Models are RL'd for code. Writing isn't native. The labs aren't going to fix it for you.
AI writing is an abstraction problem: edit the spec, not the prose. Concept before structure before line.
Two months of learning — wrote this with the system it describes:
An amazing thing about YC is that founders consistently say they get a lot more value than they expected. This is amazing because they had high expectations to take the investment. And still they underestimated it.
It’s our responsibility as users to not be complacent and stuck on closed/foundation harnesses that just help them build their moats. I’m trying OpenRouter this week.
What’s your favorite harness?
Model routing is an important thing
Controversial idea: the frontier labs will want their AI harness to be the moat, but ultimately the best case for consumers is that model capabilities flatten and commodify
Preview of the AI Harness Wars of 2027
I bet there’s a model where Tesla could make more money selling FSD to other manufacturers. I’d pay >$10K to have it on another car. And then repurpose their manufacturing for gpu satellites and Optimus.
When VCs dont have a strong enough reason to say "yes" they are going to tell you things like "i dont think thats defensible" and "idk why OpenAI wont build it"
Just your daily reminder that the reasons that a VC will give you money
is different than the reasons they say no 🙏
If I was a Chinese AI model company fighting with Anthropic and OpenAI to distill I would create or buy a large AI app and slip in distillation prompts intermingled with real AI traffic. Like how would you distinguish between real Perplexity or OpenRouter traffic vs distillation?
Every time productivity has gone up humanity has benefitted (farming, factories, computers) by any rational measure (% below poverty, life expectancy, etc). This should be the default expected outcome from AI as it delivers tremendous productivity. The doomers should have to disprove this not vice versa.
Lets focus on how we can have a smooth transition (instead of chaotic) instead of fear mongering about the end result.
Two good engineering leaders I know (CTO and VPE level) are bootstrapping apps as one person teams and making faster progress (code and revenue) than they would have a few years ago with a team of 15-20. The future is now and amazing.
Google and Facebook, unlike Microsoft and Amazon, have never don’t much performance management. When I tried to manage someone out at Google I was told it would take too long and I should just convince them to switch teams. I totally get using RIFs as a blunt, fast performance management approach. Yeah there’s better approaches but it’s better than nothing?
Founders overestimate how hard it is to become a top 100 expert on most subjects.
For most things you need to solve, there’s usually a very clear path: just go do the work, obsessively, and become friends with everyone else doing it.