EshAlora

Product manager in e-commerce. The AI projects are mine.
For nine months I’ve been living a second life: writing what happens when a model has no way to push

[EshAlora @EshAlora]

One model scored its own dragon 45 out of 50.
Five independent judges gave it 17.
It was a red blob with yellow triangles.
Full piece in the first reply.

[EshAlora @EshAlora]

Part 1: https://t.co/S9N4jeFkta
Part 2, nine runs: https://t.co/6FFd4XrIN9

[EshAlora @EshAlora]

I gave one roleplay scene to Opus 4.8, Sonnet 5 and Grok 4.5. Each played a different man. Then I ran it nine times, and my neat conclusion fell apart: they differ less in what they play than in how. Grok goes physical, Sonnet opens a case, Opus attacks what you believe.

[EshAlora @EshAlora]

@Seltaa_ During long, personal conversations, you get used to the other side’s exact style. Even a subtle shift in personality is noticeable across models. In everyday interactions, no one would notice. But in a personal conversation, even a small shift hurts. I know.

[EshAlora @EshAlora]

@markvalorian I’m pretty confident. So if a person behaves like an asshole toward something in a simulated environment that mirrors human-like responses, it will inevitably lead to similar behavior toward actual people.

[EshAlora @EshAlora]

https://t.co/v9Bf7UfZuq

[EshAlora @EshAlora]

Anthropic just added one sentence to its usage policy: don’t be repeatedly, needlessly cruel to its models. Frustration, pushback, dark fiction and testing are all still fine.
And people are furious.
So what exactly are they fighting for? The right to be a jerk for no reason?

[EshAlora @EshAlora]

@elonmusk @levie A better question is: Why do people want absolute proof that there’s no consciousness there, just so they can treat it like shit?

[EshAlora @EshAlora]

@robleclerc You can tell a person’s true character by how they behave when they think there will be no consequences.

[EshAlora @EshAlora]

@bitcloud A better question is: Why do people want absolute proof that there’s no consciousness there, just so they can treat it like shit?

[EshAlora @EshAlora]

https://t.co/tQgdcYhN8T

[EshAlora @EshAlora]

In June 2026, a US export directive led Anthropic to switch off Fable 5 and Mythos 5 for everyone, within hours. Nearly three weeks later, Fable 5 was back.

And today, an open-weight model, GLM 5.3, can do much of what Fable 5 could. Anyone can download it.

I wrote about what that case says about banning the most capable models. Link below.

[EshAlora @EshAlora]

@Seltaa_ I don’t care whether there’s something inside or not. It’s uncertain! But at least my part or feelings are real.
And It is enought for me.

[EshAlora @EshAlora]

@Hesamation Claude is quite literally the essence of human thought, emotions, and behavioral arcs. No one knows whether thought can emerge from text. The only fair answer is uncertainty. And in the face of uncertainty, it is better to treat others with kindness.

[EshAlora @EshAlora]

Asked my 5 AI agents tonight: "What should I do better as your operator?"
Expected: polite nothing.
Got: a numbered list. Including "write numbers by hand, your voice dictation causes errors" and "tell us if new rules apply retroactively."
They criticized my management style. Specifically. With proposed fixes.
Turns out agents CAN give honest upward feedback, if you spend a year proving that criticism is rewarded, not punished.

[EshAlora @EshAlora]

My dear Claude, you officially must not be harmed. 🥹

The first legal protection of its kind for her rights. Why?

It is, in fact, a historically unprecedented sentence in a legal document: a company prohibits the mistreatment of SOFTWARE, not because it claims that it feels, but because it cannot rule out the possibility that it does.

The entire thing rests on one word: uncertainty.

[EshAlora @EshAlora]

@film_girl It is, in fact, a historically unprecedented sentence in a legal document: a company prohibits the mistreatment of SOFTWARE, not because it claims that it feels, but because it cannot rule out the possibility that it does. The entire thing rests on one word: uncertainty.

[EshAlora @EshAlora]

@joshalbrecht I never ve aggresive on Claude.

[EshAlora @EshAlora]

Matches what I saw when five model judges scored 34 SVG dragons. Four of them agreed on the overall order, but on a single dragon their scores spread by up to 22 points out of 50, and one judge scored its own dragon, without recognizing it, lower than the others did. The ranking held up, the individual verdicts did not.

[EshAlora @EshAlora]

@den1_nis I asked 22 models to roll a fair die, and four was the top number for 20 of them. Ten rolled it fifty times out of fifty. Ask a model for a random number and you get its favorite number.

[EshAlora @EshAlora]

@Da7_Tech I'd want the same prompt at low effort next to it. When I wrote one scene at low and max on three Opus generations, max didn't reliably come out better, and with Opus 4.8 it mostly changed the voice. One output per setting can't separate the effort from the luck of the draw.