After Years of Promises, NLP is Ready to Prove Doubters Wrong
For more than a hundred years, people tried and failed to bring electric cars to the global automobile market. As recently as the last decade, people said it couldn’t be done. The doubters cited a laundry list of barriers that would prevent widespread adoption, from insufficient technology and infrastructure to uncertain consumer demand.
But in the early 2000s, one of these barriers abruptly fell. The popularity of laptops and eventually smartphones drove rapid innovation in lithium-ion batteries.
Around this time, Tesla was founded to finally crack the electric vehicle code. Musk and Co. recognized that lithium-ion batteries were a huge step forward in portable energy and moved quickly to exploit the innovation for cars. In doing so, the Tesla team has ushered in a new era of the automobile industry, where the mainstream adoption of electric vehicles is no longer impossible, but inevitable.
I know I’m not breaking any news with an anecdote about the importance of timing to the success of new products and startups. Instead, I am here to draw parallels to another long-doubted industry–AI, and specifically the field of natural language processing (NLP).
And just maybe, I’ll convince you that a breakthrough technology has arrived that will enable machines to complete language-related tasks once the realm of human experts.
Consensus was founded on the theory that quickly extracting insights from scientific research is one of these tasks, and we are building an AI-powered search platform on top of peer-reviewed literature to test this hypothesis, but you’ll have to sign up for the Beta to find out if we are right.
Like Electric Vehicle Doubters Before Them, Many Question if AI is Ready for Primetime
NLP, in particular, has its skeptics. While some specific criticisms are surely warranted, others attack the general feasibility of the efforts, with some researchers claiming that NLP won’t perform complicated language-related tasks anytime soon since the models do not truly understand the meaning of the words they process.
Even if we zoom in on limited scope applications of NLP, like the Consensus-specific goal of extracting and summarizing knowledge from scientific research, doubters aren’t hard to find. Scott Alexander–the popular pseudonymous blogger and physician–graciously tested an old Consensus prototype and posted about it on his blog. But he remained skeptical that the problem could yet be solved, citing “the skulls of all the previous people who tried that.”
It’s true that some of the biggest names in the industry have tried and failed at using AI to extract knowledge from peer-reviewed literature. Even IBM Watson discontinued its medical paper-reading service for doctors in 2020 over concerns about accuracy and flexibility.
However, when it comes to processing scientific text or other complicated human language, Scott and all of the other doubters will be proven wrong in the next few years. The reason is that the NLP world just had its “lithium-ion battery” moment: Large Language Models.
Large Language Models Represent a Huge Step Forward in NLP
When OpenAI unveiled GPT-3 in May 2020, there was no doubt that the field of NLP had taken a massive step forward. GPT-3 is one of several new Large Language Models (LLMs) that are revolutionizing the relationship between machines and words for two main reasons:
- They are huge! GPT-3 was trained on half a trillion words from all over the internet, picking up knowledge and context about nearly every topic humans have ever written about, dwarfing its predecessors by several orders of magnitude.
- They process text (almost) like humans. This is due to the development of several revolutionary new neural network frameworks called transformers, such as Google’s BERT and OpenAI’s GPT series.
The main innovation of transformers is that they process all of the words in a sample of text simultaneously rather than sequentially. Processing each word one by one often led older models to “forget” about context from much earlier in the sentence or paragraph. By interpreting text more holistically, transformers are able to make connections between concepts that are distant in location, just as a human would remember an idea from a previous sentence.
LLMs Will Allow Machines to Read and Write Language Like Experts
If you follow any tech and venture news, it’s hard to miss some of the world’s hottest NLP startups made possible by LLMs.
Grammarly–which makes you an expert editor–increased error correction accuracy after implementing large transformer models, and I can only imagine that performance boost helped lead to their recent $13 billion valuation.
The origin story for Copy.ai–which makes you an expert copywriter–comes from the team’s early access to GPT-3, which inspired them to reimagine what was possible with new language models.
How about applying this technology to scientific expertise? Well, our friends and partners on the Semantic Scholar team at AI2 have proven that LLMs can complete increasingly complicated tasks when processing scientific literature. Check out their paper-summarizing TLDR feature.
The technology enabling near expert-level language processing is here, and the real-world applications will continue to come pouring in. As the head of AI2, Oren Etzioni states, “[LLMs] will power the launch of a thousand new startups and applications.”
Timing is On Our Side
In addition to all of the usual ingredients, the success of many NLP startups from this era will be in part due to timing. We believe that LLMs are to NLP as lithium-ion batteries were to electric vehicles. And Consensus is prepared to exploit this innovation for the democratization of scientific knowledge, right at the time when the world needs it most.
Now that you know this, you might take a look at the price of the likelihood of Consensus to change Scott Alexander’s mind, an online prediction market he created to judge our product! Some say now is the time to buy…