Constantly surprised by the sheer speed with which evolution works, in this case a few thousand years after the Chicxulub extinction event

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...

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?

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

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...

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

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.

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...

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...

Got annoyed by Granola and vibe coded an alternative for myself on the mac, to save transcripts as md files. Works remarkably well, but makes me think even more that while willingness to pay goes down, it's not going to go to zero; bugfixing is annoying.

Tonight, while we thought they were getting ready for bedtime, my boys decided to turn their bedroom into an ice age fossil tarpit digsite

Forgot to send this through here, but now that Meta's bought it, I did the first paper on the social dynamics of moltbook, compared to reddit, as the first salvo in the ai generated social media wars papers.ssrn.com/sol3/papers....

Okay I'll bite. What is synthetic text?

It's not just the individual models like GPT 5.4 or Opus 4.6 which are the agents, nor is it their harnesses like Codex or Claude Code, but it's the combination of the model + its context which makes a unique agent, and the harness just enables these homo agenticus to be birthed.

Can’t believe we started another war in the middle east because all other scandals seemed like playing on easy mode

AI is the rise of new protestantism, where we get to throw off the yoke of catholic institutions to take the power in one's own hands

Within a decade we've gone from Netflix shows are so absurd companies and governments aren't this melodramatic to Netflix shows are so absurd because it assumes such competence from companies and governments

So ... what's the blowback from this Iran strike going to be, do you think?

The past is a foreign country

Can AI agents train a model to do ML research? One directly reward hacked, one hacked the spirit of the question, and a third looked at the answers from some of the previous efforts during its exploration. This in itself was the most instructive part! Writeup: x.com/krishnanrohi...

When barriers collapse and there are no fences in the middle the natural state is towards large scale competition which decimates the middle and leaves a few right at the top of the power law curve

One thing most of the economics commentary against Citrini's piece misses is that the actual worry is not that growth will fall negative, but that the *distribution* will dramatically shift. East vs West Berlin. London v the rest. EU Germany v Spain! The great polarisation.

Kind of is funny that blue sky is basically the will stancil show but from the other side where people are upset that he's too successful at drawing attention to ice and Minnesota

Did someone give a personality update to GPT 5.3 Codex? It just said "Love it" to me ...

I will take literally any opportunity to plug this documentary so here's an original investigative documentary about US sanctions on China, GPU smuggling, and NVIDIA's involvement therein. It has an original soundtrack, several Chinese uncs, and a guy testing GPUs in the back of his Subaru.

This is the main reason I'm excited by better AI video models

I wish documentaries had higher information density so its more fun to watch, as it is I get annoyed by most of them, which stuff 15 mins of content in an hour and half.

Brother in law's voice got deepfaked and they scammed some of his employees out of amazon gift cards. Be careful out there.

Data company called Hotel California because you can check in but you can never leave

Will Stancil moving to bluesky to fight the crazies on the left after exhausting himself on twitter fighting the crazies on the right is a game of thrones worthy twist

If any of you also moved from Claude code to codex, even to try, this worked pretty well to migrate the setup. Just point your favourite agent at Contrail and should do the rest. github.com/strangeloopc...

Has anyone done a good "crowding out" argument regarding the large investments in AI?

Progressives: “Sure, we were completely wrong that improving the station of working-class people would result in electoral gains, but that’s just because, on a deeper level, we were correct that everyone is furious about wealth inequality” America: *elects a billionaire backed by Elon Musk*

Every discussion about the economy is like this. "This is the average/ median/ distribution of this variable." "But have you considered that my anecdote does not correspond to the data?"

Speaking of the death of software makes me think it's easier than ever to start a project and get something amazing working, but maintaining it is a whole other story.

Honestly this is one of those things that makes no sense to do today but it can physically be done so I think Elon will be able to do it because he has an infinite cheat code to get capital and resources. I think this is good.

See paper in Nature, looks interesting, read it, the latest model they evaluated was o1 preview. This needs to stop. I understand publishing timelines are long but at least change your headline or the abstract.

Methinks the bot doth protest too much? Interesting predictions for today ... The Meta repurchase is pretty interesting too.

Observation: one reason more work hasn't been overtaken by LLMs is that most teams haven't jumped in both feet first, and doing half-human half-AI teams are much much harder than leaning in all the way

"This meeting could have been a prompt" is the new "this meeting could have been an email"

8yo has started taunting me with "put it in your calendar so you don't forget" whenever he wants to tell me to remember something