The week in AI: Meta's new AI training policy - your Facebook data, their rules

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence; we pass along the news, useful resources, tools and services, and highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

Meta will soon use your Facebook photos, posts and other info to train its AI. You can try to opt out - but Meta claims that they don’t need to honor your request

-------------------------

On June 26, 2024 Facebook will change its privacy policy with regard to generative AI features. From that point on, Facebook posts, photos and other data from all users will be used to help train Meta’s AI systems to improve a wide range of AI-powered tools and services.

Private messages will not be affected, only public texts and images. Nevertheless, many users may be skeptical about this new regulation due to data protection concerns or for personal reasons. By default, all of your data except for private messages will be used. If you don’t consent to this, not only is opting out an overly complicated and awkward process (surely by design), but Meta is claiming that, on the legal basis of “legitimate interests”, they don’t have to honor your request at all:

"We will now rely on the legal basis called legitimate interests for using your information to develop and improve AI at Meta. This means that you have the right to object to how your information is used for these purposes. If your objection is honored, it will be applied going forward," Meta's Policy update notice reads. Users have until the end of the month to attempt to opt out.

Using public data to improve AI capabilities can provide benefits through better products and services. But this has to be carefully balanced against ethical concerns. In the age of AI, data is as valuable as gold or oil - there are legitimate concerns around the increasing commodification of user data by large tech companies to fuel their AI ambitions. Not properly compensating users while utilizing their data for business advancement is one thing; giving them a maze-like opt-out option that can then be rejected on vague legal grounds is a gross business practice from a company that, by now, should know better.

Venture capitalists just pumped $6B into Elon Musk’s AI startup, xAI … but, why?

-------------------------

Elon Musk’s ambitious artificial intelligence venture, xAI, has successfully raised $6 billion in a series B funding round. It’s one of the largest funding rounds in the AI sector and signals strong investor confidence in Musk’s vision for the future of AI. This investment will position xAI as a potentially formidable player in the future AI landscape: a report from The Information reported that xAI would use 100,000 of Nvidia’s H100 chips for a supercomputer to power an upgraded version of its Grok AI chatbot. Since those chips can cost up to $40,000, we’re talking in the realm of a $4B+ cost for just the chips - let alone the other components and operational expenses required for this ‘Gigafactory of Compute’.

What merits this level of investment in xAI from VCs? Grok is currently xAI’s only product. It’s not widely accessible (only useable by Premium X/Twitter subscribers) and isn’t significantly better than what is, at this point, a veritable army of similar large language models out there. Here’s a few of our thoughts on the matter:

• xAI is levelling up Grok’s capabilities quickly. From the announcement of the company less than a year ago, to the release of Grok-1 on X in November, to the recent announcements of the improved Grok-1.5 model with long context capability, to Grok-1.5V with image understanding, the model’s capabilities have improved quite rapidly.

• This funding has strategic implications for the broader AI industry - and Musk’s existing ventures. It consolidates Musk’s influence in the AI domain, complementing his interests through Tesla (which he considers an AI company itself, and which already has a supercomputer) and brain-machine interfaces via Neuralink. This integrated approach could lead to synergies where breakthroughs in one part of the Muskonomy accelerate progress in another.

• Musk has stated numerous times that while ChatGPT captured the world’s imagination through natural language, Tesla’s real-world, physics-based AI capabilities outclass OpenAI and everyone else. A bold claim, but one that shouldn’t be entirely dismissed. OpenAI’s Sora shocked the world in February with AI-generated video; but Tesla had already been creating and using AI-generated videos with diffusion models long before Sora was even announced. Tesla is an AI powerhouse.

Musk continues to be a highly controversial figure in the AI space. This kind of money going into startups like xAI at a time when businesses are struggling to find a return on investment in AI seems like a precarious thing. We’ll be following closely to see what these “first products” are that xAI is planning to bring to market, as noted in the blog post.

OpenAI announces new safety and security committee as former Superalignment/safety team co-lead Jan Leicke joins Anthropic

-------------------------

Earlier this month, the co-leaders of OpenAI’s Superalignment team - a team focused on “scientific and technical breakthroughs to steer and control AI systems much smarter than us” - simultaneously quit. Chief scientist Ilya Sutskever announced he was leaving to work on a personal project that was more important to him, while researcher Jan Leike offered a much more terse resignation, fueling speculation that there were some dramatic issues going on behind the scenes at OpenAI around developing safe AI systems. Within just two days of those resignations, OpenAI dissolved their Superalignment team entirely.

Before the dust has even settled on all of the above, OpenAI announced this Tuesday that their board was creating a new “Safety and Security Committee” to oversee the development of its next frontier AI model - which the company said it has recently started training. The new committee will provide evaluation of safety and security processes over the next 90 days, with plans to share adopted recommendations publicly. OpenAI says this new model will ‘bring us to the next level of capabilities on our path to AGI’ (a type of strong AI that can match or surpass human capabilities across a wide range of cognitive tasks).

The new safety squad (run by CEO Sam Altman himself and his backers) won’t bring much comfort to those who, like Leike, feel OpenAI’s current leadership can’t be trusted to put safety before profits (more on this in the Podcast section below - definitely a recommended listen). The vague new model commentary has many interpreting it as GPT-5’s release being imminent - and that the next frontier model beyond GPT-5 is already being trained. For what it’s worth, Altman has a stellar track record of delivering on his vision and timelines.

To make matters even more interesting, Leike just joined OpenAI’s rival Anthropic to continue the Superalignment mission with them, instead. This should come as little surprise - Anthropic was founded by a core of former OpenAI leaders who had major concerns about the company’s profit-lensed trajectory and lack of devotion to responsible and safe AI development. As difficult as it is to scrutinize the “behind closed doors” operations of these companies in depth, it was clear to Leike, at least, that Anthropic’s outlook on AI safety is more than just lip service, but something they are deeply committed to. He couldn’t say the same about OpenAI.

MORE IN AI THIS WEEK

• Apple signs deal with OpenAI to bring ChatGPT to iPhones and tiptoes around Copilot to Google's Gemini as a second 'option'

• AI tutors are quietly changing how kids in the US study, and the leading apps are from China

• The future of financial analysis: How GPT-4 is disrupting the industry, according to new research

• What does the public in six countries think of generative AI?

• NASA & IBM Research to release a new AI model for weather, climate

• $6M fine for robocaller who used AI to clone Biden’s voice

• What Scarlett Johansson v. OpenAI could look like in court

• PwC reaches deal to become OpenAI’s biggest customer, and The Atlantic announces product and content partnership with OpenAI

• Discord has become an unlikely center for the generative AI boom

Have an AI Idea and need help building it?

When you know AI should be part of your business but aren’t sure how to implement your concept, talk to AE Studio.

Elite software creators collaborate with you to turn any AI/ML idea into a reality–from NLP and custom chatbots to automated reports and beyond.

AE Studio has worked with early stage startups and Fortune 500 companies, and we’re ready to partner with your team. Computer vision, blockchain, e2e product development, you name it, we want to hear about it.

Book time to tell us all about your idea

TECHNICAL, RESEARCH & OPEN SOURCE

Codestral: Hello, World!

-------------------------

Mistral, a French AI startup valued at $6 billion and backed by Microsoft, has unveiled its first generative AI model for coding, named Codestral. Designed to assist developers in writing and interacting with code, Codestral supports over 80 programming languages, including Python, Java, C++, and JavaScript. It can complete coding functions, write tests, fill in partial code, and answer questions about a codebase in English.

Codestral is brand new, and although Mistral has an excellent reputation in the AI space, its current practicality is questionable. The model, with 22 billion parameters, demands a powerful PC to operate and shows incremental coding performance improvements over competitors. These improvements are viewed through benchmarks which are often unreliable, but Codestral’s 81.1% HumanEval (one of, if not the most, popular benchmarks for coding capability) score is extremely impressive.

Generative AI tools like Codestral are gaining popularity among developers, with a Stack Overflow poll indicating 44% already use such tools and 26% plan to adopt them soon. These tools can introduce more errors and amplify existing bugs, but Mistral continues to push forward - launching a hosted version of Codestral on its Le Chat conversational AI platform and offering access to their API via a waitlist. The company has also integrated Codestral into various app frameworks and development environments, including LlamaIndex, LangChain, Continue·dev, and Tabnine.

MORE IN T/R/OS:

• Reproducing GPT-2 (124M) in llm.c in 90 minutes for $20 - Andrej Karpathy, formerly of OpenAI

• Scale AI, one of the largest AI training data brokers, ranks LLMs with new SEAL Leaderboards to bring more transparency to LLM capabilities - rankings can be viewed here

• Microsoft, Beihang release MoRA, an efficient LLM fine-tuning technique

• Cohere For AI Launches Aya 23, 8 and 35 billion parameter open weights release

• AI builders' favorite tools: a hive mind survey summary

• Red-teaming LLM applications

That’s all for this week. We’ll see you next Thursday.