EU's Historic AI Legislation

Today:

EU agrees landmark rules on artificial intelligence

The European Union has just set a world record, becoming the first to lay down strict rules for artificial intelligence. Thierry Breton, bigwig EU commissioner, is over the moon, calling it a "historic agreement". What's it mean? Basically, the EU is setting up a two-level game plan for AI. The first level is about being upfront about how general AI stuff, like ChatGPT, works. The second, tougher level deals with AI that could really shake things up.

They're putting the brakes on using facial recognition tech, except for some tight law enforcement reasons. They're also giving a hard no to AI that can mess with your mind or play on your vulnerabilities, and to any tech that tries to rate how good a person you are.

If companies don't play by these rules, they're looking at a hefty fine – up to €35 million or 7% of their global dough. But not everyone's thrilled. Cecilia Bonefeld-Dahl from DigitalEurope, representing the tech scene, is worried this might backfire, making companies spend more on legal stuff instead of tech talent.

READ THE ARTICLE ON FINANCIAL TIMES

French AI start-up Mistral secures €2bn valuation

Mistral, a fresh AI startup from France, just hit the big leagues with a whoppin' €2 billion value tag. This news is hot off the press, with the deal expected to wrap up by Friday. They've caught the eye of some big-time Silicon Valley players like Andreessen Horowitz, plus other heavy hitters like Nvidia and Salesforce. They're chipping in a hefty €400 million, mixing up equity and a bit of convertible debt. Just this June, Mistral was worth €240 million, so you can tell they're on a meteoric rise.

Founded by three brainy folks from Meta and Google, Mistral's cooking up something special with generative AI, using those big-brained language models that can whip up human-like writing in a snap. But, let's be real, making this tech ain't cheap. These guys are up against the big dogs in the US and China, but they're holding their own as Europe's bright star in this AI race.

Their secret sauce? Efficiency and cost-effectiveness. Plus, they're playing the open-source card, letting other tech whizzes jump in and add their two cents. Right now, they're not raking in dough, but they're betting on a change by year's end with a new platform for customers. Big names like Eric Schmidt and Xavier Niel are backing them, alongside Bpifrance. Everyone's keeping mum for now, but the buzz is real, with Bloomberg already sniffing around their upcoming funding round.

READ THE ARTICLE ON FINANCIAL TIMES

Google’s AI note-taking app is now available to users in the US

Google's new AI app, NotebookLM, is now up for grabs in the US. It's jazzed up with some fresh features and is starting to use Google's latest AI, Gemini Pro, for better document handling. This app isn't just your regular note-taker; it can summarize documents, pinpoint key ideas, and answer questions about your notes. Plus, it's got a cool new trick: turning your notes into different formats like outlines, study guides, or even emails and scripts.

NotebookLM now offers suggestions based on your activity. It might give you tips to spruce it up or throw in related ideas. You can also save smart responses from the app, share notes with pals, and zoom in on specific sources during chats.

Google's also pushing the envelope on NotebookLM's limits. You can now throw in up to 20 sources, each with a hefty 200,000 words. Originally showcased as "Project Tailwind" at Google's I/O event, it was first tested by a small crew. Now, anyone in the US over 18 can get their hands on it. This rollout comes hot on the heels of Google announcing its GPT-4 rival, Gemini.

READ THE ARTICLE ON THE VERGE

Enabling next-generation AI workloads: Announcing TPU v5p and AI Hypercomputer

Google's rolling out this fancy new AI system called the TPU v5p and AI Hypercomputer. This thing's like a hot rod for AI stuff - super fast, with top-notch parts for computing, storing data, and connecting everything. Plus, it's built to be green, using clean energy and saving water.

They're making it easy for developers to use. You've got access to all these high-speed tools through open software. This means folks can tweak and run AI programs more smoothly. It supports popular AI programming tools like JAX, TensorFlow, and PyTorch right off the bat. These tools help in building complex AI models, and Google's making sure they work well on a bunch of different hardware.

There's also this fancy software for training AI and getting answers from it, making it easy to handle big AI projects. They've even tied it up with Google's cloud services for better management and quick recovery from any hiccups.

READ THE ARTICLE ON GOOGLE

Paving the way to efficient architectures: StripedHyena-7B, open source models offering a glimpse into a world beyond Transformers

Together Research just dropped a game-changing set of models, the STRIPEDHYENA-7B series, shaking up how we think about AI model architecture. These include a base model (SH 7B) and a chat model (SH-N 7B), both designed to handle longer texts better and faster than the usual Transformer models like Llama-2 or Mistral 7B.

StripedHyena models are not just as good as the best open-source Transformers out there, they're also quicker and need less memory. This is thanks to a special layer called state-space model (SSM), plus some other cool tech from their research like FlashFFTConv. Compared to typical Transformer models, SH 7B can train on long texts up to twice as fast.

They've also mixed things up with what they call 'model grafting', blending parts of different models to make something even better. StripedHyena combines attention mechanisms and gated convolutions in a smart way, making it a more efficient learner.

READ THE ARTICLE ON TOGETHER AI

Long context prompting for Claude 2.1

Claude 2.1's a new and improved version of Claude, with a whopping 200K token context window, meaning it can remember stuff from like 500 pages back. It's been trained to avoid messing up or making claims without backup. Compared to its older sibling, Claude 2.0, it's 30% better at not giving wrong answers and way less likely to say something's in a doc when it ain't.

Claude 2.1 can be a bit of a stickler when it comes to answering questions based on just one sentence, especially if that sentence seems out of place. Like, in one test, it wouldn't recognize a sentence about the best thing to do in San Francisco if it felt out of context. However, if the sentence fits in well with the rest of the doc, Claude nails it.

A simple tweak in how you ask Claude can make a huge difference. Adding a sentence to the prompt that points Claude to the most relevant part of the text bumps its accuracy from 27% to 98%! It's a game-changer, especially for digging out single-sentence answers.

READ THE ARTICLE ON ANTHROPIC

Google Gemini AI Tries Outsmarting ChatGPT With Photo, Video Skills

Google's new AI model, Gemini, is shaking things up by adding photo, video, and audio understanding to its Bard chatbot. This tech is first hitting the Google Pixel 8 phones. Gemini's got cool text chat features that help with stuff like summarizing docs, reasoning, and coding. The really big deal? It'll soon handle multimedia, like reading hand gestures in videos or solving kids' puzzles.

Gemini's not just about words; it's about getting closer to how we humans experience and communicate in our 3D, ever-changing world. This leap forward highlights how fast AI is advancing. Google's not just competing with OpenAI's ChatGPT; it's integrating this tech into everyday products like search, Chrome, and Gmail. Despite these advancements, it's important to remember that AIs still have their limits and can sometimes get things wrong.

Gemini's ability to work with text, code, images, audio, and video all at once makes it a game-changer. However, be wary of promotional demos; they might jazz up Gemini's capabilities a bit. Google's still fine-tuning, especially the multimedia aspects, to ensure they're safe and responsible.

READ THE ARTICLE ON CNET

Microsoft’s Edge Copilot AI can’t really summarize every YouTube video

Microsoft's Edge browser just got a fancy new feature in its AI Copilot - it can make quick summaries of videos. But here's the catch: it's not exactly ready for prime time. The feature mainly works on videos that have already been processed or have subtitles. Mikhail Parakhin, the big cheese at Microsoft's web services, spilled the beans on this.

When someone asked about it, Parakhin said that if a video doesn't have subtitles or hasn't been prepped by them, the AI can't do much. So, the AI is basically just skimming through the text of the videos, not the actual video content. It's the same deal across Microsoft 365, like for Teams meetings or customer service calls – they all need a transcript first.

READ THE ARTICLE ON THE VERGE

The OpenAI Board Member Who Clashed With Sam Altman Shares Her Side

The situation involving Helen Toner, a former board member of OpenAI, is quite complex and has garnered significant media attention. Toner, an Australian artificial intelligence safety researcher, was involved in the controversial decision to fire OpenAI's CEO, Sam Altman. This decision led to considerable turmoil within the company, including threats from staff to quit en masse if Altman wasn't reinstated.

Toner, who has ties to the effective altruism movement, defended the board's decision to fire Altman, emphasizing that it was not an attempt to hinder the development of powerful AI systems but rather a move to ensure effective supervision of the company. She highlighted that this decision was rooted in the board's responsibility and role in overseeing the company. However, she didn't delve into the specific details of the incident that led to Altman's firing, other than mentioning it was related to a lack of trust.

Following Altman's firing, there was a backlash, and nearly all of OpenAI’s staffers threatened to quit unless he was reinstated. In response, Toner alleged that an OpenAI lawyer tried to pressure her into reversing her decision, stating that the board’s decision might lead to the company’s downfall. She maintained her stance, suggesting that even the potential self-destruction of OpenAI could be in line with the mission of ensuring that AGI (Artificial General Intelligence) benefits humanity.

Subsequently, Altman was rehired, and the company underwent a restructuring of its board. The new board now includes Adam D’Angelo, Lawrence Summers, and Bret Taylor, focusing on building a qualified and diverse board and enhancing governance procedures.

READ THE ARTICLE ON THE WALL STREET JOURNAL