AI Roundup 059: In̶f̶l̶e̶c̶t̶i̶o̶n̶ Microsoft AI
March 22, 2023.
In̶f̶l̶e̶c̶t̶i̶o̶n̶ Microsoft AI
Microsoft announced a new division, Microsoft AI, that will focus on Copilot and other consumer products. It will be run by Mustafa Suleyman and Karén Simonyan, two-thirds of Inflection AI's founding team.
Between the lines:
Microsoft is taking a lot of big bets on a lot of AI horses (the biggest being its $10B OpenAI investment). In addition to being the cofounder of Inflection, Suleyman was previously a cofounder of DeepMind.
This seems like an acquisition in all but name. Microsoft is also poaching most of Inflection's staff, and Inflection's investors are reportedly being made whole via a $650M licensing payment to bring Inflection's models to Azure.
While Inflection likely isn't shutting down anytime soon (it's raised over $1.5B), it's difficult to spin this as a positive for the company. I'm guessing in the next few years we're going see more AI companies flounder as they struggle to stand out.
Elsewhere in the FAANG free-for-all:
Apple and Google have been in talks to use Gemini to power some new iPhone features this year.
Apple quietly released a research paper detailing the training techniques behind MM1, a new multimodal LLM.
And Microsoft unveiled its first "AI PCs," with a dedicated Copilot key and Neural Processing Units (NPUs).
Meet the B200
Nvidia's GTC, the company's developer conference, was this week. The biggest news was the Blackwell B200, a bigger, faster, and more efficient GPU than the H100 (though they could cost $40,000 each).
Elsewhere in Nvidia's AI:
Project GR00T, a general-purpose foundation model for humanoid robots.
NIM, a microservices platform to streamline the deployment of AI models.
Earth-2, a platform to create AI-powered climate simulations.
And an interview with CEO Jensen Huang on the company's key decisions that led to its AI supremacy.
Elsewhere in model releases:
Reports suggest OpenAI will release the next version of GPT in mid-2024, which some are dubbing "GPT-5."
Google releases TacticAI, an AI assistant for football tactics, and VLOGGER, a model that can bring still photos to life.
Stability AI released Stable Video 3D, letting users generate 3D video from text and image prompts, as key researchers behind Stable Diffusion resigned.
And as promised, xAI open sourced the architecture and weights behind Grok.
A little more legislation (a little less action please)
Tennesse passed the ELVIS Act, aiming to protect singers, musicians, and other artists from unauthorized AI deepfakes.
Why it matters:
The new law works by updating the state's right of publicity laws to cover voices. Currently, right of publicity protections (or "likeness laws") are a patchwork of laws that vary by state.
This is the first deepfake law designed to protect musicians - the vast majority of state-level deepfake laws deal with politics and pornography.
But the text of the law has some wrinkles to iron out - it's unclear when two voices can be considered close enough to count as infringement.
Elsewhere in AI anxiety:
The UN unanimously adopted its first global resolution on AI, encouraging countries to safeguard human rights, protect personal data, and monitor AI for risks.
The GPT Store is beginning to fill up with spammy and/or problematic GPTs.
And 404 Media has not one but two stories on Facebook's algorithm: how it's boosting AI spam connected to click farms, and how AI-generated breastfeeding images can bypass its moderation.
Things happen
The BBC plans to build its own AI models.. How Google uses AI to reliably forecast floods around the world. Guiding principles for the Mormon Church's use of AI. An "ethically created" large language model. GitHub’s latest AI tool can auto-fix code vulnerabilities. ArtPrompt, a LLM jailbreak that uses ASCII art to generate harmful responses. Wired's breakdown of "Attention Is All You Need" and its coauthors. Canada and the UAE are trying to poach top European AI startups. Why is AI so bad at spelling? Inside Suno, the AI music startup looking to democratize music creation. The DHS rolls out AI pilot programs to test a variety of use cases. YouTube now requires creators to label their AI-generated content. Sakana AI releases new Japanese language models Sam Altman and Lex Fridman, part 2. Saudi Arabia plans $40B push into AI. The SEC sues money managers for bogus AI claims. Medium is revoking Partner Program status for AI-generated (and other low-quality) content. How AI companies are reckoning with elections. India drops its plan to require approval for model launches. Scientific journals are publishing papers with AI-generated text. Doctors are turning medical generative AI into a booming business.
