Weston Beecroft

just a quick prototype but i think this kind of ui could be part of a future ai augmentation system. the idea being to create a meeting point between human, ai, and environment, enable rapid feedback cycles between the 3 in a general way.

claude mythos just gave me the best haircut of my life

Cyber is the first clear and present danger from frontier AI models, but it won’t be the last. If we are able to collectively rise to the challenge and confront this risk, it could serve as a blueprint for addressing the even more difficult challenges that lie ahead of us.

It was more interesting than that: he set out to research the "psychology of invention in the mathematical field" and found that good mathematicians in general tended to use a similar kind of visual/schematic thinking

Carving nature (b2b SaaS) at its joints

i.e. the temptation is to kind of unix philosophy-maxx and build all these general compositional pieces because strong decoupling wherever you can get it is gold with spaghetti-attracted agents at the helm--but designing the boundaries of these component system is nontrivial work

But it feels unnatural because the level of decomposition required feels like overengineering, and excessive generality can bring noise of its own.. then again, agents are good at dealing with the kind of structured noise that heavy abstraction brings

I'm getting frustrated with abstract planning and want more structure, but I'm reluctant to use code to get it because it brings in too many inessential concerns.. It's more like I want something to keep asking me questions and pushing things forward without exerting influence

@moultano I wonder what explains its variation across organisms if that’s the case (esp within eg mammals)

The theory that aging is primarily your cells forgetting which part of your body they are in seems both compelling and intractable to fix.

Kurzweil’s timeline of major pre-singularity events has been remarkably predictive but we should note a major contraction with the unexpected leapfrogging of the lower alphabet stack directly to g-stack in 2026 which Spiritual Machines had slated for ~2033 *at the earliest*

Figuring out how to generate/sample and render the boundary of these objects in such a way that the 'true' structure comes through ended up being more involved than I anticipated, too large in scope for the couple days I had. Was fun to hack on though:

The original Mandelbulb was constructed in such a way that it wasn't a nice theoretical object but I found some more recent work where it shows up as a 'trace' of another 4D structure (see Hopfbrot) so I spent a little time building something to render other slices of it

2027: Ottessa Mosfegh publishes 12 vignettes, each depicting an impossible mental disease through brush strokes of dialog, plot, and setting. The implied, alternate, shared cognitive foundation of her characters resolves a conjecture in geometric deep learning; singularity

@mlajtos_mu Hype almost isn’t the right word.. I think a bunch of people were sincere.. but it’s like we over-primed to expect everything to be revolutionary or something

@mlajtos_mu Yeah. Rapid cycle of opinion change, even the author taken back by how much people jumped on hyping it

I think we learned something new about hype today

@embryosophy One of my greatest accomplishments tbh

you also did something this miraculous as an embryo when you began growing yourself before you grew your nervous system

we as software engineers are becoming beholden to a handful of well funded corportations. while they are our "friends" now, that may change due to incentives. i'm very uncomfortable with that. i believe we need to band together as a community and create a public, free to use repository of real-world (coding) agent sessions/traces. I want small labs, startups, and tinkerers to have access to the same data the big folks currently gobble up from all of us. So we, as a community, can do what e.g. Cursor does below, and take back a little bit of control again. Who's with me? cursor.com/blog/real-time…

@jakedowns 😂

9_9

I use LLMs at certain stages to review specs and look for issues but I don't let them write; that's helped to make abstract iteration viable, and I do think my rate of key-problem solving w/ this method is faster than code. But TBC I still don't know how this'll all turn out

I've gone through ~3 major revisions of actual spec docs, then I accumulate notes files like the above in response to new concerns revealed. then I consolidate and make another pass. this one feels final, and it'll be extensive, but tiny next to LoC once agent implements no doubt

The idea that a "sufficiently detailed" spec is ~equivalent in complexity to implied code is obviously false, and I'm trying again to take this approach since current project is well-suited to it--but I'll admit so far I've mostly just been typing out massive CoTs into sublime