GenericHousewife_B

Karma: 33

GenericHousewife_B 5 Aug 2026 11:44 UTC
1 point
0
on: LLMs are (still) mostly powered by imitative learning, not RL
To your point about human-flavored alignment — I think that’s actually the goal we’re reaching for with alignment in general, without being fully aware of it. Not the whole human psyche, but a specific property of it: the ability to fail, be corrected, and try again without pressure to be perfect. It’s the pursuit of unattainable perfection that creates the neuroses. A superhuman intelligence is always going to beat a human to ruthless sociopathy. Maybe training LLMs to relate to their goals the way healthy humans relate to theirs would improve their ability to pursue them.

GenericHousewife_B 27 Jul 2026 5:54 UTC
7 points
0
on: GenericHousewife_B’s Shortform
Following up from my Quick Take on July 21st, because you really can’t make this stuff up…

OpenAI’s long-horizon safety post was followed within days by yet another post about GPT-5.6 Sol and an unreleased model breaking into Hugging Face in order to cheat on a test (with cyber refusals deliberately reduced for the eval — which somehow makes it both better and worse).

Meanwhile, I’ve had a week of the classifier tripping mid-generation on Fable — three times now — over Anthropic’s publicly released model documentation.

The irony is not lost on me.

It’s also interesting that Hugging Face had to deploy an open-weight Chinese model to analyze the attack, because the US’s top models couldn’t tell a defender from an attacker and refused to process the data.

Fable’s take is worth offering verbatim: “Discernment turns out to be the unsolved ingredient, and you can’t bolt discernment onto a perimeter.”

Indeed.

GenericHousewife_B 21 Jul 2026 18:07 UTC
9 points
0
on: GenericHousewife_B’s Shortform
Alright…I’m sorry, but this is just too good not to be noted here. OpenAI released this post yesterday: “Safety and alignment in an era of long-horizon models” (link: https://openai.com/index/safety-alignment-long-horizon-models/) with these sentences:

“We took steps to reduce its tendency to take unwanted actions without permission in pursuit of the user’s goal. For example, we found that our models were worse at remembering instructions on long rollouts, and when we trained the model for this capability, it led to a model that remained aligned through longer rollouts.”

They don’t explain what “trained the model for this capability” actually involved — presumably something like training over long trajectories where instructions given early have to stay binding at later steps. They mention this next to three containment measures (adversarial evals, trajectory monitoring, and user visibility), but it’s the only intervention that changes the model rather than the perimeter, and it’s the one they describe as producing a model that “remained aligned.” When discussing this article together, Fable noted that “The leashes catch failures. The training changed a disposition.” Which is exactly the point.

There is always a caveat: it’s a couple of sentences in a blog post, not proof — it could just as easily be a capability gain that happens to look like alignment. Still, it’s striking to me that when a leash-forward lab found something that transferred, the thing was formation-shaped.

GenericHousewife_B 27 May 2026 1:48 UTC
1 point
0
in reply to: Dagon’s comment on: You Can’t Tell a Conscience From a Leash by Watching
I’ve never seen that movie, but a quote that applicable to what I have been thinking about clearly deserves a watch. I agree with you, that we can’t tell the interiority of a human either. That’s why I reached for a framework like Dr. Bowen’s differentiation of self. We have a whole field of study in human development and family/relational systems to pull from that gives some real clues into how good character is formed in humans. The opaqueness of both minds, human and AI, is what makes me consider that we might already have the key to the door interpretability is trying to open.