Literally I have arguing for 23 years that neural nets without symbolic representations would be vulnerable to hallucinations, and they have. And literally every year (often daily) I have heard promissory notes like these, and never seen a detailed technical explanation as to https://
x.com/saehtweets/sta
/saehtweets/status/1759329436007112736
…
RESEARCH
-
Neural Networks Without Symbolic Representations Vulnerable to Hallucinations
By
–
-
OpenAI’s New Game: Spotting AI-Generated Video Violations
By
–
See how many physical and biological violations you can spot in this video. Fun new game, brought to you by OpenAI.
-
Learning Causal Concepts from Rich Multimodal Data
By
–
To learn some causal concepts, which seems to be your concern, rich multimodal data can be enough, see eg https://
arxiv.org/abs/2110.10819 Either interventions or teacher corrections accelerate the process, but in some cases data (sensory streams, language) can be rich enough to remove -
LLMs Limitations: Statistical Prediction Over True Reasoning
By
–
the solutions not come from within LLMs; they are doing the wrong kind of computation, which has nothing to do with facts and reasoning, and pertains only to prediction. The statistical inference has steadily improved with more data; problems with factuality and reliability
-
End-to-End Models and Neural Network Interpretability Explained
By
–
End-to-end just means that no part of the model is crafted by hand. The models don’t do anything incomprehensible. https://
distill.pub/2020/circuits/ -
Voronoi diagrams for text embedding space visualization
By
–
today i'm making voronoi diagrams of text embedding spaces what can we do with these?
-
Long Context Capabilities Unlock New AI Potential
By
–
Yup, the amount of capabilities unlocked from the long context is hard to grasp. My tests here are just scratching the surface
-
Gemini 1.5 Pro Achieves 99.2% Accuracy on 10M Token Context
By
–
Yes, yes, and yes “For text, Gemini 1.5 Pro achieves 100 percent recall up to 530k tokens, 99.7 percent up to 1M tokens, and 99.2 percent accuracy up to 10M tokens”
-
Gemini 1.5 Pro and Sora Launch Same Day Historic AI Week
By
–
Still can’t believe we got Gemini 1.5 Pro and Sora on the same day. Definitely of the most incredible weeks in AI that we’ve seen so far
-
AI Model Accuracy Confirmed on Extended Context Testing
By
–
I didn’t believe it at first, but after testing can confirm it’s fully accurate on long context and also very detailed (at least for all the tests I ran)