'Not that innovative' ChatGPT hits 100m users

PLUS Google prepares to release its own ChatGPT, while an AI-clone of Emma Watson's voice reads 'Mein Kampf'

Feb 03, 2023

Hello all,

I’m back with Friday’s analysis on the week in Generative AI. January’s been pretty crazy. ChatGPT is two months old. Yesterday it was integrated as a paid feature into Microsoft Teams. For $7 a month, it will take meeting notes, recommend follow-up tasks and give users personalised highlights. Handy and affordable, right?

ChatGPT has also just hit 100m users. It is the fastest-growing consumer application in history. TikTok — the next closest — hit the 100m user milestone in 9 months.

Meanwhile, Google’s dialling up the heat. CEO Sundar Pichai told investors that Google’s version of ChatGPT is coming and that it will work in conjunction with Google search in “exciting and innovative” ways.

I can’t help but slightly roll my eyes when I see think-pieces debating how ChatGPT is fallible or, in the words of a certain someone, ‘not that innovative.’ These comments may be accurate in the strictest sense — but they miss the more significant point.

This is only the tip of the iceberg. ChatGPT is only one early application of one type of generative model. I predicted hundreds of millions of people would interact with (and find utility) in Generative AI-powered software, products and services this year. It looks like that prediction will come true through ChatGPT alone.

The tempo will pick up from here on in. Is anyone else finding it hard to sleep?

I’m establishing my publishing rhythm — so I’ll send out a weekly analysis every Friday in addition to my regular interviews with AI pioneers and thought leaders. As always, I aim to break it down in a way that is accessible. Get in touch if you have any ideas, comments or tips — and make sure to add me to your safe contacts list, so I land in your inbox.

Let’s get to it.

Pete @nonmayorpete

A lot has happened in January 2023 alone. If this continues, where will we be by end of year?

There is no great surprise here, but people are using AI content generation tools maliciously and in innocently misleading ways.
As new companies rush products to market, they are deploying without any safety rails, which means many new opportunities for their tech to be misused.
ElevenLabs, an AI-voice cloning company, found their open beta was being (ab)used by 4Chan users to clone celebrities’ voices saying racist and violent things — IE, using Emma Watson’s voice to read Mein Kampf.

Meanwhile, actor Robert Pattison revealed that he finds the viral deepfakes of him dancing and being generally ‘cute’ on TikTok ‘terrifying.’ Even though these videos are from clearly marked parody accounts, Pattinson’s friends have been fooled into believing they’re real.
@unreal_robert When that very song came on🕺 #robertpattinson #pattinson #dance #song
Tiktok failed to load.

Enable 3rd party cookies or use another browser

MusicLM: Google’s ‘ChatGPT’ for music

Google’s watching as Microsoft/OpenAI cover themselves in ‘generative AI’ glory. But, as I’ve already discussed, there is NO way Google isn’t working on ‘stuff’ too. We know their version of ChatGPT is coming — but they’ve got more cooking.
They gave us a glimpse into their kitchen by revealing Music LM: an astonishing text-to-music generator. Music LM far outperforms older music-generating models like Open AI’s jukebox. It can do a few minutes of complex composition and high-fidelity songs.
While Google is not opening up Music LM for anyone to use, I infer two more significant trends from this story:
- The Tech Giants are jostling for supremacy in this space. I don’t doubt that Google is sitting on a lot of exciting research — but they have been less willing to open up their models, presumably to mitigate the risks that come with that. (Note that Stability AI has already been named in two major lawsuits?) Still, they want us to know — ‘we got stuff too.’ Well, it’s working. We are watching what you do next, Google.
- MusicLM’s capabilities are significant. Stable Diffusion and DALLE disrupted imagery. ChatGPT disrupted text. Music is coming next…. and my gut tells me that it will be controversial.
  Jim Fan @DrJimFan
  MusicLM by @GoogleAI, a hierarchical text-to-audio model that generates music at 24 kHz that remains consistent over several minutes. It relies on 3 key pre-trained modules: SoundStream, w2v-BERT, and MuLan. 1.1/
  4:51 PM ∙ Feb 2, 2023
  112Likes14Retweets

Content transparency: detecting and signing AI-made content

Those anecdotes about students ‘cheating’ on college papers using ChatGPT are not a good look — Open AI has revealed that it is working on a ChatGPT output detector.
It is still early days — their text classifier is better at detecting human-written content than that generated by Chat GPT. (It’s still only about 26% accurate at the latter.)
Detecting content made by AI is difficult. First, because it’s an adversarial game: as AI detectors get better, they’ll be outfoxed by AI generators. Second, the detectors can never be 100% accurate. Even a slight chance of a false-positive or false-negative result undermines them.
Still, AI-content detectors will probably develop as one approach to dealing with the abundance of AI-generated content.
A more promising approach involves the authentication or ‘signing’ of digital content — IE, being fully transparent about the origins of content (regardless of whether or not it is created with AI.) Generative AI companies can/and should do this: Resemble, an AI voice cloning company, just released its own ‘watermarking’ tool to ‘sign’ its AI-generated audio.
OpenAI @OpenAI
We’re developing a new tool to help distinguish between AI-written and human-written text. We’re releasing an initial version to collect feedback and hope to share improved methods in the future.
openai.comNew AI classifier for indicating AI-written textWe’re launching a classifier trained to distinguish between AI-written and human-written text. We’ve trained a classifier to distinguish between text written by a human and text written by AIs from a variety of providers. While it is impossible to reliably detect all AI-written text, we believe g…
6:10 PM ∙ Jan 31, 2023
8,495Likes1,838Retweets

Is Yan Le Cun right? Is ChatGPT ‘not that innovative’?

It’s a strong disagree from me on recent comments from Meta’s AI chief, Yan Le Cun, that Chat GPT is ‘not that innovative.’
He’s right that large language models have been around for a while, and that other companies (hello, Google and LaMDA) could have released a ChatGPT-style product before OpenAI. But the point is no one did. Open AI did.
Not only did OpenAI get there first, but oh boy, did the world respond. Apart from the 100m users in two months, ChatGPT secured Open AI an additional $10bn investment from Microsoft, and it got integrated into Microsoft Office.
I would argue that ChatGPT is far more than just ‘innovative’ — but that it marks a watershed moment in the history of human/machine collaboration.
Everyone wants some ChatGPT magic. China’s Google, Baidu, is building its own version. Anthropic and Character.ai are making ChatGPT competitors. And the biggest news of all? Google’s version of ChatGPT is coming. Yesterday, Google CEO Sundar Pichai told investors that Google will soon be revealing its “newest, most powerful language model as a companion to search, in experimental and innovative ways.” I guess all the chatter about ChatGPT being an existential threat to Google got under its skin, eh?
The company is also restructuring its ‘version’ of OpenAI, Deepmind (the British AI company), into Google. Translation: Google is going in big on (generative) AI.