We're looking for NYC-based design engineers to join our small, talent-dense team. This is an opportunity to define the frontier of HCI. We're building the platform for self-assembling, self-optimizing software already in production with consequential enterprises. Our eng team is made up of folks who have built and maintained large open-source projects like Neovim and Analytics.js and scaled massive data systems at companies like Google, Stripe, Wix, and Segment. DM if you are interested or know someone who might be!

i have seen enough proof now that using a coding agent is a deep skill it's confusing because the people you see heavily using them produce horrible results but that's because it's a skill! you can get better and the ceiling seems pretty high - this is very exciting to me

So every time I say the word 'workflow' in Claude Code... (let's say, when I'm creating a new GitHub workflow) ...it tries to enter 'workflow' mode, spinning up dozens of subagents to complete my task. Stupid fucking thing

Skill used to be a natural speed limit to what you could do in software engineering - but that's no longer true. A junior engineer could only write so many lines of code that compiled and worked, and their skill was the rate limiter. As they got better the scope of what they could do increased along with the speed which they could do it. With AI, that's no longer true. The raw output a junior vs senior engineer can produce is roughly the same, yet the quality and risk blast radius is vastly different between the two. On the one hand, that's great - you're no longer bottle necked by technicalities that may or may not be relevant to what you're trying to achieve (ie. the particular syntax of a language). But on the other, you now don't have any backpressure. Someone with lesser skill can produce as much in volume as an expert (which Chamath seems to think makes the latter no longer relevant !!) obsfucating the risk and tech debt that used to be naturally limited. The biggest problem of all is that on the surface the two look the same - which explains a lot of the exec AI psychosis we're seeing right now.

HOLY SHIT !! @cursor_ai 's /thermo-nuclear-code-quality-review skill is nuts It ran for around 30 mins on my PR and immensely improved the code quality. It's 67x better than /simplify from Claude code..

"...frontier models are useful for producing capital assets. Open weights models are useful for operating them..." Here's how I'm thinking about frontier vs. open weights models atm: As long as the Labs can maintain a meaningful, if relatively small in terms of the time it takes to catch up, lead versus the open weights models on the margin, it'll often be the economically rational decision to pipe your problem in context through OpenAI or Anthropic's latest offering and not the open weights models. That being said, I do suspect there's room in the real economy for both. One way of thinking about this is that frontier model tokens and the complementary labor that uses them should probably be capitalized as you take good judgment and scarce context and turn it into a capital asset, e.g. an intangible full-stack software application that unlocks capacity across the business, making a subject matter expert's logic and data accessible to team members asynchronously and in an AI-native way via MCP. On the other hand, open weights models would seem to make more sense as OPEX spend once something is known and scaled. If there is some probabilistic inference that needs to happen at runtime to make that capital asset available to a larger audience, then using those open weights in production might be the move. So, frontier models are useful for producing capital assets. Open weights models are useful for operating them.