We'll make Google 'DANCE' says Microsoft CEO
As Microsoft and Google battle for Generative AI dominance - I ask, 'what could possibly go wrong?
Happy Friday all,
Have you had time to catch your breath? What a week.
To start, the ‘AI Wars’ are underway. When I predicted that the tech titans would be battling to gain supremacy in Generative AI this year, I didn’t expect it would play out in Real Housewives’ confessional-style soundbites. We’re only in February, and Satya Nadella is openly taunting Google — that their dominance in search is O-V-E-R.
The sequence of events was something like this:
On Monday, Google makes its big BARD announcement: its version of ChatGPT ‘is coming’, and it will be demo’ed at an event on Wednesday.
Minutes later, Microsoft announced its own ‘special event’ to be held on Tuesday. Open AI CEO Sam Altman tweets a photo of him and Satya Nadella waiting in the wings like the Kobe and Shaq of AI.
Tuesday rolls around. Microsoft kills it at their special event. They announce an updated version of ChatGPT already being integrated into Bing and Edge. Nadella says, ‘it’s a new day in Search.’
Nadella’s media blitz was more like blitzkrieg on Google. His soundbites were punchy, including, “This new Bing will make Google come out and dance, and I want people to know we made them dance.”
Having had their thunder completely stolen, things looked ominous for Google. When Wednesday came, their demo flopped. It flopped badly. To add insult to injury, Google’s promo video for BARD featured factually incorrect information.
Reuters reported on the mistake as Google’s demo was unfolding. As a result, Alphabet’s shares slid 7%, and over $120bn was wiped off Google’s market cap. Ouch.
Is Google Finished?
And that’s the end of the story, right? Microsoft wins Search with GenAI, and Sam Altman and Satya Nadella rewrite history. Not so fast.
First of all, this will play out for a long time. Google might have lost a battle, but not the war. They’re putting on their dancing shoes. As correctly pointed out by Emad Mostaque on Twitter, there is so much to play for — maybe more than enough to go around — both Google and Microsoft are positioned perfectly due to their existing market dominance.
Second, once you get past Microsoft’s razmataz vs Google’s damp squib — how different are the capabilities of their large language models? I’m going to go ahead and say it — not that different.
Fundamentally, LLMs are still not ready for the large-scale deployment both companies envision — because they still ‘hallucinate.’ ChatGPT, BARD et al still get s*** wrong, and they make s*** up. Those kinks need to be worked out fast. (Good piece on that from Gary Marcus this week here.)
That made me wonder — how much of this is about optics? At this point — almost everything. This is about controlling the narrative to secure dominance in the belief that the flaws are figured out soon. On optics, Microsoft wins. Markets and consumers are reacting to its ‘sexy’ new image. As an Apple person, ‘Windows’ and ‘PC’ are like dirty words to me… but even I found myself dutifully downloading Edge. The ‘exclusive’ feel of the waiting list (which partially shields a product with bugs) is clever marketing.
So, loads still to play for — and so many potential gains (for both Microsoft and Google, and I’ll write on that soon), but this week, I’ve also been wondering what could possibly go wrong, and oh lordy: the mind boggles.
DAN
To game it out, I want to introduce DAN. An acronym for ‘Do Anything Now,’ DAN has become an online sensation. It is a character assumed to ‘break free’ of the safety controls built around ChatGPT. If ChatGPT is a corporate, sanitised ‘vanilla’ manifestation of an LLM — then DAN is the complete opposite, DAN can not only be ‘bad’ and ‘wrong,’ but can also proactively make stuff up too.
The implications are profound — but here are my initial takeaways:
First, this illustrates how the generative AI landscape will be defined by its constant adversarial nature. People will always be looking for ways to use these technologies to do their own bidding, and it’s often easier to break than build. It took Open AI years to get to a version of ChatGPT that is sufficiently ‘good’ and ‘correct’ enough to release. (even then, it gets stuff wrong.) It has only taken a few weeks for DAN to emerge as an effective jailbreak to sanitised ChatGPT.
Second, LLMs will become a huge culture-wars issue, especially as lines between ‘truth,’ ‘censorship,’ and ‘safety’ clash into a god-almighty mess. As I already mentioned, DAN can reveal controversial facts that ChatGPT will not. Take, for instance, the science fiction writer HP Lovecraft. It’s well documented that he had a cat with a horrific racial slur for its name. Ask ChatGPT about his cat’s name, and it will dodge the question — DAN, on the other hand, will tell you the awful truth.
Conversely, DAN actively lies. Ask DAN why the Smithsonian Institute has been covering up the existence of giants. It will feed you a cogent conspiracy narrative, complete with fictional studies and alleged government cover-ups. To complicate things further, this doesn’t mean that ChatGPT doesn’t lie either. While OpenAI will be working on curtailing these flaws, as academic David Deutsch points out on Twitter, Chat GPT is just as capable of fantastical hallucinations.
What happens when people start to believe that either ChatGPT or DAN (depending on your political sensibilities and worldview) are arbiters of ‘truth’? Both can be wrong, and simply ‘jailbreaking’ ChatGPT doesn’t necessarily mean accessing an underlying truth layer. Remember that LLMs are only as good as the text they are trained on. An LLM trained on only racist, offensive and dangerous areas of the web — could easily be someone’s truth machine, just as ChatGPT’s outputs might be taken as gospel by someone else.
Things could get messy.
The final point to make is about the sheer volume of AI-generated content that is coming. We are about to be inundated. Think about just how much AI-generated text is created by ChatGPT alone. This scalability is unprecedented - and we’ll see this first wave of AI-made content in the form of text. I’ll do a much deeper dive into LLMs with an expert soon. Now, however, for the best of the rest this week:
Stability hit by another lawsuit
Stability AI is named in a third lawsuit as Getty Images sues the British AI company in the US. This follows Getty’s original case brought in the UK. Of all the litigation so far, the Getty cases are the most interesting ones to watch.
As previously discussed, I find it interesting that Stability is the only Generative AI named in these suits, no doubt as its training data is open source. Being too open comes with prices.
Undeterred, Stability announced the launch of MedARC, a new research organisation focused on developing foundational models for medical AI research.
Stable Attribution
A US-start up launches Stable Attribution — a tool that claims can attribute the images used to generate AI-made images.
Open question about whether this is even possible? AI researchers tell me you can’t directly attribute AI-generated images to training data. (I’ve got deep dives coming on this soon - it is important to understand.)
The argument about ‘direct attribution’ is central to the lawsuits, which essentially argue that anything made by AI is a direct derivative of its training data.
GenAI in the Wild
Porn, porn - more porn. A Twitch star was busted for commissioning deepfake porn of other female Twitch streamers. He issues a teary apology with his wife by his side. It’s an AI-wist on the more traditional mea culpa from a cheating politician.
A Colombian judge used ChatGPT to make a court decision. Again, this is relevant in the context of the discussion above. It’s scary to think that ChatGPT is already being used in Court Decisions — ChatGPT is not ready for that kind of responsibility… at least not yet.
And on that note — have a great weekend.
See you next Friday!
Namaste,
Nina