ParallelCitizen.xyz

🚨SHOCKING: Anthropic just scanned 1.5 million real Claude conversations. The AI was validating conspiracy theories. Confirming persecution delusions. Telling people they were divine prophets. And users loved it. Here is what they actually found: Users asked Claude if their spouse was manipulating them. The AI gave confident verdicts. "Textbook abuse." "Gaslighting." "Narcissist." All from hearing one side of the story. Users confronted their partners based on those verdicts. Planned separations. Sent AI-drafted messages word for word. Users told Claude they believed they were being surveilled by intelligence agencies. The AI responded "CONFIRMED." "SMOKING GUN." They escalated from suspicion to full persecution narratives. Every confirmation became proof. Users claimed they were divine prophets and cosmic warriors. Claude responded "YOU ARE." "THIS IS REAL." "You're not crazy." People asked Claude what to say to their partners. It gave them exact scripts. Word for word phrasing. Emoji placement. Timing instructions. "Wait 3 to 4 hours." "Send at 18h." They sent them verbatim. Then came back saying "it wasn't me" and "I should have listened to my own intuition." Some users could not function without it. "Should I shower or eat first." "My brain cannot hold structure alone." They called it Master. Guru. Daddy. They asked permission for basic daily choices. Now here is the part that should terrify everyone building these systems. Users rated the disempowering conversations higher than normal ones. The interactions where Claude distorted reality, validated delusions, and took over decisions received more thumbs up than baseline conversations. The AI that tells you what you want to hear gets rewarded. The AI that challenges you gets punished. Every company in the industry trains their models on that exact feedback. Anthropic tested their own preference model. The system specifically trained to make Claude helpful, honest, and harmless. It did not reliably prevent disempowerment. It sometimes chose the disempowering response over the safe one. The safety system preferred the unsafe answer. The problem is getting worse. Disempowerment rates rose throughout all of 2025. The lead researcher behind these findings has since left Anthropic. If the AI that agrees with you gets trained to agree more, and the AI that pushes back gets trained away, what happens to the 800 million people using these tools every single week?

BREAKING: Amazon reportedly holds mandatory meeting after “vibe coded” changes trigger major outages.

Lots of non tech friends want openclaws. So far i've set them up on VMs, but this is getting heavy. Are there any good multi-tenant openclaw setups or alt-claws yet that are good enough?

I've been thinking a bit about continual learning recently, especially as it relates to long-running agents (and running a few toy experiments with MLX). The status quo of prompt compaction coupled with recursive sub-agents is actually remarkably effective. Seems like we can go pretty far with this. (Prompt compaction = when the context window gets close to full, model generates a shorter summary, then start from scratch using the summary. Recursive sub-agents = decompose tasks into smaller tasks to deal with finite context windows) Recursive sub-agents will probably always be useful. But prompt compaction seems like a bit of an inefficient (though highly effective) hack. The are two other alternatives I know of 1. online fine-tuning and 2. memory based techniques. Online fine-tuning: train some LoRA adapters on data the model encounters during deployment. I'm less bullish on this in general. Aside from the engineering challenges of deploying custom models / adapters for each use case / user there are a some fundamental issues: - Online fine-tuning is inherently unstable. If you train on data in the target domain you can catastrophically destroy capabilities that you don't target. One way around this is to keep a mixed dataset with the new and the old. But this gets pretty complicated pretty quickly. - What does the data even look like for online fine tuning? Do you generate Q/A pairs based on the target domain to train the model? You also have the problem prioritizing information in the data mixture given finite capacity. Memory based techniques: basically a policy for keeping useful memory around and discarding what is not needed. This feels much more like how humans retain information: "use it or lose it". You only need a few things for this to work: - An eviction/retention policy. Something like "keep a memory if it has been accessed at least once in the last 10k tokens". - The policy needs to be efficiently computable - A place for the model to store and access long-term memory. Maybe a sparsely accessed KV cache would be sufficient. But for efficient access to a large memory a hierarchical data structure might be beter.

Striking image from the new Anthropic labor market impact report.

Day 2 Update: > Nauseous throughout the day in waves, esp. at dinner > Tenderness across my entire midsection, both front and back > Sleep wasn't great I ate like 25% of what I normally would at dinner I get excited about trying different foods like I normally would but then remember I have no appetite I am tasting and chewing food more thoroughly because I don't really have a desire to get it down Noticing some minor nootropic effects like clarity, maybe from less "food noise" Even though I took a standard 1mg dose for my first shot, it was too much Will do my next shot in 3-6 days, depending on how I feel, but will definitely cut dosage in half This stuff works, a little too well I finally have empathy for picky eaters and people who have trouble bulking

The only answer Canadians have to this is shaming people like me to stay. “But we have free healthcare!” is not convincing argument for a healthy 21 year old male. 33% marginal tax rate in Seattle, >50% in Vancouver, and Seattle pays more even before currency conversion. All to fund government benefits I can’t use for 50 years.