Links for 2023-12-07
Google DeepMind releases Gemini, a general AI model that outperforms GPT 4, and Cloud TPU v5p, a powerful processor to train cutting-edge AI models
https://blog.google/technology/ai/google-gemini-ai/
Exceeds current state-of-the-art results on 30 of the 32 benchmarks.
First model to outperform human experts on MMLU, which uses a combination of 57 subjects such as math, physics, history, law, medicine and ethics for testing both world knowledge and problem-solving abilities.
Achieves a state-of-the-art score of 59.4% on the new MMMU benchmark, which consists of multimodal tasks spanning different domains requiring deliberate reasoning.
Remarkable ability to extract insights from hundreds of thousands of documents through reading, filtering and understanding information.
Was built from the ground up to be multimodal, which means it can generalize and seamlessly understand, operate across and combine different types of information including text, code, audio, image and video.
Can understand, explain and generate high-quality code in the world’s most popular programming languages, like Python, Java, C++, and Go.
Powers AlphaCode 2, which excels at solving competitive programming problems that go beyond coding to involve complex math and theoretical computer science.
AlphaCode 2 shows massive improvements, solving nearly twice as many problems, and performs better than 85% of competition participants — up from nearly 50% for AlphaCode.
AlphaCode 2 requires about 100 samples to reach the level of performance of AlphaCode with a million samples, making it over 10000× more sample efficient.
Gemini: A Family of Highly Capable Multimodal Models: https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf
AlphaCode 2 Technical Report: https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf
More:
Demis Hassabis says the new multimodal model will be a foundation for rapid innovation in software agents, planning and reasoning (a la Q*), gameplay, and even physical robots: https://www.wired.com/story/google-deepmind-demis-hassabis-gemini-ai/ [https://archive.is/eH4MB]
"Along with text, images, video and code, Gemini is able to process raw audio signal end-to-end. 🔊 It can listen to and understand speech, making it not only useful for transcription but a model that has a much more nuanced perception of its environment." https://twitter.com/GoogleDeepMind/status/1732461149554094259
“The next great chatbot will run at lighting speed on your laptop PC—no internet connection required. …Every big name in consumer tech, from Apple to Qualcomm, is racing to optimize its hardware and software to run artificial intelligence at the ‘edge’—meaning on local hardware, not remote cloud servers. The goal? Personalized, private AI so seamless you might forget it’s ‘AI’ at all.” https://spectrum.ieee.org/personal-ai-assistant
“As [OthersideAI developer Josh] Bickett described, the [self-operating computer] framework ‘lets the AI control both the mouse where it clicks and all the keyboard triggers essentially. It’s like an agent like autoGPT except it’s not text based. It’s vision based so it takes a screenshot of the computer and then it decides mouse clicks and keyboards, exactly like a person would.'” https://venturebeat.com/ai/the-self-operating-computer-emerges/
“In a first, AI was able to reconstruct images from brain activity with over 75% accuracy.” https://www.sciencedirect.com/science/article/pii/S0893608023006470
Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback — “For RLHF-LMs such as ChatGPT, GPT-4, and Claude, we find that verbalized confidences emitted as output tokens are typically better-calibrated than the model's conditional probabilities on the TriviaQA, SciQ, and TruthfulQA benchmarks, often reducing the expected calibration error by a relative 50%.” https://arxiv.org/abs/2305.14975
Tiny robots made from human cells heal damaged tissue https://www.nature.com/articles/d41586-023-03777-x
“For the very first time in our history, in human history, biology has the opportunity to be engineering, not science.” — Jensen Huang, CEO of Nvidia https://twitter.com/GeneInvesting/status/1732212410587705652
“…the beaver dam is an extended phenotype of the beaver.” https://www.emilkirkegaard.com/p/taking-the-human-extended-phenotypes
Bantu expansion into Angola & Mozambique was genocidal, with at most 10% (probably less) of modern ancestry coming from the pre-Bantu groups: https://nature.com/articles/s41467-023-43717-x
Political links:
“The IDF has officially decided to flood the tunnels in Gaza. It will start doing so gradually so that it can control any side effects and so Hamas can remove hostages from there in time before they collapse.” https://twitter.com/academic_la/status/1732129584777441406
Politico reports that Deputy Russian Foreign Minister Rudenko visited Beijing in June and informed Xi Jinping that his foreign minister and relatives of top rocket force officers had helped pass Chinese nuclear secrets to Western intelligence agencies. https://www.politico.eu/article/chinas-paranoid-purge-xi-jinping-li-keqiang-qin-gang-li-shangfu/