🤖 AI Brief: US Army wants AI, Google ups their game, and the music wars continue

This is another big week for AI, with plentiful news dropping on the inspiring and concerning side.

We continue to see AI create wild stock shifts, with Palantir’s stock jumping 20% after they announced new AI tools, including a battlefield AI for military clients. 15% (🤯 ) of the world’s music is now AI-generated, according to one estimate.

But through all of this, we’re seeing glimmers of material benefits as well, including Google open-sourcing an AI-powered mouse that enables disabled gamers to play their favorite video games. Quantum computing may now come faster thanks to generative AI.

As always, I write my weekly AI memo so you, the busy reader, can rapidly digest this news and come away smarter.

In this issue:

• 📈 Google ups their AI game

• 🪖 The US Army wants to figure out AI, and Palantir is moving to cash in

• 💥 Anthropic releases Claude with 100k context window

• 🏅 Meta is winning at the open-source game

• 🎶 AI music now flooding streaming platforms

• 🧠 AI can help with quantum computing, plus other studies

Did someone forward this to you? Subscribe here!

📈 Google ups their AI game

Google held their big developer conference Google I/O this week, where CEO Sundar Pichar announced that generative AI would feature in a broad array of the company’s product. This is Google’s catchup year, and the company is now shifting to go on the offensive.

• Generative AI is coming to everything: Gmail, AI photo editing is coming to Google Photos, and Docs will now generate entire paragraphs and spreadsheets from prompts, along with helping users plan their vacation, adjust their tone, and write computer code.

• Also driving the conversation: the theme of responsibility. Google spent time here speaking to how it would combat misinformation, add watermarks to AI images, and bake in other guardrails against misuse.

• IO is now AI: “At Google in 2023, it seems pretty clear that AI itself now is the core product,” said the MIT Technology Review. AI is clearly on top of mind for striking WGA writers.

🪖 The US Army wants to figure out AI, and Palantir is moving to cash in

The DoD has released an RFI (request for information) on methods to protect its data sets for use in AI applications.

• Top of mind for them: Testing AI-enhanced systems in battlefield scenarios while maintaining data security.

• But they don’t want SkyNet, either: finding a way to demonstrate the trustworthiness and reliability of AI to users is critical.

• There’s billions of dollars at stake: Palantir this week said they had seen “unprecedented” demand for its military AI. Their stock went up 21% after it revealed their battlefield AI platform.

The use of AI in military applications has already begun (in 2021, Israel conducted an assassination with an AI-assisted gun). We’ll be watching this topic closely go-forward.

💥 Anthropic releases Claude with 100k context window

100k tokens, which translates to roughly 75k words or five hours of human reading, is a massive upgrade over Claude’s former 9k window.

• Why this matters: businesses could see massive benefits from processing long documents or retrieving information from a massive data set. GPT-4’s current limit is just 32k tokens, while GPT 3.5 is limited to 4k tokens.

• And it’s fast, to boot: Anthropic pasted the entire text of the Great Gatsby into Claude, and the model returned an answer in 22 seconds.

🏅 Meta is winning at the open-source game

Google and OpenAI are increasingly restrictive on the research they share, but Meta is taking a different approach. This week: Meta released ImageBind, an AI model capable of “learning” from six different modalities, including depth, thermal, and inertia.

• This brings AI closer to learning like humans: ImageBind gives machines an understanding of an object’s sound, their 3D shape, how warm or cold they are, and how they move.

• Meta deeps their open-source winning streak: other releases include Segment Anything, Animated Drawings, and their LLaMA LLM model – which is now the foundation of numerous open-source LLMs.

• Expect the community to move quickly: we previously wrote about open vs. closed source AI in this article – and the pace of progress on open-source was simply astounding. Expect the same here.

🎶 AI music now flooding streaming platforms

The removal of Ghostwriter’s fake Drake song was just the beginning. This week, news broke that Spotify has removed “tens of thousands of AI-generated songs” from its platform – and they’re barely scratching the surface.

• Spotify suspects foul play: most of the songs were made by a single generative AI company, Boomy, and suspicious streaming data means bots could have been used to juice royalties on these AI tracks.

• The scale is massive: Boomy claims that they’ve created over 14 million songs – about 14% of the world’s music – during its two years in existence. Expect this number to exponentially grow over time.

• Google isn’t helping: the company released MusicLM this week, which enables users to generate music from text prompts. While specific artists and vocals are forbidden, a broad array of styles can still be made.

🧪 Science Experiments

AI is helping make quantum computing possible by designing circuits

• Quantum algorithms need to be designed by hand, but it’s notoriously difficult. This could very well be AI’s superpower, much like its potential impact on drug discovery and protein folding.

• Read the full paper here.

Google introduces AI gaming mouse, open-sources code

• For gamers with conditions like muscular dystrophy, normal control devices are not usable

• Google’s tech scans the face and tracks head movements to then convert them into in-game movements. An early review called the controls “robust and intuitive.”

• Access the open-source code here.

Robotic household cleanup benefits from LLMs, Princeton/Stanford study finds

• Everyone has different cleanup preferences, due to taste, cultural background and more

• By combining an LLM with a cleanup robot, a robot was able to make remarkable decisions around where objects should go

• See the full study here.

Which open-source LLMs are good? A leaderboard now tries to provide an answer

• With dozens of open-source models releasing, it’s hard to verify performance claims. A new and ongoing study now subjects all open-source LLMs to a series of 4 benchmarks, helping provide a baseline for comparison.

• Link to Hugging Face page here.

Diffusion model can now create 3d faces for all lighting conditions from just an image

• The pace of image technology continues to be remarkable. Even this early proof of concept is quite fascinating. Full paper here.

That’s it! Have a great week!

And as always, we want to hear from you on how useful this is. The more feedback you provide, the better our newsletter gets! What else would you like to see covered?

So take the poll below — your signals are helpful!