open.substack.com/pub/strangel...
Notes from a recent trip, on why Hong Kong is a mix of Singapore and Calcutta.
Posts by Rohit
Constantly surprised by the sheer speed with which evolution works, in this case a few thousand years after the Chicxulub extinction event
Lmao! Excellent tweet.
Yes, I agree.
It's the classic conundrum with these things, at least until we start getting plentiful smart cheap grads again to grade things with.
Hello Bluesky, someone did the most wonderful experiment to see if LLMs can indeed think out of distribution. Same one I and many others have been dreaming about. The results may surprise you!
michaelhla.com/blog/machina...
That would be so funny
Because of latency or capability?
Spatial reasoning harder one because it does seem to figure some of this out. But maybe.
Writing yes
This is so cool. Does anyone have a link to the article that isn't paywalled? QT-ing because replies are blocjed.
Question: what's a list of capabilities that AI has not meaningfully progressed on from 2023 till today?
Writing well is my example, esp fiction, where it's gotten subject cohesion but the slope of the line is much flatter than eg coding. What else?
Even with a queue they need a lot more explicit scaffolding to help them coordinate
My forecasts on AI's impact on the US economy, from Forecasting Research Institute's survey. It includes both physical and digital AI in its assessments. 2050 is far enough away that there are big error bounds there but the rest feels plausible.
explore.forecastingresearch.org/participate/...
Assessing US embarrassment is like measuring the hottest day. We keep setting new records.
the FPS -> Starcraft shift is exactly right. when AI agents start managing other AI agents, the PM job moves up a level - you stop managing tasks and start designing governance systems #AIAgents #ProjectManagement #AgentOrchestration
Eat more, fight the pain
Is this the biological equivalent of repeatedly training LLMs using synthetic data
For any piece of advice there's an equally potent and opposite piece of advice. The only advice that stands the test of time are tautologies. But they only make sense once you've gone through it.
This is one of my most favourite pieces I've written yet.
www.strangeloopcanon.com/p/strange-lo...
They have 1 novel out of a new series but the expanse stands alone. You should go read the rest!
It's just a new skill
It is extremely underrated how absurdly good GPT 5.4 Pro is at making powerpoint slides. It takes a hour or two but the output is very good.
i want a dozen more expanse novels but alas they ended that series perfectly
Way I see it is if we had a clear answer it'd be priced in already
Flawed, imo, but may interest @dsquareddigest.bsky.social
I think this describes the desired state well. to the extent an org faces or can put strategic or management decisions into low-ambiguity settings, I think this could make sense.
I don't see CEO world models working so cleanly for smaller orgs. otoh, maybe it becomes essential for larger orgs.
Interesting, yes quite plausible though could also go the other way. Or works for some types of orgs and not others.
Work used to be like a first-person shooter. Managing a company full of AI agents looks a lot more like Starcraft. Which means the next big category has to be to build a world model.
www.strangeloopcanon.com/p/the-future...
Prob undercounted either way. Most people name branches and many commits aren't PRs either. It's a mess.
What can't AI do!
I know it's not apples to apples but, Github has about 20m commits where Claude is a co-author, and about 7.6m commit matches for Claude Code.
Codex usually does a branch with codex in its name, so using that, there were like 3.7m PRs where head was codex.
Closer than I thought.
I really do want the ability to use Codex for more things on my desktop. Which is why I'd made and have been using Ensign, a desktop app I made, for about 5 months.
Just flipped it to public - in case you want to try it out. Install the DMG.
github.com/strangeloopc...