The Prompt is a weekly rundown of AI’s buzziest startups, biggest breakthroughs, and business deals. To get it in your inbox, subscribe here.
Welcome back to The Prompt,
One consequence of the current AI investment frenzy is that founders of startups are seeing their ownership being chipped away as they raise enormous funding rounds at towering valuations. But AI training startup Invisible Technologies is going in the opposite direction: it’s taking out loans to buy out its VC investors and awarding stakes to its employees. In 2020, CEO Francis Pedraza revived the floundering business after a well-timed call from DoorDash, which needed help with importing menus. Now, the startup helps wrangle data dilemmas for large language model providers like OpenAI and Cohere.
Now let’s get into the headlines.
BIG PLAYS
While Chinese company DeepSeek caused an uproar last week for its capable and cost-efficient AI models, some industry insiders remain unimpressed. I spoke to founders who’ve been hard at work building similarly cheaper-to-train AI models and think the market reaction (Nvidia has lost about $600 billion in market cap since DeepSeek’s announcement) has been “misguided and misinformed.”
In response to the DeepSeek freakout, OpenAI made its o3-mini model, which is largely better than DeepSeek’s, available for free. Over the weekend, it also released Deep Research, an AI agent that will process a prompt for up to 30 minutes to browse the web and generate detailed analysis and reports. Notably, Chief Research Officer Mark Chen claimed such research capabilities, which typically cost more time and money, justify OpenAI’s move to invest billions of dollars into AI infrastructure. CEO Sam Altman posted on X: “50 cents of compute for 500 dollars of value.”
And OpenAI has seemingly flipped the Chinese AI company’s arrival into more money for the company. It’s reportedly in talks to raise $40 billion at a $340 billion valuation, CNBC reported.
HUMANS OF AI
A Nepali startup called SecurityPal is using AI to tackle the long-drawn process of security compliance for big tech players, a “necessary nuisance” as my former colleague Alex Konrad puts it. The niche company has over a hundred employees in the city of Kathmandu, who have answered millions of mundane questions for companies like OpenAI and Figma when they onboard new customers. But CEO Pukar Hamal has a more ambitious goal: to transform Kathmandu into a new startup hub, which he likes to call “Silicon Peaks.”
AI DEAL OF THE WEEK
ElevenLabs, a startup that uses artificial intelligence to generate and clone voices, has raised $180 million in a round that values it at $3.3 billion, triple from what it was a year ago. Users have used its AI audio abilities to generate 1,000 years’ worth of content, the company told Forbes.
Elsewhere, Protex AI, whose AI software helps companies like Amazon and Tesla monitor worker safety in factories and warehouses, has raised $36 million.
DEEP DIVE
Not all accidents are unfortunate. Jake Jollis would know.
Years ago, Jollis mistakenly put his personal phone number in an SEC filing for his previous startup, an online language learning company called Verbling. His phone number had ended up on Google, and he was soon inundated with scam calls and people looking for Verbling customer support long after he’d left the company to become a venture capitalist at San Francisco-based investment firm Matrix, where he wrote checks to AI startups.
“I’m just investing in AI here and on the other side I’m getting these weird voicemails, and then eventually something starts to click,” Jollis tells Forbes. That got him thinking: What if someone could pick up those calls? While Verbling wasn’t big enough to have a call center of its own with humans on the other line, an “AI receptionist” could potentially intercept the inquiries and do the job, he said.
In June 2024, Jollis started working on Answering Machine, a startup which uses an artificial intelligence to answer calls for small business owners when they’re busy. The bootstrapped company has already done thousands of calls for home services businesses like plumbers, electricians and snow removal companies. “Our goal is to bring AI to Main Street,” Jollis told Forbes.
Answering Machine is a mobile application that uses a medley of AI models under the hood, including ElevenLab’s audio generation models and OpenAI’s large language models, that have been trained on public information from the businesses’ websites as well as any additional information the customer provides. During the call, the AI receptionist collects basic details from the customer like their name, address, phone number and the service they’re requesting. It then sends an AI-generated summary as a push notification to the small business owner’s phone.
Jollis said the AI receptionist is meant to act as a stand-in for when the business owner is unavailable to take the call. “For our customers, missed calls means missed revenue opportunities… For example, the plumber is still answering their own calls when they find it convenient, and then when they have half an arm down a toilet fixing some plumbing problem and a call comes in, we take those calls.”
WEEKLY DEMO
AI: One, Jailbreaks: Zero. AI giant Anthropic has released new research on ways to prevent AI models from producing harmful content and defend them from “jailbreaks”— methods for breaking an AI model’s safety guardrails to generate content it shouldn’t.
MODEL BEHAVIOR
AI-generated books are making their way into public libraries through digital catalogues, 404 Media reported. But what makes this migration even more challenging is that as AI models get better at creating text, it’s become more difficult to distinguish real from fake. AI-generated low-quality content is already infiltrating almost every online platform, from Amazon to Ebay. One possible solution? The Author’s Guild is introducing certification for books guaranteeing that they are human-written.
