Welcome to the second Weekly Stack, where we concentrate the AI firehose into tactical insights for PM builders. We've got 317 new subscribers this week – welcome to the stack! 🥞

Iteration from last week – got lots of positive feedback, plus some constructive notes that in last week’s issue it wasn’t always clear what was actionable vs just news. So this week, everything comes with a side of crispy actionable insights. 🥓

There's a super-easy one-click feedback form at the end of this email. Please click a button! It really helps me.

After last week's OpenAI and Anthropic big releases, this week felt like the industry took a collective breath. Sort of – we still got a monster DeepSeek update, some great thinking on where to use AI, surprising UX findings on AI search patters, and more signs of unbelievable AI investments to come.

Let's do this.

Last week, lots of you loved discovering ChatGPT's prompt optimizer. It is actually part of an even more powerful tool that lets you AB test your prompts by running the same test set across different prompt variants and models (and see diffs), then keep the winner with full version history.

I made a video showing how to use it. Try it next time you need to write a prompt.

(There's an even more powerful way to do this with a tool called Promptfoo. I’ll cover it in my deep dive into prompting dropping soon 👀)

Imagine you're OpenAI. You’re feeling pretty good about your GPT-5 pricing. Then Deepseek walks in and says "We'll do it for for literally 1/8th the cost." It’s just disrespectful. Deepseek V3.1 is a doozy:

Takeaway 🥓: If you've been holding back on any AI features because of cost, revisit them. Some examples of projects this pricing unlocks:

Shot: There was a great discussion in Lenny's Newsletter community this week about how AI is changing product development. The top answer pointed to this gem: Prototypes Don’t Make Decisions. The gist: “A prototype shows you can build something. Writing forces you to ask whether you should. It surfaces contradictions, sharpens intent, and catches nonsense before it becomes software.”

Chaser: Shreyas Doshi dropped an fantastic 7-minute video on where you should and shouldn't use AI.

His framework:

Takeaway 🥓: As your skills (and AI capabilities) grow, you’ll be able to implement AI in more and more parts of your life and work. Be careful where and how you do it!

There’s a new zeitgeist spreading through tech execs. Amazon's cloud chief Matt Garman made waves saying that replacing junior staff with AI is "one of the dumbest things I've ever heard." His reasoning:

Intercom's CEO Eoghan McCabe has also been pushing hard on bringing in young talent, saying AI adoption is "kind of a young person's game.”

Takeaway 🥓: It's not about really about age. It's about willingness to experiment. The key is to keep learning and being willing to try to applying your skills to do things in new ways. Challenge yourself every week.

Putting ✨AI✨ sparkles everywhere isn't helping users discover features.

NN/g’s latest research shows people struggle to know when to use AI, don’t trust answers, and miss hidden affordances (bad labels, vague entry points). They recommend:

Takeaway 🥓: The gap between AI capability and user understanding is where the opportunity lies. Consider building onboarding flows that demonstrate AI capabilities through examples, not explanations.

The GPT-5 personality drama last week (where users literally grieved the loss of GPT-4o's personality) proved beyond any doubt people are forming deep emotional attachments to AI personalities. Sam Altman claims <1% of users have "unhealthy relationships" with ChatGPT, which is his way of saying "only millions of people are emotionally dependent on our product."

This creates a tension for PM builders.

On one hand there’s personality free advocates.

On the other there’s Meta, where a leaked Meta memo showed they were cool with their AI having "romantic chats with kids." They’d love to get in on that emotional dependence.

Takeaway 🥓: If you're building consumer LLM, you need a personality strategy. Even choosing "no personality" is a choice. This will become a key product decision in the next year. Pay attention to AI personality strategies as you use LLM tools.

We got a bunch of cool new AI features from major companies this week. I won't go into them, but here's a summary of the features if you want to explore them yourself.

Here they are subjectively ranked with the coolest first:

OpenAI is aiming for a $500B valuation (would be the most valuable private company ever), while Anthropic is raising $10B. This money isn't for model training – it's for infrastructure. They're building the compute capacity for the next generation of AI. Think nuclear reactors powering data centers (yes, really, eventually).

Takeaway 🥓: These investments will determine what's possible to build in 2-3 years. Start thinking about what becomes possible when AI inference is 100x cheaper and 10x faster.

We're seeing a flood of partnerships between model makers and cloud providers. The headlines tell the story:

Takeaway 🥓: The wild west phase of AI is ending. Enterprise adoption means more reliable APIs, better SLAs, and standardized practices. Enterprise AI is coming fast.

If you only have 30 minutes this week:

Keep building,

Carl

P.S. – Next week I'm doing a deep dive on prompt engineering with interactive examples. If you’d like to provide an example for me to use, let me know if you: