AI Update [19/09/2023]

Hi all(emaal),

For this week’s experiment, I am also adding the topics of ‘Research’ and ‘Models’ that I previously only included for fellow researchers. Not one to arbitrarily gate knowledge away, I thought I might as well include it here for those interested.

Also a small disclaimer: everything listed here is just my distillation of news I’ve come across, and as such is riddled with bias, even if I try to include varied content always keep an openly critical mind!

On to the news-

Regulation

• US to stop export of NVIDIA and AMD chips to middle east for fear of 'military end user' in China 

• US experts on AI prefer agencies or global organisation to regulate AI, with a sizable chunk saying it can't be regulated, and a small minority saying it shouldn't (or be relegated to private entities) 

• Snapshot of a draft for US bipartisan framework on a US AI Act 

Ethics

• Microsoft announces it will pay any copyright infringement costs that come from using its generative AI 'copilot' software 

• Google updated its political ad content policy to require disclosure of the inclusion of 'fictionalised depictions' 

• Opponent of Donal Trump in republican election race used AI-generated voice segments of Trump in a combative advertisement 

• Scientific article explaining the 'generative AI supply chain' and how it touches the issue of copyright 

• AI news-presenters increasingly used in Asian countries 

• Reuters report on the arms race towards autonomous military robots, from drones to submarines, and the inherent advantages over a purely conventional armoury that will keep the development going at speed regardless of ethical concerns 

Creativity & Art

• GPT-4 scores in the 99th percentile on divergent creative thinking tasks 

• Article diving into the US court's dismissal of copyrighting AI-generated art, and what it means for artists 

• Award winning piece of AI generated art failed to get copyright after last legal battle failed 

• Self publishing authors on Amazon now obliged to declare whether content was (partially) AI generated, making a clear distinction between 'AI-generated' and 'AI-supported' 

Education

• English learning platform ELSA launched their AI Tutor, a voice-based AI tutor for role-playing scenarios and generating feedback and full evaluation reports 

• OpenAI released guide for how to use GPT-4 in teaching environments, largely just including large prompt templates 

• Opinion piece including arguments in favour of integrating AI in education, its potential role for helping migrant children in the classroom, and how inaction will lead to technology only being used by the already wealthy and prosperous 

• Tips for educators on how to become more familiar with AI and integrate the conversation properly into the classroom, aimed at high schools, source material in English 

Business

• Australian pension fund invests €1.5B into European data centre 

• Very tiny transformer model found in code of Apple's new iOS update for predicting text on-device called UniLM 

• CEO of Inflection.ai (known for their Pi 'pleasant conversation' model) stating in podcast they will be training models 100-1000x larger in the coming 1.5-3 years 

• Israeli AI21 Labs growing to huge $1B+ valuation through new funding round  HuggingFace, a platform for (semi)-open-source models, datasets, and development of AI/ML programs, now valued at $4.5B 

• Coca cola used generative AI to create new flavour (I personally tasted it, kind of 'eh') 

• Consultants working at Boston Consulting Group leveraging GPT4 outperforming colleagues that didn't significantly, results of a study on various metrics and dimensions 

Health & Medicine

• Proof-of-concept online AI test for assessing severity of Parkinson's symptoms was published

• Boy diagnosed cause of chronic pain through ChatGPT, where 17 doctors spanning 3 years couldn't 

• Small scale study on GPT-4's ability to analyse medical records for diagnosing probable causes, did better than clinicians, potential for augmenting clinicians primary diagnostic process 

• Microsoft announced new semi-open-source model for solving protein folding problems given protein sequence information 

• Foundation model published in Nature Medicine for pathology detection in images 

Models

• TII in Abu Dhabi released a 180B parameter version of Falcon (twice as big as LLaMA2), trained on 3.5T tokens, as a semi-open-source model 

• New large Chinese-English semi-open-source model FLM-101B, trained progressively from a small model to a larger model for cost-effective training ($100K for 100B+ model) 

• TinyLLaMA is being trained, to be a 1.1B model trained on 3T tokens on a similar architecture as LLaMA2 but due to size more appropriate for memory/CPU-bound devices 

• Stability.AI released Stable Audio, a semi-open-source model for generating audio through latent diffusion 

• New anomaly detection model for detecting industrial anomalies, called AnomalyGPT 

• IBM published a new semi-open-source 'Mixture of Experts' (MoE) model called MoLM 

Research

• MS's new phi-1.5 model, from the seminal 'Textbooks are all you need' paper, further argues in favour of careful high quality dataset creation vs high quantity, delivering a small scale model on par with much larger models 

• Google used LLMs as prompt optimisers using what they call Optimisation by PROmpting (OPRO) 

• Scaling Retrieval Augmented Generation to planet-scale data sources through semantic hierarchies 

Etc

• Decent FT article explaining the working of generative AI through graphics 

• Overview of the various ways that AI could 'supercharge' science, from automated drug/material discovery, covering knowledge blind-spots, and automated validation studies to helping set up hypotheses 

• Wayve (autonomous driving developer) elaborating on using natural language processing in autonomous driving for explainability in training models 

See you in the next one!

Leven is mooi