This Startup Claims Its Models Fix A Major Problem With Generative AI

In 2013, May Habib found the work of Waseem Alshikh, a tech expert from Dubai, on GitHub. Alshikh was using new machine learning methods to summarize large amounts of information. Habib, who grew up in a small village in Lebanon and moved to Canada in the 1990s to escape the civil war, felt a connection with Alshikh. He had to leave Syria after illegally hacking into the government as a teenager.

After ten years and two startups, Habib and Alshikh are now leading a new wave of generative AI companies. Their startup, Writer, uses its own language models called Palmyra, named after an ancient city in Syria. These models help businesses and their employees write and edit emails, documents, ads, and summaries that follow company rules.

Most generative AI models often create incorrect information, which is a big problem for businesses. However, CEO Habib claims that Writer's latest model will "never create anything that's factually incorrect." This is because the model is designed to focus on accuracy rather than creativity.

Experts agree that the design of the model can affect how often it makes mistakes. Pranav Reddy, an investor, says that while the model's design can help reduce errors, no model can completely avoid them. If Writer's technology does make a mistake, it highlights the part of the text that is not sourced.

Writer has many big clients, including Uber, Deloitte, Spotify, and Accenture. For example, United Healthcare uses Writer’s models to review health insurance plans and write blogs or emails to explain them. Intuit uses it to create blog posts based on financial data, and L’Oréal uses it for product descriptions. Writer can also turn podcasts or videos into written content. One of its early clients was Twitter, which has not paid its bills, leading Habib to sue the company.

Writer’s models learn from public information and a company’s own data, like PDFs and style guides, totaling 30 billion parameters. Each company gets a customized version of the model, and the data is stored securely. Habib explains that the data acts like an index, allowing them to show where specific facts come from.

Writer is on the 2023 AI 50 list and competes with other generative AI startups. However, it is still behind in funding, valued at $155 million and having raised $26 million. In comparison, another tool called Jasper has raised $125 million. Writer’s founders say their main advantage is that they use their own language models, while Jasper relies on OpenAI’s technology.

Co-founder Alshikh points out that many companies only use other people’s technology without improving it. He believes that if a company does not control its AI, it cannot ensure the quality of its output.

Writer’s approach to solving the problem of incorrect information involves its design. It uses machine learning and natural language processing to understand and generate text. The system combines encoders, which understand text, and decoders, which create text.

Writer’s models are different from other models because they make two parts, called encoders and decoders, work together. When an encoder understands a question, it finds the right information from a database that the client company has provided. Then, it tells the decoder how to create a response. Habib explains to Forbes that this is why they can be accurate and use inclusive language.

Reddy, an investor, agrees that encoder-decoder models are not perfect, but they are more accurate than ChatGPT. He says, “These models usually have fewer mistakes than models that only use decoders, like OpenAI’s GPT-3.” However, he adds that these encoder-decoder models are often less creative.

Habib mentions that her clients care more about using the right words and punctuation in their content than about creating stories or poems. She says that any story included in the content made by Writer’s software is real and not made up.

Writer’s tools focus only on creating text and do not make images or videos. Newer companies, like Typeface, do offer those services. Writer can be used with many writing tools, such as Google Docs, Microsoft Word, and Figma. The company charges between $30,000 and $1 million, depending on how many words the software generates. They expect to make about $20 million in revenue in 2023.

Nvidia provides Writer with the hardware needed for computing and training data at a low cost. Alshikh says, “Our goal is to train smaller models to keep costs low for us and our customers.”

The founders of Writer are immigrants who have personal experiences with the English language and a background in machine learning. Habib, a Harvard graduate, was the first woman in her family to go to college and taught her family English. Alshikh learned English by himself at age 20 to study computer science. In 2015, they started their first project, a translation software called Qordoba, which helped companies like Sephora and Visa translate their digital content. They later developed Writer in 2020.

Habib says, “We have always thought deeply about how language helps us move forward in the world.”

CREDIT: forbes.com