***** try searching "vc-wars"

It's not officially the holiday season until every developer you know starts posting on their blog again or redesigns it. https://t.co/ORhAMCymZ5

After hitting enter on claude code w/ "please write me a working trading algorithm that never loses money, make no mistakes"

I let my 7-year-old nephew describe a video game he wanted (Mario but with dinosaurs) into Cursor with Whispr Flow and it just built it. I'm typing this by hand in an empty room because talking to my computer feels weird to me for some reason.

https://t.co/jnsfLRF6Wt

skills are fantastic for a lot of things but if you don't have a custom /deslop command in claude code that cleans things up and deletes all of the random markdown files you're torturing yourself for no reason.

Real barriers: some engineers just want to code, users can't talk to 40 people, PMs want to intermediate. Small flexible teams beat large orgs with rigid role hierarchies.

User interviews are often confirmation bias theater. Overly guided, narrow scope, designed to validate decisions someone already wants to make. Without dogfooding + real user contact, how likely are you to build something great?

The method: dogfood constantly, check with team opportunistically. Catch feedback when they mention something organically. Don't burden them with formal zoom interviews.

Example: search returned a Russia-based defunct company as relevant. Quick fix: added facets for HQ location and company status. Another: stale data from pivots. Solution: process web data more frequently. Metrics often miss this stuff.

I use our investor tools daily. Search for companies, investigate deals, participate in deal teams. Standard: I won't rely on model output I don't trust myself.

Most technical problems are people problems. Specifically: engineers not spending enough time with users or using their own product.

Something’s always happening https://t.co/HnLKqQJfpP

Here's Martin's post:

Obstacles are real: ecosystem maturity, tooling, team culture, what we teach juniors. But the direction seems set. Our abstractions may be shifting from human-friendly to agent-friendly. Does that sound like the future, or is it just an interpreted language with extra steps?

New hybrid roles will emerge: technical PMs or systems-minded engineers who write verifiable requirements for coding systems. Industry will figure out who fills these roles through experimentation.

Formal specs become the central human artifact. Implementation becomes the machine's domain. The abstraction layer for humans moves up the stack.

The near-term shift: Specs, prompts, and requirements become first-class artifacts. They merge together and become more formal over time.

This reframes the question. Formal verification is about proving code correct, yes. But it's also about keeping humans in the loop at the right altitude: one checkpoint in a larger verification and validation chain.

If AI writes most code, the dynamic shifts. The best abstraction for the writer and the reviewer may no longer be the same. You can't review a million lines of assembly. But with the right tooling, you can formally verify it complies with a spec.