AI Roundup 021: Mosaic joins the three comma club

June 30, 2023

Artwork created with Midjourney.

AI M&A

Perhaps this week's biggest news was Databricks acquiring MosaicML for $1.3 billion. Databricks is an enterprise platform for managing data, analytics and AI, while MosaicML allows companies to train and deploy generative AI models.

Why it matters:

• AI has been a white-hot space this year, with enormous valuations and, now, acquisitions.

• Perhaps more remarkable than the purchase price is the speed at which MosaicML was acquired: the company launched in 2021 and only had 62 employees.

• And the acquisitions will keep coming - this week also saw Casetext (AI tools for lawyers) acquired for $650 million and Mode Analytics (business intelligence) acquired for $200 million.

On AI fundraising:

• Normally, I don't cover AI fundraising announcements for multiple reasons. One, there are too many to keep track of. Two, raising VC funding isn't the finish line - it's the starting line.

• But Inflection AI raising $1.3 billion feels notable, because of what the money is earmarked for - building a massive computing cluster to train AI models.

• It's a sign that one big bottleneck continues to be GPUs, and that (so far) Nvidia is the biggest winner.

Artwork created with Midjourney.

People are worried about AI spam

With generative AI, we're about to enter a world with truly infinite content - but we're already seeing the problems that can cause. Moderators of user-generated sites like Etsy, Reddit, and Wikipedia are already struggling to deal with the deluge. And AI-generated "books of nonsense" are crowding out Amazon's bestseller lists.

Between the lines:

• Every marketplace is going to have to confront its policy on AI content. Valve appears to be unwilling to publish games using AI assets, whereas Unity announced a marketplace for AI models to help game developers do just that.

• Plus, brands are now discovering some of the downsides of cheap AI content - their ads are appearing on AI-generated spam sites.

• This isn't a new idea, but it bears repeating: what happens when there are more SEO spam posts and social media bots than we can deal with? What happens when the next generation of language models is trained on a version of "the internet" comprising mostly AI-generated text?

Elsewhere in AI anxiety:

• A new lawsuit against Microsoft and OpenAI alleges the companies used stolen personal data to train ChatGPT. It may seem like a frivolous case and likely won’t go far, but it pokes at an important question: should AI companies be allowed to use millions of books, comments, and images without consent or compensation?

• Housing advocates warn against the growing trend of US landlords using AI tools to screen tenants, as it risks heightened discrimination.

• In a new study, people trusted tweets generated by GPT-3 over content written by humans.

Image credit: Meta.

Meta's model cards

This week, Meta published a deep dive on how AI recommends Facebook and Instagram content, including "model cards" - the AI equivalent of a nutrition facts label.

Elsewhere in the FAANG free-for-all:

• The CEO of DeepMind details their upcoming model Gemini, a GPT-4 competitor that uses ML techniques from AlphaGo.

• YouTube is testing an AI-powered dubbing tool for creators to translate their content into multiple languages.

• Microsoft adds AI shopping tools to Bing AI in Edge.

• A profile of Amazon’s army of warehouse and logistics robots.

Things happen

ChatGPT for iOS can now Browse with Bing. The Vatican releases an AI ethics handbook. Scared tech workers are scrambling to reinvent themselves as AI experts. Making movie trailers with Zeroscope, ElevenLabs, and GPT-4. Mathematician Terence Tao on using LLMs and embracing change. Midjourney launches "zoom out" feature. Gitlab AI goes head to head with GitHub Copilot. House of Representatives restricts congressional use of ChatGPT.

Last week’s roundup