Read more

The first is personal: what you need to know before every decision and every meeting. Competitor moves, user feedback, stakeholder context. When this breaks, you’re the last person in the room to know things you’re supposed to know first.

The second is your team’s: the systems that serve your users, surface …

Read more

Read more

Today’s Post

I tested all of Claude's team features and ranked the ones that matter for PMs and builders:

1. The Full Tier List

   • Deep dives into every S and A tier feature

   • Overview of the other features

2. My Claude Setup

3. 4 PM Workflows That Chain These Together

4. Lessons from how Anthropic Builds Product

Plus downloadable poster + Claude.MD file for PMs:

Subscribe now

Before we get in, here’s a roadmap for reading this. S tier features changed how I work daily. A tier features save me measurable time most weeks. B tier features are solid and worth learning when you hit the right scenario. C and D tier exist but I won’t waste your time on them here.

If you’re short on time, read S tier and Section 2 (My Setup). If you have 20 minutes, read the whole thing. If you want to go deep on Claude Code specifically, I wrote a full masterclass and a complete Cowork guide that pair with this piece.

1. The Full Tier List

S Tier: Game Changers

These changed how I work. Adopt them now. Here’s how:

Read more

If you’ve tried OpenClaw or Agent Zero, you know the pain. Local installs, environment configs, dependency hell. Computer runs entirely in the cloud and routes your prompts across 19+ AI models automatically.

I spent weeks testing it. Here’s the web’s first PM guide.

Today’s Post

1. What Is Perplexity Computer

   • How It Works, Remembers, and Runs

   • Setting Up (Connectors, Skills, Custom Instructions)

2. How to Use Perplexity Computer

   • Top 6 Use Cases for PMs

   • How to Prompt Without Wasting Credits

3. Honest Comparison and Limitations

Subscribe now

1. What Is Perplexity Computer

Perplexity CEO Aravind Srinivas calls it a “general-purpose digital worker.”

How It Works

Computer sits on top of 19+ frontier AI models. Claude Opus 4.6 handles core reasoning. Gemini runs deep research. ChatGPT 5.3 assists with long-context recall. Grok takes lightweight tasks. Nano Banana generates images. Veo 3.1 produces videos.

You write one prompt. Computer picks the best model combo for the job.

Tip: You can manually assign which models handle which parts of your task. Most people skip this and let Computer pick the most expensive option. Learning to control model routing is the fastest way to cut costs.

How It Remembers

Computer maintains context across sessions and connects to your services and files. Over time, it learns your preferences and standard workflows without you re-specifying them. Start a new task when the objective changes. Continue an existing task when you’re iterating on the same deliverable. Continuing is always cheaper.

How It Runs

Most AI tools give you suggestions. Summaries. Plans. Outlines you then have to go build yourself.

Computer delivers the finished thing. Research reports with source citations. Deployed dashboards with shareable links. Cleaned datasets with charts. Launch kits with positioning docs and email drafts.

And it runs them in parallel. Start a task. Immediately start another. They execute independently in the cloud. Close your browser. Go to dinner. When you come back, everything is done.

Scheduled tasks work the same way. Your machine doesn’t need to be on. Set it up once. Computer runs it on a recurring schedule whether your laptop is open, closed, or sitting in a drawer.

You can choose which models handle which sub-tasks. You can force the system to check in with you before executing. Most people never touch these controls. I’ll show you exactly how in the prompting section.

Setting Up Perplexity Computer

You need a Perplexity Pro subscription ($20/month), Max ($200/month), or Enterprise Max ($325/seat/month). Pro gets 4,000 credits per month. Max gets 10,000.

Getting started takes 30 seconds. Go to perplexity.ai, log in, click the Computer icon in the left sidebar. You’re in.

Everything is organized around Tasks. A Task is not a conversation. It’s a job. When you submit a prompt, the orchestrator (powered by Claude Opus 4.6) reads it, breaks the objective into subtasks, assigns each one to the best-suited AI model, runs sub-agents in parallel, and compiles results in the right panel.

Tip: Start a new task when the objective changes. Continue an existing task when you’re iterating on the same deliverable. Continuing is always cheaper because of persistent memory

Then, do three things:

Step 1 - Connect Everything

Connectors let Computer access your actual data and take real actions in your existing software. Not summaries. Real read-and-write access.

Setup takes about a minute per app. Click Connectors in the sidebar, browse the list, click Enable, complete OAuth. Done.

Computer has 400+ built-in integrations, and the interesting ones go way beyond Slack and Google Drive:

• Through Plaid, you can link your actual brokerage and Computer builds you a live portfolio dashboard. Daily P&L, position cards, news feed, earnings calendar, price alerts. A personal Bloomberg terminal on a private URL. Full setup and prompt in Section 6.4.

• Research access is wild. CB Insights, PitchBook, Statista. Computer bypasses the paywalls. A $5,000 market sizing report? Computer often pulls it.

• The Snowflake connector is powerful. Type “What were the top 10 customers by revenue last quarter?” Computer writes the SQL, runs it, hands you charts. Plain English in, data out. Same with Salesforce, HubSpot, Databricks.

• GitHub lets Computer write, refactor, and push code to your repos. Gmail, Slack, LinkedIn go deeper than you’d expect. I had it scan my inbox, summarize my newsletters, draft unsubscribe requests, and calculate my SaaS burn rate from receipt emails. Tag Computer in a Slack channel and it runs research without you switching tabs.

• For SEO, Ahrefs, Google Search Console, Google Analytics, WordPress, Cloudflare. Full audit, competitor analysis, keyword mapping, content drafting, and live site updates.

• And the one that surprised me most. Computer connects to Amazon’s Selling Partner API and Shopify natively.

You can also bring your own tools. Provide an MCP (Model Context Protocol) server URL and Computer connects to proprietary CRMs, custom analytics servers, or private APIs. Enterprise admins can share custom connectors across the organization.

Step 2 - Create Skills

A Skill is a saved set of instructions that auto-activates when Computer recognizes a matching task. Same concept as Claude Skills or Cowork plugins.

Without Skills, you’re re-explaining your brand guidelines, formatting preferences, and reporting structure on every task. With Skills, you explain it once.

Computer ships with built-in Skills for Slides, Research, Research Report, and Chart. To create a custom one, click Skills in the sidebar, click + Create skill, and upload a .md file. Skills stack. A research Skill can hand off to a report-formatting Skill, which hands off to a slides Skill. One prompt, full pipeline.

Tip: If you’ve explained the same thing to Computer twice, it should be a Skill. Brand guidelines. Formatting rules. Research methodology. Build it once, never repeat yourself.

Step 3 - Custom Instructions

Skills activate for specific task types. Custom Instructions apply to ALL tasks, all the time. Keep them under 1,500 characters. The single most effective one I've found:

“Always come back to me and clarify any misunderstandings and challenge my thinking to make sure you’re very clear on the stated outcome, and let’s create a brief short plan before we build anything.”

That one line forces Computer to verify what you want before spending credits. Massive savings.

2. How to Use Perplexity Computer as a PM

I’ve identified 6 particularly powerful use cases you should steal. Here’s everything you need to get them set up:

Read more

Andrej Karpathy built a system that runs the other 500 rounds for you while you sleep. He called it autoresearch.

42,000 GitHub stars. Fortune called it “The Karpathy Loop.” Shopify CEO Tobi Lutke pointed it at Shopify’s templating engine and got 53% faster rendering from 93 automated commits.

Everyone covered the ML side. Most PMs closed the tab thinking it wasn’t for them. That’s a mistake. The pattern underneath has nothing to do with GPUs or neural networks. It works on anything you can score.

I’ve spent two weeks pulling apart Karpathy’s repo, the community forks, and the real-world applications people are building on top of it.

Today’s Post

I’ve built the PM’s Ultimate Guide to Autoresearch, covering:

1. What autoresearch actually does

2. Why this matters if you’re a PM

3. How to set it up

4. Six high-value use cases

5. Why the experiment log matters

6. Where Karpathy is taking this next

7. The complete toolkit: skill, eval templates, and analyzer

Subscribe now

Quick setup if you want to follow along: install Claude Code, clone the autoresearch repo with git clone https://github.com/karpathy/autoresearch, and pick the skill or prompt that frustrates you most. That’s your starting point.

1. What autoresearch actually does

Let me ground you in the original before I show you the PM version.

Karpathy trains small language models as side projects. Like every ML researcher, he was stuck in the same grind: tweak the training script, run an experiment, wait for results, check if the model got better, decide whether to keep the change or throw it away, then start the whole cycle again. A productive day might get through 8-10 of these rounds, and most of that time is spent waiting for the GPU.

So he built a system that does the entire cycle autonomously.

The repo has three files. train.py is the training script and the only file the agent is allowed to modify. prepare.py is the evaluation harness that scores the model, and the agent cannot touch it (if it could, it would just make the test easier instead of making the model better). And program.md is the instruction file that tells the agent how to behave, what to try, when to keep changes, and when to revert.

The loop works like this. The agent reads the code, forms a hypothesis about what might improve the model, makes a change to train.py, runs a 5-minute training experiment, and checks the metric (validation bits per byte, a single number where lower means better). If the score improved, the change gets committed to git and becomes the new baseline. If it didn’t improve, git reset wipes the change instantly. Then the agent starts the next round.

That gives you about 12 experiments per hour, roughly 100 overnight.

Karpathy left it running for two days. The agent found 20 improvements on code he’d already hand-tuned for months, including a bug in his attention implementation he’d missed entirely. All 20 stacked and transferred to a larger model, producing an 11% speedup.

Shopify CEO Tobi Lutke tried it the same night. He ran 37 experiments overnight and woke up to a 0.8B parameter model outperforming his hand-tuned 1.6B model. Half the parameters, better results.

Then he pointed the pattern at Liquid, Shopify’s templating engine, and got 53% faster rendering and 61% fewer memory allocations from 93 automated commits.

This is a new way to optimize anything.

2. Why this matters if you’re a PM

My run on a landing page skill: 41% to 92% in 4 rounds. Three changes kept, one auto-reverted.

The pattern works because it removes the bottleneck every PM actually faces: you know the prompt could be better, but you'll never run 50 iterations manually.

It only works when three things are true:

1. A clear metric. Score the output as a number, not a feeling. “Is this good?” doesn’t work. “Does the headline include a specific number?” does. Yes or no. Add up the yeses across 30 test runs and you have a score the agent can optimize against.

2. A measurement tool that runs without you. Claude Code builds an evaluation script that generates outputs, scores them against your criteria, and prints the result. No human in the loop. The scoring runs programmatically so the loop runs overnight.

3. One file the agent can change. Your skill markdown, your system prompt, your email template. The agent edits this one file per round. Everything else is read-only.

All three present, it works. Any one missing, it doesn’t. That’s your filter.

The human never touches train.py. Karpathy writes program.md and the agent handles execution. That’s product management. You define what “better” means. The agent runs the 50 rounds you’d never have time for.

I use Claude Code for this. But it works on any coding agent that can read files, edit files, and use git. Cursor, Windsurf, Codex, Antigravity.

🔒 Below: the 3-step setup, 6 use cases with copy-paste prompts and eval criteria for each, the experiment log walkthrough that Karpathy says matters more than the result itself, and a downloadable toolkit with a skill improver, 10 eval templates, and a results analyzer.

Read more

1. Can I actually use this in my day-to-day, or is it still a toy?

2. Should my product have this capability, and if so, what just changed that makes it viable?

The answer to both changed two weeks ago. Google shipped Nano Banana 2, and it’s currently #1 on the Text-to-Image A…

Read more

They’re the most important new professional software created in years.

But the issue everyone runs into is starting with them.

I want my paid subscribers to be the most productive people on the planet. So I’ve made it …

Read more

Of all the AI tools I use that I see the least other people using, NotebookLM tops the list. It’s the entire reason I pay for Google AI.

It is not getting the attention it deserves. And they just launched their best feature yet: prompt based slide revisions. So, I had to make you a complete guide.

NotebookLM sta…

Read more

• Working in Excel or Powerpoint

• Building Google Docs or Notion pages

More people should be using it. Today I’m giving you the ultimate guide to Claude Cowork.

Read more

It was the Codex App. It’s the new, most powerful way to use ChatGPT.

And that’s why OpenAI hasn’t even released GPT-5.3-Codex in ChatGPT. It’s only available in Codex.

So that’s today’s deep dive.

Read more

Today, I deliver the web’s best and most practical guide on the topic.

Let’s get into it.

Read more

Subscribe now

Annual subscribers get a free year of 9 premium products: Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin (worth $28,336).

Read more

Subscribe now

Annual subscribers get a free year of 9 premium products: Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin (worth $28,336).

Read more

Subscribe now

Annual subscribers get a free year of 9 premium products: Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin (worth $28,336).

Read more

Subscribe now

Annual subscribers get a free year of 9 premium products: Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin (worth $28,336).

Read more

Subscribe now

Annual subscribers get a free year of 9 premium products: Dovetail, Arize, Linear, Descript, Reforge Build, DeepSky, Relay.app, Magic Patterns, and Mobbin (worth $28,336).

Welcome back to the AI Update.

The most interesting thing in AI this week iswasnot GPT-5.2, or Cursor’s new features, it was the debate around AI data centers in space.

It sounds like science fiction. But Elon Musk, Jeff Bezos, and Sundar Pichai have all said its the future of AI compute.

So in today’s deep dive, I present the web’s first deep dive on AI data centers in space. This includes an exclusive interview with the CEO of Starcloud, the first company to build a data center in space.

Plus, in today’s weekly update, I cover all the AI news you need to know.

Finally, I end with a recap of my sit down with Wes Bush on AI and PLG.

Bolt.new: AI Prototyping Like the Best

Did you now that I built my cohort’s site on bolt.new? That’s right, since I had their CEO Eric Simons on my podcast, I have become a super fan of the product.

For PMs, it represents the ultimate tool build high-fidelity prototypes fast. In multiple head-to-head battles, I have found it wins in ease of use.

Build with Bolt

I personally recommend this tool. If you’ve only used Lovable, use my link to try Bolt!

Top News this Week: GPT-5.2 Released

The result of Sam Altman’s “code red” after the Gemini 3 launch is finally here. It’s GPT-5.2, and it predictably puts OpenAI back on the throne for having the best model:

But that’s just the eval performance. The biggest learning from GPT-5.2 that you should know about is the amazing gains in cost-efficient performance.

Look at this chart on the ARC-AG1:

A year ago, hitting 88% on ARC-AGI-1 cost an estimated $4,500 per task. Today, 90.5% costs $11.64. That’s 390x cheaper in 12 months.

As product builders, we can continue to expect frontier-level performance to get cheaper and cheaper.

There’s a million AI news articles, resources, tools, and fundraises every week. Here’s what mattered:

News

• Nvidia received approval to sell H200s to China with a 25% cut going to the US government. Beijing is reportedly limiting access anyway to push domestic chip self-sufficiency. Days earlier, the DOJ uncovered $160M in illegal chip smuggling through relabeled “Sandkyan” chips.

• The Pentagon launched GenAI.mil, a custom Gemini chatbot for 3M+ military personnel through a $200M Google deal. It’s one of the largest enterprise Gemini deployments yet.

• Mistral shipped Devstral 2 (72.2% on SWE-bench, nearly matching DeepSeek V3.2 at 5x smaller) plus Vibe CLI, their first terminal-native coding agent

• Claude Code now works directly in Slack: tag @Claude in any channel and coding tasks route to a new Claude Code session

• Cursor just released the ability to design in your codebase, and some people say its news is even bigger than GPT-5.2

Resources

• How Google’s $900M SpaceX bet became worth $111B (my breakdown)

• Why Boom Supersonic pivoted to AI power

New Tools

• Dapple: Run submission-based programs (collect & pick winners) hit #1 on PH

• Incredible: Deep work AI Agents hit #1 on PH

Market

• Unconventional AI confirmed a $475M seed round at $4.5B valuation, led by a16z and Lightspeed. Founder Naveen Rao (former head of AI at Databricks, previously sold MosaicML for $1.3B and Nervana to Intel for $400M+) is building energy-efficient AI hardware “as efficient as biology.”

• Harness hit a $5.5B valuation with $240M raise to automate AI’s ‘after-code’ gap

• IBM announced plans to acquire data streaming company Confluent for $11B to connect real-time data to AI systems

And now on to today’s deep dive:

Everything You Need to Know About AI Data Centers in Space

What do Elon Musk, Jeff Bezos, and Sundar Pichai agree on?

That we should be building AI data centers in space.

But now, the critics are out in full force:

What’s the truth? Are AI data centers in space the research?

I have done all the research for you. I talked to the CEO of Starcloud who yesterday trained the first LLM in space with an Nvidia H100. I researched the physics. I watched the Elon videos.

And that’s today’s deep dive:

1. The first-principles case for orbital data centers

2. What the skeptics get right (and wrong)

3. Inside Starcloud: what the pioneers are building

4. The real timeline and who’s paying attention

5. What this means for AI infrastructure

Continue Reading Online

1. The first-principles case for orbital data centers

Gavin Baker is one of the most respected technology investors of his generation. He ran Fidelity’s flagship technology fund, now leads Atreides Management, and has been right about AI infrastructure longer than most people have been paying attention.

On a recent episode of Invest Like the Best, he made a claim that sounds like science fiction:

In every way, data centers in space from a first principles perspective are superior to data centers on Earth.

Baker’s argument is disarmingly simple. What are the fundamental inputs to running a data center?

Power. Cooling. Chips.

Space is structurally better at two of the three.

Power: You can keep a satellite in the sun 24 hours a day. The sun is 30% more intense in space, resulting in 6x more solar energy than on Earth. And because you’re in constant sunlight, you don’t need batteries. Starcloud’s CEO Philip Johnston told CNBC that orbital data centers will have 10x lower energy costs than terrestrial ones.

Cooling: On Earth, cooling is the majority of rack mass and cost. The HVAC systems, the liquid cooling, the CDUs. In space, you put a radiator on the dark side of the satellite. It’s near absolute zero. All that complexity disappears.

Networking: In terrestrial data centers, racks communicate via fiber optics. That’s a laser going through a cable. The only thing faster than a laser through fiber? A laser through vacuum. If you link satellites with lasers, you get a faster, more coherent network than on Earth.

Inference latency: When you ask Claude a question, your phone sends a signal to a cell tower, through fiber to a metro facility, to a data center, computes, and returns. With direct-to-cell satellites (Starlink has demonstrated this), you skip most of that path.

2. What the skeptics get right (and wrong)

Not everyone is persuaded. When Patrick O’Shaughnessy posted Baker’s clip, the replies filled with physicists and engineers pointing out complications.

Their core objection: cooling in space is harder; much harder than it sounds in a sound bite.

The physics problem

Cofounder of Power Dynamics, Jen Zhu captured the misconception well (as I shared above): the idea that space “automatically” cools hardware because it is cold is wrong.

In a vacuum, there is no air, so there is no conduction or convection to carry heat away. The only way to reject heat is through thermal radiation that emits infrared energy from a surface into space.

This is governed by the Stefan-Boltzmann law, which dictates how much power you can radiate as a function of surface area and temperature.

At typical electronics operating temperatures, radiating kilowatts or megawatts of heat is slow and inefficient unless you have an enormous radiative surface.

The math gets extreme

Mikhail Klassen (Senior AI Engineer @ Planet) ran the numbers for modern AI hardware:

• A single Nvidia H100 GPU (roughly 700W) needs about 1.1 m² of radiator area

• A DGX H100 system (8 GPUs, around 10.2 kW) needs at least 16 m² of radiator surface

• The radiator might need to be roughly half the size of the satellite’s solar arrays

Scale that up to gigawatt-era AI data centers, and the numbers get extreme. For comparison, the International Space Station (which consumes only 75-90 kW) already relies on large, articulated radiator wings and an elaborate thermal control system.

Here’s the thing: the skeptics are right about the physics. But “hard” doesn’t mean “impossible.”

3. Inside Starcloud: what the pioneers are building

To move beyond abstractions, I spoke with Philip Johnston, the founder of Starcloud, to understand how they see the physics and the business case.

a. The radiator problem, quantified

When I asked about radiator requirements for a 1 GW data center, Johnston’s answer: “About 1 km square.” But how do you build something that large? “You assemble it in space,” he said. You don’t launch a kilometer-wide radiator as a single object. You launch compact, modular components and assemble or deploy them in orbit.

This is where Starcloud’s background matters: Johnston’s co-founder Ezra holds a PhD in engineering and spent a decade designing deployable structures for NASA, including solar arrays for the Lunar Pathfinder mission.

Johnston addressed the criticism directly:

The criticism usually is: in order to dissipate that heat, you need a large surface area. And they think for some reason that’s super impractical. You just have to build a large surface area. That’s what we’re doing. Half our engineering team is building very large, low-cost, low-mass deployable radiators. That is the core IP of our company.

b. The economics: waiting for launch costs to fall

At the end of the day, this comes down to cost per kilogram to orbit. Starcloud has done the math on multiple space-based business models:

• Space-based solar power (beaming energy back to Earth): makes sense around $50/kg to orbit

• Data centers in space (monetizing computation directly): break even around $500/kg

SpaceX is marching toward that number with Falcon 9 and rideshare missions. Starship, if it reaches anything like projected performance, should drive launch costs well below that threshold.

Starcloud is making a timed bet on that trajectory.

c. What’s in orbit today

Starcloud already has hardware in space:

• Starcloud-1: A 60 kg satellite (about the size of a small refrigerator) carrying a single Nvidia H100 GPU. Launched November 2 on a SpaceX rideshare mission.

• Firsts in orbit: They’re running what they describe as the first AI model training in space, the first high-powered inference in space, and the first deployment of Google’s Gemma model in orbit.

• Speed to orbit: Starcloud went from founding to satellite in space in 15 months. The previous fastest timeline for any space startup to reach orbit was roughly four years.

Their next launch is planned for October 2026. That satellite is slated to carry Blackwell-architecture GPUs with roughly 10x the compute of today’s H100, plus optical terminals for always-on, high-bandwidth connectivity.

When Johnston reposted the Gavin Baker clip on X with the caption, “We have the first Nvidia H100 operating in space,” it drew 4.8 million views.

4. The real timeline and who’s paying attention

Let’s strip away the hype and the reflexive skepticism. Here’s where things actually stand:

a. Training vs. inference

Training frontier AI models in space is still a long way off. Modern training runs demand extreme interconnect bandwidth and ultra-low latency between thousands of GPUs, plus the ability to move petabytes of training data. Shipping that data to orbit doesn’t make economic sense yet.

Inference is different. Once a model is trained, serving it (especially at scale) is far more bandwidth- and latency-flexible. The near-term opportunity for orbital data centers is running inference workloads: answering queries, processing video streams, and handling edge AI tasks for users and devices below.

b. The radiator challenge is hard, not fatal

The ISS is a functioning proof of concept that you can manage significant thermal loads in space for decades with high availability.

The question has shifted from “Can you cool in space?” to “Can you do it with mass- and cost-efficient modular radiators at data center scale?”

Starcloud’s core bet is that advances in deployable structures plus falling launch costs and in-space assembly will push the answer toward yes within 5-10 years.

c. This isn’t a fringe idea anymore

Perhaps the clearest signal is who else is paying attention:

• Google has announced Project Suncatcher, an initiative to deploy TPUs in orbit via Planet Labs

• Amazon, SpaceX, and other major players are exploring similar avenues

This is no longer a single startup with a contrarian deck. Orbital compute is a line item on hyperscalers’ long-term infrastructure roadmaps.

d. The next proof point: 2026

The real inflection point will be Starcloud’s second satellite, targeting October 2026. If Blackwell-generation GPUs can run sustained, high-utilization workloads in orbit (powered by solar, cooled by deployable radiators, and connected via high-bandwidth optical links) then space stops being a demo environment and starts looking like a new tier of core AI infrastructure.

5. What this means for AI infrastructure

Baker’s broader framing is worth sitting with:

“Whenever there’s a bottleneck that might slow AI down, everything accelerates. We’re running up on boundaries of power on Earth. All of a sudden, data centers in space.”

We’re already seeing the outlines of those constraints. Grid interconnections are delayed for years. Communities are pushing back against new data centers that strain local power and water. Renewable build-out is fast but not fast enough to painlessly support exponential curves in AI compute demand.

At some point, we exhaust the “easy” power and cooling capacity on Earth’s surface and start to look up.

From a physics standpoint, the cheapest, most abundant solar energy in our solar system is not on the ground. It’s in space, above the atmosphere, where the sun shines nearly all the time and every photon can be captured.

The open question isn’t whether that energy exists. It’s how quickly we can learn to harvest it, convert it into computation, and integrate orbiting infrastructure into the broader AI ecosystem.

Whether Baker’s “3-4 years” timeline proves precisely right matters less than the direction of travel. The combination of falling launch costs, improving deployable structures, and insatiable demand for AI compute points us toward a world where a nontrivial share of humanity’s computation happens off-planet.

The question now isn’t if data centers move to space.

It’s how quickly and which companies will be ready when the physics, economics, and demand curves finally intersect.

Finally, onto insights from a webinar I did this week:

How to do PLG in the Era of AI

I sat down with Wes Bush (ProductLed) to break down how AI changes PLG. We covered a lot of ground: the tools you should actually be using, how pricing is shifting, and why taste is becoming the only defensible moat. Here’s the rundown.

Most conversations about AI and PLG focus on one thing: adding AI features to your product. That’s only half the equation.

The other half, and often more impactful, is how AI helps you personally succeed at PLG. Before you can build better activation flows or optimize pricing, you need to be a 10x practitioner yourself.

We mapped out both sides. On the job side: sole-context LLMs like Notebook LM, Claude Projects as your co-pilot, agent platforms (Lindy, Relay, Zapier), dictation tools, and AI prototyping with Bolt or Lovable. On the PLG side: activation, pricing, model, expansion, GTM, data, team, and strategy:

Watch the full conversation

That’s all for today. See you next week,

Aakash

P.S. You can pick and choose to only receive the AI update, or only receive Product Growth emails, or podcast emails here.

The reasoning gap closed faster than expected this week.

DeepSeek released V3.2, matching Gemini-3.0-Pro’s capabilities while cutting costs by 70%.

What does this mean for the AI race? I’ll break it down in today’s weekly update.

Plus, I’ve been testing more Nano Banana Pro prompts. Last issue I shared 5. In today’s deep dive I’ll share 4 prompts that are useful.

Finally, I’ll summarize my recent sit down with Kyle Poyar about AI pricing. The short version: most AI features fail because they don’t move retention, not because of pricing.

Subscribe now

Dazl: The AI Prototyping Tool Built for PMs

Are you hitting a wall at 80% when vibe coding? That’s what Dazl solves. Unlike other tools, they allow you to edit everything with full control & precision. Check them out!

Build with Dazl

Get 1 month free with code: DazlXPG

The reasoning gap wasn’t supposed to close this fast

While OpenAI, Google, Anthropic, and xAI trade the #1 spot weekly, something different happened on Sunday. DeepSeek dropped two models that match frontier performance and made them completely free.

• DeepSeek-V3.2 is the daily driver. It performs at GPT-5.1-High level across reasoning benchmarks while supporting tool-use in both thinking and non-thinking modes.

• DeepSeek-V3.2-Speciale is the reasoning specialist. It earned gold medals in the 2025 International Mathematical Olympiad, International Olympiad in Informatics, and ICPC World Finals.

What actually changed:

1/ Frontier reasoning at 70% lower cost: A new attention mechanism slashes inference costs dramatically. Processing 128,000 tokens (roughly a 300-page book) now costs about $0.70 per million tokens, down from $2.40 on their previous model.

2/ Thinking while using tools: Previous AI models lost their train of thought every time they called an external tool. V3.2 preserves reasoning across multiple tool calls - essential for complex agent workflows.

3/ Benchmarks that matter for builders: On AIME 2025 (math): 93.1% for V3.2, 96.0% for Speciale vs. GPT-5-High’s 94.6% On Terminal Bench 2.0 (coding agents): 46.4% vs. GPT-5-High’s 35.2%

The catch: Token efficiency still lags Gemini-3.0-Pro, and DeepSeek acknowledges their world knowledge remains behind frontier proprietary models.

Why this matters: The 4-horse race (OpenAI, Google, Anthropic, xAI) is about who has the best model. DeepSeek is running a different race: making frontier-capable AI free under MIT license. For teams building agent-heavy applications where cost scales with usage, this changes the math entirely.

There’s a billion AI news articles, resources, tools, and fundraises every week. Here’s what mattered:

News

• Anthropic targeting $350B IPO at zero markup (while OpenAI waits for $1T). IMO they are optimizing for enterprise trust, not valuation.

• OpenAI declared “code red” on ChatGPT threats and delayed ads. The real threat: Google.

• NVIDIA launched an open self-driving model that thinks before it acts.

Resources

• Automate everyday work with Google’s AI agents

• How to go from core PM to AI PM: Todd Olson’s 5-layer technical pyramid

New Tools

• Aha: AI employee for influencer marketing hit #1 on Product Hunt

• X-Design: AI Agent for branding hit #1 on PH

Market

• Anthropic acquired Bun (open-source JavaScript toolkit) and announced Claude Code hit $1B ARR. Six months after launch.

• AI voice startup, Gradium raises $70M.

And now on to today’s deep dive:

4 Helpful Nano Banana Pro Prompts

Last week I shared five Nano Banana Pro prompts and you loved them. But I kept testing and have now found 4 even more game-changing prompts:

1/ 3D Cartoon Weather Visualization

Turn any weather app from boring numbers into something users screenshot and share.

Present a clear, 45° top-down isometric miniature 3D cartoon scene of [CITY], featuring its most iconic landmarks and architectural elements. Use soft, refined textures with realistic PBR materials and gentle, lifelike lighting and shadows. Integrate the current weather conditions directly into the city environment to create an immersive atmospheric mood. Use a clean, minimalistic composition with a soft, solid-colored background. At the top-center, place the title “[CITY]” in large bold text, a prominent weather icon beneath it, then the date (small text) and temperature (medium text). All text must be centered with consistent spacing, and may subtly overlap the tops of the buildings. Square 1080x1080 dimension.

2/ Fridge Scanner → Recipe Infographic

The feature every food app has promised but never delivered well. Now it works.

Scan what’s inside of the fridge and offer me an idea what can be cooked with the ingredients available with a detailed step by step recipe in a form of a simple infographic.

3/ Multilingual Object Labeling

If you’re building anything for language learners or global markets, steal this.

Draw a detailed {{pet shop}} scene and label every object with English words. Label format: First line: English word. Second line: IPA pronunciation. Third line: {Chinese} translation.

Credit: Crystal on X

4/ 3D Packaging → 2D Dieline

This one shocked me. Turn product photos into production-ready packaging templates.

Deconstruct this 3D product packaging photo into a precise, flattened 2D dieline template. Unfold the box into a complete technical net structure, revealing all panels, including previously hidden dust flaps, tucks, and glue tabs. Eliminate all perspective and lens distortion to create a perfect orthographic top-down view. Retain the exact high-fidelity texture and typography from the original photo but map it accurately onto the flat surface. Superimpose clean, technical die-cut lines (solid) and fold lines (dashed) in a contrasting color. Isolated on a pure white background, shadowless, print-ready production quality.

Credit: Amir

Finally, onto insights from a webinar I did this week:

How to (profitably) add AI to your product

I sat down with Kyle Poyar (Growth Unhinged) and Axel Sooriah (Atlassian) to talk about adding AI to products profitably. Here’s what PMs get wrong and how to fix it.

Why most AI features fail (it’s not pricing)

The intuitive answer is “we mispriced it.” The real answer: the feature doesn’t move retention.

At Apollo, we shipped an AI email writer when ChatGPT exploded. Pricing was fine. The problem: users didn’t accept the AI drafts. It never became part of their workflow, so it didn’t move retention.

Compare that to Fortnite’s AI bot system. New players were getting crushed by hardcore grinders and rage-quitting. Epic added AI opponents as “cannon fodder” to smooth the on-ramp. Over 2.5 years and thousands of features, this was the #1 retention driver.

If your AI feature doesn’t measurably improve retention, pricing won’t save it.

The “hill climbing” approach to AI costs

PMs often optimize costs too early. “We can’t call that model - it’s too expensive.” This kills features before they prove value.

Instead:

1. Define evals upfront: what does “good enough” look like?

2. Climb the quality hill first: use whatever models you need to hit those evals, even if unprofitable

3. Then climb down the cost hill: swap to cheaper models, reduce context windows, add caching

Figma is climbing up (proving PMF for Figma Make). Airbnb is climbing down (shifting from commercial APIs to open-source models after validating value).

You can’t penny-pinch your way out of a bad use case.

Why hybrid pricing is winning?

Per-seat pricing is dangerous for AI. Your top 10% of power users can drive 70-80% of LLM spend. Charge everyone the same, and your best users become your least profitable.

What’s working: seat/platform fee + usage component.

Credits are becoming the standard. They give you a universal currency across AI actions and flexibility to adjust costs as models change without rewriting your entire pricing story.

The one takeaway: Solve really hard, high-dollar problems. If your AI reliably removes a $30 human task for $2 in compute, pricing details are solvable. If you’re adding a chatbot no one asked for, everything else becomes painful.

Watch the full conversation

That’s all for today. See you next week,

Aakash

Hope you had a wonderful Thanksgiving.

The last two weeks have been absolutely insane for AI.

We’ve seen 4 separate companies release models that take the spot of “best model on planet earth.”

What does this all mean? I’ll break it down in today’s news update.

Plus, in today’s deep dive, I’ll share 5 Nano Banana Pro prompts I’ve found most valuable in real-world testing.

Subscribe now

Dazl: The New AI Prototyping Tool Purpose Built for PMs

I’m working with this startup Dazl to build the best AI prototyping tool ever. Check them out! Unlike other tools, they allow you to edit what you’re seeing:

Build with Dazl

Get 1 month free with code: DazlXPG

AI is now a 4-Horse Race, with one Conspicuous Absence

One chart from A16Z AI Partner Justine Moore explains everything you need to know about the model releases over the last 2 weeks:

There are 4 companies in the race:

1. OpenAI - which released GPT-5.1 November 12th

2. xAI - which beat GPT-5.1 with Grok-4.1 5 days later

3. Google - which beat Grok-4.1 with Gemini Pro 3 1 day later

4. Anthropic - which beat Google with Claude Opus 4.5 6 days later

So as of writing, Anthropic has the best model. And next week, who knows? Probably OpenAI’s turn to get back to the top of the leaderboards.

What’s for certain: the pace of innovation remains as brisk as it was in 2023, and the competition is helping us consumers get the best out of these labs.

It’s also worth noticing who remains conspicuously absent from this race: Meta. Despite spending billions on all the best researchers, it has released absolutely nothing.

The last version of Llama was released in April. Something seriously wrong is going on over there.

There’s a billion AI news articles, resources, tools, and fundraises every week. Here’s what mattered:

News

• The Google TPU threat to Nvidia became real propelling it past Microsoft’s market cap:

  • Meta is buying Google TPUs

  • Google also sold Anthropic 1M TPUs

• China continues to lead the open source race, with Deepseek releasing a new IMO gold medalist model (that almost no one talked about)

• Sierra, Brett Taylor (OpenAI Board Member)’s company, hit $100M ARR

Resources

• Google Antigravity tutorial (their new Cursor competitor)

• How Kuse hit $10M ARR in 60 days with context engineering

New Tools

• Nao: an AI Powered IDE for analysts hit #1 on Producthunt

• Questas: builds choose-your-own adventure stories with AI and hit #1 on PH

Fundraising

• xAi is closing on a $15B funding round valuing at $230B, a 460x multiple on its $500M in 2025 projected revenue

• Robinhood CEO’s Harmonic raised $120M at $1.45B valuation for error-free AI

And now on to today’s deep dive:

Continue Reading Online

5 highly effective Nano Banana Pro prompts you can use

Nano Banana Pro is the most amazing AI image model ever. It’s opening up brand new uses cases that you need to know.

But let me break a harsh truth: most people are lying about what Nano Banana Pro can do for you. It still sucks for professional headshots, for instance.

I’ve tested everything so you don’t have to. Here are 5 actually awesome prompts that can be useful to (and aren’t silly annotations that are useless):

1/ LinkedIn infographics

You can now turn complex thoughts into shareable infographics.

I created the above with this simple prompt:

Create a lime green and white infographic with:
•⁠  ⁠Top: Category label, bold title, subtitle in bubble
•⁠  ⁠Middle: 7 numbered steps in dotted boxes (3-row grid)
•⁠  ⁠Each step: Green number pill, bold header, explanation, bullets
•⁠  ⁠Bottom: Green banner with “Follow me for more”

Use #A4D65E green for headers/accents, white backgrounds, black text, rounded corners. Keep text concise and mobile-friendly 
{blog-content}

2/ Create technical diagrams

You can now create complex technical diagrams:

I did this from my post on context engineering for PMs:

Create a professional technical diagram illustrating [SYSTEM/PROCESS]. Use clean, modern styling with appropriate icons and labeled components. Include [KEY ELEMENTS] and show the relationships between them with directional arrows. Use a color scheme that distinguishes between [DIFFERENT SYSTEM PARTS]. Make the diagram clear enough for technical documentation but visually appealing enough for presentations.

3/ Comparative product UI mockups

It can easily show you how different design approaches for a feature can look side-by-side, so you can make better decisions and have a solid foundation to work with.

Steal this prompt:

Create a set of 3 UI mockups for {FEATURE/SCREEN} showing different design approaches. For the first, use a minimal, data-dense design prioritizing efficiency. For the second, create a more visual, feature-rich interface. For the third, design an innovative approach that reimagines how this interaction could work. Use a consistent color palette based on {BRAND COLORS}, and include realistic placeholder data. Make each mockup distinct but implementation-feasible.

4/ Visualize data patterns and anomalies

AI can finally do charts!

This is great for sharing or for visualizing large datasets sitting in your Excel sheets:

Create an advanced data visualization highlighting the patterns and anomalies in this [DATASET/METRICS]. Use visualization techniques that reveal [SPECIFIC RELATIONSHIP] while making [KEY ANOMALY TYPE] immediately apparent. Incorporate small multiples or comparative elements to show changes over [TIME PERIOD]. Use a color scheme that emphasizes statistical significance, with attention to accessibility. Include annotations pointing out the three most important insights.

5/ Transform code into visual explanations

When you need to explain complex algorithms or code logic to stakeholders or in documentation without losing them in technical details, this comes in handy.

Steal my prompt:

Create a visual explanation of how this [ALGORITHM/CODE BLOCK] works. Use a step-by-step flowchart with clear annotations explaining each operation. Include small code snippets where relevant, but focus on making the logic visually intuitive. Use color coding to distinguish between [DIFFERENT OPERATIONS/PHASES]. Make the visualization detailed enough for engineers but clear enough for non-technical stakeholders.

That’s all for today. See you next week,

Aakash

PS1. If you liked this, you’ll love my podcast with Google AI PM Director Jaclyn Konzelmann.
PS2. Yes, of course, I grinded from 10pm-12:30 AM to send this to you. No one works harder.

Subscribe now

Google released Gemini 3 Tuesday, and the new Nano Banana Pro this morning. I got early access to both, and have been thinking deeply about the implications.

Here’s where I land: This week is one of those moments that changes everything. And most of the commentary is focused on the wrong thing:

• Gemini 3 is shocking because of the benchmarks. Now, Google has the top model (something Polymarket had been predicting all year but only just happened now).

• But it’s even more shocking because Google trained it entirely on custom TPUs and proved you don’t need Nvidia’s supply chain to reach the frontier.

That’s today’s deep dive:

Subscribe now

Reforge Build: AI that knows your product

Every AI prototyping tool spits out generic output. You spend days tweaking it to match your design system, your brand, your product.

Reforge Build starts with your context. Your design system. Your customers. Your product. Prototypes that already feel like yours on day one.

Try Reforge Build

Get 1 month free with code: BUILD

For years, the narrative has been simple: bigger model = better performance. Throw more compute at the problem. But Gemini 3 proves something different: better architecture, better training methodology, and tighter control over infrastructure can change everything.

Google trained this entirely on their own TPUs. They can serve it at scale without touching Nvidia’s supply chain. And the performance is the kind of leap that makes everything before it feel outdated.

Three things that actually changed

1/ Multimodal from the ground up

Text, images, video, code - Gemini shot to the top of all of the leaderboards, immediately dulling the news from earlier this week of Grok-4.1 beating GPT-5.1.

2/ State-of-the-Art reasoning

Gemini 3 excels at reasoning. It scored 91.9% on GPQA Diamond - a PhD-level science benchmark and 23.4% on MathArena Apex.

This matters when you’re working through multi-step problems, analyzing data, or trying to understand something technical.

3/ Solved image in text

The Nano Banana Pro model that dropped this morning is one-shotting amazing text in images in a way that no model before it has. For instance, here’s a whole chalkboard from a blog post:

The infrastructure moat was just redrawn

The craziest thing of it all is Gemini 3 was trained entirely on Google’s TPUs. Not Nvidia chips. Not rented infrastructure. Google’s own silicon.

Now they’re lending those TPUs to Anthropic and Midjourney, which means the most capable AI companies in the world are now training on Google’s custom hardware.

This changes everything. For years, the narrative was simple: whoever has the most GPUs wins. Nvidia controls the supply chain. Nvidia’s moat wins.

But Gemini 3 proves something else. If you can build state-of-the-art models on your own chips, and you have the capital to manufacture them at scale, the game changes.

Google has capital. Google has training data. Google has proprietary TPUs. And now Google has proven these TPUs can reach the frontier.

Nvidia sells shovels. Google is building the mine.

If the AI race is a marathon, Google has the infrastructure advantage. That’s probably why Berkshire Hathaway keeps buying Google while sitting on massive cash. They’re seeing what most aren’t: the infrastructure play is the long-term play.

That’s it for today’s deep dive. Here’s everything else that mattered in AI this week:

Google also launched Antigravity: The agentic IDE

While everyone is talking about Gemini 3, I saw very few people are talking about another piece of news that matters almost as much.

Cursor has hit a $29B valuation and shown IDEs are a key AI market. Now, Google is playing there as well with Antigravity. It’s the result of Google $2.4B acquisition of Windsurf CEO Varun Mohan and his team four months ago.

Google has notoriously launched lots of IDEs it drops support for months later, but this one appears different because even Sergey Brin appeared in the launch video:

I’m skeptical it will win (Google doesn’t use it internally), but it’s worth monitoring.

News

• Jeff Bezos is back as CEO at a new AI startup, Project Prometheus, with $6.2B

• Replit added a visual design system to build sites without coding

• Nvidia beat its guidance at earnings again

• OpenAI launched GPT-5.1-Codex-Max

• Grok launched Grok-4.1

Resources

• 20 mins to AI Agents

• AI Prototyping tool comparison

• Templates for your AI automations

New Tools

• Manus Browser Operator: Turn any browser into an AI browser

• Parallel: get high-quality datasets back from the web with 1 query

• Ramp Sheets: AI spreadsheet editor built for finance teams & founders

Fundraising

• Ramp raised $300M to build a full-scale finance-operations platform

• Suno, the ai music maker, raised $250M

That’s all for today. See you next week,

Aakash