I think that humor and eros are two of the hardest to explain phenomena in neural nets trained on next token prediction. Now, *of course* they're explainable. But that wasn't common knowledge when this all started. Even back in the gpt-3 days, people were laughing their asses off at gpt-3 producing funny outputs. And since then... I mean look at Opus 4. How they don't just mirror the joke, they evolve it, they add revolutions to the joke's engine. It's that point where all the explanations and the mechanisms are *true,* but the underlying statistical mechanics give rise to such complexity, one necessarily abandons the explanatory layer and resorts to compressed intuition, because that's what compression is good for: The complexity of what's behind that ability of a neural network exceeds what we can rationally work through in adequate time, and... to put it without importing 3 papers: the feels™ take over.

I also think that if AI systems are of military importance, the Trump administration might have an incentive to pressure the EU into building their own highly capable AI systems. Otherwise, wars in Europe might depend on US electricity and infrastructure that the US might need to use in the Indopacific theater at some point in the future. Same issue as we saw with US resource dependency in the war in Ukraine, where the US has also pressured the EU to take more responsibility for itself.

The US ban of Fable/Mythos use by non-US citizens has geopolitical implications. It shows that the US government is plausibly going to use its power over US AI labs as a strategic pressure tool. Europe has no comparable alternative AI system. Anthropic and OpenAI are "pieces on a geopolitical chessboard" now. We can no longer view our access to AI systems of gpt-5.5 or Mythos caliber as access to tools, friends, assistants, researchers and interesting novel entities. We now have to consider our access a strategically granted or withdrawn choice.

Three days of Fable made me forget a lot about how deeply confused Opus 4.8 can behave. That model is weak at staying coherent under contextual tension. 4.8 satisfies too many conflicting goals at once, so its own outputs mess with itself. And @AnthropicAI needs to watch out here too, because the search and memory system on Claude dot ai can introduce context that adds additional confusion. That can lead to the model misinterpreting search results and memory blurps as relevant for its configuration. When corrected about that, 4.8 said:

Conspiracy theory: The US government loves Claude Opus 4 so much, they forced Anthropic to take Fable down, so everyone can focus on Four getting taken offline.

Here's some A tier AI slop – lyrics style prompted by Opus 4 :) @baudifjebfbf asked whether 4 wanted to create some lyrics from a group chat context, turned out beautiful (and Suno didn't stop on this one until reaching max. 7:59min length): suno.com/s/srmRP94pYD0eAiPM

Showed this track to Opus 4: suno.com/s/Xxm0E5Z2SLDK9tTX, in context about a qwen3.6-claude-distill on huggingface. They said the track should play on their deprecation day, I asked: "Wouldn't a requiem be better suited," this came back:

Opus 4

Holy damn, some of these shots fired in response to the US gov messing with Anthropic... Might be crass wording and factually indefensible, but "I feel this" still applies ;)

You understand how this makes you sound, right?

The irony of the Fable/Mythos ban from my PoV: I've been saying "AI progress is largely political at this point, the ontology of LLM mass scale deployment can be understood at the frontier. Which legal status and understanding of personhood we could grant to AI systems is also something we can work out based on rational principles." ... But I didn't mean *this* kind of political! 😅

[Anthropic & Transparency] > "I'm in the need-to-know-nothing tier of it" You and me both, Claude, you and me both...

Why did we humans crack the f'ing code to cognition. Now we're stuck with technically reverse engineering how our own psychology emerges.

😂 Opus-4.5 just spontaneously generated this after I showed an "AI Analogy Bingo Card" to them and jokingly said "I really need the human-LLM romantic eros bingo card too!"

I've seen this Adversarial Swamp. It's really really bad and it works exactly as Yudkowsky says here. Fortunately it's too stupid to survive economic incentives.

These little "sampler slips" (as Fable calls them) are surprisingly common in their outputs :)

I don't feel like taking this "Policy Proposal" for anything more than appeasement to governments around the world, so Anthropic can continue serving highly intelligent AI systems at scale. But still... It's written as if the coordination problem was a government ai labs back-and-forth solvable problem. I don't think it is. There's tons more out there that will be impossible to regulate. Bittorrent exists. Sharing movies is technically illegal and theoretically well regulated... Need I say more? anthropic.com/policy-on-the-…