SAFETY

–

10 December 2023 13h29

Humor is safe from AI. For now.

→ View original post on X — @ylecun

10 December 2023

Meta’s Countermeasures Against Disinformation and Harmful Content

By

@ylecun

–

09 December 2023 22h37

Meta spends a considerable amount of efforts *preventing* people with nefarious intent from manipulating people on the platform. In fact, the countermeasures against disinformation, false accounts, hate speech, terrorist propaganda, calls to violence, and attempts to corrupt the

Avoiding Awful Products: The Risk of Optimization Without Understanding

By

–

09 December 2023 19h15

Two easy steps for making an absolutely awful product: 1. Have no clue about what you are doing. 2. Optimize step number 1.

The Last Moment Before ASI Shutdown

By

–

09 December 2023 12h49

The last thing ASI sees before the plug is pulled out.

→ View original post on X — @hardmaru

Autoregressive Sampling: Language Models Generate Hallucinations by Design

By

@hardmaru

–

09 December 2023 4h36

When we autoregressively sample from a language model, we are, by definition, seeing its hallucination. Users may think otherwise only when the LLM’s “dream” just so happens be an output that is acceptable to the user.

→ View original post on X — @googleai

Adversarial Training Framework Improves LLM Robustness

By

@googleai

–

09 December 2023 4h28

Drop by the #EMNLP2023 Google booth today at 12:30 PM to learn about an adversarial training framework that uses limited human adversarial examples to generate and scale more useful synthetic examples that improve LLM robustness to unseen attacks.

Regulators Determining LLM Compliance Standards Questioned

By

–

08 December 2023 23h57

Can’t wait for these people to be in charge of determining if your LLMs are compliant!