AI Roundup 052: AI, EO, DPA
February 2, 2024.
AI, EO, DPA
The Biden administration plans to use the Defense Production Act to force tech companies to inform the government when they train AI models above a compute threshold.
Between the lines:
These actions are one of the first implementations of the broad AI Executive Order passed last year. In the coming months, more provisions from the EO will come into effect.
OpenAI and Google will likely need to disclose training details for the successors to GPT-4 and Gemini. The compute thresholds are still a pretty murky area - it's unclear exactly when companies need to involve the government.
And while the EO was a direct response from the executive branch, Senators on both sides of the aisle are eager to take action on AI (and Big Tech more broadly).
Elsewhere in AI regulation:
Bipartisan senators unveil the DEFIANCE Act, which would federally criminalize deepfake porn, in the wake of Taylor Swift's viral AI images.
The FCC wants to officially recognize AI-generated voices as "artificial," which would make AI-powered robocalls illegal.
And a look at the US Copyright Office, which plans to release three very consequential reports this year on AI and copyright law.
FAANG free-for-all: Meta
Meta released the 70B version of Code Llama, a beefier, more advanced version of its existing Code Llama model.
Why it matters:
70B models are (usually) much more sophisticated than their 7B counterparts. With this newest release, Code Llama is now the largest open-source coding model and is closing the gap with GPT-4.
Most programmers lack the hardware to actually run the 70B model. But it's a powerful new tool for AI-coding startups and the engineers they serve.
Ultimately, these state-of-the-art releases are clearly part of Meta's overall strategy, which Mark Zuckerberg detailed on an earnings call this week.
Elsewhere in AI model releases:
The Allen Institute for AI released its Open Language Models (OLMo) and the dataset used to train them.
A Reuters report details how China has approved over 40 models for public use in the last six months, including 14 LLMs just last week.
And after some speculation, Mistral's CEO confirms a "leak" of a new model that could finally be an open-source rival to GPT-4.
FAANG free-for-all: Google
While Meta is releasing swathes of AI models, Google is productizing them; the company debuted several new features that incorporate its wide variety of in-house AI.
What to watch:
MusicFX, an upgrade to MusicLM, the music-generating tool that was released last year. MusicFX can make songs and loops up to 70 seconds long.
And Google Maps is getting new LLM-powered features to suggest shops and restaurants when users ask.
Elsewhere in the FAANG free-for-all:
Tim Cook says that Apple is spending "a tremendous amount of time and effort" on AI features that will be announced this year.
Amazon launched Rufus, an AI-powered shopping assistant trained on its product catalog.
And Microsoft added safeguards to its AI image tool after it was used by 4chan to create fake images of Taylor Swift.
Artificial Ignorance is reader-supported. If you found this interesting or insightful, consider becoming a free or paid subscriber.