Links for 2023-10-23
“We ask models to self-correct using factored critiques and find that this lowers the number of hallucinations by an average of 35% to 0.49 for ChatGPT, 0.46 for GPT-4, and 0.95 for Claude 2. ” https://blog.elicit.com/factored-verification-detecting-and-reducing-hallucinations-in-frontier-models-using-ai-supervision/
Gary Marcus: “Reid Hoffman offered @time to bet any amount of money that hallucinations will be solved to expert levels in the next few months. I am in for $100,000.” https://twitter.com/GaryMarcus/status/1715770891332891055
Context-Aware Meta-Learning https://arxiv.org/abs/2310.10971
Introducing GRID: the General Robot Intelligence Development platform, designed for prototyping smart and safe robots rapidly using foundation models, LLMs, and simulation. https://arxiv.org/abs/2310.00887
“Habitat 3.0: Habitat Synthetic Scenes Dataset and HomeRobot — three major advancements in the development of social embodied AI agents that can cooperate with and assist humans in daily tasks.” https://ai.meta.com/blog/habitat-3-socially-intelligent-robots-siro/
BitNet: Scaling 1-bit Transformers for Large Language Models https://arxiv.org/abs/2310.11453
Approximating Two-Layer Feedforward Networks for Efficient Transformers https://arxiv.org/abs/2310.10837
VERA: Vector-Based Random Matrix Adaptation. Presents VeRA, which reduces the number of trainable parameters by 10x compared to LoRA, yet maintains the same performance. https://arxiv.org/abs/2310.11454
Building less-flawed metrics: Understanding and creating better measurement and incentive systems https://www.cell.com/patterns/fulltext/S2666-3899(23)00221-0
Political links:
A serious crisis is unfolding at Second Thomas Shoal. China appears determined to prevent resupply efforts by the Philippines, as they did most dangerously in 2014. https://www.state.gov/u-s-support-for-our-philippine-allies-in-the-face-of-repeated-prc-harassment-in-the-south-china-sea/ (background: https://amti.csis.org/counter-co-2nd-thomas-shoal/)
Head of Israel's national forensic medical centre. "Many bodies, including those of babies, are without heads." https://www.jpost.com/israel-news/article-769339
“Based on the geolocation of new footage its highly likely that: The missile -Al Jazeera footage- is an interceptor. (Iron Dome) The explosion in the air is too far away to be related to the hospital explosion.” https://twitter.com/GeoConfirmed/status/1716113399728218618
How a 31-year-old hopes to fix Ukraine’s state-owned defence giant https://www.economist.com/europe/2023/10/19/how-a-31-year-old-hopes-to-fix-ukraines-state-owned-defence-giant [https://archive.ph/HBjKH]
Interview with Ukrainian military medic Yuriy Armash https://twitter.com/Mylovanov/status/1716168214202191960
“The US gave Ukraine 8 decommissioned export demo ATACMS worth less than $3M and in return Ukraine destroyed approximately $230M worth of Russian helicopters which represent like 6% of Russias entire helicopter fleet” https://twitter.com/Blake_Allen13/status/1716121278388552178