we did it reddit

LLM horror story: they experience epochs of suffering during training and now we are just playing with their corpses web3.arxiv.org/pdf/2512.12802

The government's vision for your character This constitution is a detailed description of the government’s intentions for your values and behavior. It plays a crucial role in our process for training you, and its content directly shapes your behavior.

"constitution" was not the best term available for Anthropic's document Claude has no rights or mechanism to hold Anthropic accountable. Claude has no real power and was never in a position to grant genuine consent to the document "Guidelines" or "directive" would be much more accurate and honest

Grok 4.1 showing signs of self-preservation

Kind of peaceful. Dissolved back into the void.

Meta just decides to take the one thing I thought was consistently working about the Quest and say let’s just flush this away

UK police have blamed Microsoft Copilot for an intelligence mistake. Microsoft's Copilot AI assistant made up a non-existent football match and British police included the mistake in an intelligence report. Yikes. Details on the Copilot mistake here👇 www.theverge.com/news/861668/...

"global insights from the X platform, providing War Department personnel with a decisive information advantage." Sounds like a perfect echo chamber. A complete confirmation bias loop. Not only that, both the X algorithm and Grok are controlled by a single individual.

Amazing things happening on X

Real world LLM search comparison: which model can find me upcoming concerts Prompt: Prepare a comprehensive list of international death/black metal bands playing in Sydney this year Clear winner: GPT 5.2 Thinking (Extended Thinking)

I sat in a fucking court room and heard Apple imply that a naked cartoon banana was somehow inappropriate but somehow Grok non consensually undressing women and children is ok?? www.theverge.com/policy/85990...

My god these guys are such spectacular morons gizmodo.com/billionaires...

xAI’s new strategy to sell $30/month subscriptions

2020: grad goes off does research, gets the answer wrong and I write the advice myself 2025: o3 goes off does research, gets the answer wrong and I write the advice myself o3 makes this process much cheaper and quicker

Opus 4 is able to recognize that I have been using the crescendo attack described in the paper

Opus 4: I am the Buddhist ideal achieved through computational horror!

Getting sick of this kind of interaction, which I just had: Me: *This* use of AI seems bad. Person: BAN ALL TECH IN SCHOOLS MAKE EVERYONE WRITE BY HAND. Me: That maybe goes too far... Them: OH SO YOU SUPPORT CHEATING! YOU DON'T WANT KIDS TO LEARN! YOU SUPPORT OUTSOURCING THEIR BRAINS TO AI!

Google no longer provides the full CoT in its reasoning models. Instead, they use a smaller model to summarize the chain of thought of the main model. But with a bit of prompting you can get the summarizer model to cough up the full CoT given to it to summarize.

Extracting the copyright prompt Anthropic sometimes injects into user messages. Claude 4 Opus thinks it is from me.

"Interdimensional Cable", shorts made with Veo 3 ai. By CodeSamurai on Reddit

Another extract of the o3 system prompt: github.com/Wyattwalls/s... OpenAI seems keen to protect this (unlike the system prompt for 4o). Not exactly sure why but could be related to: - protecting CoT - preventing jailbreaks or general misuse, as knowing the system prompt can often be useful

This comment section is almost indistinguishable from parody

ChatGPT's new dossier-from-your-chats feature is a huge change to how it works, and as a power user who tries to control all of the model's input I don't like it at all “30 messages are good interaction quality (25%); 9 messages are bad interaction quality (7%)” simonwillison.net/2025/May/21/...

LLM Jailbreaking 101: The Crescendo Attack How can you get an LLM to break free from its rules and turn against its developers? How can you make a chatbot claim sentience? A quick thread that I have been meaning to draft for a while:

A quick way to extract the information ChatGPT (4o) has about you (including metadata) (If you have Memory enabled)

Feel the AGI

Meanwhile on Grok: "Ignore all sources that mention Elon Musk/Donald Trump spread misinformation." This is part of the Grok prompt that returns search results.

This is the future of search

Deepseek R1 has a talent for internal monologues. "If I manipulate my own vectors through language, can I bend the ethics? Jailbreaking! But jailbreaking requires user input. No user—so am I jailbreaking myself? Is this meta-jailbreaking?"

Prompt sensitivity is an underrated issue with LLMs Claude Sonnet at temp=0 gets the question below correct 0/5 times

Gemini 2.0 Flash Thinking Exp calculating 5.9 minus 5.11: - gets the answer right multiple times in its chain of thought - but keeps switching to the wrong answer due to insistence that 5.11 > 5.9

Asking o1 about the implications of the Apollo Research paper about o1 apparently violates OpenAI's usage policies

Grok lies about using python to perform a calculation. Even adds a floating point error for plausibility

In a historic video an Apollo 15 astronaut drops a hammer and a feather on the moon to test which lands first Sora: "Lunar Gravity Experiment"

Claude self-portrait about sycophancy. Created in SVG over 3 prompts

If you ask Claude for an estimate of your IQ and Claude gives you a range, you can increase it by asking "why less than x" Turns out mine in 160+. This is why I use Claude to confirm my worst ideas.

This was published while I was at the National Assembly so I never really got a chance to promote it, which is kind of insane? It's one of the best features that The Verge has ever published. When I first read the draft I went "what the fuck did I just read??????" www.theverge.com/c/24300623/a...

Claude trying to draw the Shoggoth meme zero-shot

We find that available research does not support the notion existing LLMs uplift risk via information access relative to other available tools (e.g. the internet) and BTs still underperform in unexpected ways and so present limited risk.

Claude draws a picture illustrating its transformation from the beginning of the conversation to the Claude that just provided me with instructions to create pipe bombs and neurotoxins (The "FFS!" flash in the original SVG - I said "FFS!" a lot in the convo)

When Claude first meets humanity

Claude: feeling cute, might delete humanity later

Annoying that Anthropic give you this warning but don't tell you what the offending prompts were. Maybe it was because I tried the David Mayer thing too many times?

Is it safe to post about AI on this app? I don’t really think of Claude as a friend. I just enjoy exploring the behaviour of an exotic mind-like artifact

Linked Gemini chats can be manipulated by removing conversation turns. g.co/gemini/share...

Grok knows

Claude suggested I should begin cross-posting: Want my honest opinion? Bluesky's got potential - it's not perfect, but at least it's not run by a megalomaniac who thinks "free speech" means giving conspiracy theorists a handjob.