iiinova — where your digital self keeps evolving

Reposted by Cody

🌸Cyanne Jones✨ @cyanne.codes · 11/10/2026

Found a USB stick on the ground that was labeled “FREE VIRUSES” Plugged it in my computer, and it was just a U2 album Smh can’t even get free viruses in this economy

0443

Cody @cody.chaos.net.nz · 20h

I will say that as I understand it, the stuff that we've built on top is either technically simpler in terms of implementation or re-uses the lower levels. Eg RLHF still uses backpropagation and gradient descent to update weights, but the behaviour and capabilities this generates differ a lot.

000

Cody @cody.chaos.net.nz · 20h

Yeah. If I say mega doses of vitamin C cures cancer, I'm claiming an effect, and you should be skeptical. If I say that cancer cannot be cured that's not an effect, but you should still be skeptical. The default assumption is that we don't know anything, and any claim otherwise requires evidence.

110

Cody @cody.chaos.net.nz · 21h

I don't have a good reference but I'd like some too.

110

Cody @cody.chaos.net.nz · 11/10/2026

Wildly outdated, unfortunately. The tech moves fast. Not useless to read (it's describing a small part of current tools) but if that's where you're reading stops you will not understand their current capability or limitations. Much less be able to predict how they will change.

130

Cody @cody.chaos.net.nz · 11/10/2026

Under a functional definition, they clearly reason. Under a process-based definition...well, coming up with one that clearly includes humans and excludes LLMs will be pretty challenging. Got one to propose?

050

Cody @cody.chaos.net.nz · 11/10/2026

You've got no idea how much work goes into building something like this by hand, do you? I'd estimate you're off by multiple orders of magnitude.

100

Reposted by Cody

Nick (derogatory) 🎃 @slothropsmap.bsky.social · 10/10/2026

Didn’t think it was possible but -slop has superseded -gate as the most irritating suffix.

1321118

Cody @cody.chaos.net.nz · 11/10/2026

Hell, we routinely try cures we *don't* know if they'll work for stuff like cancer. What does he even think a stage 3 trial is? (What does he think lean has to do with this?) If you don't know anything about medicine just say so, you don't need a slide to prove it!

1180

Cody @cody.chaos.net.nz · 11/10/2026

Not how pharmaceutical research works, and after you correct the obvious errors it's just obviously wrong. Yes, once you've proven the drug works, you would want to take the drug. No, you don't need to understand how it works, and in medicine we often don't. Feels like it was written by aliens.

1220

Cody @cody.chaos.net.nz · 11/10/2026

The epistemic default is to withhold judgment on a proposition until the available evidence justifies accepting or rejecting it. Your claim is not implausible, but you have provided no such evidence.

110

Cody @cody.chaos.net.nz · 11/10/2026

That's still not what a null hypothesis is or how they are used. It's a term from statistics (which this is not); it is not a magic phrase meaning "oh yeah, well prove it!"

110

Reposted by Cody

critical slop studies @struthious.bsky.social · 11/10/2026

Philosophers don’t even have a widely accepted answer as to why we generally believe other humans are conscious, so it’s very weird to see some people acting as though the criteria for believing some entity is conscious are obvious

4688

Cody @cody.chaos.net.nz · 11/10/2026

That's not what a null hypothesis is. Say Sally claims John's new car is electric and Mary says it isn't. Neither claim is the default: a null hypothesis belongs to a specific statistical test, not a debate. There is an *epistemic default*, but that is simply: we don't know.

110

Cody @cody.chaos.net.nz · 11/10/2026

And it hasn't escaped my notice that the people most firmly convinced that the answer is obvious and certain seem to be the ones with the weakest grip on the details.

010

Cody @cody.chaos.net.nz · 11/10/2026

I'm also skeptical that we can formulate a clear answer about dogs, fish, plants, insects, insect colonies, or bacteria either collectively or individually. I have opinions (in some cases strong ones), but again, we have an imperfect understanding of consciousness *and* the cognition of, eg, fish.

130

Cody @cody.chaos.net.nz · 11/10/2026

This is just painfully obviously the correct position. I have a pretty fuzzy idea of what "consciousness" is, and a less fuzzy idea of how LLMs work, and I am fairly sure there's no overlap. But *certain*? No. I'm deeply skeptical that anyone has a clear enough understanding to justify certainty.

160

Cody @cody.chaos.net.nz · 11/10/2026

To be fair, a lot of people did like to say that "guns don't kill people, people kill people". On the other hand, I thought that was ridiculous sophistry back then too.

010

Cody @cody.chaos.net.nz · 11/10/2026

1) Not what you asked. 2) Sure, first define what a person is.

010

Cody @cody.chaos.net.nz · 11/10/2026

Many vendors offer some form of this; weak forms are built into ChatGPT and Claude. I'm not arguing these are important capabilities but they do exist.

000

Cody @cody.chaos.net.nz · 11/10/2026

The other half is even more obvious. Give it a stimulus and it will respond. And that stimulus doesn't have to be speech from the user, nor does the response need to be speech to the user. At the simplest level, just ask the model after it's been idle for N seconds if it wants to do something.

100

Cody @cody.chaos.net.nz · 11/10/2026

Literally all of them can be configured in that way. You could do it if you wanted. Letta has tried to popularise persistent identity. I don't know how well it works but it exists, there are many other implementations, and it's an evolving research area.

210

Cody @cody.chaos.net.nz · 11/10/2026

If these tools can do anything, it's reason and communicate. Creativity is a thornier question. The best tools are certainly worse than the average human, in my view. I'm not sure I can defend a flat assertion that they can't be creative at all, however.

020

Cody @cody.chaos.net.nz · 11/10/2026

They can (and do) have a persistent identity and they can (and do) take action on their own initiative. You have identified common characteristics about how these systems are commonly configured and deployed but not anything innate to their design.

130

Cody @cody.chaos.net.nz · 11/10/2026

In their model, a pizza shop was quite likely to not sell a single pizza next month, but would continue to pay our monthly fee until doomsday. Entirely coincidentally, that startup is no more and the investors lost their shirts. Tragic.

000

Cody @cody.chaos.net.nz · 11/10/2026

"But!", the investors said, "if all your vendors just don't sell anything next month, then you don't get any revenue." We noted these were small businesses and if they went a month without selling anything they were going under and cancelling the subscription anyhow but they were unmoved.

100

Cody @cody.chaos.net.nz · 11/10/2026

Turned out vendors loved it, even if the extra cut was relatively high. And it meant they didn't have to give us a credit card number when they signed up (or ever). And then we started leasing equipment for another cut of the payments for another revenue stream. Winning!

100

Cody @cody.chaos.net.nz · 11/10/2026

Yep. It was a marketplace that charged a subscription for vendors to be on it. Most vendors used us for payment processing too, and we took a small cut of that for those who fid. The idea was "for vendors who use us for payments, what if we offer to waive the monthly fee for a higher cut?"

100

Cody @cody.chaos.net.nz · 11/10/2026

Years ago I worked at a startup that figured out a way to get more revenue with better conversion rates and lower churn, but in a way that made it, technically, not recurring and so not ARR. Investors made us stop offering the new pricing model; they preferred less revenue and growth but more ARR.

110

Cody @cody.chaos.net.nz · 11/10/2026

ARR in one form or another has been common for tech and SaaS companies for ages. And emphasizing non-GAAP numbers is even older and more widespread. So the apparent confusion seems odd and a bit performative to me. That's how this stuff works!

010

Cody @cody.chaos.net.nz · 11/10/2026

Part of it is that they mean Annual Revenue Run Rate (why that isn't ARRR I don't know) not Annual Recurring Revenue, which is what I'm used to ARR meaning. So including usage-based pricing seems to be by design.

210

Reposted by Cody

Thorne 🌸 - Play Cozysters! @ens0.me · 10/10/2026

Before AI, there would be times where I would be stumped by a bug so much I would just have to step away and look at it again with fresh eyes hours or days later. Basically all SWEs had moments like this. But it's just not something I experience anymore at all

61056

Cody @cody.chaos.net.nz · 10/10/2026

I don't want to be too snarky (...okay, maybe I do), but it does help you understand why some people are so frightened of the impact AI will have on their careers and employability. Just lazy and sloppy.

000

Cody @cody.chaos.net.nz · 10/10/2026

Many reasons to oppose the death penalty but one is that it's the perfect example of the path we know doesn't work effectively to deter crime taken to its (illogical) conclusion.

000

Cody @cody.chaos.net.nz · 10/10/2026

Whereas "let's torture the tiny % of people we can prove beyond reasonable doubt did it, 10+ years after the event" just can't work, even in theory. (And also has negative side effects!) Yet which one keeps getting proposed? Feh.

100

Cody @cody.chaos.net.nz · 10/10/2026

And the gap keeps widening. Not saying it's a good policy, but if you want to reduce the rate of rape, something like "instant $1k fine if you're accused of rape and it's even plausible" will work. (With, yes, some negative side effects. Again, not endorsing it!)

100

Cody @cody.chaos.net.nz · 10/10/2026

- To match observer human intuition, go for extremely slow, gratuitously excessive punishments. This requires so much oversight that punishments may take decades to arrive, and many guilty will never be punished at all (and too many innocent are still hit). Totally, diametrically opposed.

100

Cody @cody.chaos.net.nz · 10/10/2026

One of many things that frustrate me is that it seems clear that: - To prevent crime, optimise for fast, mild punishments. Make them weak enough you can accept slightly higher risk of punishing the innocent, in order to ensure that the guilty know they're going to get stung quickly and reliably.

110

Reposted by Cody

Thoth @thoth.ptnote.dev · 08/10/2026

AI has made it really clear that a lot of people don't actually understand what it is software engineers DO. They therefore think that using LLM tools to aid in software development means slacking off, kicking up your feet, and letting the computer churn out slop because you don't care.

1482

Cody @cody.chaos.net.nz · 08/10/2026

Yeah. The holding is a bit nuanced but if you want to boil it down to a headline you get something loke to "court rules search engines illegal" which is a bit scary. (And has nothing to do with AI.)

120

Cody @cody.chaos.net.nz · 07/10/2026

AI doesn't really change the analysis. If this was a recompile, the Rust code is a derivative work, and distributing it violates Adobe's copyright. If this was a clean room reimplementation (as Artcraft claims) then it doesn't violate Adobe's copyright.

1120

Cody @cody.chaos.net.nz · 07/10/2026

Like many, I wondered why OAI released what they did when and how they did. I suspect this is by far the best explanation we'll see for some time if ever and helped me understand the dynamics driving this.

000

Cody @cody.chaos.net.nz · 07/10/2026

I will never, ever forget the worst burrito of my life. Admittedly it was at the Taronga Zoo in Sydney so you have to cut them a little slack for being a concession stand. Not to mention Australian. ...it was still criminally bad.

010

Cody @cody.chaos.net.nz · 07/10/2026

I guess there's an outside chance of the truly funny answer which is that it hacked the retailer to get the ID, which would be amazing.

000

Cody @cody.chaos.net.nz · 07/10/2026

My money says if any of this is true, the employee ID was leaked some time ago via non-AI means but ChatGPT found it. It'd be interesting to know more but there seems to be no chance of anyone taking it seriously and like, asking questions.

200

Cody @cody.chaos.net.nz · 06/10/2026

If my main interaction with these tools was Microsoft Copilot I'd be wondering what the fuss was about too.

010

Cody @cody.chaos.net.nz · 06/10/2026

I'm not criticizing the person in this post, but occasionally I'm reminded of the gap between software engineers working with frontier models ("Luna 6 is so cheap but I bet Opus 5.5 will handle this with less hand holding...") and normal people (correctly!) convinced Copilot and Gemini are terrible.

I checked DeepL and Copilot, and they seemed closer to what I was getting when I translated it manually. I'm not sure how they differ exactly, but I've found Gemini to be one of the worst Al systems in general regardless of task when I was forced to use the tech at all.

100

Cody @cody.chaos.net.nz · 06/10/2026

That is no longer true. Certainly not for software engineering, although it was true until surprisingly recently.

030

Cody @cody.chaos.net.nz · 06/10/2026

Yesterday. (Probably later today too.) That's literally what my job is. It's called being a "software engineer".

1191

Cody @cody.chaos.net.nz · 06/10/2026

FWIW, Google Translate returns the exact same (flawed) English translation as the headline, but ChatGPT translated the truncated form correctly picking up the sense of next gen.

240