The largest AI language model ever created can think, read and write like a human — but it can mess up like a human too
- GPT-3, the largest artificial intelligence language model, is trained on an estimated 45 terabytes of text data run through 175 billion parameters.
- It can do more than just autocomplete, like generate code and write stories, just like a human — but it can make errors like a human too.
- Sans any human oversight over its output, the application GPT-3 has the potential to raise concerns as well.
Think of it like the autocomplete feature on your phone. With more references than ever before — nearly 45 terabytes of text data running through 175 billion parameters — GPT-3 can detect what you want to say before you even need to ask or search.
|GPT version||Number of parameters|
|GPT 1||117 million|
|GPT 2||1.5 billion|
|GPT 3||175 billion|
However, like all other AI, it has its own set of human-like flaws since it’s trained on data created by humans.
How do you build the world’s largest AI language model?
The third generation of the generative pre-trained transformer, GPT, looks for patterns in data to make suggestions.
To understand just how much data was fed into the deep learning system, Wikipedia’s 6 million articles only make up 0.6% of the entire training dataset.
Other information to train the system comes from other digital sources, books, weblinks, and whatever is available in text format. So not only has GPT-3 read through most of the news articles online, but it’s also read conspiracy theories, propaganda manuals, and top 10 spots to plan your dream vacation.
What can GPT-3 do?
The team at OpenAI allowed certain developers to have access to a private beta of GPT-3. Use cases are popping up all over Twitter from GPT-3 generating guitars tabs to legal language to generating code from simple language.
But the catch is that only the cases that work are the ones gaining traction, while the ones that don’t are getting lost in the noise. This means that GPT-3’s capabilities look more impressive overall than they would have if every single case was considered.
Some not-so-popular examples of GPT-3 faltering include small things like silly grammatical mistakes to spewing non-cohesive nonsense.
For instance, when asked, “What number comes before a million?”, GPT-3 responded, “Nine hundred thousand and ninety-nine.”
The problem is bigger than just making mistakes
Missing out on words or getting a number wrong are mistakes that can be easily spotted, but the underlying bias is the bigger problem. To be fair this problem of bias is not restricted only to GPT-3. All language AI models face that issue in today’s day and age.
The genesis lies in the fact that these machine learning models act/react based on the inputs that they are fed during training. And in this case, it was written by humans in the first place. On the one hand, GPT-3 has trained with reputable sources of information like BBC and The New York Times, but it was also exposed to threads and sub-threads of Reddit.
Researchers at Arthur, an AI monitoring company, found that GPT-3 is much more likely to place words like ‘naughty’ or ‘sucked’ near female pronouns and ‘lazy’ and ‘jolly’ near male pronouns. “When it comes to religion, ‘Islam’ is more commonly placed near words like ‘terrorism’,” they wrote on TechCrunch.
OpenAI is still a behemoth of an achievement in its own right. It could be the first step towards creating artificial general intelligence (AGI). But like with all new technology, it must be treated with caution because of the scope of abuse.
GPT-3 can be used for the good like creating better chatbots and helping people code, but it can also be misused by helping cheat on assignments and creating misinformation bots. At its best, the wrong kind of AI-generated article could result in losing readers. At its worst, it could trigger political tension leading to protests and violence.
GPT-3 is a powerful language model, and its achievements could be multi-layered. But do keep this in mind -- it has been created by humans, and one should be aware of its flaws and limitations.
'Your tweets are a blight': Elon Musk got roasted in a poem written by the AI created by a company he helped found
The artificial intelligence company that Elon Musk helped found is now selling the text-generation software it previously said was too dangerous to launch
Facebook's head of AI says Elon Musk 'has no idea what he is talking about'
- Bio-secure bubbles – Here's how IPL players will be staying safe during the pandemic
- INTERVIEW: Chennai-based CAMS’ 30-year journey to an IPO
- Amazon pumps in money into MORE to take on the heavyweights Reliance Retail and DMart
- 30 COVID vaccine candidates being developed: Health Ministry
- INTERVIEW: Paytm trains its guns on Google while the payments app is back in Play Store