In statistics, bias is deviation of a model from the structure indicated by the data. In political discussions, bias is generally not a lack of knowledge, but a distortion of a model by normative preferences (suppression of data, possibilities, interpretations one does not like)
RESEARCH
-

12B model could memorize entire English Wikipedia given bit density
By
–
English Wikipedia is 29.4B characters, with ~1.5 bits of information per character. So that means, at 3.6 bits per parameter, a 12B model could memorize all english Wikipedia, it seems wild!
-
Why We Lost Our Vision for the Year 2100
By
–
Our great-grandparents dreamed constantly of the year 2000. When did we stop dreaming of 2100? Why is retro futurism cool, but we have no modern futurism?
-
We are all midjourney: temporal perspective on progress
By
–
you're now closer to the year 2050 than the year 2000 – a reminder that we are all midjourney
-
Grok’s Epistemological Challenges: Wikipedia Bias and First Principles
By
–
Grok is using sources like Wikipedia, which are not always neutral (Wikipedia articles often get captured by activists when it comes to contested topics). First principles reasoning and sound epistemology are hard to achieve for now
-
LLMs Are Context Sensitive Not Averaging Mechanisms
By
–
LLMs are not averaging, they are context sensitive. All ideas that are sufficiently represented in the training data can be reproduced equally well, regardless of their relative prevalence.
-

Building Aligned Reliable AI Through Domain Expertise
By
–
Building aligned, reliable AI takes more than prompts, it takes domain expertise and human feedback. @AnthropicAI http://
snorkel.ai/LaunchEvent -
AI-Generated Content Quality and Shallow Thinking Concerns
By
–
Agreed — but this doesn't look at all like any AI-helped translation/rewriting of thoughtful human ideas that I've seen. And the content is really the shallowest possible take in this case.
-
Academic Publishing Accessibility for AI Research Papers
By
–
This looks fascinating – will dig in. Any plans to put it on arxiv? Publishing only on the corporate site makes it pretty hard for discoverability, paper libraries like Zotero, AI-powered research, etc.