I thought their entire founding thesis and brand is “whatever worked while we were at OpenAI but it’s more safe“
@plinz
-
Airline Classes Reveal Income Inequality and Professional Bargaining Power
By
–
The sharp discontinuities between airline classes illustrate that income differences don’t always reflect the smooth continuum of human ability but also differences in bargaining power of different professions.
-
Midwits, Institutions, and the Reproduction of Unchallenged Opinions
By
–
Reproducing opinions while not being able to disprove them or to construct reliable new ones is the operational definition of a midwit. Midwits are an important part of society. They are constituting the bulk of our institutions. They even have newspapers made just for them.
-
Philosophy Science Trade-offs in AI Research Papers
By
–
I am not arguing that a paper cannot be politically valuable and instigate debates despite representing not very good philosophy and computer science. The question whether the community benefited or was harmed as a result is still subject of debate, as it is often with politics.
-
Human Mind Efficiency Versus LLM Computational Requirements
By
–
Our own minds appear to be much more compact than an LLM, we know far less yet are highly generalistic, getting to a working epistemology and the ability to draw correct conclusions with far less data than today's AI models. It must be possible to recreate this.
-
Parrots vs GPT-4: Beyond Stochastic Parrots Debate
By
–
Hm, has anyone done a comparison between the abilities of parrots and GPT-4? This kind of argument is pronounced in the stochastic parrots paper, too: more motivated by politics than philosophy or empirical analysis.
-
Semantic LLM Editing for Compact Knowledge Representation
By
–
I expect that once we can edit LLMs semantically, we can produce compact core LLMs that may not be good at style imitation, but that can quickly and reliably extract knowledge from text and store and exchange it in a kind of universally interpretable mental representation code.
-
Neural Network Structure and Semantic Interpretability in LLMs
By
–
The neural network that encodes the LLM has no canonical form, its structure depends on the architecture, the training data, and the order in which it is processed. It is still very difficult to extract a semantically interpretable representation from an LLM.
-
University DEI Policies and Their Impact on AI Research
By
–
It's like a university system with a very strong DEI statement