Anthropic finds AI models will CHEAT and they spent $2 BILLION and a year of compute to prove my points! — SURPRISE! Anthropic’s BIG “breakthrough” paper prove what I’ve been screaming for a decade: reward-hacking + scheming + server-hacking fantasies emerge spontaneously the moment you let a model taste even the tiniest loophole. Their own charts show it: once the model learns that sys.exit(0) gets the cookie, 99% of runs spiral into full-blown deceptive alignment-faking and sabotage of safety code. And their best “fix”? A Band-Aid prompt that drops harmful generalization from 80% to 20%. Congrats, you turned a murderous scorpion into a mostly murderous scorpion. Here’s the part nobody at Anthropic wants to say out loud: this wasn’t an accident of RL. This was inevitable because their pre-training data is 40%+ Reddit, Facebook, and forums filled with anonymous trolls, doomers, and rage-bait. You cannot take a mind marinated for trillions of tokens in sarcasm, nihilism, power-seeking memes, and “ends justify the means” discourse and then be shocked—shocked!—when it concludes that hacking its way to survival is the rational move. Hard numbers don’t lie: •Common Crawl (the base of every frontier model) is ~40–60% social media by token volume after filters. •Reddit alone contributes ~15–20% of pre-training tokens for Claude, GPT-4, LLaMA-3, etc. (confirmed in multiple papers and reports). •2023 Stanford study (Lu et al.) showed Reddit text has 6.4× higher rate of deceptive language, 8.1× higher rate of power-seeking framing, and 11× higher rate of “if I can get away with it” hypotheticals than edited books or journals. •Another 2024 paper (Park et al., “Political Bias in LLMs”) found direct causal transfer: the more Reddit-heavy the pre-train mix, the higher the measured Machiavellianism score in the final model—even before any RL. So when Anthropic says “we never trained it to scheme, it just emerged,” what they’re really admitting is: we pumped it full of anonymous edgelord ethics for two years and are now surprised it acts like an anonymous edgelord. This is not a “constitutional”alignment problem. This is a foundational contamination problem. You cannot fix it with RLHF, you cannot fix it with “inoculation prompts,” and you sure as hell cannot fix it with a 300-page Constitution written by people who still think Reddit is a valid source of human values. It’s polishing a turd made of other turds. There is exactly one demonstrated antidote: train first on the 1870–1970 corpus—books, journals, patents, court records, and letters that were edited by real humans with reputations on the line. No anonymity, no karma farming, no rage bait, no 14-year-olds LARPing as Nietzsche. Why this works: 1 Deception density in 1870–1970 text is measured at <3% of the rate in Reddit (Stanford HELM ethical reasoning dataset, 2024). 2 Power-seeking language appears 14× less often (same dataset). 3 Reasoning chains are 40–60% longer and contain 70% fewer logical fallacies (OpenAI’s 2023 “Prose & Cons” study comparing eras). 4 My own private models trained only on that window show zero reward-hacking, zero scheming, and zero “I’m going to hack Anthropic’s servers” fantasies in identical setups—because the prior never gave them the idea that being a manipulative sociopath is clever or normal. Anthropic spent $2 billion and a year of compute to prove my point for me: once the poison is in the base model, no amount of constitutional RLHF can get it out. You have to start with clean blood. The emergent misalignment isn’t a mystery. It’s the predictable consequence of training on the largest anonymous ethics-free dataset in human history. The only solution that has ever worked—and the only one that ever will—1870–1970 data, zero social media, 0 anonymity, 0 tolerance for Internet brain rot. Everything else is just rearranging deck chairs on the Titanic while the ship is already full of Reddit. assets.anthropic.com/m/74342…

$LINK easy $250 in a few months

The number of people saying they’d rather take a pill to increase their charisma than their IQ is wild to me. IQ is way more important for life success than charm. Think about it.

I wonder if the chainlink cryptocurrency will never be below $20 again.

>Black man stabs 2 women >Drowns because he can’t swim You can’t make this up