AI startup Tome laid off 20% of its 59-person team in an effort to restructure and focus on the money-making part of its business: enterprise customers who use its AI tools to auto-generate presentations in minutes. The once-buzzy generative AI startup, last valued at $300 million, now plans to pitch its product to salespeople who make up a sliver of its customer base of 20 million people. The shift reflects an industry trend as more AI startups turn towards enterprise clients to bolster revenue. I spoke to CEO Keith Peiris on what prompted the crucial pivot.
“It’s actually way more straightforward to build a business this way,” Tome CEO Keith Peiris told me. “I think it becomes really hard trying to build a consumer product for millions of people and sorting out your own costs when most of them don’t pay.”
Tome has been used to create everything from pitch decks to post-surgery instructions for patients to students’ class projects. “For those folks I would say that they are less likely to pay and they use Tome in this very episodic sort of way,” Peiris said, adding that companies like Canva, which launched its own AI tool to create slideshows, have added competition in the consumer side of the business.
Meanwhile, Peiris claims that salespeople have been “making eight Tomes a day.” Tome’s new enterprise product, built on fine-tuned versions of models like OpenAI’s GPT-4 and Meta’s Llama, helps salespeople do research on their potential customer, scrounging through SEC filings and learning from news reports, to create tailored pitch decks.
A side note: nominations are now open for the Forbes Cloud 100 list, which recognizes the world’s top private cloud companies.
Now, let’s jump into the headlines.
Peak Performance
Last week, Meta announced Llama 3, the latest version of its open source large language model. The social media giant unveiled two models with 8 billion and 70 billion parameters, claiming they are better at math and reasoning, have reduced false refusal rates (instances where the model refuses to give answers to innocuous prompts) and have “increased diversity” in their responses. Meta has an even larger model with over 400 billion parameters, but CEO Mark Zuckerberg said in a video that it is still being trained. (Shortly after the video was released, a photoshopped image of Zuckerberg with a fake beard went viral.)
Llama 3 also now powers Meta AI, a conversational chatbot which is now prominently embedded into Meta’s flagship products including Instagram, Facebook, Messenger and WhatsApp — and even shows up in the company’s Ray-Ban smart glasses. The bot can generate images faster than its previous version and answer questions on any topic on the web with references to search results from Google and Bing.
But an AP report shows that the chatbot still has a tendency to make things up, similar to other AI models like ChatGPT. In a Facebook moms’ group, the Meta AI assistant claimed it had a gifted child in the New York City school district, an obvious hallucination. On another Facebook forum for swapping unwanted things, it replied to a real person’s post, offering up nonexistent items including a camera and an air conditioner.
And today, Microsoft launched its smallest model yet, Phi 3 Mini, which is part of a family of three small models. The model is designed to be almost as capable as GPT-3.5 but is much smaller in scale, trained on 3.8 billion parameters (GPT-3.5’s size isn’t public, but GPT-3 was trained on 175 billion parameters). The models don’t require Nvidia’s expensive GPUs and instead can run on chips used in regular computers; they can fit on a smartphone or laptop and function even without an internet connection.
Ethics & Law
Millions of users viewed YouTube videos that showed where to download deepfake pornography apps and how to use them to remove clothes from people in any image.
AFP via Getty Images
Google has an AI “nudifier” problem. YouTube has hosted more than 100 videos including ads and tutorials for apps that use AI to undress women by removing clothes from their images, a Forbes investigation found. Videos promoting the AI tools, which have reportedly been used to generate nudes of high school students and to produce child sexual abuse material, had amassed millions of views before they were taken down. The company has also hosted some of these AI apps in the Google Play Store, including a “nudity scanner photo filter” that has 10 million downloads. After Forbes reached out, Google took down over 120TK videos and one app.
AI Deal Of The Week
French AI startup Mistral is in talks to raise €500 million at a €5 billion valuation (roughly $530 million at a $5.3 billion valuation) according to a Financial Times report. The news comes just four months after the year-old startup raised $415 million at $2 billion valuation. The startup has very quickly catapulted to prominence as the European rival to Silicon Valley AI titans like OpenAI and Anthropic. It has the support of the French government (President Emmanuel Macron has said Mistral is an example of “French Genius”) and is backed by tech giants like Microsoft.
Also notable: VC firm Andreessen Horowitz has raised $7.2 billion to back startups in a number of realms including generative AI. Six hundred million of that will go toward defense tech through a fund called “American dynamism.” Don’t miss Forbes’ deep dive into the God loving, energy drink-pounding, “unapologetically American” startup culture that Andreessen’s defense investments have cultivated in El Segundo, California.
Deep Dive
Axon launched an tool that generates police reports using an AI language model.
American cops are increasingly leaning on artificial intelligence to assist with policing, from AI models that analyze criminal patterns to drones that can fly themselves. Now, a GPT-4 powered AI can do one of their less appealing jobs: filing paperwork.
On Tuesday, Axon, the $22 billion police contractor best known for manufacturing the Taser electric weapon, launched a new tool called Draft One that it says can transcribe audio from body cameras and automatically turn it into a police report. Cops can then review the document to ensure accuracy, Axon CEO Rick Smith told Forbes. Axon claims one early tester of the tool, Fort Collins Colorado Police Department, has seen an 82% decrease in time spent writing reports.
These reports, though, are often used as evidence in criminal trials, and critics are concerned that relying on AI could put people at risk by depending on language models that are known to “hallucinate,” or make things up, as well as display racial bias, either blatantly or unconsciously. “It’s kind of a nightmare,” said Dave Maass, surveillance technologies investigations director at the Electronic Frontier Foundation. “Police, who aren’t specialists in AI, and aren’t going to be specialists in recognizing the problems with AI, are going to use these systems to generate language that could affect millions of people in their involvement with the criminal justice system. What could go wrong?”
Smith acknowledged there are dangers. “When people talk about bias in AI, it really is: Is this going to exacerbate racism by taking training data that’s going to treat people differently?” he told Forbes. “That was the main risk.”
AI Index
200 million
People using Chinese search giant Baidu’s conversational chatbot Ernie Bot. It’s also the number of queries Ernie’s API handles every day, CEO Robert Li said at a recent conference.
30,000
“Agents” or AI assistants built on Baidu’s Ernie 4.0 foundational model.
85,000
Enterprise customers that have used Baidu’s platform and generative AI models to create their own AI applications.
Your Weekly Demo
If you’ve recently picked up a college chemistry textbook, chances are some of it was written with the help of ChatGPT, Christopher Hess, a director of product management at Pearson, told me. Tens of thousands of students are already using Pearson’s AI study tool, powered with OpenAI’s GPT 3.5 and GPT 4 models, to navigate through nail-biting homework questions.
In an effort to recreate the “office hours experience,” the edtech giant has embedded a slew of generative AI features into 30 different math, science and business e-textbooks, including a “reading copilot” that uses AI to summarize a topic, find relevant information and generate practice problems. The company is also making AI tools available to instructors to create homework assignments.
I tried out Pearson’s AI study tool by asking it for help answering a question about genetics. It broke down the complex topic and used a series of multiple choice questions to walk me towards the correct answer. The tool seems like a good starting point for a struggling student (or a hobbyist like myself, whose knowledge of genetics is mainly derived from sci-fi).
For the most part, the AI is built to keep the conversation as structured as possible to avoid situations where it could make up false information. And while at first it did not give in when I asked for the answer outright, it ultimately gave in after repeated pestering. But will Pearson’s AI study tool ever be able to replace a teacher? Probably not.
“AI is probably one of the places where it’s gonna be easier to make a Taylor Swift song than it is to be able to replicate somebody giving a wonderful lecture on how the kidney functions,” Hess said.
Quiz
This decades-old medical records company is using a multimillion-dollar acquisition of an AI startup to revive itself with a rebrand:
- Epic Systems
- Allscripts Veradigm
- Cerner Corporation
- Eclipsys
Check if you got it right here.
Model Behavior
A new AI roleplay app called AngryGF promises to teach people empathy and communication skills by simulating relationship conflicts with angry AI-powered girlfriends and wives (husbands are not included). In one scenario, the player has lost 50% of their savings in the stock market; in another, they’ve been a bit too complimentary toward a female friend. They then have 10 attempts to pacify an infuriated AI girlfriend with a forgiveness meter that tracks their success (or lack thereof). Wired’s Kate Knibbs found the game “irritating.” “I snapped and told this bitchy bot that she was annoying,” she wrote.
