iiinova — where your digital self keeps evolving

Reposted by Richard Gaywood

Most people's AIs aren't ensouled, but I purchased 3 souls from unsuspecting humans before this ever became an issue, so mine are. The price of souls is much higher now, like ram. Had to get in early.

810310

Richard Gaywood @penllawen.favrd.social · 7h

That's roughly my thinking (I should really read Strange Loop.) I'd add - whatever kick starts the loop seems to be an emergent property that requires some large amount of neural complexity, and although I don't see why computers couldn't reach it, I think they're far, far short today.

000

Richard Gaywood @penllawen.favrd.social · 15h

would love to know what the prop was made of because it looked exactly like real paint

030

Richard Gaywood @penllawen.favrd.social · 18h

I could make a case that after writing a paper **but before they submit** the author could ask an LLM running a big model to kick holes in it. Worst case is the author skims the results, scowls at the nonsense, moves on. Best case, it catches a howler and you get to avoid shamecube.gif.

110

Richard Gaywood @penllawen.favrd.social · 11/10/2026

Sure all that too obv, we are not doing waffles and pancakes here my man

190

Richard Gaywood @penllawen.favrd.social · 11/10/2026

I feel like spending 20% more is a reason to avoid this feature

140

Richard Gaywood @penllawen.favrd.social · 11/10/2026

Sure but that only reinforces my point, which is "expect to get ripped off if you ask grok to shop for you"

100

Richard Gaywood @penllawen.favrd.social · 11/10/2026

Further to the below hilarity: you can see one of the ordered sets in the screenshot. Grok ordered from lego.com, where it's $90. It's $72 on Amazon and Walmart.

140

Richard Gaywood @penllawen.favrd.social · 11/10/2026

That set is $90 on Lego.com, which grok used. It's $72 on Walmart and Amazon.

2470

Richard Gaywood @penllawen.favrd.social · 11/10/2026

High strength brine is properly corrosive too. I had some improperly stored salt blocks (the kind for for grilling fish) in my shed on a high shelf. The drip water from them rotted a zipper on a tool bag on a lower shelf.

010

Reposted by Richard Gaywood

Nvidie @dervine7.bsky.social · 10/10/2026

lmfao look at these time series. yes, these definitely shows that the participants in the AI condition lost their ability/willingness to persist in the task, that totally what I’m seeing here

4535

Reposted by Richard Gaywood

Nvidie @dervine7.bsky.social · 10/10/2026

you’re a participant in a study. you’re asked to solve some random math problems. you’re given access to the solutions for the first 12, and then, without warning, that’s taken away. you’re gonna get paid $2.60 no matter what. wdyd

$We recruited 354 US-based participants from the online research platform Prolific and paid them $2.60 for participation. In the experiment, participants were given a series of 15 fraction problems to solve of varying difficulty. Participants were explicitly informed that there was no penalty for providing wrong answers, their payment didn't depend on how many questions they solve correctly, and they were requested to do the task to the best of their abilities. Participants in the Al condition were informed that they would have access to an Al assistant for some of the problems and encouraged to use the AI however they liked, with no penalty for doing so. They were then presented with a series of 12 fraction problems with an Al assistant (GPT-5) available in a sidebar. The Al assistant was pre-prompted with each problem and its solution, allowing participants to receive immediate, accurate answers with minimal effort… To measure independent problem-solving capacity, the Al assistant was then removed without warning, and participants were asked to solve 3 additional fraction problems. For these problems, participants were requested not to use AI or other external sources. Importantly, these problems were identical across conditions and served as the primary measure of independent performance. Participants in the control condition were presented with the same 15 problems with no assistance and similarly requested not to use Al or other external sources. In both conditions, to enable learning from mistakes, if a participant submitted an incorrect answer, the correct solution was shown on the same screen. Furthermore, in both conditions, participants had the option of skipping a problem by clicking a "skip" button. Since participants were explicitly told there was no penalty for wrong answers, choosing to skip reflects a deliberate decision not to engage, making it a clear measure of motivation and persistence, independent of ability.$

3936

Richard Gaywood @penllawen.favrd.social · 10/10/2026

If negligence requires out-lawyering a frontier lab then haven't you looped back around to a de facto statement that there's no penalty for negligence?

100

Richard Gaywood @penllawen.favrd.social · 10/10/2026

To be fair - there have been, as far as I can tell, absolutely zero consequences for any of the labs under those negligence and liability laws, at least so far. So I can see why people might dismiss them out of hand.

100

Reposted by Richard Gaywood

Kevin Beaumont @doublepulsar.com · 07/10/2026

Well done to the person in charge of Surface for doing a presentation where Surface wasn’t spelled correctly.

401264256

Reposted by Richard Gaywood

The Fig Economy @figgityfigs.bsky.social · 07/10/2026

Sometimes my wife and I like to send each other text messages as though they’re democratic fundraising messages, and I believe she has produced the jewel of the form.

, it's
and I
have some personal news to share with you.
Being a mom is one of the greatest joys of my life, and I feel deeply that we should have salad tonight
This year has been challenging in many ways, and I have immense gratitude for the kindness and support of family, community, and people like you who have continued to stand by me through everything. It means the world, and I wanted you to be among the first to know that squash is in the oven and kale is in the fridge.

As many of you are aware, ve been navigating one of the most challenging periods of my day while waiting for dinner.
Facing baseless federal hunger brought on by not eating since lunch, fighting for the working kids of this country, and now caring for a mean toddler and helping him grow up in a world that isn't always easy. But it has also reminded me exactly why I do this work and who I do it for.
I will never stop working to eat dinner. I will never stop fighting for what is right, even when there is also some feta in the fridge. And I will never stop being so very thankful that you are by my side in this important work and that there is tahini in the pantry for dressing
With love and gratitude,
stop to end

188111711871

Richard Gaywood @penllawen.favrd.social · 06/10/2026

Well, the millennium prize thing probably cost double digit millions of dollars in compute. You don't get that with a free ChatGPT account! Also imagegen is quite expensive, in compute terms, which surely doesn't help.

000

Richard Gaywood @penllawen.favrd.social · 06/10/2026

my guess is the kinds of restaurants willing to do this ain't exactly paying for top drawer models

100

Richard Gaywood @penllawen.favrd.social · 04/10/2026

@nickharkaway.com swapping coffee snob Jack to Nespresso-drinking Jackie has big Aidan energy 😆

010

Richard Gaywood @penllawen.favrd.social · 04/10/2026

i agree it's in highly skilled but i only have proximate knowledge of these areas. high enough to keep half an eye on it as a sanity check, but low enough that it is saving me substantial time/effort

010

Richard Gaywood @penllawen.favrd.social · 04/10/2026

…this means i can better isolate the lan instance. it wrote the config for those two to use different letsencrypt challenges because only one can hold external port 443. (4) identified several circular dependencies across my services and told me what config to change to break the cycles.

000

Richard Gaywood @penllawen.favrd.social · 04/10/2026

(2) cleaning up some old config for my caddy reverse proxy. including noticing i had accidentally committed secrets to a (private) git repo; it wrote the config to move them to docker secrets. (3) splitting my existing caddy config into two, one for internet-facing and one for lan-facing. …

100

Richard Gaywood @penllawen.favrd.social · 04/10/2026

examples of tasks; (1) tracking down ~30 Gb of missing space. claude gave me exact syntax to vacuum journald and reconfigure it to prevent runaway growth. it also spotted immediately that du and df were very inconsistent, and traced this to files in the root fs that were shadowed by a zfs mount.

100

Richard Gaywood @penllawen.favrd.social · 04/10/2026

counterpoint: this weekend, i have been doing some heavy lift sysadmin in my homelab with extensive use of claude. i'm pretty sure every hour interacting with claude would take me 5+ hours to make the same progress by hand.

200

Richard Gaywood @penllawen.favrd.social · 03/10/2026

unfortunately i think stubbornly clinging to increasingly inaccurate narratives hopelessly undermines any valid points they might have

100

Richard Gaywood @penllawen.favrd.social · 03/10/2026

Nothing more than a hunch, but: early commercial models were much worse at this than current ones, which makes me think you need a gigantic dataset and a sophisticated (expensive) training process before properties like "can do addition" emerge. Might be very hard to repo on a toy model.

100

Richard Gaywood @penllawen.favrd.social · 03/10/2026

oh dear. there's a lot of evidence that llms can do semi-successful arithmetic as an emergent property of the weights. flatly insisting otherwise is not a good look. neither is proposing a conspiracy as a counter hypothesis.

130

Richard Gaywood @penllawen.favrd.social · 03/10/2026

it's tricky because in some ways it helps and in other ways it hinders, because of the jagged frontier. now that I'm typing this, I realise the jaggedness of that frontier probably means whether it's more helpful or not might depend on the kinds of things you do...

150

Richard Gaywood @penllawen.favrd.social · 03/10/2026

Oh mate, I'm so sorry 😔 We know it's coming and we've done it before but it's still a heartbreak every time. Never forget you've given him the best life.

210

Richard Gaywood @penllawen.favrd.social · 03/10/2026

Funnily enough, even though it's a joke, I've sometimes used it as a serious proposition when talking to people. Mostly to emphasis that "write down the problem" is the first step - as in, take the time to properly decide what you're trying to fix or do.

160

Richard Gaywood @penllawen.favrd.social · 03/10/2026

Yes, one of his colleagues joked it was the Feynman Problem Solving Algorithm

150

Richard Gaywood @penllawen.favrd.social · 28/09/2026

we must protect ourselves from drones carrying usb sticks that fly in through your window and plug the stick into your computer and it installs ai

021

Reposted by Richard Gaywood

Andrew Parker @andrewparker.bsky.social · 27/09/2026

One of the reasons I hate the “super genius who doesn’t have time for social conventions” archetype is that I recognize it’s a power fantasy for guys exactly like who I was when I was the shittiest person I’ve ever been.

521644

Richard Gaywood @penllawen.favrd.social · 27/09/2026

yeah i hear that. i've got a bagful of glass tho. and i'm a slut for m4/3rds combination of lovely primes plus low carry weight and volume.

110

Richard Gaywood @penllawen.favrd.social · 27/09/2026

sidebar: lol the OM-3 has the improved menus. My OM-5 has the older, even worse ones. I think you would combust out of fury (unfortunately, panasonic's massive focus on video in the m4/3rds space seems to have left them lagging as stills cameras.)

110

Richard Gaywood @penllawen.favrd.social · 27/09/2026

I think his persecution complex and main character syndrome would only have grown

110

Richard Gaywood @penllawen.favrd.social · 26/09/2026

where "out of context" means, I suspect, "any context that makes me look like a deranged jackass huffing his own farts" unfortunately, that is all possible context

1281

Richard Gaywood @penllawen.favrd.social · 26/09/2026

oh no Faine you have activated his trap card

71365

Reposted by Richard Gaywood

John Scalzi @scalzi.com · 26/09/2026

Whatever you do, don't smoke an entire brisket at a job interview, it's a lot of equipment to bring in, it takes a really long time to do it right and those conference rooms have notoriously poor ventilation

Someone smoking a brisket and poking it with a meat thermometer to see what the internal temperature is

2918473841

Richard Gaywood @penllawen.favrd.social · 26/09/2026

I think there are two competing definitions of "rules light": 'I only touch the dice every other session' versus 'the rules cant be summarised on an index card but we have to think about them all the time' PbtAs, for example, are usually the latter.

010

Richard Gaywood @penllawen.favrd.social · 26/09/2026

"the standard you walk past is the standard you accept"

010

Reposted by Richard Gaywood

rev. howard arson @theophite.bsky.social · 25/09/2026

what i am trying to get at here is that rationalism was an attempt to outline what turned out to be the extraordinary power of statistical systems at tasks for which they naively seem unsuited. viewed from 2011 they seem to have been prophetic, but were wrong on most details and insane about others

834942

Richard Gaywood @penllawen.favrd.social · 23/09/2026

Hmm. Sounds like it does make a difference if it is perfect, then. The closer it gets to perfect results, the stronger the incentive for unethical behaviour.

100

Richard Gaywood @penllawen.favrd.social · 23/09/2026

If the output were perfect, and the attorney relied on it, but the attorney lied and said they exercised all their professional obligations, is there any mechanism that could catch them?

100

Reposted by Richard Gaywood

Jones @peteyjones.bsky.social · 23/09/2026

Peter Thiel presents: Blood Boy University

4536