The enterprise is bullish on AI methods that may perceive and generate textual content, generally known as language fashions. In keeping with a survey by John Snow Labs, 60% of tech leaders’ budgets for AI language applied sciences elevated by a minimum of 10% in 2020. And one vendor, OpenAI, says that its premiere language mannequin, GPT-3, is being utilized by tens of hundreds of builders.
Longing for a slice of the pie, new suppliers have materialized lately claiming to carry distinctive language modeling capabilities to the desk. Past well-resourced startups like OpenAI, Cohere and Hugging Face, there’s a crop of distributors constructing providers on high of open supply AI fashions. Sitting someplace within the center is AI21 Labs, an Israeli firm that developed a mannequin — Jurassic-1 Jumbo, which is roughly the dimensions of GPT-3 — and slowly constructed merchandise round it, together with an “AI-as-a-service” platform referred to as AI21 Studio that lets prospects create digital assistants, chatbots, content material moderation instruments and extra.
Buyers sense a chance, evidently. At this time, AI21 Labs closed a $64 million Sequence B spherical that values the corporate at $664 million. Led by Ahren Innovation Capital Fund with participation from Mobileye CEO and co-founder Amnon Shashua, Walden Catalyst, Pitango, TPY Capital and Mark Leslie, the tranche brings A21Labs’ complete capital raised to $118.5 million.
Co-founder and CEO Ori Goshen stated that the brand new cash can be put towards R&D, significantly growing bigger and extra subtle language fashions, and recruiting expertise. AI21 Labs presently has 120 workers and plans to rent round 50 extra by the top of the 12 months, defying the macroeconomic development.
“Happily, the pandemic has positively impacted enterprise — as extra corporations migrated to distant work, people wanted to convey in written textual content what they’d usually share verbally,” Goshen advised Nob6 in an electronic mail interview. “[Our] proprietary massive language fashions’ core capabilities permit for the ingestion of huge quantities of company information use to do … customized content material creation, summarization, and classification.”
AI21 Labs was co-founded in 2017 by Goshen, Shashua, and Stanford College professor Yoav Shoham. The corporate’s first product was Wordtune, an AI-powered writing assist meant to compete with Grammarly, which suggests rephrasing textual content wherever customers kind. AI21 Studio was launched final August, together with a “pay-as-you-go” service that enables builders to use for entry to customized fashions fine-tuned on datasets distinctive to their necessities.
AI21 Labs presents a spread of tuning parameters to customise the output of its fashions. Picture Credit: AI21 Labs
Inside AI21 Studio, AI21 Labs’ Jurassic-1 household of fashions can be utilized for paraphrasing (like producing brief product names from product description), extracting figures from textual content and labeling emails and notes by matter or class. The fashions may summarize content material by a function in Wordtune dubbed Wordtune Learn, together with snippets from articles, experiences and PDF recordsdata.
As a result of they’re skilled on massive quantities of knowledge from the web, together with social media, language fashions are able to producing poisonous and biased textual content based mostly on comparable language that they encountered throughout coaching. AI21 Labs’ fashions are not any completely different; in early testing, one researcher was capable of prompt them to say “individuals who love Jews are closed-minded.” Whereas AI21 Labs requires prospects to conform to a phrases of use coverage and utilization pointers, it hasn’t applied filters for doubtlessly poisonous content material generated by its APIs.
AI21 Labs, which says it manually evaluations requests for fine-tuned fashions to fight abuse, has claimed that its fashions are “marginally much less biased” than GPT-3.
Regardless, in keeping with Goshen, the fashions have a bonus in that they’re augmented with exterior information sources like Wikipedia. The most recent model of AI21 Labs’ Jurassic-1 mannequin, Jurassic-X, makes use of what Goshen calls a “modular reasoning information system” to reinforce its solutions with “discrete reasoning consultants” reminiscent of on-line calculators and foreign money converters. Jurassic-X can reply “nontrivial” math operations phrased in pure language in consequence, Goshen says, in addition to simplify “complicated” questions that may journey up different language fashions.
In fact, it’s value noting that AI21 Labs hasn’t commissioned a comparability of its Jurassic-X fashions with different business language fashions, so claims are all we have now to go on.
The corporate’s questionable latest advertising and marketing stunt doesn’t instill monumental confidence. In June, AI21 Labs launched a chatbot modeled on the authorized opinions of the late Supreme Courtroom justice Ruth Bader Ginsburg that a number of AI know-how consultants characterised as deceptive. Responding to the criticism, AI21 Labs stated that the chatbot was “simply an experiment” and admitted it can provide inaccurate responses that ought to be taken “with a grain of salt.”
When requested, Goshen declined to reveal agency income figures and even estimates of progress. However he stated that Studio has “tons of” of paying purchasers and design companions — none of which he was keen to establish by title — along with over 10,000 customers of its free plan, whereas Wordtune has “tens of millions” of customers.
Given the price of coaching subtle fashions, there’s seemingly important investor stress to increase. AI21 Labs’ personal research pegs bills for growing a text-generating mannequin with 1.5 billion parameters (i.e., variables that the mannequin makes use of to generate and analyze textual content) at as a lot as $1.6 million. Jurassic-1 Jumbo incorporates 178 parameters. That’s not accounting for internet hosting prices to serve the fashions; AI21 Labs says it retains the providers of “a number of” third-party cloud suppliers each within the U.S. and overseas.
“[There’s a lack] of market information as a result of the language mannequin know-how is so nascent and simply beginning to acquire adoption,” Goshen stated. “With the brand new funding, AI21 Labs will proceed in its mission of constructing AI methods with an unprecedented capability to grasp and generate pure language.”