😺 The AI Act

PLUS: this *free* new model beats ChatGPT-3.5

Welcome, humans.

We accelerated hard in 2023, not publishing on only three federal holidays to honor your well-deserved PTO. Heads up, though: this is our final week of publishing for the year. Time to recharge and come back firing on all cylinders in 2024!

Here’s what you need to know about AI today:

The EU finally agreed on a draft for The AI Act.
Mistral AI raised $415M at a $2B valuation!
Grok rolled out to all X Premium+ subscribers in the US.
Google released its AI note-taking assistant.

Here’s the first serious attempt to regulate AI.

yikes

The EU finally agreed on a draft for The AI Act, the first major piece of AI legislation.

It’s essentially a laundry list of do’s and don’t, with the general gist of ensuring “high-risk” AI companies don’t exploit consumers or, worse, destroy the world.

A key focus of the Act is banning AI uses like internet facial scraping, social scoring, or biometric categorization— think Black Mirror-esque scenarios.

However, the juiciest part is that there are new regulations for big foundational models.

Until now, the inner workings of major models like GPT-4 have been a complete black box.

Under the new EU regulations, these large foundational models will need to:

Cough up technical documentation.
Offer detailed summaries of their training data.
Comply with EU copyright law, which some argue is a complete non-starter.

Non-compliance could result in hefty penalties: up to €35 million or 7% of global turnover…

PSA: These rules primarily target *closed-source models*.

Open-source models, aka AI that’s freely available to build with, were given “broad exemptions”, a win for companies like Meta and European startups Mistral and Aleph Alpha.

Why it matters: Silicon Valley often doesn’t react well to regulation, and the prevailing view is that the AI Act is, at best, overly broad and vague and, at worst, an abuse of authority that will kill AI innovation in Europe and shift it entirely to the US.

One question we have: will companies like OpenAI really act on previous threats to cease operations in the EU altogether with the AI Act set to pass?!

FROM OUR PARTNERS

The AI model strategy every company should be adopting.

Between you and us, the “just use OpenAI” days are limited (or, dare we say it, over!).

Startups and enterprises alike are now using both closed models like GPT and open-source ones like Llama and Mistral.

GPT where it’s good, open-source models for everything else. It’s cheaper, guards against vendor lock-in and gives you insurance when cough there’s drama at your vendor.

And it doesn’t have to be complicated, either.

OctoAI’s new Text Gen Solution is the fastest and easiest way to assemble multiple open-source LLMs under a single API endpoint.

Click here to read more and get FREE credit!

Open-source is catching up to closed-source.

This weekend, Win #1 for open-source—AI that’s publicly available—was regulatory “exemptions”.

Win #2 is that Mistral AI, which makes open-source models, just dropped a model that rivals ChatGPT-3.5 (and beats LLaMA 2).

Win #3 is that this model is accessible to anyone who wishes to use or build upon it right from their laptops (they literally tweeted out the torrent to the full model).

Win #4 is that Mistral is competing with ChatGPT-3.5 with only 22 employees in seven months and just raised another $415M at a whooping $2B valuation!

Why it matters: Open-source models are innovating extremely fast, though they haven’t quite caught up to the level of OpenAI’s (super closed) GPT-4 or Anthorpic’s Claude.

Around the Horn.

The FTC is evaluating whether Microsoft’s investment in OpenAI violates antitrust laws.
xAI’s chatbot Grok is now available to all X Premium+ subscribers in the US.
Anthropic published research on how to prevent AI from discriminating.
BlackRock will start releasing AI tools to its clients in January.