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.
Elon Musk has built a gigantic supercomputer in Memphis, which he calls a “gigafactory of compute.” Operations began yesterday, Musk posted on X. Named Project Colossus, the data center houses 100,000 Nvidia H100 chips and is being used to train xAI’s ChatGPT competitor Grok. But multiple city council members told Forbes they were sidelined from decision-making conversations about the facility and have lingering concerns about its effect on the city’s already strained power grid and water supply.
The new data center will use 50 megawatts of electricity–enough to power about 50,000 homes– and will use 1.3 million gallons of water every day for cooling. But despite the company’s verbal promises to improve the city’s infrastructure with new facilities, Musk’s past track record on similar pledges gives the city reason to worry.
Now, let’s get into the headlines.
PEAK PERFORMANCE
Today, Meta released its most adept AI model, Llama 3.1, which is also the largest open source AI model with 405 billion parameters. The Facebook parent claims the model outperforms top AI models like OpenAI GPT-4o and Anthropic’s Claude 3.5 Sonnet across math, translation and reasoning tasks. The model was trained using 16,000 of Nvidia’s H100 GPUs and now powers Meta AI assistant, which Mark Zuckerberg claims will surpass usage of ChatGPT by the end of the year. In an open letter, he also made a case that open source AI models are safer and more transparent than close sourced systems, making them overall better for humanity.
Meanwhile, open source AI model maker Mistral AI launched two new small-sized academic models, Mathstral and Codestral, which excel in specific tasks like solving math problems and generating code.
BIG PLAYS
OpenResearch, a nonprofit backed by OpenAI CEO Sam Altman, released the results of a three-year-long study that probed the benefits of universal basic income in an experiment where 3000 people were randomly given a $50 or $1000 monthly stipend with “no strings attached.” The concept, which has garnered support from other tech visionaries, stems from the implications of a future where artificial intelligence has eliminated a significant number of jobs. Broadly, the study found that free cash allowed room for flexibility where people could be more selective in job searches, get medical care and help relatives with their own bills.
AI DEAL OF THE WEEK
As it competes with Nvidia, chip giant AMD acquired Helsinki, Finland-based AI research company Silo AI for $665 million, the largest takeover of an AI company in Europe since Google bought DeepMind in 2014 for $400 million, Forbes reported. A research lab with 300 PhDs, Silo AI has built an open source large language model that can interpret and produce text in English as well as Nordic languages.
Also notable: Autonomous tractor company Monarch Tractor raised $113 million at a more than $500 million valuation, the largest funding raise to date for an agricultural robotics company.
DEEP DIVE
AI chatbot misses its target
Last month, Target officially plunged into the AI bubble. Like many of its peers in the retail industry, the company went with something simple and accessible — a ChatGPT-esque bot. But instead a consumer-facing bot for customer support, as is typical, Target’s was focused on providing information for its employees.
Target CIO Brett Craig enthused in a press release that the “transformative GenAI Technology” would arrive in all of the company’s nearly 2,000 stores by August. “Help AI,” would be a sort of “store process expert and coach,” helping new team members learn on the job. It would also “answer on-the-job process questions, coach new team members, support store operations management and allow teams to “work more quickly and efficiently.”
But that hasn’t proven to be the case.
Target employees told Forbes that “Help AI” struggles to provide decent answers, is frustrating to use and is more distraction than anything else. Many described it as a waste of company resources.
“This tool feels like something [Target] can say: ‘Oh look at we’re so innovative,” one Target employee who wished to remain anonymous for fear of professional reprisal told Forbes, adding that Help AI felt like a bad version of OpenAI’s ChatGPT, but with more restrictions. Said another: “We call it the shitbot because it gives shit answers.”
Worse, when a Texas employee asked Help AI how to handle an active shooter in the store, it told him to confront the attacker if he was near a weapon, specifically suggesting he use a baseball bat.
“It’s just not a good idea, just go hide,” the employee told Forbes. “If you’re in a break room, stay there, don’t go find a weapon.” Department of Homeland Security guidance suggests “evacuate,” and “hide out” as first and second options before taking action “as a last resort.”
Target spokesperson Brian Harper-Tilado did not respond to questions about the chatbot’s guidance around an active shooter situation. He said in a statement that the company is “committed” to making employee jobs easier” as a way to “better serve our guests” and was open to feedback about the bot.
Read the full story in Forbes.
YOUR WEEKLY DEMO
Don’t know if an image or video of a person is real or an AI-generated deepfake? Research derived from technology used by astronomers to study galaxies suggests analyzing the eyes in the image and comparing the reflections of light in the right and left eyeballs. If they match, it’s most likely an authentic image. If they don’t, it could be a fake.
QUIZ
This AI model is the best overall, according to LYMSYS Chatbot Arena, a popular crowdsourced platform for ranking artificial intelligence products.
A. GPT-4 Turbo
B. Llama 3
C. GPT-4o
D. Gemini 1.5
Check if you got it right here.
MODEL BEHAVIOR
Elon Musk posted a video of an “AI fashion show” where artificial intelligence imagined world leaders sporting garb that seemed to reflect their widely perceived identities. The video shows AI-generated representations of Jeff Bezos, Kamala Harris, Donald Trump and Barack Obama among others, walking down a runway.
