Why the rush for Artificial General Intelligence (AGI)?
Welcome to the fast-paced world of generative AI.
Some of us are still attempting to fully comprehend the phenomenon known as generative artificial intelligence (gen-AI), which gained significant traction in November 2022 with the introduction of ChatGPT.
Since then, we have seen job losses, a chip company’s valuation reaching USD 2 trillion, and a messy corporate upheaval in public glare, while new prophets of this age on social media keep warning people that if they don’t learn prompt engineering, they will soon suffer the fate of the dinosaurs.
When people are both thrilled and terrified at the same time, it doesn’t say much about the path that lies ahead.
While the ink is still relatively wet on generative AI, a few firms have already flipped the page to Artificial General Intelligence(AGI).
Let’s briefly retrace the steps.
How did we get here?
Large Language models have enabled computers to react in a manner similar to that of humans and, in some cases, even to converse like them. OpenAI trained ChatGPT, a foundational AI model, by scraping billions of bytes of data from public websites, leading to an ongoing lawsuit against them. Next, this trained model was fine-tuned by people sitting in Kenya for a period of time by prompting the model with human like questions in a process called reinforcement learning with human feedback (RLHF) until it developed the ability to converse with humans.
Since then, the floodgates have opened, and other foundational models from other companies, such as Gemini from Google, Llama I and II from Meta, and a whole ocean of foundational models, have sprouted up on platforms such as HuggingFace.
Countless models are being added daily for language translation, text-to-image, and now text-to-video as well.
In fact, some individuals spend countless hours reviewing these models on HuggingFace daily.
The ecosystem is now on auto-pilot.
So what is AGI then?
Artificial General Intelligence (AGI) refers to the ability of artificially intelligent systems to breach levels of human cognizance.
Human-level cognizance the key here.
In order to comprehend this, we must recognize that humankind’s greatest ability has led to our planet’s dominance. After all, a million years ago, we made it through the savanna among larger animals. When early humans discovered fire and mastered its use, they used it to defend themselves against more formidable animals, and eventually subdued them. We humans learn from the experiences around us, and then we have the ability to master things. Our built-in human-level cognitive systems have been extremely helpful.
Presently, advanced AI systems cannot do this. They lack the in-built cognition that we have. Even unsupervised learning systems, which sound like they can do it alone, still need human intervention.
Take, for instance, self-driving car crashes. To make the point here, I will use one of the first ever self-driving car crashes by a Tesla back in 2016.
In this case, the self-driving vehicle was traveling along the highway without incident until a trailer crossed its path. Sunlight reflecting across the trailer’s body at that moment blinded the car’s front cameras and sensors, obscuring their view of the trailer in front. The vehicle struck the trailer head-on, driving as if there was nothing ahead of it.
After retraining the model, the vehicle’s software was updated to handle cases like this.
Point: The self-driving car will not be able to learn by itself here, unless the model that it uses for object detection is fed cases again with these scenarios and trained on what the reaction should be. A lot of human intervention is necessary for this, even if you may not want to believe it.
Complicated cases require more data for training. It looks like we may need some more time to get to the point where self-driving cars can drive like humans.
Why the rush?
Following the fallout from the OpenAI debacle in November of last year, Microsoft CEO Satya Nadella met with Sam Altman and declared that the two would collaborate on AGI. The board of OpenAI felt very uneasy about the company’s rapid push towards artificial intelligence, sparking Ilya Sutskever’s coup.
The most ridiculous outcome from this whole OpenAI debacle was that instead of treating this as a wake-up call for guardrails on GenAI, both OpenAI and Microsoft have doubled down their efforts on AGI and are going full speed ahead.
Now, a nuclear-like race to create the first artificial general intelligence has begun. Similar to the Manhattan Project, which required enormous tracts of land and specialized infrastructure in order to produce the first atomic weapon, hyper-scalers like Meta, Google, Microsoft, and others are already spending millions of dollars to establish server farms that will house several GPU racks.
Why do you think NVIDIA’s stock price is going through the roof?
So, who’s building AGI
Until now, the following giants have joined the race to build AGI: -
- Meta: — Meta recently announced that they’re going to build open-source AGI. They have until now been able to seize a sizeable portion of the GenAI developer community with their open-sourced Llama models. In fact, Meta is scaling up its data centers specifically for this.
- Google Deepmind: — Google and their subsidiaries have been actively involved in AI for an extended period. After all, the transformer architecture has been their biggest contribution to genAI until now. In fact, they have been doing it for so long that they have come up with a definition of the six levels of AGI.
- Microsoft & OpenAI: — Since their GenAI strategies are interlinked, I intentionally grouped these two together. Microsoft has invested a significant amount in OpenAI, which serves as the foundation for their OpenAI services on the Azure Public Cloud. Microsoft’s investment will continue to flow towards quality research being done by Altman and his team.
- Anthropic: — This is an interesting breakaway from OpenAI in 2021 from two senior executives who felt that OpenAI was not prioritizing safety and governance appropriately. Their management structure has been completely different than OpenAI, and they claim to be doing AI development a lot more responsibly..
The aforementioned are just a few of the companies rushing headlong into building the first computer with a human-like conscience.
However, there are two schools of thought that prevail about getting to AGI.
The skeptics: — AI veterans like Yann Le’Cunn have openly said that a completely newer approach is needed to get to AGI. The fact that the transformer architecture has deceptively gotten us to human-like speech should not fool us into believing that it will take us all the way there.
The Brute-Force optimists: — Then there is this other group that thinks that to get to AGI, all they need to do is keep training foundational models with more and more data and parameters, until they somehow turn human!!
So, who will get there first, the skeptics or the optimists?
