AI Roundup 032: The Insight Forum

September 15, 2023.

Artwork created with Midjourney.

The Insight Forum

On Wednesday, major tech and AI leaders visited Washington for the AI Insight Forum. Organized by Senator Chuck Schumer, the 22 guests (including Altman, Gates, Musk, Pichai, and Zuckerberg) discussed risks and opportunities of AI behind closed doors.

Between the lines:

• AI regulation is a matter of when, not if. Lawmakers are intent on new rules for AI companies.

• And CEOs, if anything, seem to be pushing for that. Many have said that Congress needs to step in to stop them from building something dangerous.

• But some lawmakers are wary of regulatory capture, as tech companies seek to shape new legislation.

AI Roundup 015: Mr. Altman goes to Washington

Elsewhere in AI regulation:

• The White House says eight more companies signed its voluntary AI pledge, making fifteen in total.

• California Senator Scott Wiener proposed that AI models meet transparency standards above a certain compute threshold.

• And the Financial Times argues that disparate AI regulation approaches risk throttling the industry with red tape.

Théâtre d'Opéra Spatial, created by Jason M. Allen/Midjourney.

Théâtre d'Opéra Spatial

In another blow to AI-generated art, the US Copyright Office has rejected an AI image that won an art competition. Unlike previous rejections, this image had "at least 624" prompt tweaks and was altered using Photoshop.

Why it matters:

• The copyright cases working their way through the legal system could upend how we use AI.

• At one extreme, agencies and courts could decide that all AI-generated content is ineligible for copyright - a dealbreaker for most companies.

• At the other end, copyrighted AI content could mean paying licensing fees to artists/authors whose work went into the training data.

Elsewhere in AI copyright:

• Another group of writers is suing OpenAI over copyright claims. This time, it's because of books in ChatGPT's training data.

• Amazon will require its eBook writers to disclose whether their books contain AI-generated content after complaints from the Authors Guild.

• And the Washington Post examines how Google, Meta, and Microsoft use our data to train their models, whether we like it or not.

Artwork created with Midjourney.

Return of the Gemini

Google has reportedly given a small group of companies early access to Gemini, an LLM intended to compete with GPT-4.

What to watch:

• The bar is high: GPT-4 is still the most capable chatbot available across most benchmarks.

• And it's not just Google competing. Meta plans to train a new LLM early next year that aims to be as advanced as GPT-4.

• But the chatbot launches continue - this week saw Roblox's Assistant and Salesforce's Einstein Copilot.

Elsewhere in the FAANG free-for-all:

• Microsoft unveils EvoDiff, a novel protein-generating AI framework.

• Amazon announces AI tools to help sellers write better product titles and descriptions.

• And Google pledges $20 million towards the responsible AI fund.

Things happen

Stability AI launches Stable Audio, its text-to-music model. ChatGPT beats Wharton MBAs in an idea contest. A guide to building RAG-based apps. RETFound, AI that predicts heart failures from retinal scans. OpenAI buys ChatGPT.com. Adobe Firefly is now generally available. 53% of US adults say AI misinformation will impact the 2024 election. Nvidia's TensorRT will double the H100's inference performance. Bluebird, the UK’s AI-powered air traffic control. ChatGPT uses 500ml of water for every 5-50 prompts. The IRS is using AI to catch tax evasion. Nvidia's AI supremacy is only temporary. Against LLM maximalism. Glass Health is building AI for medical diagnoses. Texas billionaire creates AI clone, charges $55/meeting. ChatGPT solves a boy's diagnosis after 17 doctors failed. AI sex worker ads are flooding Instagram and TikTok. How to run a competitive AI startup fundraise in 2023. What if AI treats humans the way we treat animals?

Last week’s roundup