AI Roundup 065: The gpt2-chatbot mystery
May 3, 2024.
The gpt2-chatbot mystery
This week, a mysterious model named "gpt2-chatbot" appeared on the LMSYS leaderboard, demonstrated state-of-the-art capabilities, and then disappeared.
Between the lines:
The disappearance isn't actually that strange - LMSYS allows companies to test unreleased models on a temporary basis.
Many suspect it was an unreleased OpenAI model, based on some tokenizer quirks.
It would probably be weirder if it wasn't an OpenAI model - that would mean another organization has developed a GPT-4 class model, seemingly out of nowhere.
Elsewhere in AI anxiety:
The National Archives banned its employees from using ChatGPT.
Microsoft reaffirmed that police departments cannot use its enterprise AI tools for facial recognition.
And Catholic advocacy group Catholic Answers "defrocked" their AI priest "Father Justin" after some hallucination issues.
Executive function
In the six months since President Biden's AI Executive Order, dozens of federal departments and agencies have been staffing up and marching forward with AI-related changes.
Here’s the latest:
The CFTC appointed Ted Kaouk as its first Chief AI Officer - Kaouk is currently the regulator's Chief Data Officer.
NIST launched a new platform to assess generative AI with plans to release benchmarks and help create deepfake detection systems.
And the Department of Homeland Security is creating a new Artificial Intelligence Safety and Security Board - including a lot of industry CEOs.
Elsewhere in AI regulation:
Google urges the Department of Labor to update immigration rules, lest the US lose out on valuable IA talent.
On Twitter, California State Senator Scott Wiener offers a defense of his proposed AI safety bill, SB 1047.
And Politico examines major loopholes present in state laws designed to prevent AI bias.
Satya's spending spree
Microsoft CEO Satya Nadella has been on a tour of Southeast Asia this week and has announced plans to invest $1.7B in Indonesia and $2.2B in Malaysia - the money will go towards infrastructure and training.
Why it matters:
AI isn't just causing a boom in Silicon Valley - top companies are now investing billions around the world to try and keep up with demand.
For example, Google plans to invest $2B in Indiana and $1B in Virginia to expand its data center footprint.
Investments of this size can certainly warp incentives - we've already seen smaller nations adopt AI-friendly regulatory stances in the hopes of attracting more outside cash.
Elsewhere in the FAANG free-for-all:
GitHub unveils Copilot Workspace, an AI-powered dev environment to brainstorm, build, test, and run code with natural language.
Google has agreed to pay News Corp $5M-6M per year to develop AI-related content and products.
And Amazon makes its Q assistant generally available with new tools for app creation.
Things happen
AI is helping automate one of the world's most gruesome jobs. Sam Altman's advice for aspiring entrepreneurs. Nvidia updates its ChatRTX chatbot to support voice queries and new AI models. Atlassian launches Rovo, its new AI teammate. How good is OpenAI’s Sora video model — and will it transform jobs? Sam’s Club’s AI-powered exit tech. AI is making Meta’s apps basically unusable. The FT and OpenAI strike a content licensing deal. Inside the rise of Jesse Lyu and the Rabbit R1. Second AI safety summit faces tough questions, lower turnout. Google Quantum AI. Nurses say hospital adoption of "half cooked" AIis reckless. It's time to switch to Pareto curves. Anthropic's new Claude mobile app. Another 8 newspapers sue OpenAI and Microsoft. Men use fake livestream apps with AI audiences to hit on women. OpenAI is rumored to be launching a new search engine product. Apple targets Google staff to build artificial intelligence team. AI photo geolocation. MIT and IBM's new model to detect crypto money laundering. Facebook's Zombie Internet.
