Links for 2023-05-11
PaLM-2 is a next generation large language model with improved coding, multilingual and reasoning capabilities. It will power over 25 new Google products and features, bringing the latest in advanced AI to benefit people.
It excels at advanced reasoning tasks, including code and math, classification and question answering, translation and multilingual proficiency, and natural language generation better than our previous state-of-the-art LLMs, including PaLM.
On MATH, PaLM 2 is competitive with the state-of-the-art performance achieved by the dedicated Minerva model. On GSM8K, PaLM 2
outperforms Minerva and GPT-4 while on MGSM, it surpasses the state of the art even without self-consistency.
Read more: https://ai.google/discover/palm2
Paper: https://ai.google/static/documents/palm2techreport.pdf
GPT-4 but 98% cheaper:
Our experiments show that FrugalGPT can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost.
Paper: https://arxiv.org/abs/2305.05176
“Announcing Helion’s first customer: Microsoft. We expect to start producing electricity in the world’s first fusion power plant by 2028, dramatically shortening the timeline for commercially viable fusion energy.” https://www.helionenergy.com/articles/announcing-helion-fusion-PPA-with-microsoft-constellation/
“Helion, which is backed by OpenAI founder Sam Altman, committed to start producing electricity through fusion by 2028 and target power generation for Microsoft of at least 50 megawatts after a year or pay financial penalties.” [Wall Street Journal] https://archive.is/5G6ty
Personalized Robot Assistance with Large Language Models: "We show that robots can combine language-based planning and perception with the few-shot summarization capabilities of large language models (LLMs) to infer generalized user preferences that are broadly applicable to future interactions. This approach enables fast adaptation and achieves 91.2% accuracy on unseen objects in our benchmark dataset. We also demonstrate our approach on a real-world mobile manipulator called TidyBot, which successfully puts away 85.0% of objects in real-world test scenarios." https://tidybot.cs.princeton.edu/
Researchers detect and classify multiple objects without images https://www.optica.org/en-us/about/newsroom/news_releases/2023/may/researchers_detect_and_classify_multiple_objects_w/
Multi-Space Neural Radiance Fields https://zx-yin.github.io/msnerf/
Control 100,000+ HF models by talking to Transformers and Diffusers. Fully multimodal agent: text, images, video, audio, docs... https://huggingface.co/docs/transformers/transformers_agents
Training Stable Diffusion from Scratch for <$50k with MosaicML https://www.mosaicml.com/blog/training-stable-diffusion-from-scratch-part-2
ChatGPT Fever Has Investors Pouring Billions Into AI Startups, No Business Plan Required [Wall Street Journal] https://archive.is/nN4GV
Stable Diffusion Deepfake - De-Aged Harrison Ford https://www.reddit.com/r/StableDiffusion/comments/13d9ahv/stable_diffusion_deepfake_deaged_harrison_ford/
Existential espionage: How intelligence gathering can protect humanity https://thebulletin.org/2023/04/existential-espionage-how-intelligence-gathering-can-protect-humanity/
“Ramsar .. has the highest measured background radiation levels in the world at up to 200 times higher than the global average. .. residents actually had lower rates of cancer than the control groups.” [Quillette] https://archive.is/vvemM
"Over the last 50 years... evidence has accumulated that only about 10% of school achievement can be attributed to schools and teachers while the remaining 90% is due to characteristics associated with students." https://www.cambridge.org/core/journals/spanish-journal-of-psychology/article/abs/education-and-intelligence-pity-the-poor-teacher-because-student-characteristics-are-more-significant-than-teachers-or-schools/F73D9839D7E1398C33A117CB37E084E8
'This study used a large data set of 3,636 U.S. patients with high blood pressure, and showed a 99.86 % match between cluster-analysis assignment and self-classification into white, African American, East Asian, or Hispanic.' https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3951706/
“There are numerous issues with the idea that Asian culture explains Asian-American achievement like” https://twitter.com/cremieuxrecueil/status/1653888337151680512