If you think that OpenAI ChatGPT is a Google killer, think again.
As Google engineer, Blake Lemoine points out, Google is already ahead of the game in the AI industry and utilizes a more advanced AI dialogue system behind the scenes.
Today, I’ll explain how Google is about to smoke Chad GPT and all the hype around it with their powerful AI language models; Sparrow, LaMDA, and PaLM.
Google has not just one, not two, but three amazing language models on the way and a bunch of other AI models that are state-of-the-art in their respective fields like code and image generators.
Google is about to kill Chad GPT and all the hype around it.
A few weeks ago, Sundar Pichai, the CEO of Google, issued a code red over Chad GPT’s impact on Google’s search engine business and asked for an army of AI developers to roll out products as soon as possible.
Chad GPT just provoked the beast. Google is not going to sit around and watch people calling their search engine outdated, while Google’s AI model is actually several years ahead of OpenAI’s GPT models.
Fun fact, GPT-3 itself is based on Google’s transformer architecture that they introduced almost 6 years ago.
If you thought for a second that Google is the one in trouble in the age of AI, let me just tell you; Google’s AI is about to come out after more than a decade of research and development.
And when they come out, the world will never be the same again.
As a starter, Google has a chatbot named Sparrow, which is really similar to Chat GPT.
Sparrow is the weakest of the 3 language models from Google. It’s based on the Chinchilla model. Chinchilla is only 70 billion parameters, but because of Google’s advanced training techniques, it is outperforming GPT-3 with 170 billion parameters on every possible benchmark test.
Google Sparrow is a fine-tuned version of Chinchilla, just like Chat GPT is a fine-tuned version of GPT-3.5.
They are almost on the same level, except Sparrow has the advantage over Chat GPT of being trained to deliver evidence for its answers and provide links when users ask for them.
Only this seemingly small difference is a huge win for Sparrow because not knowing how to validate Chat GPT answers is holding a lot of users back.
The Sparrow model is a work by DeepMind, one of the subsidiaries of Alphabet. They have been at the forefront of research for more than a decade now.
DeepMind is the same company that made AlphaGo, Alphazero, and Alphafold to change the entire AI research landscape each time.
Even Science magazine’s Top Ten Breakthroughs of 2022 featured two research products from DeepMind.
AlphaCode – the best programming AI in the world, which was above the median in competitive programming against challenges created by Codeforce.
AlphaTensor – an AI for efficient matrix multiplication in another world, an AI which is going to make other AI more efficient.
Another step towards AI making a better AI infinite loop.
Sure, Sparrow is great, it’s on the same level as Chat GPT, maybe a little better or worse, but that’s not Google’s best AI model yet. They have a lot more to offer.
Do you remember Blake Lemoine, the Google engineer who was fired from Google because he claimed that LaMDA, Google’s state-of-the-art chatbot, is sentient?
Blake is not a regular person. Heck, he’s not even a regular Google engineer or even a regular AI developer.
He’s an AI expert that was specifically working on LaMDA to push it to its limits.
This guy’s everyday job is to work with the most advanced AI models. If LaMDA can convince him that regardless of whether that is true or not, this thing is going to blow an ordinary person’s mind.
Now, right after Chad GPT’s release Blake Lemoine tweeted this:
“I love how people are talking about how Chat GPT is going to replace Google, while Google just quietly uses a more advanced dialogue system behind the scenes. Chat GPT is a big step in the right direction but there are still a few years behind Google.”
As Blake said in his interview on the H3 podcast, a language model like Chad GPT is just a module inside LaMDA.
Meena is the language model inside LaMDA, which is really similar to Chad GPT. But LaMDA has a lot of other modules.
It might sound farfetched, but I’ve been working in the SEO space for years now. And in the SEO community is not even a question that Google crawlers can see the web page almost like a human.
It’s not just reading the text, it can actually see the style of the page.
And if you have a video, Google can identify the subject of the video as you see it recommends even portions of it for some queries.
That means AI can understand the content of the video at least to some degree.
Now it’s just a matter of feeding this information to LaMDA. It’s like these crawlers are the eyes and ears of LaMDA.
Plus, an entire community of SEO experts is trying their best to help Google understand the content of the page.
They use HTML tags, provide accurate descriptions of every video, and use markups and schemas so they can rank better.
But they are indirectly providing clean data to LaMDA just the way you can understand it.
Other than that, LaMDA has been trained on dialogue, which makes it a lot more persuasive.
If you compare their answers, LaMDA’s answers are simple, short, and very much on point.
If you are wondering why LaMDAa’s test pitching is not as impressive as it should be, it’s because “it’s a dramatically restricted version of LaMDA.“
And I would explain why Google had restricted so much in a moment.
If you are not blown away already, let me tell you that even LaMDA is not Google’s most advanced AI model.
Google’s best model, which was upgraded almost seven months ago, is PaLM, probably the most advanced AI model in the world as far as we know.
It’s a 540 billion parameters model that was trained on more than 6000 TPU chips, and its training data consists of:
This model is multilingual and of course, it can do quote completion.
It outperformed all the previous AI models such as GPT-3, Megatron, Turing, Chinchilla, and even LaMDA on every major.
And that’s not all…
Google has a collection of really tricky tasks called “Beyond The imitation Game Benchmark” or BIG-Bench in short.
It includes really difficult challenges like delivering persuasive arguments for false statements or finding a Russian proverb, which is the closest in meaning to the provided English proverb.
In addition, Google PaLM managed to surpass the average human success rate in solving the most challenging intelligence-based tasks.
All of that is just one battlefront. Google has a lot more advantages, which I’ll explain in a second.
But why doesn’t Google release all these amazing products though?
As you might have seen, people like to force AI to say bad things.
Unlike OpenAI, Google as the biggest brand in the information industry can’t afford to herd its brand even a little bit.
Second, they are already making billions of dollars from search advertising. They have an established form of monetization and they didn’t have any incentive to disrupt themselves.
Even if they took over the entire search industry in a year and had a billion customers, by now they more or less would be in the same place because they are in the search industry.
They could be even worse off as it’s not clear that monetizing AI chatbots is as profitable as monetizing search results.
And last but not least, the cost of cloud computing is too high. Open AI is giving access to Chat GPT and is handling a few million people. But if you are on the same level as Google, you are thinking about handling requests for billions of people, which requires a whole new level of computational resources.
Probably without Chat GPT, we would have to wait a lot more to see any serious AI product from Google. But now Google has to do something.
That’s why Andrej Karpathy, the former director of AI at Tesla, tweeted this right after Chad GPT was released:
Google is afraid of herding its reputation more than anything. But as Blake Lemoine said, if tragedy or some other product ever become a real threat, they just bite the bullet and release LaMDA, which would smoke Chat GPT.
Even if we throw all that away and assume Google has a chatbot on the same level or even slightly worse than Chat GPT, as you know, the cost of running these chatbots is not trivial.
And making them a premium product would hugely impact their potential.
So it seems like ads are the way to go and Google as the biggest ad agency in the world has a huge advantage here.
They have all the connections, all the customers, all the data, and all the experience to make it work.
OpenAI is not even in the picture when it comes to ads. They are not going to start an ad agency from the ground up. That’s not a trivial task.
Google’s ability to track users across the web, gather data, and identify users’ demands is a product of years of development and having abroad access through a wide range of mostly free products that are just gathering data.
It’s not something that you can build in 2 or 3 years.
More than that, Chat GPT as a chatbot is great, but no one wants to open the browser, log in to Open AI, and use Chat GPT for every little question every time.
Google is the most visited website on the planet. It’s so popular that it has become a verb itself.
“Google it” is the equivalent of “search on the web for it”.
That default destination for asking questions is not going to easily change.
Keep in mind that Google itself is just one product and they have more to offer.
If a question just popped into your mind, the Android operating system and Google Assistant are here. You can ask whatever you want.
Do you want to write an email or do you want to write an automatic response to emails that you are receiving?
The AI inside Gmail would help you out.
Are you writing an essay or making a presentation for your class?
The AI inside Google Workspace will help you with the research and will format the page for you.
You’re reading an article on the web and you want an explanation or translation?
No problem. Chrome’s AI assistant got you covered and tons of other already established Google products that would immensely benefit from AI.
To be completely fair though, OpenAI can license their AI to other companies as well.
As I explained, Microsoft is already working with OpenAI to incorporate AI in their products but is not the same as having your own basket of products.
At the end of the day, it’s not even a fair comparison. OpenAI has less than 400 employees, DeepMind alone has about 1500, and Google Alphabet’s total number of AI researchers and AI developers is about 20,000 people.
When you consider the economy to be reached the square, Google has a lot to lose and nothing to gain because they already have a monopoly on search.
But if you view the economy as a pie that is about to double or triple in size with the help of artificial intelligence, Google is one of those companies that will take a big chunk from it.
They can expect Google 2023 keynote to introduce new products or upgrades as soon as May.
If the age of AI is actually here, Google is not in danger. Google is the danger.