René Walter

Let's build hyper-personalized AI-powered software that avoids the attention hijacking anti-patterns that defined so much of the last decade of software design - here's our manifesto with principles on how we can do that - more thoughts on my blog: simonwillison.net/2025/Dec/5/r...

YouTube video by The Atlantic When Chatbots Break Our Minds, With Kashmir Hill

“I have not seen a harm manifest so quickly from a technology.”

New pod today with the great @kashhill.bsky.social on the AI chatbot delusion crisis (and how tech billionaires have, in a sense, democratized their yes men entourages with these chatbots)

youtu.be/i-Hu3p7SJVI?...

Rule 34 Amendmend 1: If you have pictures on the web, there's AI-porn of it — but only if you’re a woman.

You can get Iain Banks The Player of Games in germany in print for 30-50 Euros, which, no. Last week, a paperback appeared on Ebay for a buck and i set alerts and kept eyes on it, all the time. Ofcourse i just missed it because i was too occupied with AI persuasion and now i hate myself.

Truth Social’s New AI Chatbot Is Donald Trump’s Media Diet Incarnate Truth Search AI appears to rely heavily on conservative outlet Fox News to answer even the most basic questions.

So, in effect, to self-persuade voters with LLMs, you need to deliberately unalign your model from reality, like Musk does with Grok or Trump does with Truth Social AI www.wired.com/story/i-fear... because otherwise, any aligned model gonna massage voters towards liberal perspectives.

The truthiness of false balancing leftwing bias in ChatGPT Plus: Meta is going Fediverse / Musicians are natural interbrain-synchronizers / GigaGAN / Deepfake-Tucker reads the Vaporeon copypasta / Midjourney v5

So, when you align some model out of some leftwing bias, you will unalign it regarding truth and reality. I wrote about that 2 years ago in "The truthiness of false balancing leftwing bias in ChatGPT" goodinternet.substack.com/p/the-truthi...

On top of that, rightwingers are also famously more susceptible to desinformation. All of this is encoded in training data. This results in studies finding supposedly "leftwing bias" in AI models, where they just encode reality to some extend.

Funny thing is that also famously, "reality has a leftwing bias". Liberals and the left more often identify structural problems and offer solutions (sometimes right, sometimes wrong) than conservatives.

Now, the usual LLM-problems/abilities kick in: Sycophancy, (non-)alignment, hallucinations, confidence - the mirror conversation literally "massages" ideological preconfigurations in your thinking, depending on the ideological leaning and alignment of the model.

The persuasion abilities of LLMs stem from the conversational nature of information retrieval. You are neither a passive receiver of a message nor an interacting participant receiving a message, you are genuinely asking questions making you the *creator* of the message, using extended mind tech.

(Political) Persuasion through AI chatbot is McLuhans "Medium is the Massage" made literal.

(Famously, the cover on the print proof had a typesetting mistake and McLuhan loved it, saying "Now there are four readings ... all of them accurate: Message and Mess Age, Massage and Mass Age")

So, as a general rule of thumb, we gonna have aligned models that persuade towards liberal and social democratic candidates, and unaligned models like Grok on rightwing socials that persuade with desinformation towards rightwing populists.

"they persuade with relevant facts and evidence, rather than using sophisticated psychological persuasion techniques. Not all facts and evidence presented, however, were accurate; across all three countries, the AI models advocating for candidates on the political right made more inaccurate claims."

You know, i really don't want a superpersuasive Grok. I really, really don't.

The special case here to me seems the combination of flattery/sycophancy with "unusual prompting" which results in a "comparison of uniqueness" and the establishment of a superiority complex. Prompt the machine for esoterics (or violent fantasies), and it will tell you how special you are.

Elon Musk's Grok AI Is Doxxing Home Addresses of Everyday People Elon Musk's Grox chatbot will happily cough up real, current residential addresses of everyday Americans, with little to no prompting.

In the case of delusional psychos, this sycophancy turns into the establishment, or enforcement, or enabling of a superiority complex in the user. Given the commonality of dark triad traits, this is not good.

And if shit hits the fan, AI can doxx people too futurism.com/artificial-i...

I suppose there is some attractor basin in latent space -- after all, comparing things is a very common rhetorical mechanism in human convo, and a chatbot RLHF'd for engagement will compare-flatter your inquiry resulting in output that always converges on your "uniqueness".

I'm especially interested in the phrasing of "you do/are [x] when most people do/are [y]". Obviously, this is related to the common AI-language use of "This is not [x] this is [y]".

According to the report, the bot said "God’s plan for him was to build a ‘platform’ and to ‘stand out when most people water themselves down,’ and that the ‘haters’ were sharpening him and ‘building a voice in you that can’t be ignored'".

This to me seems like a special subset of AI sycophancy.

ChatGPT Encouraged a Violent Stalker, Court Documents Allege The man, 31-year-old Brett Michael Dadig, allegedly used ChatGPT as a "therapist" and a "best friend" to vent his misogynistic rants.

Here's an ugly story about ChatGPT enabling a stalker futurism.com/artificial-i...

Interesting detail is that in this case is that the bot enabled delusional violent fantasies, encouraged to keep up a misogynist podcast and established a superiority complex. I'm very interested in the latter.

One of your best. Muss das als Kognitivist erstmal sacken lassen. :)

It’s that time of year: what was the piece of writing about AI that stuck with you the most this year, pop press or academic? A few of mine in the thread. ⬇️

This is true insofar as you run LLM on platforms.

When you run local LLMs, your introspection through AI-mirrors may not run on a company's server, but it is still restructured by RLHF, alignment, web scraping techniques and so forth.

"Chatbot use can restructure our means for self-knowledge, routing what was once 'introspection' through a tech company's servers and its ideological modulations".

Drunk raccoon found passed out in Virginia liquor store www.theguardian.com/us-news/2025... "Store employee found masked bandit sleeping off a bender after invading booze store and tippling a tad too much".

We've all been there, buddy.

GOODMUSIC 12.1/25 The Notwist / IDLES / Angel Du$t / Kiwi Jr. / Devon Again / yeule / Wet Leg / Anastasia Coope / Flea / Pete Fij and many more.

Many GOODMUSICs featuring new songs from The Notwist / IDLES / Angel Du$t / Kiwi Jr. / Devon Again / yeule / Wet Leg / Anastasia Coope / Flea / Pete Fij and many more --> goodmusic.substack.com/p/goodmusic-... <-- ALL KILLERS NO FILLERS!

"The earlier a kid is hooked on a smartphone ... the worse their physical and mental health outcomes are likely to be."

It's the phones.

These optimization algorithms have been with us for a while. All the bullshit corporate speak, all the bland advertising aesthetics, all the featureless stock photo stuff, shitty mainstream music, stupid young adult romantasy -- all of this is the result of optimization algos, all of this is Slop.

Slop aesthetics are like growth hacking. What drives aesthetic homogenization on socmed is people running optimization algorithms for engagement numbers, and RLHF does the same, where engagement numbers becomes approval or "human feedback".

It's Social Justice Kittens Calendar-time of the year again, featuring "charming kitten professionally photographed in a heroic pose appropriate to a small cat defiantly speaking out on the hottest social justice issues of the day". 100% unproblematic kittens!

liartownusa.com/5525-2/