Maybe this is an unfair comparison, but Apple saying they’re going to release the AI stuff “in the Fall” sounds insane when you contrast it with Anthropic releasing like 10 things a week
It’s like criticizing runners for running too many miles. Yes, there may ultrarunners who are running hundreds of miles and breaking down there bodies. But, for most people, they’d do a lot better if they ran more.
At the extremes, Goodhart’s Law makes tokenmaxxing look silly. But the average employee right now doesn’t know how to use AI properly. They don’t know how to spend that many tokens. So tokenmaxxing is actually a good policy.
I know a few people inside Big Tech, and I'm sorry to report but Token Psychosis is real.
Their bosses are hyperfixated on AI usage and tracking who uses the most tokens, so they're essentially running agents 24/7 on do-nothing tasks just to use more tokens. Incredible stuff.
Finally killed Asana by rebuilding my to-do list as a live artifact in Claude then using scheduled tasks to scan my slacks, email and calendar each day to propose to-dos
Infinitely, unbelievably better and more effective
More and more workflows being sucked into Claude
Wanted to give them the benefit of the doubt, but Opus 4.7 has been pretty bad for me. Takes at least 2-3x the tokens to complete an equivalent task as 4.6, so I run out way faster than before. The results aren’t significantly better either, certainly not 2-3x better
@buccocapital It seems like teams there are just blasting out random products and features in like a week from start to finish. I’m guessing they didn’t know about it until very late
Anthropic vs OpenAI so far is one of the shining examples that product still actually matters. Not just momentum, business agreements, brand recognition, etc
@karpathy I’ve been wondering how this will be used in a work context. People should be able to leverage their personal knowledge bases to do their work without surrendering what they’ve built up over time to the place they’re working for at a given moment.
Skills are amazing, but the more I try to share them with others, I’m convinced that the best thing to do instead is share a CLI and tell them to have their agent make their own skill in a way that makes sense for them
it can be tricky to share skills between harnesses since the features of those harnesses can be quite different like you said (e.g. does it have subagents, does it have askuserquestion, what is the frontmatter, etc)
but I do think skills as a concept will still be a winning move in the end, the harness can create skills for you and the progressive disclosure of skills is a killer use case. your slash commands are already quite close to that, maybe if you let them live in a folder?
@buccocapital I’ve been thinking about this problem. There should be a way for a worker to share their context temporarily for work, but not have to give up their entire knowledge to a company permanently.
It should be like how some trades own their own tools that they bring to work
Hand-typing notes + Obsidian CLI + Claude Code is an overpowered combination right now. I doubt these surveillance apps like Granola can even approach it
At @Google, we are moving from a writing‑first culture to a building‑first one.
Writing was a proxy for clear thinking, optimized for scarce eng resources and long dev cycles - you had to get it right before you built.
Now, when time to vibe-code prototype ≈ time to write PRD, PMs can SHOW not tell. Role profiles are blurring, creativity and building are happening in parallel.