Welcome to this week's AI Weekly Update! We're here to give you a quick look at the newest and most important happenings in the world of artificial intelligence. Our goal is to help you stay informed about the latest AI breakthroughs, research, and discussions, whether you're an expert or just curious.
ChatGPT is gradually introducing plugins from other software providers to study their real-world use, impact, and safety. This rollout aims to address user demands for plugins, which can greatly expand use cases. Starting with a small user base, the plan is to expand access as more is learned. The goal is to build a community that shapes the future of human-AI interaction. Plugin developers can create plugins for ChatGPT using provided documentation. Early collaborators include companies like Expedia, FiscalNote, Instacart, KAYAK, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier.
Meet In The Middle Training Paradigm
Researchers focused on developing training programs for language models have released a paper detailing a new perspective on generating training data: Meet In The Middle. Until now, the vast majority of language models have followed a sequential path, where each word or token is generated based on a likelihood or confidence interval about how well it fit after the previous word. Meet In The Middle training takes a new approach, by defining the beginning and the end, and asking the model to fill in the middle.
Wolfram Alpha + ChatGPT
We look back at early experiments in 2010 involving the generation of Wolfram Language code from natural language. While limited without modern LLM technology, these efforts proved useful for creating symbolic structures. Today, the powerful collaboration between LLM technology and the Wolfram Language, as seen in ChatGPT + Wolfram, marks a historic moment. For the first time, the statistical and symbolic approaches to AI are being combined, paving the way for a new paradigm in AI-like computation and unlocking immense potential.
Talk With Your Documents (or Books!) using AI
New applications featuring "Chat" or "GPT" in their names are being announced daily, such as ChatPDF, ArvixGPT, and GPT for Sheets and Docs. These applications can chat with documents that are over 100 or even 1,000 pages long by converting them into a numeric representation (vector embedding) and storing them in a vector search engine. When users chat with the document, the system uses Approximate Nearest Neighbor search (ANN) to find and return similar text. The benefits of this approach include smaller prompts and accurate responses from ChatGPT. Additionally, to ensure correct answers, users can direct ChatGPT to their organization's knowledge base, which guarantees answer accuracy based on the quality of the knowledge base.
Google releases Healthcare-focused AI tools
Google unveiled a range of AI-driven healthcare tools, including an enhanced chatbot called Med-PaLM 2, which scored 85% on medical exam questions, placing it at the expert doctor level. This performance marks an 18% improvement from its previous iteration and surpasses other AI models like ChatGPT. However, Google acknowledges the limitations of AI in healthcare, as Med-PaLM 2 had "significant gaps" when tested against criteria such as scientific factuality, precision, medical consensus, reasoning, bias, and harm. The company plans to collaborate with researchers and the global medical community to address these issues and explore how the technology can improve healthcare delivery.
An ER Doctor's Take on AI
An emergency department doctor tested ChatGPT's performance in real-world medical situations by feeding anonymized medical history and symptom notes of 35-40 patients into the AI. Using a specific prompt for differential diagnoses, ChatGPT successfully identified the correct diagnosis in about 50% of the cases, which is considered decent but not sufficient in an emergency room context. The AI struggled to identify crucial diagnoses in some situations, like missing an ectopic pregnancy in a young woman who did not initially know she was pregnant. This highlights the limitations of ChatGPT in healthcare and the importance of experienced human intervention for accurate diagnoses.
AI Generated Movies Coming Soon!
The Writers Guild of America is working to prevent software companies from undermining writers by using artificial intelligence to produce written content. The proposal allows studios to use AI to generate original scripts but human writers who polish or re-write the script will be considered the first credited writer. The Guild is proposing that AI should be treated as a tool rather than a writer on the payroll to preserve writers' working standards. The Guild is currently in the bargaining phase to agree on the definition of "literary material" and "source material" and how they will protect writers from losing their share of writing credits or residuals to AI.
Northeastern London Commemorates Philosophy and AI Graduates
Northeastern University London recently held its postgraduate commencement ceremony where 25 graduates were awarded degrees. The ceremony included speeches from graduates, faculty, staff, and friends of the university. Among the speakers was British philosopher and author Anthony Clifford Grayling, who praised Northeastern's program for operating at the cutting edge of contemporary concerns. The university's master's program in philosophy and artificial intelligence, which was launched three years ago, was the first of its kind in the world. 19 of the 25 postgraduates received degrees in this program, while the remaining six earned degrees in other programs, including philosophy, digital politics and sustainable development, and AI with a human face.
Musician Sandee Chan and Taiwan AI Labs Founder Reveal AI-Sung Single
Award-winning musician Sandee Chan and Taiwan AI Labs founder Ethan Tu revealed at a press event that Chan's latest single, "Teach Me the Ways to Be Your Love," was sung by AI. The song, released on March 14, uses "I Learn to Love" as a metaphor for the process of AI learning to sing, and features jazz pianist Martin "Musa" Musaubach. The revelation has sparked discussions on the role of technology in the arts. The song was produced by combining Chan's past album sound files with the work done by AI Labs personnel over one year. An integrated audio-visual creation platform that can automatically generate music is set to be launched soon.
AI-controlled drone may have attacked forces in Libya
A drone that attacked forces loyal to Libyan commander Khalifa Haftar last year may have been controlled by artificial intelligence (AI), according to a UN report. The drone, which fired on Haftar's forces during a retreat, was a Kargu-2 quadcopter produced by Turkish military tech company, STM. The report stated the AI systems powering the drone had a "fire, forget and find" capability, and could attack targets without data connectivity. There is no definitive proof that the AI drone killed anyone, but if it did, it would represent an historic first for the use of autonomous weapons in warfare.