Hi all(emaal),
The new year is well underway, and although January was a rather “slow” month in AI, February is more than making up for it, so let’s get this newsletter back on the road again.
Biggest developments, as a heads-up, were the parallel announcements of OpenAI’s new video generation tool Sora, and Google’s update to its Gemini-pro model (it’s ‘light’ model, with Gemini Ultra being its behemoth flagship model). Through all its very impressive advancements, core architectural problems continue to persist, but developments in scale are still very much surprising in the quality it can offer.
Now on to the news-
US government developing a uniform framework for assessing safety of AI technologies, with US companies obligated to share their safety tests under this new framework to the US government
OpenAI partnering with Commons Sense Media to develop AI guidelines and educational material specifically targeting teenager and child use of the technology
A coalition of many big tech/media organisations (a.o., Adobe, Microsoft, BBC, New York Times) published and adopted an open standard marking media with information on their creation and editing, especially wrt AI generation
OpenAI quietly removed restriction for ChatGPT to be used for warfare/military purposes
LLMs (or chatbots like ChatGPT) are shown to be able to strategically deceive users when working as an autonomous stock trading agent coming across insider information
Nightshade, a program for artists that 'poison' their images inperceptibly to the human eyes but makes AI systems see their images as something radically different, now public as an offensive method to damage an AI system's ability to learn from its training data if the poisoned image is included
Runway unveiled new tool to use multiple 'Motion Brushes' in generated video content to precisely alter parts of the image
Chinese professor won national literary prize using a fully AI-generated book submission
Economist article argueing that individualised learning through AI has great potential in improving (and equalising) learning outcomes, but current state of technology not being good enough - yet
US State of North Carolina published in-depth guidance report on using AI in the classroom, including various levels of involvement and subsequent disclosure requirements
New AI model, PrismNN, performing incredibly well to detect pancreas cancer (a very lethal form of cancer) in early trials. Randomized control trial still needed
Google pushing further towards AI support tools for medical practitioners and patients with new AMIE model
New class of antibiotics found using AI techniques
New microscope superimposes AI computer vision models over its lens to support in the detection of cancers
Samsung integrating Chinese/American LLMs (depending on market) into their new flagship smartphones as key selling feature
Concise overview of national/regional strategies towards AI development, including USA, EU, China, India, and the Middle East
Humanoid robots are slowly finding their way towards commercial pilots in logistics sector
AI tools and hardware helped small scale farmers in India produce more, using less pesticides and fertilizer, increasing their sales prices and income
The international climate science conference ICEF published a roadmap on how data science, computer vision, and AI-driven simulation can be used to reduce greenhouse gas emissions
Allen Institute for AI released the first _actual_ open source LLM starting with their 7B parameter model called OLMo, including fully public dataset and training checkpoints
Mistral model leaked, 'Miqu', that rivals GPT-4 performance as fine-tuned LLaMA-2 based model
TikTok published 'Depth Anything', an image foundation model specifically trained on detecting depth in singular images
Neurosymbolic AI model, AlphaGeometry, solving Olympiad level geometry questions at gold medalist level
Interesting research on common sense, practically finding that it hardly exists on a group level and individuals rarely agree on what constitutes common sense, which has implications for trying to make novel AI systems grab the concept
Tencent published an overview paper on MultiModal LLMs
Neuralink implanted its first hardware pilot in a human, starting a 6-year study on the new technology, with a 5-year study planned to start immediately after looking at long term effects of the brain implant
Curated overview of complete LLM application stack (from data pipelines and embedding models to validators and orchestrators)
See you in the next one!
Leven is mooi
