Neuro-symbolic paradigm strikes again! If in #AlphaGeometry, the creative language model (System 1) suggests insights for the reliable symbolic engine (System 2) to complete a proof, we see that pattern again in #AlphaProof. The language model suggests key proof steps in a
SAFETY
-
AlphaProof Solves Hardest 2024 IMO Problem, AI Dominance Expands
By
–
AI is beating me at the things I love, one step at a time (coding, StarCraft, mathematics, …). AlphaProof solved the most difficult 2024 IMO problem (P6). Answer doesn't fit in this tweet
-
Synthetic Data Debt: Model Quality Degradation Risk
By
–
7/ watch this space closely. My prediction is that model developers who are not careful about training on synthetic data with no information gain will find their models getting steadily stranger and dumber over time. Synthetic data accumulates a debt with the model that must be
-
Synthetic Data Training Risks Model Collapse Long Term
By
–
3/ This core idea is very important to pay attention to: Synthetic data can create a short-term boost in eval results, but you will pay for it later with model collapse! You accumulate debt with mangling the model that starts invisible, and is very hard to repay.
-
Model Collapse in AI: Recursive Synthetic Data Training Risks
By
–
1/ New paper in Nature shows model collapse as successive model generations models are recursively trained on synthetic data. This is an important result. While many researchers today view synthetic data as AI philosopher’s stone, there is no free lunch. Read more
-
Meta releases Llama 3.1 with new trust and safety tools
By
–
Build safe and responsible experiences with our latest Llama trust and safety tools. Along with Llama 3.1, we released Llama Guard 3, Prompt Guard and CyberSecEval 3. They’re available as part of the model download on our website and you can find more details here
-
Rule-Based Rewards for AI Safety and Capability Alignment
By
–
Rule-based rewards (RBRs) use model to provide RL signals based on a set of safety rubrics, making it easier to adapt to changing safety policies wo/ heavy dependency on human data. It also enables us to look at safety and capability in a more unified lens as a more capable
-
Aleksander shifts focus to new AI safety research project
By
–
this is wrong. aleksander is working on a new and v important research project, and joaquin and lilian are taking over the preparedness team as part of unifying our safety work. aleksander will continue to support preparedness work in various ways.
-
Meta releases Llama 3.1 with Llama Guard 3 and Prompt Guard safety tools
By
–
In addition to Llama 3.1 models, we’re also releasing new resources to support trust and safety to enable developers to build responsibly from the start. Llama Guard 3 brings improved performance and multilingual support. Prompt Guard is a multi-label model to help developers
-
AI Risk: Misconceptions About Statistical Nature of AI
By
–
IMO, the biggest risk from #ai is people thinking it is something it isn't. Right now, it is sophisticated statistics with little conceptualization. (Humans work by stats too btw, we just aren't as good at it. And we have an advantage of concepts, but that's another path.)