SAFETY - AI Dynamics

AI and Books Enable Knowledge Access, Reject Doom Scenarios

By

–

29 November 2023 13h48

Printed books and AI systems both serve to enable better access to knowledge. They both make people smarter, more creative, & more productive.
That's intrinsically good.
That's the reality. Scenarios of AI-fueled doom are imaginary bullshit.
New forms of medieval obscurantism.

→ View original post on X — @ylecun

29 November 2023

We Must Embrace AI’s Uncertain Future with Optimism

By

@rafiksmati

–

29 November 2023 8h44

Le journal @LaTribune résume ma pensée sur l’IA : « Nous sommes montés dans un véhicule qui va de plus en plus vite et dont on ne sait pas où il va, s'exclame Rafik Smati […]. Mais il se veut optimiste.» Optimiste ? À vrai dire, nous n’avons pas le choix. Soit nous embrassons

→ Voir le post original sur X — @rafiksmati

29 November 2023

AI’s Role in 2024 U.S. Elections: Synthetic Media Concerns

By

@stanfordhai

–

28 November 2023 22h04

Less than one year until 2024 U.S. elections, how is AI shaping political campaign ads? @stanfordhai deputy director @russellwald discusses the real concerns around synthetic media. https://
abcn.ws/46B5V3G

→ View original post on X — @stanfordhai

28 November 2023

Will AI destroy humanity? Debate Le Cun vs Hinton

By

@dfintelligence

–

28 November 2023 19h24

-> Will AI destroy humanity? The great debate between Yann Le Cun and Geoffrey Hinton currently. AI in the new crazy video? We're covering the tech news live now!!

→ View original post on X — @dfintelligence

28 November 2023

OpenAI Hiring Teams for AI Safety Research

By

@lilianweng

–

28 November 2023 19h24

We have various teams working on AI safety at OpenAI. Let us know if you are interested!

→ View original post on X — @lilianweng

28 November 2023

Effect not from pre-train mimicry but system message updates

By

@goodside

–

28 November 2023 18h05

Effect is similar but my instinct is it’s not from mimicking pre-train data. There are updates to the system message for things like refusing to recite recipes and lyrics found online; I think it’s interpreting these as a more general preference against long quotation.

→ View original post on X — @goodside

28 November 2023

Calibrated LLMs must hallucinate; RLHF reduces calibration

By

@goodside

–

28 November 2023 7h38

“Calibrated Language Models Must Hallucinate” — Hallucination rate from a pre-trained LLM ≈ proportion of facts seen only once in training. RLHF reduces hallucination but makes LLMs less well calibrated as models. Real text has news; LLMs need to unlearn that fact to please us.

→ View original post on X — @goodside

28 November 2023

Reducing LLM Hallucinations Through Confidence Detection

By

@whats_ai

–

27 November 2023 21h53

Why Hallucinations happen with LLMs?

Paige Bailey discusses how language models are less likely to have hallucination. She also covers a nice approach where models are able to spot when it has low confidence in answering the question or ask follow up question to ensure… pic.twitter.com/eLzhVZ4TBY
— Louis-François Bouchard 🎥🤖 (@Whats_AI) 27 novembre 2023

Why Hallucinations happen with LLMs? Paige Bailey discusses how language models are less likely to have hallucination. She also covers a nice approach where models are able to spot when it has low confidence in answering the question or ask follow up question to ensure

→ View original post on X — @whats_ai

27 November 2023