Screenshot
In recent years, AI21 Labs has been hard to miss. The AI startup of Amnon Shashua, Uri Goshen and Dr. Yoav Shaham that brought us, among other things, the super-useful Chrome extension Wordtune. Now AI21 Labs presents its new development that wants to compete with the mega-popular ChatGPT.
Spice up your text
AI21 Labs announced today (Tuesday) a new feature called Spices that comes to the text editor in Wordtune and will allow you not only to play with the tone in which you wrote your texts, but to build long texts in a few clicks of a button by “seasoning” them with elements that it will create for you The language model on which the tool is based – Jurassic.
All you have to do is install Wordation, enter the text editor (clicking on the plug-in icon will offer you the option to get to it easily) and start writing. Even if you only wrote a few words, you can upgrade your text with various spices such as adding a reasoned explanation to your claims, clarifying the point by using different words, expanding your claims on a certain topic, giving examples, writing a counter argument to your original argument, adding analogies, Examples and even statistical facts. It is very important to note that unlike ChatGPT, the new tool will not write all the text for you and you will not be able to use Spices if you have not written a single word or by a simple prompt.
A very interesting part of the new development presented today by the Israeli AI company is the fact that, unlike ChatGPT, which is based on the huge language model GPT-3.5, the model on which Spices is based will not only give you answers based on the past information that was trained on it in the past. The new tool can connect to the network and add up-to-date information to your texts, including links to the information sources from which it took this information. This is an interesting step, and very unusual in the field – as part of which the team that worked on the development is also publishing an article on the subject today.
Source: AI21 Labs
Dr. Yoav Levin, who headed this development team and who led the research underlying the article explained in a conversation with Gigtime the choice of developing a language model that also connects to the network. “In our eyes, this is exactly what can turn the use of language models from something cool that looks like it can provide value, to something that is widely used, and realizes the crazy potential that we are starting to see in these models,” says Levin. According to him, when examining long answers that are produced by strong models that do not link to external sources – such as ChatGPT for example – you can often find that they make up facts: “This thing becomes even more ‘dangerous’ the better the model, the more we trust it, and then a fact is not Right can easily pass under our radar.”
Levin points out that his team has developed a method that makes it possible to reinforce any language model with the capabilities of providing additional context by connecting to a network. “Our development includes a network that pulls relevant documents from trusted sources, and presents them to the language model during its text generation. This both substantially improves inaccuracies and inherently solves the provenance issue – we know which source the language model relied on, literally during the production of every sentence “.
According to him, one of the advantages of this method is that there is full control over the sources they allow the model to access. The training itself is done only to enable the fetching of the documents from any source to which they direct the system. The identity of the system’s sources is always under control, to make sure that the source does not extract “fact” from dubious sources at best or those spreading outright lies at worst. Levin emphasizes that “there is still a lot of important research work to clear up such phenomena, but the progress is enormous once we have moved from a ‘bubbly’ language model, you will know what he practiced and why he says what he says, to a model that inherently produces a fact only if it is currently Read it in the source that is in our control.”
The new add-on for Wordtune is available for use starting today and unlike the limited number of corrections that the add-on allows you in emails or anywhere else, if you are on the free plan, Spices is available without limitation to all users. The company said in response to Giktime’s question that it will remain so for the foreseeable future, but of course it is impossible to know when and how it will change.