The AI Daily Brief
Posts
Biden Meets with AI Safety Experts as White House Plans "Decisive Action"

Biden Meets with AI Safety Experts as White House Plans "Decisive Action"

Plus a textbook-trained LLM beats GPT3.5 with 1/1000th the training data

Nathaniel Whittemore
June 21, 2023

The AI Breakdown First Five - Tuesday June 20, 2023

Today on the First Five:

5. Textbook-Trained LLM Beats 1000x Bigger Models
4. Nvidia Chips Cost Double in China’s Chip Underground
3. A New Turing Test for Chatbots?
2. OpenAI to Launch an AI App Store?
1. Biden Meets AI Safety Community in San Francisco

5. Textbook-Trained LLM Beats 1000x Bigger Models

Maybe bigger isn’t better? There has been a lot of chatter lately about whether ever-larger training sets are the right approach. Phi-1 is a new LLM that achieved better HumanEval than GPT3.5 and others at less than 1/1000th the size of its training dataset. The secret? Training it on textbooks.

New LLM in town:
***phi-1 achieves 51% on HumanEval w. only 1.3B parameters & 7B tokens training dataset***
Any other >50% HumanEval model is >1000x bigger (e.g., WizardCoder from last week is 10x in model size and 100x in dataset size).
How?
***Textbooks Are All You Need*** https://t.
— Sebastien Bubeck (@SebastienBubeck)
1:17 AM • Jun 21, 2023

4. Nvidia Chips Cost Double in China’s Chip Underground

The furious quest for computing power is on everywhere. In China, the market has been hit by US actions to deny the Chinese actions to advanced chips, and Reuters is reporting that an underground market has developed that sees Nvidia chips going for double their normal price.

Inside China's underground market for high-end Nvidia AI chips

Psst! Where can a Chinese buyer purchase top-end Nvidia AI chips in the wake of U.S. sanctions?

www.reuters.com/technology/inside-chinas-underground-market-high-end-nvidia-ai-chips-2023-06-19

3. A New Turing Test for Chatbots?

In 1950, Alan Turing devised a test for understanding if a computer could think like a human. Now, DeepMind and Inflection.ai founder Mustafa Suleyman argues the test should be updated and replaced with a challenge which is won if an AI agent can turn $100,000 into $1m.

A prelude of a few new concepts from my book - one is ACI - artificial capable intelligence.
Read more about it in @Bloomberg today:
— Mustafa Suleyman (@mustafasuleymn)
8:04 PM • Jun 20, 2023

2. OpenAI to Launch an AI App Store?

A huge number of companies are building AI models on top of OpenAI. According to a conversation with developers last month in London, OpenAI is considering building a marketplace to allow them to sell those models directly. Aquant and Khan Academy are examples of interested sellers.

🥱Tired: ChatGPT Plugins
😀Wired: customized ChatGPT models i.e. trained on proprietary data
Will this be OpenAI's App Store?
@aaronpholmes@jon_victor_
— Amir Efrati (@amir)
3:13 PM • Jun 20, 2023

1. Biden Meets AI Safety Community in San Francisco

Last month, Biden dropped in on a meeting between VP Kamala Harris and AI CEOs in the White House. Yesterday, President Biden held a meeting with a number of AI safety experts in San Francisco. The meeting underscores the White Houses promise of forthcoming “decisive actions.”

I am looking forward to this meeting with @POTUS on the dangers of AI and what we can do to prevent harms already impacting everyday people seeking mortgages and housing, in need of medical treatment, encountering workplace surveillance, & more @AJL@AJLUnitedp
— Dr. Joy Buolamwini (@jovialjoy)
3:08 PM • Jun 20, 2023