Just recently came across the preprint of “Learning a Generative Meta-Model of LLM Activations” by Liu et al. (https://arxiv.org/pdf/2602.06964) Which seems to be a work on using generative techniques for steering a model persona. Interestingly it seems to be the most recent work supervised by Alec Radford. Couldn’t really find a public discussion on that paper and would be interested in peoples opinion that are more in the field.
fencebuilder
Karma: 11
Reading some of the recent writings by Dario Amodei (https://www.darioamodei.com/essay/the-adolescence-of-technology#3-the-odious-apparatus) or things he has said in recent interviews (e.g. his recent interview with Dwarkesh Patel), he still seems to be highlighting the risk of China in regards to authoritarianism and misuse of AI and strongly contends that US hegemony in AI is different.
Given the recent events, there seems to be a risk that needs to be addressed that any AI company can essentially be coopted by any government to perform directly harmful or longer term irresponsible tasks (unaligned SOTA AI in military adjacent use is essentially nightmare fuel or anyone worried about alignment). And from a european perspective the core american institutions seem increasingly fragile. Is there even a reason to single out countries like China in such a discussion, as it doesn’t look like the US is being a particularly responsible actor in regards to AI safety?
Not so far at least. I did notice that certain prompts work better for certain use-cases and nudge the model more towards a personality that seems to respond to question at the correct depth.
Did notice, that at least for explaining things in presentation slides, Gemini 3 Flash is almost equivalent and is much faster and cheaper.What topics are you trying to cover? I am currently mostly trying ML/Linalg Math, these might be easy for current models.
Using Gemini 3.1 Pro for personal learning has been a revelation for me personally. I was just uploading some PDF slides on ML Math and the model is perfectly able to reference plots/visual elements in the slides and explain formulas in Slides. Knowing that I learn better by myself rather than in classes, it’s like having an endlessly patient and knowledgeable tutor for my personal learning.
This usage also seems to be quite token efficient, a normal multi-turn session might be more like 50k-500k tokens.
I know that AI-based tutoring is a large market at least in mainland China, but not sure how explored this field is otherwise.
Do you have a reference for the geophagia in Buddhism? I can’t seem to turn up anything on the internet, but maybe it’s not well documented in English?
Funny enough I have been to lesswrong events and felt like the author and probably looked like the random people the author describes to others.
When I join these events I always feel like I am missing a ton of context and interaction with the concepts discussed.
As far as I know, our track record of deliberately driving species extinct that are flourishing under human conditions is pretty bad. The main way in which we drive species extinct is by changing natural habitat to fit our uses. Species that are able to flourish under these new circumstances are not controllable.
In that sense, I guess the questions becomes what happens, when humans are not the primary drivers of ecosystem change?
Did you try both opus and sonnet 4?
fencebuilder’s Shortform
The current acquisitions by OpenAI (io, windsurf) point towards a pivot to commercialization, that at least in my understanding, would be superfluous in very short ASI timelines. Why focus on building consumer AI hardware, if self-improving AI is around the corner?
Thanks for the clarification! Not in the field and wasn’t sure I understood the meaning of the results correctsly.
What is your opinion on the Low Probability Estimation paper published this year at ICLR?
I don’t have a background in the field, but it seems like they were able to get some results, that indicate the approach is able to extract some results. https://arxiv.org/pdf/2410.13211
In the worst outcome this effectively slows down US-based AI development, while motivating a lot of non-US citizens to move to more open development environments elsewhere. Most of the promise of AI is still unrealized and I believe premature restrictions can cause the opposite effect, as one can potentially see in the chip export restrictions to China.