fix what is not broken.
The piece was so good that it surprised me. New employees want to improve something in the system because they feel pressure to prove themselves. But the pressure to prove yourself early leads work in the wrong direction. I liked Elena Verna's point that early on, you should make what is already working work even better rather than try to prove yourself, and her advice to write down your perspective on the company between days 30 and 90.
(10) OpenAI's Windsurf acquisition falls through; Windsurf CEO joins Google DeepMind
OpenAI's acquisition of AI coding startup Windsurf fell through, and tensions between Microsoft and OpenAI were cited as the cause. Microsoft reportedly did not want OpenAI to secure AI coding technology through Windsurf.
As a result, Windsurf went back on the market, and Google paid $2.4B (about KRW 3.3 trillion) to bring talent including Windsurf's CEO into Google DeepMind. Google did not acquire equity in Windsurf at that time. (It gained no control.)
This is similar to Google's character.ai deal, Microsoft's Inflection deal, and Meta's Scale AI deal. It appears to be a way to strengthen their position in the AI race without government regulatory scrutiny.
However, the companies acquired in this way have not ended very well. character.ai no longer has the same momentum, Inflection withdrew from the consumer AI business, and Scale AI is losing customers such as OpenAI, Microsoft, and xAI. Attempts to avoid government regulation seem to be causing greater harm to AI companies.