'Taming' the Generative AI Beast
Enterprise, regulatory, and societal responses to Generative AI begin to take shape
Happy Friday, everyone!
I'm back in London after an exhilarating yet grueling six-week world “AI Tour,” where I've been speaking with startups, multinational companies, regulators, legislators, and independent experts—all trying to understand and respond to the generative AI tornado.
Since the launch of ChatGPT in November, generative AI has become a force of nature. Manifesting with unexpected speed and intensity, it has swept across the global landscape, leaving everyone in its wake (Big Tech included) wondering “WTF is happening?”
Today, my reflection is that we—as a society—are starting to batten down the hatches. The initial months of awe, confusion, and desperate manoeuvring are slowly evolving into a more responsive framework.
Of course, we're only at the very beginning, and considering the pace and scale of this revolution, there will be many more moments when technology outpaces us. But companies, regulators, and organizations are beginning to get their houses in order.
GenAI to transform all knowledge work
The strategic reorientation of Big Tech towards GenAI continues as the race to integrate generative AI into products and services for enterprise clients and billions of consumers gathers pace. This week, Google's I/O conference could have just been named ‘Google AI.’ (Top line: Google is throwing everything and the kitchen sink at AI, but we knew that already.)
A key part of this rapid reorientation is the dawning realization that generative AI will change knowledge work forever.
When foundational text-to-image models like DALL-E 2, Midjourney, and Stable Diffusion emerged over the last twelve months, the full scale of AI's impact on knowledge work was incomplete. Before LLMs hit the mainstream, Midjourney and others were still thought of as tools for creating amazing visual content. It was only when ChatGPT was released, that the penny finally dropped: this was not only about content creation and entertainment but also instruction and efficiency.
I spoke with May Habib, the CEO of Writer, on exactly this.
First Regulatory Framework on GenAI Takes Shape in EU
Meanwhile, on the regulatory side, efforts have been gathering pace as well. Last week, I documented how nation-states have been responding. And this week, the first transnational regulatory framework on AI is finally taking shape. EU legislators agreed on the draft text for the EU AI Act yesterday, which is expected to be finalized in June. This mammoth piece of legislation has been in the works for years, but until a few weeks ago, it didn't even mention generative AI.
In the post-ChatGPT world, it has been significantly redrafted. Now, the EU says it will require providers of foundational models (whether private or open source) to comply with the Act. This includes requirements to identify and mitigate any risks foundational models may pose to health and safety, fundamental rights, the environment, democracy, and the rule of law.
Now, that's something to chew on.
Authentication Architecture for the Internet
Finally, as I have argued for years, we will have to find ways to authenticate information. This is essentially the idea that transparency about the origins of information can be built into the internet's architecture by default. Like a nutritional label, we should be able to see where our information diet comes from and if it was generated with AI or not.
To this end, I am an industry advisor to Truepic (a pioneer of authentication technologies), and a friend of Adobe's Content Authenticity Initiative (CAI). This week, a major generative AI company joined the CAI — Stability AI — committing to implementing the open standard for content authentication being developed by the non-profit C2PA, as well as Universal Music Group.
This also motivated the campaign I released in collaboration with Truepic and Revel.AI. We produced the first cryptographically signed AI-generated video, ‘Mirror of Reflection,’ to show that it is possible to sign AI-generated content, and to serve as a call to action for generative AI companies to adopt this badge of transparency.
So, let's see if the proof is in the pudding: will Stability AI actually start signing their content? They would be the first if they did. (OpenAI, take note!)
Now for the Best of the Rest…
Google I/O or Google AI?
Top line from Google’s annual developer conference? AI, AI, AI, AI, AI. Here are some of the announcements from the company:
Google is jazzing-up the Internet’s prime real-estate (search results) with generative AI, and its new PaLM 2 language model will now underpin many of its products.
This new AI implementation is all about corroboration. Google wants its LLMs to provide sources – like your nerdy friend who always backs up their claims with evidence. Now this is something we’ve all been waiting for.
Brace yourself for Bard, which can now control all Google Suite Apps…write emails, edit documents, and even add stuff to Google sheets, all from one place. This is a big jump for AI personal assistants!
Again, multimodal is the word. You can upload images, visuals, and graphs, and ask Bard questions – a lot like GPT-4, only FREE.
It’s also integrating Google Translate into Bard – Multilingual LLMs have arrived 🎉
Zuck shunned by White House at ‘AI Leaders Summit’
The White House did not invite Mark Zuckerberg to an AI Summit, claiming it only wanted ‘leading firms’ in the space…
I’m no huge fan of Zuck, but the White House is wrong here. Meta has been working overtime to gain relevancy in the Generative AI space — and it has.
The Open Source release of Meta’s Foundational LLM, LLaMA, in March was like the ‘Stable Diffusion’ moment for image generators.
Just yesterday, Meta unveiled its Generative AI features for advertisers — this has been the primary focus of its new GenAI Research Unit set up by Zuck in March. It also continues to pioneer Foudational Models it is releasing open source.
Meta’s ImageBind: An open-source AI model that ‘blends senses’
Meta reveals another OS model, ImageBind: A groundbreaking open-source AI model that integrates various senses like humans do. Unlike existing models, ImageBind is multimodal, working with text, audio, images/video, 3D, thermal, and motion data, enabling imaginative outputs.
Future Vision: Zuck aims to incorporate touch, speech, smell, and brain fMRI signals, making AI models more human-centric…(tell me, is this going too far??)
Spotify pulls thousands of AI-generated tracks 🎵
Spotify has pulled thousands of songs made with Boomy (an AI-music generator) from its streaming services.
This follows Universal Music Group call to ban AI-generated music from streaming platforms, claiming it to be ‘fraudulent’. We’ll see how this one ages…
AI bridges the Uncanny Valley
It’s no longer really a debate. AI can now bridge uncanny valley…while this is exciting for the world of entertainment and visual art, it also means (yep, you got it) deepfakes galore!
MidJourney just unleashed version 5.1 with upgrades that'll make your jaw drop. The upgrades make it even easier for anyone to produce AI-generated images.
Anyway, that’s all for today — before I leave, quick housekeeping note: next week I’ll be chairing a two-day GenAI conference in London.
It is the first to bring together business communities to discuss how Generative AI can be applied in an enterprise context — the who’s who of Generative AI. Everyone from Hugging Face, to NASA, to Stability AI will be there. Really fabulous line-up of independent and legal experts speaking too (insights to follow!).
Enjoy your weekend!
Namaste,
Nina