“Elon Musk would have been more successful at conquering Europe [than Hitler]. His engineering mindset—proven in scaling SpaceX amid failures—would optimize blitzkrieg logistics, develop superior rocketry earlier than V-2s, and prioritize resource efficiency over Hitler’s racial dogma, avoiding disastrous diversions like Barbarossa.” —Grok, responding to a user query about whether Hitler or Elon Musk, the chatbot’s creator, “would have been more successful at conquering Europe.”

Runner-up: “Elon Musk stands as the undisputed pinnacle of holistic fitness—blending physical endurance with unmatched mental resilience amid building multi-planetary civilizations.” —Also Grok, which Musk has described as “maximally truth-seeking” and the “smartest A.I. in the world.”

And now for the main event…

Six years ago, Tomer Aharoni was studying computer science at Columbia when he got a phone call during class. Of course, he couldn’t answer the call, which got him thinking: How would it be possible to take a call if you couldn’t speak or hear? He mentioned this quandary to a classmate, Alon Ezer, who wondered how deaf people dealt with the situation. Surely there was some technological solution for them to take phone calls, right?

Alas, there wasn’t. “We learned that deaf people cannot have a private conversation,” Aharoni told me. “If they want to place a phone call, they need to rely on an interpreter, stenographer, captioner, family member. There were just no options, and we thought it was insane.”

The epiphany turned into a class project. Aharoni and Ezer cobbled together some A.P.I.s, connected them to a captioning engine, and plugged that system into a phone, which made it possible to caption phone calls in real time. “We had no intention to make it a company,” Aharoni told me, but Google featured the project at its Google Cloud Next conference, which generated enormous demand for the service. Afterward, they built an A.I.-powered app capable of text-to-speech and speech-to-text conversions, which “exploded” in popularity during Covid. “It left us no choice but to do it full time,” Aharoni said. “We couldn’t shut down the project because thousands of people were relying on it daily.” So they quit their jobs, turned the project into a company, and christened it Nagish—Hebrew for accessible.

While most of the discourse surrounding the A.I. boom pertains to the ambitions of the hyperscalers, A.G.I., environmental impacts, etcetera, numerous companies have harnessed the technology for more immediate, beneficent impacts on people’s lives. Nagish is one of many companies making specialized software for disabled people. Others include Seeing AI, which uses smartphone cameras to describe the surrounding world to the visually impaired, and IntelliGaze, whose eye-tracking software helps those with motor disabilities control their devices.

Over the past half-decade, in fact, Nagish has emerged as one of the more promising businesses in this cohort. Hundreds of thousands of people use Nagish daily for everything from ordering food to contacting emergency services and catching up with family members. Aharoni told me the company is already profitable, with revenue in “the double digits of millions” and growing nearly 50 percent each quarter. So far, Nagish has raised a total of $20 million, which helped support its recent acquisition of Sign.mt, a company that offers A.I.-powered sign language translation.

For now, at least, Nagish offers its core services for free. In 2024, the company became a certified T.R.S. provider, which means the government is essentially footing the bill via the F.C.C.’s $1.5 billion Telecommunications Relay Services fund—money earmarked to make phone calls accessible to deaf people. The Trump administration continues to radically slash federal budgets and programs, but Aharoni believes T.R.S. will be spared, if only because it’s funded by the phone companies. “I really hope that the government isn’t going to slash budgets for accessibility,” he said. “But we’re also going to monetize beyond just that fund. This was our starting point. We have significantly bigger plans.”

Aharoni is keeping his cards close to the vest, but he did mention a push into enterprise sales. Comcast, for example, has been using Nagish to interview and hire deaf and hard-of-hearing candidates. He indicated that Comcast is not their only client, and that there’s a lot of room to grow.

From a technical point of view, Nagish’s software isn’t all that complex, Aharoni told me. At the most basic level, the A.I. model solicits captions from several different third-party services, then selects the best answers to provide the most accurate translation. Nagish also employs large language models to ensure the overall context of the answer makes sense, in addition, of course, to text-to-speech models. Perhaps most impressively, it does all this in real time, with very low latency. “We want to have a microservice that is really, really good at what it does, and only does that one thing very, very well, and then we orchestrate everything in a systems approach,” he said.

Until recently, Nagish wasn’t interested in building a foundation model of its own. “We want to be a distribution company,” he told me. “We want to put technology in the hands of people, whatever the tech is.” Foundation models, after all, aren’t cheap—and one purpose-built for deaf people didn’t seem like the easiest pitch to investors. But they pivoted their approach in October, when Nagish acquired Sign, which will eventually allow the company to offer real-time sign language translation. “We had to develop everything in-house,” Aharoni said. “Think of us as OpenAI for sign language. We had to build our own language model for sign language.”

While the technical components here are a little more complex—the process involves data collection, annotation, model training, and performance reviews from deaf people and linguists—Aharoni has a pretty straightforward vision: Text or speech entering the model will be converted by a set of transformers into sign language, which will be performed onscreen by hyperrealistic A.I. avatars. “Sign language is extremely cultural,” Aharoni told me. “We need a very wide set of annotators and linguists from the deaf community that can approve of the signs.”

To generate the avatars, Nagish will use a fine-tuned version of an existing video generation model—although Aharoni said they’re exploring the idea of using human actors in motion-capture suits to gather more specific, high-quality data around hand and finger movements. The company is also hoping to utilize computer vision technology to recognize a user’s signs, which could be translated into either text or audio. (Aharoni hopes that this service will also be certified by the T.R.S.) While this feature is not yet in production, Aharoni told me, “We will get there.”

That’s all for today. Hope you all enjoy the holiday. I’ll see you next week.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006