A new app from the makers of the Jurassic-1 large language model, AI21 Labs, claims to be able to help writers craft pages of text much like ChatGPT, but unlike the OpenAI-built chatbot Wordtune Spices can cite its sources and has access to the internet to get the most recent data available before generating text.
One of the biggest complaints raised against ChatGPT, and the wider GPT-3 large language model from OpenAI, is the fact it often generates “realistic looking but factually inaccurate” information and that it often completely makes up sources when asked to back-up data and claims.
This is something the company is working to resolve, explaining that ChatGPT is currently a “free research preview” – but with the news there are plans in the works to open API access to ChatGPT for Microsoft Azure customers, the question of accuracy is high on the agenda.
Israeli startup AI21 Labs has updated its popular Wordtune tool, which offers alternative wording for sentences and paragraphs, with a new feature called Spices that draws on its Jurassic-1 large language model to write the entire paragraph from a simple prompt. It is available through the Wordtune Editor.
Breaking down AI generation
The ChatGPT alternative is more granular in nature, working paragraph-by-paragraph and giving the writer a variety of options including an explanation of a given subject, expanding on a topic you’ve started writing about, specific examples to put your argument or subject in context.
It is also capable of adding analogies, definitions, facts, statistics and background detail. On top of this, it can include entire blocks of text such as a list of citations or a breakdown of terms used in the copy. While all of this is possible with ChatGPT, Spices breaks it out line-by-line and doesn’t rely on the user knowing to ask for specific types of content.
Wordtune Spices is designed to “enhance the writing process for a multitude of use cases, from professional documents to university essays to creative work” rather than replace the writer, said AI21 Labs. “Using a choice of 12 cues, Spices works alongside a writer to assist in the writing process, generating a range of textual options to add to and enhance sentences.”
It solves the “realistic looking but inaccurate information” problem by introducing a verification process for facts. It attributes all of its sources, providing users with a link back to where it found the information used when it generates text.
This, says AI21, solves one of the major problems with large language models, in that it ensures the original source of a piece of information is given credit. This will be particularly useful for students writing an essay or a business drafting a report as they will be able to back-up claims with links to the original data or claim.
“Unlike other generative text platforms, Spices does not replace the writer in the writing process, but, instead, serves as a ‘co-writer’, working alongside the writer to offer suggestions of additional complete sentences that improve and enhance the text that is being written,” the company said in a statement.
It comes with three main features: Core Writing helps users create the main messaging and argument of the text, including by producing a counterargument and explanation. Additional Writing helps users enrich the core idea with analogies and statistical facts and finally Awesome Writing adds “zing” to the text with jokes and inspirational quotes.
Professional writing cues have also been included offering solutions for legal and healthcare use cases to create “industry-specific texts” based on live and current data. “Our mission at AI21 Labs is to change the way people read and write using AI, while focusing on empowering – not removing or replacing – the writer,” said Ori Goshen, co-founder of AI21 Labs.
“Spices is a toolbox that melds the best that both man and machine can offer, working alongside writers as a source of inspiration for better, more efficient and more compelling writing, while ensuring that writers themselves have the space and freedom to best express their thoughts, insights and information. The future of writing starts today and we are excited to share it with you.”
Source citation
To back-up the ability to cite its sources, AI21 has published a research paper that explains its work on Retrieval Augmented Language Modeling (RALM) methods, demonstrating that pulling in relevant documents from a body of text can mitigate inaccurate text generation.
The methods they developed can also be used on other large language models, the company explained, demonstrating its use on GPT-2 and other similar models. “The method we publish can be applied to any off-the-shelf language model to make its text much more reliable, a simple API access to the language model is enough,” the company explained.
Yoav Levine co-Chief Scientist of AI21 Labs told Tech Monitor that Spices has been designed to find a relevant source before generating text, then it generates the text conditioned on that source. “This creates a situation that is qualitatively different than how language models (LM) operate today.”
He said a vanilla language model would generate text in a blackbox manner due to the training corpus – a large collection of structured text – used when building the model. These can be outdated or include bias. ChatGPT was trained on data only relevant up to 2021.
“A retrieval augmented LM would inherently produce a fact only if it’s currently reading it in a verified source, which you as a user get to look at in parallel to the generated text,” said Levine. “It also enjoys the additional merit of source up-to-dateness: since we’re taking our facts from external sources it doesn’t matter when the LM was trained, it has now access to later sources!”
Spices draws from Jurassic-1, the LLM built by AI21 with some 178 billion parameters in its largest form, putting it on a par with GPT-3 from OpenAI, but with the addition of using RALM techniques to draw in live data to back-up claims. “To access the web in real time and attribute the generation with the relevant web page or web source is a powerful feature,” said Goshen in an interview with Tech Monitor.
“It’s also supporting recent information, not just things from that when the model was trained on so you can actually you ask for statistical facts and know they will be up to date,” he explained. “It’s a very innovative and unique way to blend into the writer’s process.”
Spices is available through the Wordtune Editor.
Read more: Who are the LLMs competing with ChatGPT?
