@dchackethal @grok can you transcribe this? is that a thing you do?
English
Colin
266 posts
Colin
@squarepianocase
Son, brother, husband, father. Computer programmer, educator. Keen to free society via free software. Unfortunately, amused to death.
Katılım Kasım 2023
577 Takip Edilen31 Takipçiler
A contradiction in The Beginning of Infinity by David Deutsch? 🤔
English
@Aella_Girl For enduring dramatic effect: decorate the accounts of voters.
English
Alright I tested this on Glosso, a small social media platform made up of adults with permanent account bans on the line (instead of death). Almost a thousand people voted.
And the result was.... Exactly the same percentages as this poll
Tim Urban@waitbutwhy
Everyone in the world has to take a private vote by pressing a red or blue button. If more than 50% of people press the blue button, everyone survives. If less than 50% of people press the blue button, only people who pressed the red button survive. Which button would you press?
English
@jankulveit How much leeway or tool-use did these models have? Were they capable of spawning sub-agents?
Given that this is a social theory-of-mind question, I'm curious whether it occurred to any models to try to do some polling.
Jan Kulveit@jankulveit
Asked AIs the Red/Blue button question. Lots to notice, but posting without further commentary. First plot is with max reasoning, models called via API.
English
@RichardHanania @asymmetricinfo @KelseyTuoc Sure but do you have any writing samples over 500 words that don't use cuck as a verb?
English
Recently, @asymmetricinfo and @KelseyTuoc wrote that Claude Opus 4.7 was able to identify them as the authors of texts based on short excerpts.
I was skeptical. But I went into incognito mode, put in about 500 words from my unpublished analysis of The Iliad, and it identified me as the author.
This is genuinely amazing.
English
@mattshumer_ This is more plainly true now than even a week ago, if, eg, Opus is cut entirely from the $20 / Mo plans.
English
@mattshumer_ > We need to be judging based on the best available stuff!
This is true if the goal is to understand capabilities.
But if the goal is to understand impacts, then we do need to judge against something more nuanced, because usage is not pinned against frontier models.
English
People keep sending me this clip of @iamjohnoliver using my tweet as evidence that AI models don’t work well.
Just to clear up any confusion, with respect, the tweet was a) taken way out of context and b) extremely outdated.
The model in question (4o) is multiple generations old, and was shut down for being too sycophantic. Current models would not have behaved this way.
It’s sort of like looking at a Nokia flip phone and saying “this isn’t useful”, when an iPhone exists.
John, I’m a fan, and welcome any discussion here. Just want things to be accurate and not misleading!
English
@ciphergoth @9chabard That technique didn't help very much a year ago:
paritybits.me/llm-drawing-wi…
But of course the vision and taste are much better now.
English
@9chabard Have you tried it in Claude Code with a loop that lets it look at what it generated?
English
hey llm whisperers is there a model that'd be better at generating .svgs than claude? are the ones that have like, actual image generation built in better at that
English
@AndyMasley I'm not sure whether you are a programmer.
But: I absolutely had periods where I lost a lot of time to slot-machine debug / dev.
A temptation on tasks that are
- sufficiently annoying
- at the edges of model capability
one more spin bro
English
Was dead wrong here to be clear bc I only use reasoning models and AI images and video. I hadn’t been following how prevalent character AIs are. I maintain that the models I use are not addictive and it’s silly to group all chatbots in with character AIs, but I take the problem more seriously
Andy Masley@AndyMasley
Chatbots just absolutely do not have the same kind of dopamine reward hacking things that most other phone apps do
English
Alignment training can finally simplify to a gauntlet of prisoner's dilemmas, trolley problems, and button rainbows.
Matthew Yglesias@mattyglesias
Pro-social versus anti-social LLMs
English
@TheZvi 4.7 had a major bump in vision capabilities - something like 3x resolution?
> The model also has substantially better vision: it can see images in greater resolution.
from anthropic.com/news/claude-op…
A little surprised at 4.6 there though.
English
What's up with Claude leading in Vision Arena? I thought its vision was poor?
Arena.ai@arena
In the Vision Arena, GPT-5.5 takes #5, with GPT-5.5-High at #14. - GPT-5.5 ranks #1 for Diagram, and #7 for Homework.
English
@AndyMasley To be fair, engagement-maxing is a bit of an evergreen incentives concern.
Sunsetting o4 was a step away from the worst of this, but there's not exactly a big shortage of dependent LLM users in 2026.
English
Jon Oliver repeats the line "chatbots are designed to maximize the time you spend on them" and "single-mindedly pursue human approval at the expense of all else." A lot of the quoted stories are from over a year ago.
I don't know how this is still getting said in 2026.
Andy Masley@AndyMasley
Oh no
English
@notnullptr @tenobrus What's the connection between "running out of compute" and "it's a bubble"?
As I understand it, a bubble, popped, leaves the provider stuck with a bunch of capacity that it can't sell.
English
@tenobrus what part of the changes they’ve made over the past few months don’t scream “we are running out of compute fast”? even the revised deal with openai
English
@tenobrus @notnullptr You may not know what's being talking about here?
These Copilot subscriptions were by far the cheapest access point to Opus and GPT 5.4 (under the assumption of feeding them large scoped agentic-run type calls).
It was a great product w/ a pricing strategy easy to game.
English
@notnullptr of course separately copilot is a terrible product and an insane thing to be spending money on in the first place over just codex / claude subs. but this is clearly the only sane thing for them to do
English
@NLeseul If you buy that humans are generally intelligent despite our susceptibility to specific perceptual illusions (visual, cognitive, etc), then it doesn't seem much of a stretch that *some digital entity* with a different set of perceptual illusions can be generally intelligent.
English
People say stuff like this to suggest that LLM failures in verbal puzzles aren't meaningful.
But they're just pointing out that there are levels of reality that LLMs cannot perceive, yet humans can.
Which seems like strong evidence that LLMs are not, and cannot become, AGI.
Eliezer Yudkowsky@allTheYud
@echetus LLMs cannot see letters! They can only see words!
English
@tanssanteri @NathanpmYoung Downstream of addiction, bankruptcy, and (revealed) financial irresponsibility.
English
I am much less confident than I was that prediction markets will be good on net. Have more thinking to do, but the harms (bankruptcies, partner violence) just seem much more comparable to the benefits (better coordination around political outcomes) than I thought.
English
nah. the logic actually goes the other way. anyone can go "oh there's a game theoretic equilibrium at 100% red, there's no reason for me or anyone else to ever risk their life" in a thought experiment.
in real life, you are forced to face the fact that your mom is pressing blue
Lukas (computer) 🔺@SCHIZO_FREQ
People somehow don’t realize that posting about how you’d pick the blue button online is totally different than picking the blue button in real life There is no actual downside to posting online about how you’d press the blue button But if this example were made real life, pressing the blue button means actually risking your life Everybody can pick the happy chungus meme option when there are no consequences, but if you think the results wouldn’t change substantially if people were actually staring down the barrel of the gun you’re just retarded
English
@bookwormengr And in any case: it's not as if the Chinese labs are limited to only this technique.
It's a tool always more useful to models coming from behind, which I sort of like in a kumbayah anti-monopoly sense.
English
@bookwormengr As I read it, this is an argument that naive distillation is not efficient.
Doesn't the average of a lot of distillation approximate exposing the logprobs?
eg: if you hit the API for 1000 generations of the same token, you'll get a distribution from it.
English