CoS role ->

This one is a personal ask. As we grow quickly, I need someone who can be an extension of me – a Chief of Staff/Ops working across the biz to help us scale. For this role, I need someone who can:.- Accelerates our recruiting process to match growth.- Instill processes for

Design Engineer -> Full Stack Engineer ->

AI researchers might be scarce…. Unless you’re paying $20m+ like Zuckerberg. We’re on a different race: hiring engineers who are excellent coders and also think like designers, PMs, and system architects. We’re hiring a few - see comments. This sort of engineer solves for

A new role is emerging: Forward Deployed Engineer. I’m a big fan (and need one). FDEs span what would traditionally be separate roles for sales, solutions, or support engineering. They drop those barriers and are tasked with helping prospects and customers through the entire

->

We’re going to see “engineer” in job titles like never before. AI is forcing non-technical positions to think about systems and processes. It's all about scaling. That’s why we’re hiring 'engineers' across every discipline:.-Forward Deployed Engineer.-Design Engineer.-DevEx

OpenAI really took the "one more mode" tweets to heart. Seems like Deep Research will be a button instead.

Chrome, Cursor, Slack, Superhuman, Figma, Linear, and Zoom. That's it. No local DeepSeek running in the background.

When "founder mode" takes up >128gb RAM. Here's my stack:

So this is probably the weakest part of the AI models today. And it's not tested much in the benchmarks. Human creativity/intuition/empathy will be at a premium.

There's a whole set of domains where this is true: what's a good product to build? does a UI look good? what's a good government policy? . A correct solution is subjective or hard to measure. And defining the problem is in essence part of the problem itself.

Hints at the larger question: how do you teach a model to be good at subjective tasks?. With marketing content, it is consumed and evaluated by humans. By definition, the model can't tell what's good/bad without trying it on humans. Consuming the internet is also not a solution.

This means:. 1) you don't need humans to create content to train the model.2) you have a way to teach it how to think methodically. There was a big gotcha: the "<thinking>" steps were gibberish to humans. They had to tune it with human supervision so it made sense (to us).

On its own, the model learns that doing more <thinking> steps and trying many approaches before outputting an answer is the best approach. I.e., it learns to reason. This is possible on math/coding problems because they are well defined and have factual answers. 2*3=6.

The paper explains that the LLM is *not* taught how to think. Instead, the LLM is given a math or programming task and given free reign on how to come up with an answer. The LLM is then graded (trained) on whether or not the final answer it arrives at is correct.

Noticed LLMs are pretty meh at writing marketing or sales content? The reasoning models don't solve that. We'll probably have an AI math olympiad PhD software engineer before there's an AI SDR or AI content marketer. DeepSeek's paper gives a hint as to why 🧵.

Time to buy a mac mini to run AI agents locally. Combine this with vision models and you have an Operator that can run on device. Great stuff by the Block team. Uses Anthropic's MCP.