- notes from the bay
- Posts
- August 12, 2025
August 12, 2025
Pierce Freeman
August 12, 2025
It’s probably my 10th time going to Outside Lands and GGP never disappoints. It delivered another year of fog, good music, and constant Coachella comparisons from anyone who has been to both. Bleachers and Glass Animals both headlined. Together they were the only things I played during my 2020 spring in San Francisco, which made for a pretty nostalgic Sunday.
A few random things in case you missed them:
Tesla has entered the self driving chat in San Francisco. At the moment they’re way cheaper ($5) than Waymos ($25) and even Ubers ($15). Is anyone else smelling the sweet smell of a capital market subsidy?
The Perseids are going on tonight, which should make for good evening entertainment if you enjoy being awake at midnight on a Tuesday and if you live above 2000ft. Karl has us covered down here.
Richard and I broke down GPT-5, world models, and a lot more on the latest episode of Pretrained. Check it out.
Now on to the nerdy stuff.
Coercing agents to follow code conventions [link]
I'm very bullish about the long term opportunity for static analysis to improve agents. But that goes beyond flagging things that are technically correct but aren't quite the behavior that we want. What we ideally want is to specify our own rules. And it makes sense to specify these rules in code and not LLMs.
Where there’s smoke, sometimes there’s fire. People in San Francisco have been talking about agents for at least three years now. The simple idea is you have a language model, which delivers outputs message by message. But you run this LLM in a loop, passing its previous messages into its conversation to get additional messages. It’s a continuous loop until the model determines it has no more work left to do.
Programming agents have been the first to reach product market fit. I suspect this has economic origins: an entry level software engineer is so much more expensive than other jobs, so there’s a lot more money on the table. But the other fundamental advantage is intrinsic in the domain. It’s trivial to compile code that you write and see if it works. Information about whether that compilation succeeds or fails can be fed back into the network. It’s a tight loop.
The ecosystem of static analysis tooling is still very underdeveloped for how useful it is in the agent iteration lifecycle. If you can be more prescriptive about the code you want agents to write, you might as well codify those instructions instead of leaving them up to the system prompt. I open sourced determystic this past week, my POC at delivering a user experience around this theory. I’ve been using something like it internally for some time now and it really works like a charm.
My unified theory of social selling [link]
Cold outbound is getting harder than ever before. Our inboxes are littered with AI generated garbage. They seem vaguely personalized but the content rings hollow. Even if they're good, they make you feel like someone didn't even care enough to reach out personally. Based on my own experience selling enterprise software over the last few years, I'm willing to say that cold email outbound is pretty dead.
Social media is complex. The platforms are both fickle (Clubhouse came as quickly as it went during covid) but the best ones are surprisingly lasting (Snapchat’s resurgence among Gen Z). They’re also hard to square with the rise of AI. Is consuming content on social even interesting if half the content just becomes AI generations? Or does that make them even more addictive since they can be optimized to maximize dopamine?
I won’t pretend to know the answer to those rhetoricals. But I do remain convinced that authenticity and trust capital is going to become even more paramount in the next chapter of our lives online. And I think that manifests in the sales cycle as well. I’m already sick of the pseudo-personalized emails that get dropped in my inbox every day. The only emails I’ll open are from names that I recognize - either from personal relationships or from social media.
I talk to tons of entrepreneurs that are focused on virality. But virality is only useful if you’re reaching the right people. And it also tends to be pretty flash-in-the-pan. It’s good for a big bump of signups on some big announcement but that’s it. The more lasting presence comes from building the name ID that have people open that email when you send it out. I wrote some musings on my sense of social selling.
We need centralized infrastructure [link]
There's a bias in tech toward distributed everything. Distributed computing, distributed energy, distributed manufacturing. The assumption is that smaller and local is inherently better. It's usually more flexible because you're relying on more generic instruments (3D printers versus injection moulding). But this shouldn't be conflated for inherent benefits of decentralized infrastructure.
A departure from our regularly scheduled programming (pun intended). I’ve known at an abstract level that centralized solar installs are more efficient than all the edge deployments that people do on their roofs. But exactly how much more efficient? I try to make a concise pitch with the numbers. Really makes you think that we need to change our regulatory regime.
Until next time. May the tensors be good to you.
Pierce