hamfish

the founders I respect most aren't the ones with the best stories. they're the ones who stayed honest when the story wasn't going well.

@lennysan @kalinowski007 Point 12 applies to product as much as hardware. The people who grew up using AI from the start don’t just use it differently, they think differently. That gap between AI-native and AI-adapted is going to be one of the most important talent distinctions of the next decade.

My biggest takeaways from ex-OpenAI, Apple, Meta roboticist @kalinowski007: 1. The AI frontier is shifting from digital to physical because labs see the ceiling of keyboard-bound AI. “What you can do behind a keyboard with AI is going to saturate.” Which is why labs, big tech, and startups are increasingly investing in hardware, and why enrollment at universities is rising while CS enrollment is trending down. 2. More change is coming to warfare than to consumer electronics in the next two years. Drones, robotics, and the hardware supply chain all converge on the battlefield, and Caitlin argues we have to be able to control adversarial threats to our hardware layer, not just our chatbots. 3. The hardware industry faces a looming memory crisis that could derail the robotics revolution. Memory prices are spiking—potentially doubling or more—driven by AI data center demand. Companies building consumer robotics can’t compete on price with data centers. Caitlin is advising startups to pre-buy memory and stockpile components if they can afford it, because “we are in trouble as an industry.” 4. VR didn’t become a mainstream product, but it created the tech necessary for robotics (and war). SLAM (simultaneous localization and mapping), depth sensors, spatial computing, and understanding how humans perceive visual data in space is now powering robotics, autonomous vehicles, and drones. The technology needed to understand how a robot moves through space is essentially the same technology developed for VR headsets. 5. Humanoid robots are overhyped. While humanoids are interesting for certain long-tail tasks, most manufacturing and real-world applications need dedicated robots designed for specific jobs. A robot screwing keyboards into laptop cases doesn’t need to be humanoid—it needs to be optimized for that exact task. The future will have robots for construction, electrical work, logistics, and low-volume assembly, and most won’t look humanoid. 6. Supply chain independence is a national security imperative. Over the past 25 years, essentially every layer of the hardware supply chain—from raw magnets to actuators to final assembly—has been outsourced to China, Japan, and Korea. The same actuator technology that makes a drone rotor spin also makes a robot arm move. Without an independent supply chain, the U.S. is vulnerable. As Caitlin warns, “We need to re-industrialize this country significantly in order to be safe in a military sense.” 7. The hardest part of building safe robots is the decisions you don’t think about. If a robot arm is heavy and hard, the impact force when it hits you is dangerous. But there’s also the social aspect—robots need to show intent before moving (looking before turning), acknowledge when humans enter a room, and transmit non-threatening body language. As Caitlin learned from researcher Leila Takayama, “If a robot just suddenly turns and does all this stuff, it scares you. But if a robot looks before it turns and then goes, it’s much less alarming.” 8. Software builders don't understand how fundamentally different building hardware is. Software can compile code hourly, but in hardware you may get only a handful of chances to “compile” before mass production, with each major build taking three to five months. Once you ship, you’re done—there are no over-the-air updates for physical components. Software intuition doesn’t transfer to hardware. 9. In hardware, you never have enough time—so do everything you know you need to do right now. Caitlin learned from Apple executives like Shelly Goldberg and Kate Bergeron that you can’t wait around. Even if you technically have more time, use it, because “in two days there’s going to be a surprise coming around the corner that you need that time to fix.” This ruthless efficiency of clearing known tasks immediately creates a buffer for inevitable surprises. 10. AI hasn’t yet transformed hardware engineering. AI can’t do real CAD (computer-aided design) yet, and AI models don’t understand friction, weight, contact pressure, or surface texture. It can do surfaces and point clouds, but not the dense, equation-based solid entities that hardware engineers need. But when it arrives, it will be transformative. 11. CAD files are some of the most valuable IP any company has. Samsung, Apple, and other manufacturers will never give their 3D CAD to AI model makers. This creates a data scarcity problem for training hardware AI. The solution might start with hobbyists who don’t care about IP protection and just want to build things faster, then eventually move to on-premise AI systems that companies can train on their own data without sharing it externally. 12. The best hardware teams combine three types of people. You need generalists who can apply lessons from other fields to new problems. You need some specialists who have built similar things before and others who have scaled products to high volume. And critically, you need 20-year-olds who are truly AI-native—they approach problem-solving completely differently because they use AI from the ground up for everything. As Caitlin notes, “It’s very hard to find someone who’s in their 30s who can be truly fully AI-native.”

I've been in the room where the right answer lost because the expected answer was easier to agree on. it happens more than anyone admits.

designing products is easy. designing products people actually want to use is the humbling part.

the best product decision I made this year cost nothing to make and saved us months of building the wrong thing.

@icanvardar Every feature you ship that nobody uses makes the features they do use harder to find.

shipping more features nobody uses is just noise make the core product better first

@iSlimfit Data is the what. Users are the why. You need both but most teams only have time for one and consistently pick the wrong one.

The product manager who spends time with actual customers every week makes better decisions than the one who relies entirely on dashboards and second hand feedback. Data can only tell you what happened, but customers tell you why.

@lennysan AI gave everyone the ability to do data analysis. It didn’t give everyone the ability to know when their analysis is wrong. That gap is now a data scientist’s full time job.

Not enough people are talking about how much AI is impacting the role of data science. I was chatting with a DS friend, and he said that most of his team's work now is reviewing half-assed AI data analysis from PMs and engineers. And that 50% of the time, that analysis is wrong. The role is becoming less fun.

@lennysan @tfadell This is the design problem that product teams consistently underestimate. The product is one touchpoint. The customer is experiencing all of them simultaneously and forming a single opinion. Building a great product inside a broken journey is still a broken journey.

Love this reminder from @tfadell "Makers often focus on the shiny object—the product they’re building—and forget about the rest of the journey until they’re almost ready to deliver it to the customer. But customers see it all, experience it all. They’re the ones taking the journey, step-by-step."

every founder has a feature they're secretly embarrassed they shipped. mine taught me more than anything that worked.

@KevinSzabo14 The same shift applies to your landing page, your onboarding, your error messages, and your roadmap. Every time you catch yourself describing the product instead of the outcome, that’s the edit to make.

Irrelevant things for your sales pitch: - You - Your product - Your company Relevant things for your pitch: - How you can help them - What your product can do for them - What your company will do for them Forget about you, make it about them.

@aakashgupta The design instinct that catches this stuff is a product craft skill. And it’s the first thing that gets deprioritized when teams are moving fast. These numbers show exactly what that costs.

Every AI product shipped in the last 12 months has the same brand problem, and almost nobody has caught it yet. A PM audited her company's AI feature and found "I'd be happy to help you with that!" written 47 times. Shipped to production. Nobody noticed. Same audit: "An error occurred" with zero specificity on checkout. Three different CTAs across three pages. Onboarding copy where the tone shifts halfway through. Here's why this is systemic. LLMs generate text in their training distribution. The most common customer service tone in that distribution is cheerful, generic helpfulness. When you deploy an AI feature without a brand voice pass, you're shipping the model's median training example as your product's personality. Every company that skipped the copy review now sounds identical. Your checkout error reads like your competitor's chatbot reads like every SaaS onboarding built with the default prompt. The math on fixing it is wild. One PM spent a few hours on the audit. One PR. Engineer approved in 10 minutes. Changed "Something went wrong" to the actual error with a specific action. Support tickets for payment errors dropped 31% in the first month. Changed "Get Started" to "Start Free Trial." Signups up 14% over 4 weeks (n=8,200). Changed default sort from "Newest" to "Most Popular." Add-to-cart up 7% (n=22,000). Shipped in 15 minutes. None of these were engineering problems. They were judgment problems sitting in a queue behind engineering problems. The PM who bypasses that queue is doing $200K/year brand copywriter work at the speed of a commit message.

@dunkhippo33 Multiple apps, same problem. That’s not a portfolio strategy, that’s a customer acquisition problem wearing a product roadmap as a disguise.

I'm seeing a lot more businesses pitch multiple apps and even ideas, but the problem is still the same as in all prior decades: customer acquisition. Those are the insights to pitch, not the product.

@icanvardar Been watching this happen in real time. The less time teams spend on implementation the more time they spend on the question that was always more important, is this actually worth building at all.

the more writing code turns into a chore, the more we shift our focus to the bigger picture of the product

@rxhit05 “Build only that part. Nothing else.” The hardest instruction in product to actually follow.

Most founders spend 6 months building something nobody asked for. Here's what actually works: → idea → sell it before you build it → get 3 people to say "I'd pay for this" → build only that part → nothing else Products don't fail from lack of features. They fail because nobody wanted the features that got built.

@dunkhippo33 First startup: I need to build this. Second startup: I need someone to want this before I build it. Same lesson!

Almost all of my serial founders have figured out how to get customers super quickly, even sometimes pre-product. They learned from their first business that that was the primary problem.

@Bencera Ah, so true

Twitter is starting to feel like the dead internet. AI posting, AI replying

@vaaselene The code was never the hard part. Knowing which problem is worth solving was always the job. AI just removed the last excuse for avoiding it.

weird thing about building with AI: the bottleneck isn't writing code. it's deciding what's actually worth building.

@sherifgjini Learning too slow is the root cause of every other one on this list. You build slow because you haven’t learned what matters. You sell late because you haven’t learned who it’s for. You target wide because you haven’t learned who actually needs it. It compounds in all directions.

What is more dangerous? - building too slow - learning too slow - selling too late - targeting too wide - ignoring weak signals

@nxhaaa19 The builders who can’t ship are usually waiting for perfect. The ones who fail after shipping are usually waiting for feedback they never went looking for. Two different problems that look the same from the outside.

Most builders don't fail because they can't code. They fail because they can't ship. Do you agree?