Safe is the new open

If Rowling was a coder, she'd be doing waterfall

Getting our new platform online, ensuring compatibility for our pre-existing customers and building our new infra automation was a tough grind but it's now done. While we're working on the next set of capabilities, I'd love to hear what you'd like to see us do in the platform.

microsoft rn #Ignite2024

Best protection against piracy is a product that's constantly being updated.

It's conceivable that an unrestricted BLT would be able to do things like "remove the license validation from this app" ai.meta.com/research/publi…

you can just do things now

If model weights are like a giant jpeg of the data on the internet, then hallucinations are a compression artifact. The more parameters in the model, the less lossy it will be.

This articulates well what it means to really lean in to ai for coding. Asking the llm for help might 2x you but letting it build entire capabilities like another human is where the tantalising prospect of 10x'ing your throughput lies. Personally, I can't commit to that (yet).

Do-nothing scripting: the key to gradual automation blog.danslimmon.com/2019/07/15/do-…

never thought about it like that,

LLM's should be trained on commit diffs to deal with the iterative coding quality deterioration problem.

This is a great observation

Sometimes I take a moment to appreciate how absurd all of this is.

This feels like the beginning of AI hype correction

Curious whether this could also mean that broad alignment could be achieved through narrow alignment.

> be andrej > coin "vibe coding" > internet breaks > casually mention you don't use it for stuff that's important

One of the bottlenecks for adoption of AI chat in enterprises is that the AI doesn't have a standardized way to access business data. That's changing with MCP (model context protocol) - a standard that's being rapidly adopted. We're a few weeks away from shipping an MCP server

Context is all you need. That really is the biggest bottleneck for AI in enterprise rn

> describe gpt 5 in one word > incremental