Losing Ian Goodfellow to DeepMind is the dumbest thing Apple’s ever done
There are a lot of machine learning developers in the world. But only one has been mentored by both Andrew Ng and Yoshua Bengio, invented a new kind of artificial neural network, contributed to or led research at Google Brain, OpenAI and Apple, and still has fewer than 40 candles to blow out on their birthday cake.
And Apple just let him walk out the door and right into Google’s offices where he’ll soon be working for the DeepMind research team.
His name is Ian Goodfellow. And letting him walk away from your California-based, Silicon Valley elite, ultra-progressive technology company because he disagrees with your in-office work demands is a scenario that’s so dumb I can’t believe GPT-3 didn’t come up with it.
To put this into a sports analogy, this is like letting Tom Brady or Michael Jordan leave your team over a disagreement between them and the team owner on how towels should be folded.
Let Goodfellow and his team work from wherever they want. If they think they can code a better machine learning model from the International Space Station, you should probably look into building a rocket.
Let’s back up a second and put things into perspective. I can sense our C-suite readers reflexively prepping a “good of the company” or “everyone is a rockstar at our organization” speech in their head in response to this opinion piece.
I don’t think anyone deserves special treatment at Apple. Or any other company, for that matter.
But Ian Goodfellow’s contributions to the field of machine learning cannot be overstated. Office hours are a pretty silly thing to lose any talented developer over, and it’s even more ridiculous to let your ML director walk because you think in-person smiles are important.
Losing Goodfellow is a huge blow to Apple for two reasons:
- His talent can’t be easily replaced
- His work at DeepMind could put him in direct competition with Apple
Goodfellow’s greatest claim to scientific fame is as a part of the team that invented the generative adversarial network (GAN).
A GAN is a neural network that learns how to create content by trying to fool itself and, eventually, humans.
Any time you hear about an AI that can generate text, write poetry, create images, or produce its own original music, you’re almost certainly hearing about a GAN.
The cool thing about GANs is that they work by pitting two neural networks against each other. Without GANs, humans would have to fine-tune every generative iteration — like trying to do coarse sanding with fine paper.
But a GAN has one network that’s creating and another that’s discriminating. The second network is essentially a bouncer that blocks a lot of useless output before it even has a chance to manifest.
Artificial general intelligence (AGI)
After Goodfellow et al. invented the GAN, he continued his work at Google Brain where he helped solve computer vision and ML security issues. And from there he did a tour at OpenAI, the Elon Musk and Microsoft-funded AGI think tank where some of the world’s brightest minds try to figure out how to invent and control a human-level artificial intelligence.
This is important to note because there’s really only one other company on the planet as deeply invested in AGI and full of well-known talent as OpenAI, and that’s DeepMind.
When Ian Goodfellow became the director of machine learning at Apple, many of us in tech journalism were surprised. It seemed like a huge loss for the Google Brain team, but it made sense for Goodfellow (it seemed like a well-deserved promotion) and, based on what we know about Apple’s AI programs, it didn’t seem like something that would come back to bite Google in the butt.
Fast-forward to today, and Google looks like the smartest player on the board as it re-welcomes Goodfellow to the fold.
DeepMind is, basically, Google’s version of OpenAI. Where OpenAI seems to be a bit more focused on ensuring an AGI doesn’t rise up against us, DeepMind is more fixated on creating a generalist AI that can do anything a human can do without needing to be retrained over and over again to gain new abilities.
This is something that could come back to bite Apple in the butt.
Siri, Siri, wherefore art thou Siri?
About five years ago, if you wanted to make a joke about virtual assistants or anthropomorphic AI, you had to invoke Siri. That’s the only named AI that all of my readers were familiar with in 2017.
Now, I’m better off using Alexa for name recognition. But nobody’s forgotten about Siri. At least not yet.
DeepMind is on to something huge with its new GATO AI system. No, I don’t think it’s on the path to AGI with GATO (or, really, anything else it’s currently doing, but that’s a debate for a different article).
But I do think that GATO could be extremely marketable if DeepMind can overcome the problem of giant-scale models and bias.
Imagine Siri, but a version of Siri that could do a thousand different tasks on your behalf. Right now, our virtual assistants essentially do web searches and open apps for us. It might feel like Siri can do hundreds of different things, but telling you what time it is, how many messages you have, and what the capital of Nebraska is, are all pretty much the same task.
I’m talking about a version of Siri that could control a robot capable of washing your dishes, while simultaneously identifying high-weed areas in your front lawn, while also generating a completely original cartoon for your kids to watch based on your specific prompts, and so on and so forth.
Currently, it would be an impressive feat for a team of AI devs to create a system that could do all of that in a simulated environment. The challenge of letting a generalist AI loose in the homes of random consumers is far more immense.
But what if DeepMind pulls it off? What if, instead of Siri or Alexa, it’s Google Assistant that becomes the world’s first AI assistant capable of actually assisting in your day-to-day life?
If DeepMind and Google manage to turn the played-out, boring, 2-D idea of what a virtual assistant is into something that could feasibly start to look like a real life-assistant, everyone’s going to forget about Siri. And Alexa. And any other “assistant” that can’t do what GATO can.
I’m not so sure DeepMind can pull it off, but I am sure that the odds increased by a large margin the moment the company inked a contract with the GANfather himself.
It’s your business, Tim
At the end of the day, who knows what really went down at Apple. Maybe Goodfellow wasn’t happy, or maybe Apple wasn’t.
There’s no guarantee that DeepMind’s work will ever interfere with what Apple’s trying to accomplish, though most of what everyone’s trying to accomplish in the field deals with some of Goodfellow’s ideas on deep learning.
And it’s also worth mentioning that big tech poaches talent from each other all the time. Let’s not forget that Goodfellow left Google twice, once to join OpenAI and the second time to join Apple.
But the timing of his joining the DeepMind team is quite exciting. He’s reportedly onboard as an independent researcher. That sounds a lot like he’ll be given anything and everything he needs to do his best work.
Maybe Apple CEO Tim Cook has good reasons for letting his star quarterback leave to join a rival team in the middle of the playoffs. It’s hard to see from our vantage point outside of the Cupertino company’s walled gardens, but it’s possible.
No matter what, it’s an exciting time for the field of AGI research. There’s no telling what Goodfellow and the DeepMind team can accomplish together.
Make sure to sign up for the Neural newsletter so you don’t miss a thing.