The First Big AI Failure Just Took Place. About Time.
When Huge Capital and No Real Product Converge
It finally happened. We have our first real AI company debacle.
Inflection, a company invested in by the likes of Bill Gates, Eric Schmidt, and Microsoft, among others, has become the first Generative AI relevant company to get flushed down the toilet.
Their most important product was Pi, a ChatGPT rival that focused on being a friendly and ever-supportive AI companion.
Now, they are being dismantled by Microsoft and it’s becoming an AI studio, a terrible ‘end’ for a company that had raised an astonishing $1.3 billion last year alone.
But this is also a stark reminder that the AI space is at complete risk of consolidation around the main incumbents, to the point that most AI developments one day could be traced down to a single set of two or three players.
Therefore, with news like this, there’s much more at stake than meets the eye.
This insight and others have mostly been previously shared in my weekly newsletter, TheTechOasis.
If you want to be up-to-date with the frenetic world of AI while also feeling inspired to take action or, at the very least, to be well-prepared for the future ahead of us, this is for you.
🏝Subscribe below🏝
We Reached the End of Pi
Let me start by saying this, I didn’t call this debacle, I liked Pi, so I am as surprised as anyone on this.
But what did actually happen?
An Inflection Point
Since the dawn of ChatGPT, multiple companies have rushed into the arena to build a competitor and lure some money and customers.
One such company is Inflection, founded by Mustafa Suleyman, co-founder of Deepmind (now part of Google), Karén Simonyan, creator of AlphaZero, and Reid Hoffman, co-founder of LinkedIn.
Pretty strong roster if you ask me. And investors thought the same, as they rushed in as soon as they became a company.
The mission was simple, creating an LLM not focused on answering humanity’s most complex questions, but simply being a companion to humans. A model that would endlessly listen and support you all the way.
This concept came to fruition with Pi, an LLM that was unequivocally much nicer, more empathetic, and more supportive than ChatGPT, Gemini, or Claude.
Try Pi here.
The experience of talking to Pi is actually pretty nice if you ask me, so I thought that this thing could find its customer base and flourish into one of the main AI companions of our time.
Turns out, I was very wrong.
Until Something Sticks
When big corporations and venture capital firms have too much cash they can bear, they will basically throw their money by firing from the waist until something sticks.
And here came Microsoft
In other words, they invest on the premise that ‘for every 10 investments I make, 9 will fail but the outsized returns of the one that succeeds will cover the expenses of the other 9’.
It’s no surprise that investors and companies alike will rally when three reasonably seasoned founders get together and launch a company.
And if one of the co-founders serves on the board of Microsoft, it’s no surprise either that Microsoft led the investment round.
But, as it turns out, what sounded like a great idea was actually not that good. Even though Inflection had money and a good team, their product really had a tough time following up with the ‘big guys’.
Now, a few months later, Microsoft’s CEO Satya Nadella announced that two out of the three co-founders, Mustafa and Karén, would join Microsoft to create Microsoft AI, a new area in the company that would focus on improving and serving customer-end AI systems like Copilot or Bing to Microsoft customers.
In turn, Inflection has now become an ‘AI studio’, as acknowledged by the company, meaning that it will no longer focus on creating Pi or other AI companions, but a platform for custom LLM training, fine-tuning, and testing, with “no immediate changes to the service” in the foreseeable future, meaning that they are not scraping Pi, as they will serve it through an API, but almost.
But why did Pi fail, and more importantly, what does it mean for us all?
A Rich People Game
Although we don’t know for sure what went wrong, it’s obvious: at the end of the day, it was just another product looking for a problem to solve instead of actually solving one.
In the case of Inflection’s Pi, it seems nobody cared about it, plain and simple. In fact, when looking at a16z’s recent update on the top-100 generative AI products, Pi is nowhere to be seen.
I mean, not one single mention, literally.
And if we consider that Pi’s product range, AI companions, are without a doubt the stickiest GenAI products today, with companies like Character.ai having enormous traction, even more so than productivity-based tools like ChatGPT, something’s off with Pi.
Maybe, it simply wasn’t good enough.
Additionally, one of the areas where Pi was lagging was the context window, one of the most critical aspects an AI companion must have.
What is the context window?
The context window is the maximum sequence length a virtual assistant can work with at any given time. It’s like its workspace. Thus, for AI companions that need to learn over time of your experiences and stories, this context window must be huge.
In Pi’s case, from my own experience, describing its context window as ‘poor’ was an understatement.
In fact, with their latest model release, Inflection 2.5, they didn’t even disclose its context window, which tells you all you need to know.
I asked the actual LLM and it responded that it had a 4,096 token window. That is inexcusably short by today’s standards.
And the real problem here is that, based on the recent innovations that are driving longer context windows, like Ring Attention, money rules, as the length of your maximum sequence is proportional to the amount of GPUs you have at your disposal.
Also, to create a different product you need unique data, which has also turned AI training into a rich people’s game, as providers are licensing data by paying fortunes to providers like Reddit.
Thus, if the product doesn’t stick, and it also needs loads of money to compete, it’s no surprise that Microsoft has lost faith and decided to unofficially scrap the company and absorb most of its talent.
Now, as 2023 closed with up to $50 billion invested in AI start-ups, this event is simply the start of a larger trend that many people feared:
Consolidation.
AI and open-source
In my view, this event, besides being a wake-up call for companies to see what happens after raising billions from industry incumbents while being incapable of getting traction, is also a wake-up call for all of us.
In fact, other companies like Anthropic, Cohere, et al could soon share the same fate and be de facto absorbed by their large investors if the stretch between ChatGPT and Gemini starts to grow larger and larger.
Adding insult to injury, it’s not only about capital and hardware consolidating, it’s also talent.
Besides some honorable mention at some universities, most AI talent is already part of OpenAI, Microsoft, Google, Meta, Nvidia, Tesla, Apple, or Amazon’s rosters.
And the few remnants that remain are mostly already invested by some of the ones above, such as Mistral by Microsoft, or Figure AI by Microsoft, Intel, or NVIDIA.
If the trend continues, soon, most AI breakthroughs will take place in the labs of the aforementioned companies.
It must be pointed out that companies like Meta or Apple are playing the open-source game, but that could always change.
And when thinking about OpenAI, Microsoft, or Google, there’s no way they switch to open-source, it’s too late by now as there is too much at stake.
But when dealing with such a powerful technology as AI, consolidation is never a good thing. Moreover, AI isn’t only consolidating at the company level, but also geographically, as most, if not all of such companies are US-based.
Seeing this could induce other nations into speedrunning their efforts and turn the entire space into a mindless race — or war — if it isn’t that way already.
But as AI is poised to disrupt every industry and job known to man, having less than 10 companies — all for-profit entities, naturally — having so much weight and power over it can’t be good, because if the trend continues, its benefits won’t look out for our society, but to the benefit of those companies’ shareholders.
For those reasons, Inflection’s failure is terrible news, and just simply clarifies what was latent but obvious all along, cutting-edge AI and talent are more and more constrained to a smaller set of corporations, so Meta’s and other companies’ role in democratizing AI is as important as ever.
On a final note, if you have enjoyed this article, I share similar thoughts in a more comprehensive and simplified manner for free on my LinkedIn.
If preferable, you can connect with me through X.
Looking forward to connecting with you.
