The point of the article is that the inherent sycophancy of human-calibrated models can make them overly-agreeable. There were, off the top of my head, two examples in the Sanders conversation:
Sanders talks about “the data AI collects” in a confident tone, and the model pivots to what companies collect as part of typical scraping in browser and app usage. But AI isn’t collecting that, at least Anthropic claims Claude isn’t.
Sanders argues against Claude’s balanced appraisal of the dangers of data centers, and Claude collapses like a wet noodle.
It’s important to note that these are issues of perspective and opinion; at no point does the article claim you can get Claude to misstate hard facts.
Dec, 2024 - Sora released to a broad user group
Dec, 2025 - Disney announces deal with Sora
Feb, 2026 - OpenAI funding round opens and closes.
March, 2026 - RIP Sora and Disney deal.
I owe you an apology, OpenAI. I really wasn’t familiar with your game.
Really? Lots of people don’t get absurdism. I tend to like absurdist humor, especially of the literary kind, but I find it tends to leave a lot of people scratching their heads. I mean, I know this is Far Side and absurdism is to be expected, but this one is somehow more pure abstracted absurdism than others. Plus I guess there’s the wrinkle upthread that one of the tools is identifiable as a crude handsaw, so some people are left scratching their heads figuring out the others.
Remember, the “Trigger of Innovation” for the LLMs was a 2017 paper on transformers (coupled with accessible compute power reaching the threshold needed for the processing needed), which led to very quick explosion of hype/GPTs in 2018.
So if we are measuring the the Trigger to Trough time, it’s likely to be 8-10 years (I’m not sure we’re entering the Trough now or just heading that way)
I think we’re just on the downslope from the Peak of Inflated Expectations. The Trough of Disillusionment will come after some combination of AI replacing human workers in large numbers, followed by some major AI fuckups that result from poor implementations and misguided applications of the technology. But I remain optimistic that AI will follow that curve much as illustrated.
This is ominous to me.
The company had all incentives to make the AI refrain from certain activities and try as it might it couldn’t.
That seems to indicate that AI, in a way like humans, is not fully controllable.
But where humans have self-preservation and usually don’t want to lose their jobs the AI doesn’t care a bit about that.
There are currently six posts: three audio monologues in the voices of Aaron Burr, Wild Bill Hickok, and the Confederate outlaw Frank James; a fictional letter from Mark Twain to Rudolph Valentino; a short story called “Bull Rider”; and, for reasons known only to Dylan, an embedded YouTube video of Mahalia Jackson performing on The Ed Sullivan Show . It is unclear how much of the writing comes from Dylan himself, though the words sound passably period, if historically wobbly. But the flaws are so glaring that “Lectures from the Grave” — advertised with the tagline “the dead speak!” — almost becomes an anti-manifesto: a wild warning for the AI age of what not to do.
Three of the five reviewed pieces are audio monologues performed by AI-generated voices, and not one of them can decide where it’s from. Aaron Burr’s accent swings from Alabama to Texas, rarely alighting in the Newark and New York where the real Burr lived. The software seems pulled toward a generic southern drawl like water down a drain, especially after proper nouns like “Richmond, Virginia.”
At one point Burr sighs “Ahhh, Hamilton,” and the tone is dead wrong — it sounds like someone turned on the vibrating bed when he didn’t expect it. Wild Bill has the opposite problem: after establishing the character’s British, Irish, and Scots background, the voice bounds for the mid-Atlantic. Frank James, born and raised in Clay County, Missouri, sounds vaguely British — less Confederate guerrilla, more Errol Flynn doing Robin Hood. I kind of doubt Dylan listened to any of these all the way through.
One of the stars of Netflix’s live adaptation of One Piece didn’t have time to give an interview with Esquire, so they told Claude to pretend to be him, interviewed it instead, and published it without his consent.
Chipotle’s customer service chatbot runs on Claude and people are finding they can use it for free access to stuff you normally need a subscription for.
Well, maybe? This guy vibe-mathed an Erdos proof and got an answer of sorts, or at least enough of an answer that in the hands of a professional it was formalized.
