The stronger the models, the pricier the PM? Start with the data from 101,934 launches

Lenny Rachitsky’s podcast recently featured Tulsee Doshi, who is the product lead for Google Gemini and other frontier AI models. The episode is 58 minutes long. On the surface it is about PMs, but in fact it is about a more specific question: as models can do more and more tasks, where should product teams put their effort?

Tulsee’s first judgment is direct: when there is more that can be built, the scarce ability is judging what is worth building. She describes the new baseline for a PM as making something substantial—a prototype, a solid data analysis, or an MVP that can ship. Just handing a PRD to engineers and waiting for results is not enough in this environment.

When she hires PMs, she looks at a few things: they play with models every day, have their own use cases for what makes a model good or bad, can stress-test capabilities themselves, and are technically close enough to data and evals. She also mentions that this field changes every two weeks, and a roadmap written two months ago needs to be reviewed and rewritten.

How do you judge a new model? Tulsee’s advice is to test it with work you care about. Choose tasks where the model has already shown a hint of capability but is not yet perfect by your standards. That way differences between models stand out, and you can see how they improve.

She compares model training to Snape’s potions class: you put the ingredients in and are clear about the expected outcome, but you still have to accept a bit of magic. The team’s job is to find where the model unexpectedly shines and turn that into something users can use.

On the roadmap, she gives a 60-20-20 ratio: about 60% on what works today, 20% on what is close to usable, and the last 20% on directions that are still a little out of reach but where early signs are visible. The numbers can change; the goal is to have something you can double down on when the model upgrades.

Evals also do not need to wait for a big project. Start with five prompts and act as your own rater. If 20 prompts can distinguish model A from model B, that is already very useful. The hard part is not the tooling; it is writing down the good. One reader said that rating the same output twice makes them argue with themselves.

Here is a counterintuitive piece of data. X user Barely Bernie analyzed 101,934 Show HN launches since 2023: launch volume grew 3.7x, but breakthroughs grew only 1.3x. In 2023, one in every 22 launches was a breakthrough; in the last 12 months it became one in every 62. 62% of recent launches had zero comments.

AI has not automatically solved the problem either. Among 17,149 recent launch titles containing AI, LLM, or agent, the breakthrough rate is 1.1%; for the other 30,957 it is 1.9%. AI accounts for 36% of launches but only 23% of breakthroughs. Of the 350 launches whose titles asked for feedback, criticism, or ideas, none since 2025 has reached 50 points.

Put this data together with Tulsee’s view, and they do not conflict. The cost of building has fallen, but bad products also reach users faster, so judgment and distribution become more expensive. The valuable part of a PM is not the title; it is whether they can define things well, whether they can call a stop to things that should not be built, and whether they can translate model capabilities into something users are willing to use. One reply said a PM who can tell the team what to stop building is worth hiring. Others with development backgrounds said the boundary between product and engineering is thinning, and critical thinking does not belong only to PMs.

Tulsee also mentions burnout: teams work long hours, but what really drains people is churn, chaos, and chores that feel like a waste of time. She hopes AI tools reduce chores and let people spend more time thinking and creating. Lenny mentions that his own technical labor survey found the same thing: longer hours do not necessarily equal burnout.

Google’s strategy is another layer of context. Tulsee says Google covers the full value chain, from custom TPUs to products with billions of users, and its goal is to place models along the Pareto curve of cost and quality—not only frontier LLMs, but also serving search users, software engineers, and cyber defenders. She says explicitly that Google wants to win the SOTA model race.

A year ago people were still discussing whether the PM role would disappear; now the same group is discussing why PMs are valuable again. The title has not changed, but the work has. Treating this interview as a PM recruiting ad would miss the point. It is more like a product work spec for a time when model capabilities change quickly: roadmaps need room for post-upgrade models, evals start with five prompts, and PMs need to get hands-on and make something substantial. As for whether the title will disappear, the replies are already arguing about it. Some say PM principles need to be learned; some say the role is merging. The common thread is that judgment is becoming harder to hide.

The data has limits: HN points measure attention, not revenue, and the sample skews technical. But 101,934 launches at least show one thing: being able to launch does not mean you are worth seeing.

Original interview tweet: https://x.com/lennysan/status/2119823537666183440

Podcast video: https://youtu.be/qGf2Yvan4Wg

Show HN data: https://x.com/bare_knee/status/2119818608000033207