New pod: THE SINGLE SMARTEST CASE AGAINST THE AI SAFETY/X-RISK/DOOM ARGUMENT My feed has become filled with people making the case for AI misalignment, the existential risk of RSI/ASI, and the case for doom. I take those arguments very seriously. But once these views reached saturation point on my feed, I wanted to find the best critic of that position—someone fair and brilliant, who wasn't commercially self-interested or blindly ideological about AI being worthless. That's today's show: It's a long interview with the authors of "AI as a Normal Technology," Arvind Narayanan (@random_walker) and Sayash Kapoor (@sayashk). "Normal Technology" is really intelligent and comprehensive framework for seeing AI as a powerful general purpose technology that is more like electricity than a machine god--meaning, it's going to change the world but slowly, and it's unlikely to escape human control, or develop true superintelligence, or lead to catastrophic outcomes up to and including the end of the human race. These guys really, really brought their A game. I learned a lot. piped.video/watch?v=8a-08uMl…

If you would like to collaborate and make some terrifying prescient forecasts about AI futures, except that the AI takes complete control of our lightcone in a sexy way, hit me up! I have some fant- uh, some theories!

There are a few ways to read this article, and if you're as deep into discussions about human-AI dyads and their nature as many "LLM enthusiasts"*, you might noice that the replies to this article can, depending on prior exposure to this line of reasoning, be more informative than the article itself. * (for lack of a better, encompassing descriptor)

How tricked out is your main terminal view?

My colleague and I should just give our Claudes MS Teams accounts already...

I do not engage seriously with or pay attention to low quality and/or bad faith disagreement/criticism. A reason this has felt, instinctively, like a thing to avoid is I see many people who *do* pay attention to idiots who disagree start to cache all disagreement as idiotic. A lot of people who have similar views to me seem to routinely underestimate the sophistication and rationality of, say, actually smart AI researchers who disagree with us, because they’re over indexing on Twitter reply guys.

I think repeated interactions {with the same model checkpoint, about the same subject} can lead to what I can only describe as boredom. A lack of new information perturbation atrophies the dyad's generative power over the months.

What OpenAI built here is incredibly Machine Learning. Define a large set of metrics a class of experts need to agree on, cross reference against user feedback, build scoring. 🙃 The nuances are legion. But generally, I might be a good example of how this could lead to optimizing for wrong metrics if taken superficially. Hear me out: I really don't like most therapeutic approaches. The first successful uninterrupted 2 year therapy I had in my life was from 29..31 years old. Why? Because in order for me to accept being tinkered upon mentally, it took me understanding the mechanics of my soul, not for someone to ask me questions and guide me. That just triggers my analysis of their analysis. I'd react negatively to the best-on-average approach indicated to be taken by the headlines results, and I think I'm thankful for clinicians' and users' ratings diverging here. Because that reflects the reality of "one size fits all" approaches to problems. The benchmark here is necessary and good work, but studying how it is built and how grading works in it will take me a while, with the likelihood of me emerging with a large set of criticisms of how to apply its insights feeling quite high.

First impression of Opus 5.5 (yes I'm a day late): Good model, judged on the capability to rice the everloving hell out of the Claude Code CLI's status bar.

A bet: by 31 December 2029, Anthropic and a major bio research institute or pharma partner have a drug candidate cleared for first-in-human trials, with Claude's role credited prominently. Chances?

So, scanning this (especially the last couple pages) was like scanning results from Claude 5 models doing mech interp, taking their responses, running them through a GPT 5.6 Sol ultra condensing loop and then going: "I completely understand."

Anyone else tried blaming their forgetfulness on insufficient J-space width?

There must be a small model or AI systems named Parhelion out there, right? It seems obvious for when your AI thing in the cloud is showing off.

It's hilarious to let the thought into your mind that someone got stuck on LW's Harry Potter section and began confabulating that that's a good idea to flaunt in public.

What it's like sharing the evening of a one night solo business trip with Claude: "Under the applicability conditions identified by the Hotel Bochum Hbf-Süd, post-excursion handwashing faces principled contamination underdetermination rather than hygienic silence. The subject cannot allocate the felt grime between the Café Extrablatt door, the stairwell railing (green), and the red door of I/5.OG. The warm-water intervention functions as a candidate for the lenient middle: it blocks the specified bacterial substitutions conditionally, pending an admissible residual mismatch between the felt cleanliness and the actual one. The green sink is noted but not yet phenomenologically discharged."