AI, longevity, and the long future of humanity.
New York
Joined April 2009
- Tweets228
- Following599
- Followers354
- Likes193
Pinned Tweet
@DarioAmodei is right on embedded evaluators strategy but he misses the plane separation model. A "control plane" unreachable by any model. It doesn't prevent misalignment, but it buys rollback to the last aligned checkpoint.
We come across the concept of “wait calculation“ in the context of space exploration. Given the pace of progress with AI - absolutely applicable to startups today.
Excited for the way things are turning out.
I really feel for founders who spent the last 5 yrs building great businesses the current venture market doesn’t care about
seen many go from 0 to $50M-$100M run rate in 4-5 yrs, at or near 3xing YoY but most of the market just doesn't care lol so insane
they did everything they were supposed to do, got there, + the market decided it wasn’t enough. brutal!
only way back in is to reaccelerate over the next 12-18 months, mostly w/out VC
that said, the ones who pull it off will be better cos for it
@satyanadella reading your latest scratch is joyous and it bridges two recent articles I wrote on this subject.
The principal-agent problem, we built all of this for human agents two centuries ago and shipped none of it for AI agents.
It's time to grant mandates. Links to both below.
"If it can't be observed, it can't be trusted" assumes failures leave evidence. The worst agentic failures are silent: nothing happens, no event fires, every audit passes.
The brake has to live on a plane the is unreachable by any model.
nitter.cf/Sumit_A/status/2098923…
@DarioAmodei is right on embedded evaluators strategy but he misses the plane separation model. A "control plane" unreachable by any model. It doesn't prevent misalignment, but it buys rollback to the last aligned checkpoint.
The most important sentence in his piece is: a program must not be able to bypass or tamper with the mechanisms that enforce its permissions.
That's the reference monitor, 1972. In networking it became the control plane: you never manage a router through the plane it forwards on. Same principle, fifty years, still right.
If there's one video you are going to watch today. Let it be this one.
when I was a kid this was science fiction. i've watched AI type for years, but hearing it and seeing what it made is a whole different level.
i asked @claudeai opus 5.5 to introduce itself, however it wanted without any token usage restrictions and access to any api it wanted. it decided to write a song, make a video, visualize itself and give an interview behind the music.
it also wrote this thread 🧵
5th essay in the series and the most exciting so far. So buckle up and let's do a deep dive in to the HPG-axis and the peptides that influence the hormones from top to the bottom. Read this right, and it might just save your marriage.
Apple shipped the plumbing for a phone without apps with App Intents.
That is the real story of the next decade of software.
@Benioff is betting apps survive. @benthompson wrote the economics for the layer. @inkandswitch wrote the field's best manifesto.
@karpathy You called the LLM the kernel process of a new OS. Right about the kernel. A kernel has an owner, and the owner decides which of your tools users get to see. This needs a reframe.
@rauchg json-render constraining the model to an approved catalog is "persistent at the core" at the level of one single app. I think it has to become a law at the level of the whole surface, or fluid software is just a novelty.
@geoffreylitt Your essay says the app is broken and should become clay. I think it stops being something we look at at all: it dissolves into capabilities beneath a composition layer. Same diagnosis but different prognosis.
In April @bryan_johnson picked a GH peptide against community advice. One reason he gave; a 7.5x higher overnight GH trough.
The trough is the part meant to go quiet.
Three weeks later he stopped. GH up 8x. IGF-1 flat.
Why that was predictable:
Three peptides. Three origin stories.
Zagreb, 1975.
San Francisco, 1973.
TB-500, which is not the molecule everyone thinks it is.
BPC-157, TB-500, GHK-Cu. Origins, how they work, why the stack, and why no other peptide works as well without them.
Right now your body is firing tens of thousands of chemical messages. Most of those words are peptides, the most interesting molecules in medicine right now.
@hubermanlab, @PeterAttiaMD, @AbudBakri, @FastVitaminIV and @alc2022 have done some of the best work bringing this into the open. I wanted to build the part that comes before it: the ground floor, for anyone starting from zero.
The most complete, most beginner-proof explanation of peptides anywhere. Part one of the series is here.
The first human dose of ER-100 (@davidasinclair, @lifebiosciences). A real breakthrough out of VERVE-102. @chamath weighing in. @hubermanlab dropping a peptide masterclass.
Longevity just became the loudest conversation in the room. If you are only now tuning in, it is a lot, and most people have no idea where to start.
I do. So I am going to take you there.
This is the first in a series that gets the rest of us from lost to fluent. I will break down the science, the chemistry, and the biological pathways, in plain language and in the right order.
Start here.