- AI Time to Impact
- Posts
- . . AI: New startups, new levels of performance, new regulation ideas (12.12.23)
. . AI: New startups, new levels of performance, new regulation ideas (12.12.23)
Marshall Kirkpatrick
December 12, 2023
Ok team, today's been a real solid day of important AI news: New startups! Benchmarks topped! Regulation debates! Today's got it all.
Welcome to our new readers, and thanks to those who introduced them. I just marvel at the thought of the new people who subscribe here each day doing work utilizing these incredible new tools, ideas, and inspiration that the folks specializing deeply in AI are working on and that we can read about here.
For today’s newsletter we looked at more than 20,000 posts online to find the following 6 stories that rose to the top of our weighted analysis of AI community engagement (and my editorial judgement). I hope you find at least a few of them useful! Some may be important in the short term, others in the long term.
Now here’s today’s news…
Marshall Kirkpatrick, Editor
First impacted: AI developers, AI researchers
Time to impact: Short
Last month I wrote about Microsoft's powerful Medprompt strategy and I couldn't stop thinking about it. Now Microsoft's AI team has shared on GitHub their new approach using a modified version of Medprompt with GPT-4, which they say achieved a record score of 90.10% on the MMLU benchmark (Massive Multitask Language Understanding), a test for LLMs' general knowledge and reasoning. This score surpasses the results of Google's Gemini Ultra, according to the team. [Steering at the Frontier: Extending the Power of Prompting - Microsoft Research] Share by email
First impacted: AI researchers, AI product/service consumers
Time to impact: Long
Jeremy Howard, co-founder of Kaggle and fast.ai, and Eric Ries, creator of Lean Startup and the Long-Term Stock Exchange, have launched Answer.AI, an AI R&D lab. Supported by a $10m investment from Decibel VC, the fully-remote lab aims to create practical end-user products based on AI research breakthroughs, with a focus on making AI more accessible. [Answer.AI - A new old kind of R&D lab] Share by email
First impacted: Data Analysts, Business Owners
Time to impact: Medium
New startup Essential AI is developing AI products aimed at increasing productivity by automating workflows, according to the company. They say their technology will increase the speed of data analysts by a factor of 10 and enable business users to make independent data-driven decisions, while also identifying major risks and suggesting improvements in supply chain management, with support from March Capital, Thrive Capital, AMD, Franklin Venture Partners, Google, KB Investment, and NVIDIA. [essential.ai] Share by email
First impacted: AI technology developers in France, European AI industry stakeholders
Time to impact: Medium
French President Emmanuel Macron and Meta's Yann LeCun have expressed concerns over the EU's proposed AI Act, stating it could potentially stifle innovation, particularly in France, a significant contributor to AI in Europe. The AI Act is currently under development, leaving room for further discussions and potential amendments. [EU’s new AI Act risks hampering innovation, warns Emmanuel Macron] Share by email
First impacted: U.S. government bodies, AI sector regulators
Time to impact: Medium to Long
MIT has released a series of policy papers, including "A Framework for U.S. AI Governance: Creating a Safe and Thriving AI Sector," which suggests that existing U.S. government entities can regulate AI tools within their respective domains, emphasizing the importance of understanding the purpose of AI tools for appropriate regulation. The policy also proposes the creation of a new, government-approved "self-regulatory organization" (SRO) agency, akin to the Financial Industry Regulatory Authority (FINRA), to oversee AI and address legal issues where AI capabilities exceed those of humans. [MIT group releases white papers on governance of AI] Share by email
Early Stage: Big news from early stage companies
First impacted: Graphic Designers, AI Developers
Time to impact: Short
Segmind has launched two new open-source text-to-image models, Segmind-VegaRT and Segmind-Vega, which the company claims are the fastest and smallest for generating high-resolution images. According to Segmind, VegaRT can produce a 1024x1024 resolution image in less than 0.1s, and both models, which are smaller than Stable Diffusion 1.5, are available for commercial use. [Introducing Segmind Vega: Compact model for Real-time Text-To-Image] Share by email
That’s it! More AI news tomorrow!
-Marshall