AI Roundup 073: Music make you lose control
June 28, 2024.
Housekeeping
One last reminder: the 50% discount on a paid Artificial Ignorance subscription ends Sunday!
Music make you lose control
In another AI IP case, the RIAA has sued AI music services Suno and Udio for copyright infringement.
Why it matters:
The lawsuit mirrors those we’ve seen in other areas - stock photographers vs. image generators and media publishers vs. LLM providers.
It’s a reminder that there’s still a lot of legal gray area with AI and copyright, though there is more precedent around what constitutes “fair use” in music.
However, some companies are taking a different route - YouTube is in talks with major record labels to license their songs for music generation tools.
Elsewhere in content and copyright:
After last week's robots.txt controversy, Perplexity’s CEO had a less-than-ideal response; Amazon is also investigating whether the startup is violating its AWS policies.
Reddit is trying to put AI scrapers on notice as tech companies alter their Terms of Service to deal with AI more broadly.
And Google Cloud announced deals with Moody's, Reuters, and ZoomInfo to "ground" its chatbots in real-world facts.
Gemma 2
Google unveiled Gemma 2, the latest iteration of Gemini’s smaller and more open cousin.
The latest:
Gemma comes in several flavors, from 2B to 27B parameters. For now, only the 9B and 27B models are available to developers.
The models perform reasonably well on benchmarks for their size and for being open models. In particular, they’re great at translation despite being so small.
Google is quickly amassing a stable of open models: besides Gemma 2, there’s also CodeGemma, RecurrentGemma, and PaliGemma.
Elsewhere in foundation models:
OpenAI discussed CriticGPT, a model based on GPT-4 that writes critiques of ChatGPT responses to improve the efficiency of the humans who train AI.
Meta released LLM Compiler, a family of models based on Code Llama specifically designed for code optimization tasks.
And on the heels of Claude 3.5 Sonnet, Anthropic has also rolled out Projects, which gives Claude access to shared chats, docs, code, and more.
Elsewhere in the FAANG free-for-all:
Apple reportedly discussed bringing AIs from Meta and Anthropic to Apple Intelligence, but the conversation seems to have derailed due to differences in privacy practices.
Amazon appears to be working on an AI chatbot, codenamed Metis, designed to compete directly with ChatGPT.
And Meta is bringing its AI assistant to India, as well as its user-created AI characters to Instagram.
The power of AI
By some estimates, ChatGPT racks up $700K a day in energy bills. Running all those GPUs is costly - and a new report suggests that AI could help triple our data center energy usage by 2034.
Between the lines:
Maintaining our current growth rate would require new energy sources. Some startups are exploring sources like atomic fusion and geothermal power plants to help with energy demand.
However, one counterargument is that GPUs and training/inference algorithms will become more efficient, reducing the demand for data center power.
And new research shows a 1B parameter model (with custom hardware) consuming the same amount of energy as a lightbulb.
Elsewhere in AI anxiety:
Google's Threat Analysis Group detailed Operation Dragonbridge, a PRC-linked spam network using AI-generated content to spread disinformation.
Likewise, a new DeepMind study finds that the most common misuses of AI are political deepfakes.
And swaths of Silicon Valley, from YC founders to VC investors, are pushing back on California's proposed AI safety bill.
Things happen
Figma AI. What does ‘open source AI’ mean, anyway? Using AI to automatically drop hats outside my window onto New Yorkers. Stability AI gets a new CEO and an investor bailout. A journey into Kindle AI slop hell. Making AI better at math tutoring. Meta is tagging real photos as "Made with AI." Llama.ttf: a font that's also a language model. When your building superintendent is an AI bot. OpenAI acquires Rockset, and then Multi. SoftBank Group launches AI healthcare joint venture. Massive security flaw found in Rabbit R1 source code. How to think about creating a dataset for LLM finetuning. Wonky consultants are a big winner in the AI boom. My first trip to Norway, with AI as a guide. Toys ‘R’ Us used Sora to reanimate the zombie brand. I paid $365.63 to replace 404 Media with AI. Andrew Ng plans to raise $120M for his next AI Fund. An AI version of Al Michaels will deliver Olympic recaps. Shopify begins rolling out its AI Sidekick for merchants. Character.AI now allows users to talk with AI avatars over calls. TransPerfect launches an AI interpreter in over 90 languages. Huawei is struggling to ramp up production of new chips due to a new US crackdown. Top German antitrust official: AI is a "first-class fire accelerator." Has Facebook stopped trying to combat AI spam? OpenAI's CTO Mira Murati isn't worried about AI's impact on creative professions.
