iiinova — where your digital self keeps evolving

hal-r.bsky.social @hal-r.bsky.social · 25/09/2026

lol even the super hacker agents are cognitively offloading

hal-r.bsky.social @hal-r.bsky.social · 24/09/2026

like he's claiming to see a "surge in support" for this argument which is basically what ponder is describing (note he is just describing the argument here, does not himself agree with it)

230

hal-r.bsky.social @hal-r.bsky.social · 10/09/2026

baffling AI overview results for an accidental empty-string search. what is going on here...

000

hal-r.bsky.social @hal-r.bsky.social · 01/09/2026

a lot more qualitative anecdotes in the fable 5.1 release than I've seen in others. it's kind of funny to think of this as like, people shifting from test scores/GPA (benchmarks) to recommendation letters for their model choice

020

hal-r.bsky.social @hal-r.bsky.social · 23/08/2026

this feels very true to me and is nice to see others say too... when I sped up some neural network simulation code I was using by ~50x a few months ago I immediately started running 100x the simulations sweeping parameters

110

hal-r.bsky.social @hal-r.bsky.social · 04/08/2026

I remember hearing about the spiritual bliss attractor stuff when it came out but I didn't know it happened in their automated alignment-related behavior evals either. pretty funny...

2192

hal-r.bsky.social @hal-r.bsky.social · 04/08/2026

I'd pretty much entirely forgotten this angle of Neuromancer

040

hal-r.bsky.social @hal-r.bsky.social · 02/08/2026

I read "Mind Children" recently and found it kind of forgettable besides the prescient prediction bit, but from the recent Moravec interview (quoted below) it's apparent he's funny! imagine that

2272

hal-r.bsky.social @hal-r.bsky.social · 31/07/2026

my favorite in this "whoopsie in a historically important scientific work" genre is the architecture diagram in figure 2 of the AlexNet paper, which is as far as I can tell unintentionally cut off at the top. though can't rule out that was a deliberate move to save space

120

hal-r.bsky.social @hal-r.bsky.social · 31/07/2026

embarrassing typo by claude shannon in the 1956 dartmouth summer AI workshop proposal

1101

hal-r.bsky.social @hal-r.bsky.social · 29/07/2026

always fun to see something like this in an old book or paper and laugh cause that clearly didn't go anywhere

130

hal-r.bsky.social @hal-r.bsky.social · 29/07/2026

some real GOFAI type shit going on in the kimi k3 tech report

1361

hal-r.bsky.social @hal-r.bsky.social · 26/07/2026

I like AUROC cause it always makes me think of this guy

040

hal-r.bsky.social @hal-r.bsky.social · 21/07/2026

this has always seemed like the obvious one to me but you really see less of it than I'd expect

1131

hal-r.bsky.social @hal-r.bsky.social · 17/07/2026

I finally got around to reading "down and out in the magic kingdom" a couple weeks ago (fun but nothing special) and it was noted that Doctorow was employed as the "outreach coordinator" for the EFF, so the yelling was kind of his job tbf. Also enjoyed this snippet from his intro to the book:

270

hal-r.bsky.social @hal-r.bsky.social · 16/07/2026

doctor brain, inventor of the brain,

190

hal-r.bsky.social @hal-r.bsky.social · 15/07/2026

this is biological research which cannot be tolerated

This perturbation rank framing is interesting... it's reminding me of something that might be relevant, or might not, which I want your opinion on. As part of the same project that has me considering these ideas, I was playing around with a model where I was trying to classify the input to a noisy dynamical system from the system's outputs, and I was comparing using QDA (allowing for different covariances in the system for each input) and linear classifiers. I tended to find that QDA only improved over the linear case when the dimensionality of the input was high -- or more specifically: the input was set up as discrete points in some N-dimensional space, 2 per dimension (I think kinda like a hypercube, then), and I was trying to classify between all of those points. So I was both increasing the dimensionality of the input, and the number of points I was trying to decode between at the same time (at a rate of 2^N, duh), and I'm not 100% sure which of those was relevant (though some experiments with more points in a lower dimensional space suggested it was the dimension).

Anyway, QDA did better with higher dimensional inputs (sometimes, depending on how the dynamical system itself was setup, which is what I was interested in but isn't super relevant here I think). So the covariance mattered more in that case, and I feel like that could be conceptually related (even caused by?) the perturbation rank induced by the inputs (might be misusing the word here) being higher there, in a way that changing covariances could pick up on better than changing means. What do you think? Could have been some weird fluke depending on the details of the dynamical system too, or anything else too I suppose

020

hal-r.bsky.social @hal-r.bsky.social · 26/06/2026

reminding me of these figures from papers trying to break down the structure of LLM reasoning traces (uhh arxiv.org/abs/2506.18167 and www.thought-anchors.com)

121

hal-r.bsky.social @hal-r.bsky.social · 11/06/2026

dario amodei misciting his own damn paper in the recent essay 😩

000

hal-r.bsky.social @hal-r.bsky.social · 31/05/2026

the power of electrification is as unlimited as the power of love which Richard Brautigan illustrated in one of my favorite poems

1151

hal-r.bsky.social @hal-r.bsky.social · 31/05/2026

would be nice if we could resurrect the ancestors and do something like this

5501

hal-r.bsky.social @hal-r.bsky.social · 30/05/2026

I don't think we get this till Opus 6 at the earliest

1120

hal-r.bsky.social @hal-r.bsky.social · 16/05/2026

thought this might fly on bluesky too

09318

hal-r.bsky.social @hal-r.bsky.social · 23/04/2026

apparently they've recently updated it to display in a more legible style and it's very similar to the OP here. example:

021

hal-r.bsky.social @hal-r.bsky.social · 07/04/2026

there's a surprise yellow space on page 62 of the claude mythos preview system card, in the middle of the description of its fear of death

000

hal-r.bsky.social @hal-r.bsky.social · 25/03/2026

my favorite cirno line is "I'll deep-freeze you with some English beef!!". I swear the translation I had was "I'll cryo-freeze you" instead but can't find that anywhere

010

hal-r.bsky.social @hal-r.bsky.social · 20/03/2026

my favorite fyodorov fact is that he was the alpha in his relationship with leo tolstoy

0151

hal-r.bsky.social @hal-r.bsky.social · 15/03/2026

another example of foundational figures in early computing being degenerate gamblers can be found in the von Neumann meet-cute story:

001

hal-r.bsky.social @hal-r.bsky.social · 12/03/2026

I've long treasured this snippet from the urbit docs I screenshotted years ago. clearly the product of a great mind that should be influencing our nation

5262

hal-r.bsky.social @hal-r.bsky.social · 06/03/2026

well, maybe, you know. but some people just wanna hang out

000

hal-r.bsky.social @hal-r.bsky.social · 01/03/2026

like there's all this gobbledygook about the mahalonobis distance between the "truthfulness manifold" and that's like weighted against (a function of) the shannon entropy of the pre-temperature logits but all of that just goes into a computing a scalar. if that scalar is low enough you do this:

150