Former Army officer develops offline AI for military use as Pentagon funds tech giants

Last month, the announcement that four artificial intelligence companies, OpenAI, Anthropic, Google, and xAI, had secured a government contract with the Pentagon drew the attention of one CEO who expressed skepticism about the proprietary models developed by these tech giants.

The companies, known for their chatbot services powered by large language models (LLMs) such as ChatGPT, Claude, Gemini, and Elon Musk’s Grok respectively, each received $200 million in government contracts.

These funds can be used by the Department of Defense to support warfighting and national security challenges across federal, state and local levels.

TRUMP’S AI ACTION PLAN IS ‘CRITICAL’ FOR NATIONAL SECURITY, CONGRESSMAN SAYS

“Right now, when you use ChatGPT or Grok or Anthropic Claude, you’re connected to the internet and every time you prompt it or ask it a question, you’re giving them your data,” Tyler Saltsman, CEO of EdgeRunner AI, told FOX Business.

Saltsman, a former Army officer with eight years of service in the Reserves, contributed to Operation Atlantic Resolve in 2017 as a logistician. Throughout his career, the CEO has worked in technology, having previously served at Amazon Web Services and Stability AI, a company known for advancements in text-to-image models.

EdgeRunner AI is a technological agent operated offline and isolated from unsecured networks in order to safeguard sensitive data and reinforce security.

ELON MUSK XAI STARTUP ROLLS OUT GROK-2 ARTIFICIAL INTELLIGENCE IN RACE AGAINST CHATGPT

“What we’ve done at EdgeRunner is we’ve taken these models, these LLMs, we’ve compressed them, and we fine-tune them on military data,” Saltsman said. “So now, they’re so small that they can live on a chip, and they don’t require big data center servers to host them.”

Saltsman champions EdgeRunner AI as open source and told FOX Business “you can see the guts of the models” if requested, whereas code bases and architecture for other chatbots are unattainable, according to the CEO. He added that EdgeRunner AI models can rest right on military armour or equipment, including battery packs, laptops or phones.

Chatbots like ChatGPT and Grok are trained on data from sources including websites like Wikipedia, news articles, books and consumer interaction.

NVIDIA CEO SAYS TRUMP’S AI PLAN WILL ‘FUNDAMENTALLY CHANGE’ US’ POSITION IN YEARS TO COME

“The problem with ChatGPT and frontier models like Anthropic and Grok, they’re too generalized, and they’re not specific to military expertise,” Saltsman told FOX Business.

“What we’re doing is we’re starting from the ground up and crafting these models to think like a military expert. More importantly, that captures our culture as an American warfighter because we’re very different from the rest of NATO.”

EdgeRunner AI, in partnership with Meta, is trained on over 30 billion tokens, or words, of military doctrine, military history both in and outside the United States, military tactics, training manuals, field manuals and philosophy, including Greek mythology, Roman Legionnaires, Bushido Code, principles of Jesus Christ and ideologies that shaped Western civilization, according to Saltsman.

AMERICA ‘GOING TO WIN’ AI RACE AGAINST CHINA, TRUMP DECLARES AS META EXECUTIVE APPLAUDS STRATEGY

“If you were to prompt ChatGPT with the same prompts you use with us, ChatGPT would say ‘You’re breaking the terms of service, or we’re sorry, we cannot support that kind of request’, and so, it’s unusable for the military, which is what we’re solving for,” Saltsman said.

He added that when certain chatbots are prompted, they process and transmit answers over the internet, heightening the risk of interception.

“There’s a risk of bad actors, let’s say, injecting malicious material within these AI’s,” Saltsman said.

Hallucinations and unstable data output, including a tirade of antisemitic social media posts from Grok to X in early July, have prompted some users to mistrust the technologies developed by Silicon Valley firms.

TRUMP’S AI PLAN IS PAVING THE WAY FOR ‘DISRUPTIVE INNOVATION’ IN THE US, EXPERT ARGUES

“It’s because these models have been trained on all over the internet and there’s a lot of junk and bad data on the internet,” Saltsma said.

“I think that the biggest problem right now is that AI’s been this sort of this cool niche product, but it’s not really solving anyone’s business problems because, again, it’s too generalized and also, it’s too biased. The military can’t use AI that is anti-Israel and pro Hamas or pro Iran because they have been heavily politicized to the left and that’s wrong as a warfighter, because we’re supposed to be an objective fighting force. We’re not left or right. We’re Americans, and we need AI that’s indicative of that.”

EdgeRunner AI can be used to write operations orders, risk assessments, memos to generals asking for support and more, saving hours and even days of time, according to Saltsman.

BILL MCDERMOTT ENCOURAGES US TO ‘WIN,’ ‘LEAD’ IN AI: ‘THERE’S NO OTHER CHOICE’

“As an officer, I spent 80% of my time doing this planning and paperwork,” he said. “It’s just very monotonous. Now, with generative AI, this can be done instantly, and I can spend the vast majority of my time just working with my soldiers. This is something that increases battle readiness and, as we become better, we increase the probability of winning the fight.”

EdgeRunner AI is programmed to aid service members based on specialty codes.

“If someone were to go down, or things were to go sideways, these AI’s can help us better have cross-MOS capabilities,” Saltsman said. “So, if I have a bespoke machine gunner, for example, and he or she goes down, that AI can actually help me understand how to use that niche weapon system.”

On June 16, OpenAI revealed in a blog post that the company would launch OpenAI for Government with plans to unlock new capabilities, including ChatGPT Gov, that directly serve public servants and enhance governmental workers’ capabilities.

“They could come up with something that’s completely open source with the transparent datasets and the models, but I just don’t see it because that’s effectively their secret sauce,” Saltsman said.

“There are superpowers that we have that they don’t have,” he added. “I’m not saying that nobody should use OpenAI. I think ChatGPT has a place.”

EdgeRunner AI is live in an overseas deployment with the United States Special Operations Command and, as a newer company, Saltsman is proud of the contract.

“They were happy with the fact that it never needs internet connectivity,” he said. “We never really thought what life could be like with no internet. And so now, the bigger vision is creating a J.A.R.V.I.S. to your Iron Man.”