John Zeratsky

Screw it, I've decided to start using em-dashes again. I think we're all mature enough to handle the cognitive dissonance.

I am now one of those weirdos who uses Obsidian for notes. The basic reason is that I wanted to keep my notes in text files. After I switched, I grew to love the Daily Notes feature — especially the iOS widget that provides one-tap access to today's note.

The "uni-context" is a really powerful concept. It does, indeed, explain most everything about the world today. Thx @DKThomp @AgnesCallard derekthompson.org/p/a-philosophe…

Launching is useful not because it exposes your idea to people, but because it creates news, and news brings attention, and attention is a social primitive that can be used for many purposes — including exposing your idea to people.

@Brock_Heinz Perhaps. Depends on whether engineers code to think or coding is more of a production exercise. Wdyt?

Will a similar statement about writing code be apt some day in light of LLM-based development?

@narsagna welcome to the light side

Just found out how to type an emdash as a human: option + shift + hyphen

@Stammy and built by some great guys, @nazarinhamid and @rcarver !!

main thing i like about aphera is that it's more culling-focused, editing is keyboard-driven and focused on minor tweaks and using pre-defined "looks". sliders are hidden by default, which gives you fewer decisions to worry about. you can even edit while you're in a grid view. it's aphera.co - im not affiliated but really liking it after dealing with Lightroom for a decade

📍Kona, Hawaii 📷 Shot on Leica, minor edits with Aphera

I’ve been asked a few times recently what I think about open-weight models and the criticism of foundation model labs from people like Chamath, Karp, and others. I think a lot of that criticism is directionally useful. But I also think the conversation gets compressed too quickly. It turns into open vs. closed, cheap vs. expensive tokens, privacy vs. no privacy. My main view is that model choice matters, but it is only one part of the decision. The more important question is what sits around the model, and how much of that layer a company understands or controls. A few thoughts: - Companies are starting from very different places. Some are just trying to get started. Some are already using AI and trying to cut costs. Some care a lot about privacy or control. Others mostly need something people will actually use. For a company just getting started, adoption may matter more than the perfect architecture. - You have to separate the model from the application. Claude Code and Codex are not the same thing as the LLM. They are products built around an LLM. Much of the value is in the harness: the connections to tools, applications, context, permissions, and workflows. - That is what lets the model fit into how people already work. It is also why I think the labs are right to build products around their models, not just release models. - But a good generic product is not the same thing as what a company eventually needs for its own work. The labs can give you a strong starting point. They usually cannot give you the full version that understands your business, your workflows, your data, and your standards. - One thing that does not get enough attention: the harness can be opaque too. People focus on whether the model is open or closed, but the application layer can be closed as well. - What instructions is it running? How does it choose tools? How is context selected? How do those rules change over time? - Those questions matter if real work is going to depend on the system. There are more transparent ways to build harnesses around both closed and open-weight models. The question is not only “which model are we using?” It is also “do we understand what we are putting around it?” - I do not think every company needs to build this from scratch today. But I do think harness engineering will become a real skill. Most companies do not have it yet. Over time, more companies will want to own, inspect, or at least understand that layer. - Another reason this matters is model flexibility. Ideally, testing a frontier model, smaller model, or open-weight model should be closer to a configuration change than a rebuild. That is not always clean, especially when a closed model is tightly coupled to its own product, but it is the direction companies should push toward. - To do that, you need evals. You need a way to know whether the workflow still works when you swap the model. Companies that get good at this can move faster, use the best model for the task, avoid overpaying where a smaller model works, and make decisions based on results instead of model vibes. - The frustration with foundation model labs is real. Token costs can be high. ROI can be hard to measure. Data concerns matter. A lot of companies are still figuring out what AI should actually replace, speed up, or change. - AI is not just another SaaS tool. It changes how work gets done, who can do it, and how teams are structured. It also has a different cost profile: upfront learning and process change, plus ongoing usage costs. - None of that changes my view that companies have to figure this out. Used well, AI creates real value. The hard part is that the value often shows up as faster iteration, avoided work, better leverage, or different team structure. Those things are real, but they can be hard to measure early. - Control matters, but not equally for everyone. Some companies will care a lot about where data goes, what tools the model can touch, and who can inspect the behavior. Others will care more about speed and adoption. That is why the answer will vary by company. - Over time, the model itself will matter less for many internal use cases. Not zero, but less. Open-weight models are already good enough for a lot of internal work. Many use cases do not need the best frontier model. They need a good-enough model in the right system. - But “good enough model” does not mean done. You still need to define the problem, provide the right context, design how people will use it, catch mistakes, and define what good output looks like. That is real work. So I do not think the takeaway is “foundation model labs are doomed” or “open weights win.” Some teams will start with closed products because they are easy to adopt. Others will use open weights because they want more control. Both can be rational, but in either case, the model choice is not the strategy. The strategy is knowing what you want AI to do, what context it needs, what tools it can touch, what rules it should follow, and how you know whether it is working. That is why the harness matters. It is where the model turns into something useful. It is also where hidden behavior can creep in. And if you understand that layer, you can test models instead of arguing about them in the abstract. Over time, more companies will want to understand it, shape it, and own more of it themselves.

True greatness cannot be captured in a headline, essay, interview, speech, or book. There will always be a gap between what we know about greatness, and what greatness actually means and does. In that gap is art, and alpha, and complexity, chaos, the supernatural, and indeed, life itself.

We just opened applications for Character Labs G7. The deadline is not until September, and the program starts in October, but we do give priority to early applications and sometimes invest in teams before the program begins. So, if you're at all interested in working with us (or even curious), you might head over to character.vc/labs and check out Labs. We'll also be sharing a lot more about what's changing for G7 in the coming months!

And this is why writing must survive

@paigefinnn And so often the domains are those you would NOT expect to be so interesting 😅

one of the reasons I love partnering with domain expert founders who have faced the challenges they're solving themselves (spent time on the factory floor etc) is their stories during a fundraising pitch are so wonderfully alive, specific, and full of interesting DETAIL.

Speed is a powerful advantage. You're not the only one who has it, but that doesn't make it less useful. Whether you're building, investing, selling, hiring... going fast will improve your odds. But there are two tricks: 1. Knowing when to use it. E.g. don't be hasty in deciding who to hire, but once you know, move fast to close them. 2. Staying fast after you achieve success. It's incredibly rare, so your advantage will compound if you're able to maintain your speed.

@nickbaumann_ You visiting Wisconsin?

SFO --> ORD last night

@Brock_Heinz That's a great point. I think you're right that it's largely a challenge of personal/leadership growth.

@jazer I think that one of the challenges is that after "Crossing the Chasm", startups have found PMF and it's then time to evolve from pioneers to settlers. That shift is radical and it's why often you see founders drop out of CEO role around that point.

It can be hard for startups to break past $4–6 million in revenue. Obviously, plenty do it, and others would be thrilled to get there at all. But I've seen a curiously large number of companies who get stuck around this range. I wonder why that is?

When showing someone your product, think about tours, not blueprints. Walk them through the experience of using it, and lead them to the a-ha moment. Don't click through each tab and explain what's there. It's kind of like showing someone your house... you don't pull out the floorplan and give them an overview of the rooms. You welcome them at the front door and show them around.