Bye bye Bard
Google is doubling down on Gemini - both the name and the model. Previous AI product lines Bard and Duet are being rebranded, while the most advanced version of the Gemini model (Gemini Ultra) finally went live.
What to watch:
Bard is now Gemini, and is coming to a new standalone Android app and the Google iOS app. Gemini (the product) will have multiple tiers, including a paid plan that's powered by Gemini Ultra (the model).
Early impressions point to Gemini Ultra being a genuine GPT-4 competitor but not necessarily a major step forward. So far, the benchmarks would seem to agree.
That said, it's heartening to see Google finally give OpenAI some real competition. Going by Sundar Pichai's recent comments, the CEO understands that generative AI could represent the evolution of search.
Elsewhere in the FAANG free-for-all:
Apple dabbles in AI, releasing MGIE (an AI-based image editor) and preparing to acquire Brighter AI (technology to anonymize face and license plate data).
Copilot gets a new look and model, while Microsoft partners with Semafor on an AI-assisted breaking news publication.
Mark Zuckerberg details his next steps in AI - learning from the vast corpus of Facebook user data, which he estimates to be larger than Common Crawl.
Deepfakery
After weeks of headlines on deepfake scams and abuse, leading AI companies are taking steps to try and alleviate the problem. The primary effort is the Coalition for Content Provenance and Authenticity, or C2PA.
Between the lines:
The C2PA is both an organization and a set of technical standards for identifying synthetic content. The organization was founded in 2019 by Adobe, Microsoft, and several news outlets; the standards aim to watermark and fingerprint AI content via metadata.
This week, Google joined the C2PA steering committee, while OpenAI committed to using CP2A watermarks for images generated by DALL-E.
Meta, which is not a part of the C2PA, is promoting similar standards from the Partnership on AI. It has also said it will require that users disclose when realistic posts are made with AI.
Elsewhere in deepfakes:
Two Texas companies were behind the AI Joe Biden calls that urged New Hampshire residents not to vote.
OnlyFake is an underground marketplace that uses neural networks to instantly create fake IDs, with impressively realistic results.
And 4chan, after kicking off the Taylor Swift AI fiasco, is now using AI to clothe scantily-clad women instead.
So the FCC won't let me be
This week, the FCC ruled that robocalls featuring fake, AI-created voices are now illegal.
Why it matters:
While individual senators are proposing AI regulation - last month another bill, the DEFIANCE Act, was authored - Congress as a whole is not making much progress.
As a result, government agencies are increasingly bypassing Congress to make something happen. In this case, the FCC interpreted an existing law to cover AI robocalls.
Likewise, the FTC is using its antitrust responsibilities to scrutinize recent multibillion-dollar investment deals from Microsoft, Amazon, and Google.
Elsewhere in AI regulation:
The Biden administration taps Elizabeth Kelly to lead the new AI Safety Institute.
The Centers for Medicare & Medicaid Services (CMS) informed insurers that they could not use AI to deny health care coverage.
And British regulators are being urged by OpenAI and Deepmind to speed up the safety tests of their newest models.
Things happen
OpenAI may be working on two types of agent software. Meta's Fundamental AI Research Lab is made up largely of women. Indonesia's elections are testing the limits of OpenAI's political usage guidelines. Confessions of an AI clickbait kingpin. 3 students used machine learning to recover text from a charred, 2000-year-old scroll. Smaug-72B, a new winner on the Open LLM leaderboard. College admissions experts now say "tapestry" is a red flag for AI-written essays. Training data for the price of a sandwich. The UK government is budgeting £100M for "responsible" AI R&D. Sam Altman is in talks to raise as much as $7 trillion for an AI chip initiative. Roblox adds real-time AI translation. Birdcast, software that uses AI to predict how birds migrate. Copilot heads to the Super Bowl. Insatiable demand for AI chips has forced Huawei to slow smartphone production. AI-generated Grandma porn is flooding the internet. US police departments are using AI to watch millions of hours of body cam footage. How Quora died. Hugging Chat Assistants are open-source alternatives to GPTs. Neal Stephenson was right about our AI age. Geppetto, an open-source AI companion for your Slack teams. AI can now master your music. Better Call GPT: comparing LLMs against lawyers. A $349 pair of multimodal-AI glasses. How to keep your art out of AI generators.
Congratulations on the impressive milestones. You deserve it!
Charlie, congrats on 6500! I know it's not a particular milestone, but I just wanted to say: well done, dude. You deserve it.
I researched today's article via Gemini. It's a half step behind GPT4 for research, but truly only a half step behind, and it adds tremendous value in tandem with GPT4 if you can just go back and forth to verity info. Image gen wise, it's still like a newborn, very limited in what it can do, but I will be closely watching this space.
If you or Dan Nest need to hear from a power user who does research and image generation every day, I might be able to report back from time to time.