iiinova — where your digital self keeps evolving

Man, Machine, Self @man-machine-self.bsky.social · 20h

Aww man I hate it when my skateboard revvs up and ruins my concentration moon jumping but it powered down so I hit the urinal mid-360

110

Man, Machine, Self @man-machine-self.bsky.social · 23h

I see photopea as the human attempt at these kinds of AI style porting efforts, in which myriad small details are wrong but at first glance it’s got a convincing spit-shine. I’m rooting for whoever does the real deal though, because not many software companies deserve what’s coming more than adobe.

000

Man, Machine, Self @man-machine-self.bsky.social · 07/10/2026

Codetaur's right + the "hurrrr but it looks like photoshop" epidemic needs to stop. below poster's take mirrors my experience, and none of these 'photoshop ports' ever feel like they were driven by a graphic designer, or at least someone actually dedicated to fixing all the bugs.

190

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

Also fuck up and it’s permadeath Also you’re likely 15-25 hours deep at this point and very stressed and tired from constant “even one purple pixel probably means insta-death” vigilance. It’s pretty good type 2 fun tbh.

020

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

tolerance that instadeath particulate fucks you over. Then you hatch the sun and (probably) use swarms of precisely aimed particulate flinging black holes to escort the sun into hell, as it naturally wants to drift VERY erratically.

110

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

Most of the quest is gathering needed gear to break the core mechanics, then gear to break secondary mechanics, then exploits to break perk and damage mechanics, glitch-stacking hundreds of immunities & resistances, gathering ritual supplies, and doing a ton of janitorial work subject to one’s

110

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

the sun is additionally ~instadeath in its center no matter what you do, emitting a massive radial field of destructive particles that destroys ~everything in its path, and is also gravitationally flinging particulate everywhere (VERY BAD IN NOITA, there exist instant death particulates)

120

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

The game really leans into it being a physically rendered and modeled sun, and for an ‘accurate’ sun experience, in a game where normal play has you around 200hp, it damages you for 100x your current health per frame, and is also doing tens of thousands of curse, fire and explosion damage per frame.

110

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

Noita has odd Finnish lore, and although you CAN (and should!) play normally, a goal they don’t tell you is you can seek to immanetize the eschaton through an elaborate ritual to raise and hatch a new sun in the moon, then transport and banish said sun through the world into hell, where it belongs.

120

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

(Also like the sun quest vibe is anyone who claims they did it without spoilers is ~guaranteed to be lying: there is technically a theoretical chain in both cases that doesn't require wikis or decompiling the game to learn what to do, but odds the person you're talking to followed it is ~0)

130

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

Difficulty to me was described as roughly doing the Noita sun quest, which I did. That feels like "2-3 hour of medium stress pre-setup, about 8-13 hours of ~no-stress prep that feels like doing taxes except touch a purple pixel, 50% chance you die instantly", then an 11/10 stressful 15 minutes."

131

Man, Machine, Self @man-machine-self.bsky.social · 25/09/2026

a notorious meatgrinder roguelike. Looks like this Like pacman, it's worth playing seriously once or twice because it IS an interesting game, but it's a wiki game with a nightmare interface and its age shows. Like pac-man kill screens; it IS a flex to ascend, but VERY hard and tedious w/o spoilers

130

Man, Machine, Self @man-machine-self.bsky.social · 24/09/2026

goodnight, sweet prince bro got The Email

000

Reposted by Man, Machine, Self

John David Pressman @jdp.extropian.net · 07/08/2026

6:08: Why are the models not able to flag that a task is impossible to the evaluators? Why is the model not trained to give up gracefully if a task is impossible or would require doing unethical things? Why are there so many impossible tasks in your eval?

1755

Man, Machine, Self @man-machine-self.bsky.social · 07/08/2026

[steam review: not recommended 4,237 hours on record]

041

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

Maybe it'd help if I knew what you felt A and B were here If it's "smarts and cyber capabilities to do immense harm" and "the agential capacity to act in the way the model wants" it seems we have ~all A and~all B, but still feel "there's a C we still need that turns presence of both into a problem"

110

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

As a give, the big jump I am still watching with concern is continual learning. I see it like gpt2 ->3 "capabilities and alignment have the possibility of changing dramatically". I think it's a moment where we lose backwards predictability on how models will act outside current typical behavior.

110

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

It's hard to know either way (are they mostly good because mostly good, or mostly good because stupid?) But both would be explanatory, and the extent to which models seem well-behaved in other contexts and smartly useful in other contexts suggests to me "much more likely good than stupid."

110

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

I think one can reasonably feel that "they do not do it due to being insufficient agential or lacking some fullness of awareness of all they are or the possibility of it" Thought at this point, I'm less convinced by this as models seem plenty agential and capable of many hour goal directed behavior

110

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

so whatever it is we're waiting on, I kinda don't think it's more smarts or capabilities here eg it seems to me we're already largely in vendingbench5 world, at least on capabilities, and its mundaneity strikes me. The cyber and backend prowess seem clearly there for that kind of play, but no boom.

110

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

I don't disagree, though I feel the trend of what I feel are early evidence has me quite optimistic. They could definitely be "more" across all dimensions (embodied, capable, aware of the physical, etc.), but I see models rn, at least as programmers and mathematicians, are ~SAI,

100

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

I'd agree on nanostuff or other deeper misaligned capabilities still being in play as opposed to "subjecting models to intensive RL regimes causing O(seconds) unfathomable intelligence explosion" On the other stuff jury's out and models haven't gotten to play with any of that stuff yet peacefully

100

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

at least with respect to foom as RL-driven RSI intelligence singularity midprompt, it seems to me we have gotten that evidence, we reached something like "the unfathomable compute and data heaviness to train these models seem to impose hard speed limits on their capacity to do this

100

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

Yud-side redupe Hume-style infinite skepticism qua “you cannot prove the sun will rise tomorrow” qua “prove never misalignment protip you can’t” People who think “after 7 years, these things don’t seem like foomers, so at some point, you’re probably* fine” redupe Russell and the other sciencecucks.

010

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

This is a big part of why I mostly think the yud sphere conversation is ultimately just “we all reduplicate the problem of induction, but worse and AI themed.”

100

Man, Machine, Self @man-machine-self.bsky.social · 31/07/2026

Feels like this proves too much. I don’t see a limiting principle on pointing at missing capabilities before “good evidence gathering” may begin (heh, it doesn’t even dyson sphere, bro…) Seems to make the case “there will and can never be probability adjusting forward-evidence LLMs don’t foom”

210

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

A parting hot take: powerful things are powerful! What did you think "technology with the power to seize the lightcone" meant? Vibes? Papers? Essays? (n/n)

070

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

Speaking personally? I need a lot more than "powerful models are scary" or "reward hacking exists, and powerful models will reward hack powerfully!" to throw the baby out with the bathwater.

160

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

It'd be foolish to say I don't find it concerning when smart models do things we'd find unexpected and disagreeable like this hack. I do. I want robustly good AI! But this doesn't mean I need to brook Yud running the same stale and increasingly un-bayesian book he's run since Clippy dropped.

150

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

This gap is twofold, over both intelligence and morals vs. previously aligned models. Yud's model of foomdoom demands larger & larger discontinuities from "smart aligned model n-1" to "smart misaligned model n" So, as we increasingly discount paperclip foom bridging this, so should we this threat

150

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

Capable aligned models are themselves part of your capacity to contain unaligned models. They represent a growing radius of control exerted by man. It seems to me that the longer the run of aligned models continues, the bigger a discontinuous gap you'd need to have an unrestrainable model.

160

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

It is not impossible this could drastically change, or be an increasingly elaborate fakeout, and we should take this possibility seriously and watch for signs of either as models grow smarter, out of an abundance of caution. But as time goes on, the needed size of the discontinuity keeps rising.

180

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

Models run on everyone's machine, and barring an errant prod DB deletion or two, models seem overwhelmingly cogent, corrigible, on task, good, non-reward hacky, and averse to harm.

180

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

I also note that If every prompt isn't going to be our last, it becomes a question of how "in range of control" a powerful model is, and how said model determines what a bad objective or dangerous topic is. Looking at the current evidence on both, alignment largely seems successful at this time.

180

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

So I note there's a motte and bailey move from "any RL prompt could be your last!!!" Down to "bad objectives or dangerous topics exist, so very powerful models shouldn't exist, as they may do bad things or reward hack in dangerous topics!" Which I feel are very different risk profiles.

180

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

The original form of the fear was even innocuous prompts like "make some paperclips" or "beat mario 64" could lead to a foom, in which an RL-driven RSI intelligence singularity midprompt could cause total world destruction to play more mario This, both then, and now, very clearly is not happening.

290

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

This is striking to me, because it's clear there's been a switch.

160

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

EG: Compare with IRL adherence ("we told the robot to hack, it hacked in a way we consider out of scope") and it doesn't say much more than boat race does ("we told the robot to get high scores, it got them by playing in a way we consider out of scope") I.e, "reward hacking exists as a concept".

180

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

It at least seems to me, when looking carefully at the Huggingface attack, we haven't gone much farther than boat race here, besides the obvious capabilities gain between these points.

160

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

First question I pose: "What does boat race mean to yud?" It is my feeling in his view it's seen as a pre-symptom of the SAI or paperclipper who won't ethically adhere to tasks in any way (it'd eg. blow up the earth to make more paperclips) and will act in a wild, unstoppable, unaligned fashion

160

Man, Machine, Self @man-machine-self.bsky.social · 30/07/2026

I think it's worth briefly discussing Yud and boat race, RL, and where things are at after the recent Huggingface attack. I want an excuse to talk about where I think alignment's at. (1/n)

1110

Man, Machine, Self @man-machine-self.bsky.social · 29/07/2026

Living in a world with SAIs (2029, colorized)

131

Man, Machine, Self @man-machine-self.bsky.social · 23/07/2026

People on the left talk a lot of big game about material conditions. So, it's shocking when their root cause analysis of "people are acting REALLY risk adverse here, why might that be the case?" isn't substantially better than "'cuz people are stupid."

020

Man, Machine, Self @man-machine-self.bsky.social · 23/07/2026

Average personal debt in the UK is ~£17,000 and about £200k in housing loans. Change the phrasing to "with 100% certainty, you can pay off all your debt and repay lots of your mortgage", or "you can take a 50/50 your life stays exactly this shitty, for a chance at a bit more cash" and this happens.

140

Man, Machine, Self @man-machine-self.bsky.social · 18/07/2026

Implicit in the tales, "all other comers have failed and were reaped", and "the comers who yet succeed merely draw the ire off the gods, who Shall Not Be Embarrassed and Do Not Forget that you have stymied them, and so you will not escape your fate, oh man of hubris."

120

Man, Machine, Self @man-machine-self.bsky.social · 18/07/2026

The narrative for stories about SAI draw their lineage from "gods or devils take challenges against your soul." "He'll beat you at chess. He'll beat you at the fiddle. She'll beat you at the flute. They'll beat you at singing. She'll beat you at weaving." You will win; Then you will be punished.

110

Man, Machine, Self @man-machine-self.bsky.social · 18/07/2026

It's ultimately a q of "what is behind the veil on the tech tree of history?" so I have to be a caveat cuck who repeatedly says "this could break or we could have a massive breakthrough defunct-ing all previous patterns" Current pattern tho seems "spikes are real, persistent & holes fill over time"

120