Cloud Hosting

While OpenAI is propelling toward artificial general intelligence, AI that's smarter than human beings, Writer cofounders May Habib (left) and Waseem Alshikh (right) hold a different opinion. If you can unplug it, it's not AGI, Habib says.

In 2013, May Habib was browsing through GitHub when she came across the work of Dubai-based tech executive Waseem Alshikh, who dabbled in then-nascent machine learning techniques to summarize large blocks of information.

Habib, who grew up in a small village in Lebanon before immigrating to Canada in the 90s to flee the civil war, was immediately taken by the similarities she shared with Alshikh, who was also forced to leave his native country, Syria but for a different reason: as a teenager, he had illegally hacked into the countrys Ministry of Interior.

Ten years and two startups later, the duo is at the cutting edge of a wave of generative AI companies that have merged into the mainstream as the rudimentary techniques Alshikh was experimenting with a decade ago have been supercharged by advancements in transformers and deep learning models. Their generative AI startup Writer uses its own large language models called Palmyra (named after an ancient Syrian city) to let enterprises and their employees write and edit content such as emails, documents, ads and summaries, all of which will adhere to a companys editorial guidelines.

Unlike the vast majority of generative AI models out there that hallucinate, or spout incorrect information a major issue for businesses that are incorporating the technology CEO Habib claims that the latest version of its language AI model will never create anything that's factually incorrect. Thats due to the models architecture, which is designed to prioritize accuracy over creativity.

Thats technically plausible to an extent. Model architecture definitely does have an impact on hallucination rates, says Pranav Reddy, an investor at Conviction who previously worked at generative AI search engine startup Neeva. But there is no model structure that guarantees that they dont hallucinate, he continued. In the rare scenario that the companys technology does hallucinate, Habib says Writer highlights the portion of the text that has no source.

Customers, which include giants like Uber UBER , Deloitte, Spotify and Accenture ACN , seem sold so far. United Healthcare is using Writers HIPAA-compliant models to examine the fine print in health insurance plans and then write blogs or emails to explain those plans to health providers and patients. Intuit INTU has it writing blog posts based on financial data, and LOral uses it to write product descriptions and push notifications. Writers customers can also feed the software a podcast or a video and it can repurpose it into written content. (One of Writers early clients was Twitter, which hasnt paid its bills, says Habib, who sued the Elon Musk-owned company in late February for defaulting on payments.)

Writers models are trained on public information as well as companys own data such as PDFs, editorial style guides and brand words, together forming 30 billion parameters. Each company gets its own customized version of the model, and the data is stored on the companys cloud. The data is actually used like an index and we have this path to explainability that shows where specific facts came from, Habib says.

The company, which is featured on the 2023 AI 50 list, directly competes with other generative AI startups on the list that are increasingly targeting the enterprise sector. But its still catching up to its competitors in terms of funding. Valued at $155 million, Writer has raised $26 million in venture funding from Insight Partners, Upfront Ventures and others. In comparison, copywriting AI tool Jasper is a unicorn with deep pockets of $125 million in venture capital. Writers main distinguisher, according to its cofounders, is that it uses its own large language models (unlike Jasper, which is built on OpenAIs GPT-3.5) and trains those models on company-specific data to get more accurate results.

Most of the companies in the market are only playing in the application layer, and they don't even do it well. They literally use someone elses technology and put a UI on top of it, says cofounder and CTO Alshikh. Jasper is an example. Its just a reseller. As an AI company, when you don't control your AI itself, how are you going to control the quality of output?

The San Francisco-based startups strategy to address generative AIs tendency to make up facts has to do with how its built. On the backend, Writer uses machine learning, natural language processing and transformers to understand text and generate more of it. Writer combines encoders, which is the part of a transformer that is good at understanding text, and decoders, the components that predict and generate text.

Unlike other models, Writers models make encoders and decoders talk to each other. Once an encoder understands the query, it grabs the information from the database uploaded by the client company and then informs the decoder how to generate a response based on it. This is how were able to be accurate and write in an inclusive language, Habib tells Forbes.

Reddy, the Conviction investor, agrees that while the encoder-decoder models arent perfect, they are better than ChatGPT, at least when it comes to accuracy. Encoder-decoder models do tend to have much lower hallucination rates than decoder-only models, which is the OpenAI GPT-3 structure, Reddy says. The tradeoff you pay for this is that these (encoder-decoder) models also tend to be less creative.

Habibs says her clients are more concerned about using the right wording and punctuation in their branded content than being able to generate stories, sonnets and poems. Any story included in content written by Writers software is a real one and not fiction, she says.

Committing to its name, Writers tools only produce text and do not generate visual content like images and videos a service that newer enterprise-focused generative AI startups like Typeface do offer. It can be plugged into many writing and content creation tools including Google Docs, Microsoft Word and Figma. Writer charges customers a platform fee anywhere between $30,000 up to $1 million. Pricing is also based on the number of words generated by the software. Through this business model, Writer expects to book an estimated $20 million in 2023 revenue.

At a relatively low price, Nvidia provides Writer the physical hardware (GPUs and CPUs) for computing purposes and training data, Alshikh says. Our goal today is to train smaller models as much as possible to keep the costs cheap for us and our customers, he tells Forbes.

The immigrant entrepreneurs say that Writer is an amalgamation of their personal experiences with the English language and their business foundations in the machine-learning realm. Habib, a Harvard graduate, was the first woman in her family to attend college and taught her family how to speak English. Alshikh taught himself how to speak English at age 20 to learn computer science. This prompted them to start their first venture in 2015: a machine learning based translation software called Qordoba that helped companies like Sephora and Visa translate their digital content to other languages and dialects, before evolving it into Writer in 2020.

The medium of language to move ahead in the world has always been something weve thought about really deeply, Habib says.

Go here to see the original:
This Startup Claims Its Models Fix A Major Problem With Generative AI - Forbes