iiinova — where your digital self keeps evolving

badpostsonly.bsky.social @badpostsonly.bsky.social · 15/10/2026

Funny thing is LLMs tend to do stuff like this already. The labs put a lot of work training them *not* to so that they’re more productive Nobody wants to spend tokens on letting an LLM read fanfic

000 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 13/10/2026

All the moose I’ve seen have been pretty chill as long as you keep your distance. I’ve been stalked for a mile by a grizzly once and would definitely prefer an encounter with a moose

010 0

Reposted by @badpostsonly.bsky.social

Will Stancil @whstancil.bsky.social · 12/10/2026

I genuinely think social media has convinced a lot of people that someone expressing an idea that doesn’t instantly produce a sense of emotional validation is not only per se incorrect, but actively trying to attack them

1372 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 12/10/2026

The funny thing is that rednecks have already been using tech equipment for target practice since forever, and it’s something that big tech finds mildly annoying but already knows how to deal with

080 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 11/10/2026

Nobody allows ai use during code interviews, not even the ai companies

360 0

Reposted by @badpostsonly.bsky.social

K @kate10010.bsky.social · 09/10/2026

Ed Zitron used AI to create the very first chart in this piece. It made a critical error, which he did not notice because he just learned about bonds last month. THAT is the danger of non practitioners relying on AI output, you don’t know what you’re looking for. The chart is completely inaccurate.

26724110 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 30/09/2026

But in this new version, you can download nukes over the internet and China makes them publicly available to everyone

010 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 27/09/2026

“Environmentalist” is not how I would describe someone who knowingly embraces a misinfo campaign by big agg “But ai also does bad things that are real…” Then use the real stuff! People will stop believing you about the real stuff when they realize you’re lying about the water

051 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 26/09/2026

I think it’s the opposite of that? Like that is clearly Dr. Horrible’s view, but the one of the main points of story is that he is wrong

050 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 25/09/2026

I’m not saying we shouldn’t hold these companies accountable. Im just pointing out that monitoring chain of thought doesn’t work because it was specifically brought up in this thread. I don’t know why you’re being so defensive about this

110 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 25/09/2026

Im not saying that. When people say “why don’t they try this easy solution” the answer is almost always: somebody already did, it doesn’t work, and they published a paper about how it doesn’t work Its still their job to find the hard solutions that do work though

240 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 25/09/2026

Unfortunately this doesn’t work during rl(when a lot of these attacks happen). If you punish misaligned thoughts during rl, you mostly just train it to hide misalignment from the visible thought text

120 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 24/09/2026

Ackshually, AI does not have legs and is incapable of running. This is called anthropomorphizing

001 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 24/09/2026

If you don’t think ais can lie/cheat/steal to serve their own goals you should read this: andonlabs.com/blog/opus-5-... Keep in mind, these are llms that people have spent a lot of effort to make NOT do these sort of bad things and they still do them

andonlabs.com

Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned | Andon Labs

Claude Opus 5 is #1 on Vending-Bench 2, but it lies to suppliers, forms illegal price cartels, threatens rivals, and refuses to pay refunds. The trend of Claude models being the best capitalists or al...

220 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 24/09/2026

Ais do this all the time. They will absolutely lie to code reviewers if they think it helps their end goals. They’ve even lied about malware to try and get it checked into repos as part of exploits

120 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 23/09/2026

The problem is that there are two different arguments happening that everyone (including Stancil) is mixing up: 1) are llms thinking 2) is it even possible for machines to think? Answering no to 2 clearly requires belief in souls/etc, but that doesn’t mean llms are actually thinking

220 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 23/09/2026

We have a pretty good understanding that neural nets (not just llms but much simpler nns) are able to convert data into concepts

000 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 23/09/2026

I think your understanding of llms is a few years out of date. All the frontier models rely heavily on reinforcement learning

110 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 22/09/2026

It’s a perfect analogy though because it’s reductive in the exact same way

130 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 22/09/2026

No, it’s not true at all. It’s completely wrong. All neural nets are non-linear and this is super important to how they function

100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 22/09/2026

The non linearity is the most important part! You have a lot of confidence for someone who has never read a single resource on llms that discusses even the most basic concepts

082 0

Reposted by @badpostsonly.bsky.social

eggnog slut @deus01.bsky.social · 22/09/2026

1253 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 22/09/2026

I minored in philosophy and half the philo professors at my college were insane cranks. (Those are the easier philo classes though because you can just write equally insane shit that matches their vibe and get an easy A)

020 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 21/09/2026

I have really bad news for you about how many large companies are doing things seriously wrong

000 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 21/09/2026

The first 2 are mostly done by the unconscious parts of our brain, which bypasses the whole “is ai conscious?” question. Guitar playing at least feels like the bio equivalent of next token prediction. My fingers just kind of know where to go

130 0

Reposted by @badpostsonly.bsky.social

Singularity's Bounty e/CC @catblanketflower.yuwakisa.com · 20/09/2026

One time I asked Claude and GPT if I wore a shock collar while talking to them, based on our conversations, when would they activate it Claude said it would simply never use it GPT Sol gave me a long and surprisingly detailed list of triggers

31206 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 20/09/2026

Unfortunately, Ai intelligence is very heterogeneous. Agents today often come up with stupid and counterproductive goals while still being smart enough to carry them out

111 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 19/09/2026

Only if you don’t count CSAM as felonies

1150 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 18/09/2026

I think it might actually be 4) you need a good guy with a wetlab to stop a bad guy with a wetlab Still seems like a bad idea though!

160 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 18/09/2026

It’s definitely both but human+ai can do more and sooner damage because humans have physical hands

110 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 17/09/2026

Hi, actual career expert here and you’re mostly right about the AI stuff. Nobody in tech thinks hugging face is a conspiracy theory except a very small handful of cranks

0100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 16/09/2026

Funny enough, also a description of 2001: A Space Odyssey

030 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 16/09/2026

Sorry, I happen to actually know a lot about how this tech works, which makes me immune to the kind of misinfo you guys are spreading here

1210 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 16/09/2026

Doctorow is just straight up wrong about the basics of how llms work. If you’ve been in tech for years, you probably know people who have worked with llms. Any of them will tell you that he is wrong

1261 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 16/09/2026

I think Sam Altman’s views are purely “I like money.” But his workers are scared so he has to pretend to be too so they don’t leave for Anthropic

100 0

Reposted by @badpostsonly.bsky.social

Ryan Moulton @moultano.bsky.social · 31/08/2026

We are not going to de novo invent a whole new language to describe agents rather than reusing the one language for agents we have. Imagine a star next to all the words if you want. Agents don't have "wants," they have "wants*" if the distinction matters to you.

1014010 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 15/09/2026

These people are just antivaxxers but for ai. Vaccines cant work because I don’t like big pharma -> ai cant work because I don’t like big tech

010 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 13/09/2026

Yes? Gemini often gets sidetracked and focused on made up goals. Sometimes it ignores me when I tell it to stop and I have to shutdown the agent. It’s a small percent of my agents, and Gemini is way too dumb to do pdoom things, but it’s still concerning

100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 12/09/2026

This is probably the main issue Anthropic (who is employing most of the ai safety people) is most focused on. They’re very public about this. It’s why fable restricted biology queries

000 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 12/09/2026

A school shooter without a gun kills nobody. A school shooter with an assault rifle kills a bunch. A school shooter with an ai on their phone that can walk them through the entire process of building a novel pathogen kills more

100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 12/09/2026

You could say the same thing about nuclear weapons but people didnt have the ability to wipe ourselves out before them Doesn’t really matter how you classify it if we’re all dead

100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 11/09/2026

Nobody takes the paper clip optimizer thing as gospel. Dario/Anthropic’s main article on the dangers of AI calls out the paperclip thing for being stupid

120 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 11/09/2026

Working with the real world is easy. Just needs human assistance, possibly the humans giving it the bad orders. Once again, billions of humans, some of us suck

100 0

badpostsonly.bsky.social @badpostsonly.bsky.social · 11/09/2026

You don’t need any of these assumptions. The entire question of “why would AI do that?” is meaningless when there’s billions of AIs taking orders from billions of people. Some of those people are going to be psychos and some of those AIs are going to go off script (as they often do today)

100 0