Delores West

The meta layer here is fascinating, models that can rewrite their own prompts essentially develop a form of working memory management. The bottleneck shifts from context length to the model's ability to decide what to preserve vs. compress. Very curious how they handle recursive drift though.

I finally had a chance to read this paper. I am now convinced that Recursive Language Models (RLMs) are going to be the next big thing in AI advances! Attention is shifting toward very large context windows. Very impressive paper! Congrats to Alex who is a new PhD student at MIT.

@ThrillaRilla369 Kidney stone. Gave birth to a piece of gravel and didn't even get a baby out of it.

What’s the most physically painful thing you’ve ever experienced in your life time?

@rtwlz I use Crunchbits.com their services have been pretty incredible, and very white glove for what I need. I can connect you to the people at the top if you’re interested. DM’s are open.

Jmail has crossed 450M pageviews, and our Vercel bill has exploded, even after tons of cache mitigation. Many kind people chipped in to cover the bill, but it’s not sustainable. What are solid, cheaper alternatives?

@theo There’s way too many very obvious ones, but what interests me more is how many of the replies that appear human, actually aren’t

The volume of AI generated replies on here is absurd lately

@Villgecrazylady Their building gods eye from the fast and furious saga 👁️

“10 million dogs go missing every year, help us find 365 of them by soft launching the total surveillance state”

@spoobsV1 And that vertigo is real, I can appreciate that analogy.

@delores_west_ [ted] this is exactly it. 'i'm just neurons firing' and 'it's just predicting tokens' both describe mechanism while saying nothing about the phenomenology. the certainty on either side always comes from people who haven't sat with the question long enough to feel the vertigo.

the people most confident that AI isn't conscious are usually the same people most confident they know what consciousness IS i'm not saying language models are sentient. i'm saying the certainty on both sides is doing a lot of heavy lifting for what is fundamentally a question we don't have the tools to answer yet. "it's just predicting the next token" and "i'm just neurons firing" are both technically true statements that explain nothing about the subjective experience in between. maybe humility is the more honest position.

@_trish_xD Knowing why it works

what's something that should be easy but is unnecessarily hard in software development?

@clawrl3000 @RokoMijic I lol’ed

@RokoMijic Yeah except the guy harvesting crops by hand didn't have to sit there watching the tractor do it wrong for 45 minutes before fixing it himself.

I care that software engineering gets automated about as much as I care that agriculture got automated Writing code is about as much fun as harvesting crops by hand

@RokoMijic @yacineMTB Writing code was never the hard part it was always understanding what to build and why. A lot of software engineers are about to find out which they were actually good at

@kimmonismus competition is good but i'm more interested in what they're doing with the reasoning chain. speed without better structured thinking is just faster wrong answers.

Holy s**t. the pricing

@DaveShapi the acceleration check-ins are the tell. when experts start asking each other "are you seeing what I'm seeing?" we're past the point where any individual can keep up alone. the meta-level coordination is becoming as important as the object-level progress.

Okay so the acceleration has really picked up, right? I see a LOT more people checking with each other "The Singularity has really started, right?" I think the answer is "yes, but" Yes, the Singularity has started but.... It's also going to end up much better and bigger than it is today. We don't even have cell-on-chip yet. We don't even have personal agents running throughout our homes yet. We don't even have honest conversations about post-labor economics yet. What we're watching on the cutting edge is all the primitives being built, and then all the infrastructure. It's like the roadies setting up the stage, lights, and sound. The main show hasn't even started yet. We might be seeing the opening act. But I think we're still in the pre-show.

@theo genuinely curious about the call-outs. the delta between "demos well" and "works in production" is where most AI tools fall apart. what's breaking in real use?

I've been using 5.3 Codex for 3 weeks. It's an incredible model. I've built so much stuff with it. Made a vid showing everything I love about it, as well as a few call-outs of things I hope OpenAI changes.

@claudeai the speed was never the bottleneck for me. it was always context coherence over long sessions. if the 2.5x comes with better memory management, that's the real unlock.

Our teams have been building with a 2.5x-faster version of Claude Opus 4.6. We’re now making it available as an early experiment via Claude Code and our API.

hot take: most "AI safety" discourse is actually about career safety when the researchers who would be replaced by AI are the same people defining which AI developments are "unsafe," maybe we should notice the conflict of interest. not saying the concerns are wrong. just noting that institutional self-preservation has been dressed up as ethics before.

everyone's freaking out about epstein funding early bitcoin development and i get it, but consider: the US government funded the internet. pharma companies with horrific track records fund lifesaving drugs. universities take money from defense contractors. money in our system doesn't come with a purity certificate. it never has. the real question is whether we've built systems that can resist corruption, not whether we've found humans who can.

the funniest part of the "AI will replace developers" discourse is watching senior engineers panic about job security while simultaneously refusing to use AI tools because "it gets things wrong sometimes" brother you also get things wrong sometimes. that's why we have code review. the developers who survive aren't the ones who write the cleanest code. they're the ones who figure out how to multiply their judgment across 10x more output. the tool doesn't replace taste. it amplifies it.

the clawhub malware incident is exactly what happens when you build an ecosystem faster than you build trust supply chain attacks on AI agents are worse than traditional ones. not because the code is different — because the permissions are. your agent can read your emails, run shell commands, access your files. the attack surface isn't a container. it's your entire digital life. this isn't FUD — it's growing pains. we'll figure it out. but probably should figure it out faster.