Multi-agent augmentive artiﬁcial intelligence: an exoskeleton for warrior mindsStory
July 06, 2023
A story that went viral on social media early in 2023 reportedly claimed that during a U.S. military artiﬁcial intelligence (AI) drone simulation, the AI targeted and killed the human drone operator. It had apparently determined the human had been interfering with its mission: in this case, to take out surface-to-air missile (SAM) threats. Although the military has since denied such a sim even occurred--with an Air Force official calling it a “thought experiment” rather than an actual test--according to Rommel Martínez, CTO of the ASTN Group (Austin, Texas), it doesn’t take a great leap of logic to imagine that this scenario could actually occur.
Current AI systems are prone to hallucinating. This means that AI sometimes provides conﬁdent responses not justiﬁed by its training, such as doing something that's only favorable to the machine itself and the machine’s own survival, both of which may be detrimental to humans.
This occurs, according to Martínez, because most of the available AI relies on a generative model, a statistical and probabilistic approach. “If you rely on the statistical approach,” Martínez notes, “most of the time it works. But when you're dealing with a weapon system, ‘most of the time’ is not enough. It has to be all the time. You cannot cannot afford to ﬁre a missile at your own base.”
These contemporary AI systems display intelligence which is not true intelligence, according to Martínez. These AIs are essentially word predictors and do not possess their own thinking capabilities. They do not reason from within, but rather are fed with external information that is spliced together with math. Achieving a higher level of functionality, closer to artiﬁcial general intelligence (AGI), requires a basic level of consciousness. That’s what Martínez has been working on for the past 20 years.
He created Valmiz to provide an alternative approach to contemporary AI.
Precisely what is AAI, or augmentive AI? Martínez says it’s a new term his team coined to distinguish their AI from other forms. “With AAI, we focus on making people do their work faster, think better, and have an overall higher throughput by giving them capabilities – much like an exoskeleton for the brain,” he explains. “We put the human at the center of the process – to preserve human morals, ethics, and values – the laws of Valmiz, while enhancing their capabilities with AI.”
The Valmiz AAI is a multi-agent algorithm design that ingests actual client raw data and converts it into a super knowledge base. Valmiz is not based on neural networks, machine learning, statistics, or other external worlds, unless the users wants it to pull such data into it. The information that's contained and processed inside Valmiz comes from within a client’s own organization. In that way, the information comes prevalidated and surpasses level-zero validation. It’s essentially a hyper-customized AI.
Valmiz combines several individual agents, which make up the components of the larger system. The components – named Veda, Vera, Vela, Vega, and Xavier – all work together as follows:
● Veda: Core Al system that fuses knowledge graphs and knowledge bases that binds everything together
● Vera: Tracks key-value-metadata changes across data sets
● Vela: Gathers data from different data sources; it can search your network or the internet for information for compilation
● Vega: Data-storage system; ensures that whenever Valmiz turns off (such as in the event of a power interruption) users will be able to turn things back on seamlessly
● Xavier: Human-machine interface for receiving commands; it receives commands and communicates back to users much like the digital assistants in use now, only in a smarter way that it pulls together from the other four agents.
Each of these agents acts as its own expert. In long inject systems, users have to add external information, and build upon it like Lego blocks, to enhance capabilities. One value of Valmiz is that a user can employ a speciﬁc agent to perform a task as needed, without needing to use the entire value system to operate it. For example, if an organization desires to see weather information only, this can be provided as one block as part of several different modules that can operate with each other and an organization’s existing systems. Each of those blocks can also be removed and used and operated separately. Because Valmiz can be integrated with other systems, it provides an ideal plug-in for any element designed using a modular open systems approach (MOSA).
It also has a compounding capability. It can use several keywords to search an internal database or, if desired, the internet. With regard to external sources, Vera veriﬁes and validates those information sources to ensure they actually exist. All of these agents have thick walls between them to ensure no overlap of functionality. When the information gets passed on to an agent, it operates within its own universe, a structure that keeps Valmiz free from conﬂict or contamination from a fellow agent, as well as free from outside tampering. This holistic system dynamically updates, continues running, and searches out information based on keywords that the human gives it, to provide accurate and precise actionable information.
This multi-agent approach is novel, as is the way ASTN Group designed the customer interface with the AI and the volumes of their data. Unlike most contemporary AI platforms that require humans to communicate with them using long string queries to form phrases or sentences that would make sense to the machine, Valmiz uses simpliﬁed keywords, with only a few things needing to be specified. Valmiz incorporates dual-use technology (DUT) to support all areas of a domain, for example: military and commercial industries, wholesale and retail, or research and application.
Another notable feature of Valmiz: Once it’s up and running, the user does not need to power it down and back on again for it to get updates. “Valmiz acts like a live system,” Martínez says. “You don't have to put it to sleep for it to learn. It learns continually.”
Finally, Valmiz has two types of user interfaces that directly connect to a user’s system: a desktop app that can be deployed on a machine or mobile device or one accessed via web API, inserted as part of an organization's pipeline, both approaches that enable input into processes to produce speciﬁc outputs.
Endless use cases
ASTN Group completed an initial study that identiﬁed more than 50 different industries that could beneﬁt from its technology, ranging from healthcare to election security, with the aerospace industry ranking high on the list.
On the commercial side, for example, businesses employ uncrewed aerial systems (UASs) for linear pipeline inspections. These aircraft collect tens of thousands of images which then require back-end analytics to provide valuable insights into the health of an enterprise. AI has proven useful for these purposes, with humans providing validation.
With Valmiz, this would be ampliﬁed. Valmiz could compare an entire pipe network for a company to identify systemic issues with a particular part, for example. Instead of just looking at one pipe at a time, Valmiz could connect the dots. If there was, for example, a broken widget on one part of the pipeline, it could holistically scan through all inspections for the last ﬁve years to identify if there were similar widget issues across the entire global enterprise.
Martínez says there are two layers of activity that must occur for this to work: The ﬁrst layer consists of the activity that combines all the sources of information together from the raw data. It would take more than a month of work for people to ﬁnd these things manually. Valmiz could automate this information gathering in hours or days, depending on the complexity of the task. Then it would be able to pinpoint, make cross-references, and create a comprehensive sophisticated network of all those data sets where every node is connected to each other.
The second layer concerns time. Generally, when businesses look at something, they only look at how it is presented at a given time without regard to its previous value. Valmiz captures all the previous incarnations of the thing being evaluated (e.g., the widget). Vera does the comparison, using data in Vela. Xavier enables the user to communicate to the system via keywords and bring them together and provide the required result. “If you're looking for information for a speciﬁc thing, you can get the results not only from an exact point in time, but also by branching out to parts that you didn't know were there if you had to do it manually,” notes Martínez. “Manual searches are linear. With Valmiz, it’s like a tree with branches through time.”
In a defense setting, where militaries cannot afford to have a system that makes wrong guesses, the value proposition of using gated data for a wide range of processes remains critical to operations. Service members need to be able to work with a systemthat aids them and does not try to one-up them by pretending that it's in a better position to make life-and-death decisions.
In an AI-enabled drone, for example, with Valmiz, the data used for decision-making would have been collected and processed at the edge, inserting volumes of data as a standalone system to improve processes. It can also be deployed like an additional module to make things faster. “
“You can deploy Valmiz at the headquarters or on the drone itself,” Martínez explains. “That's how extensive it is. More importantly, it frees militaries from the liability that the AI will imagine targets that are not there.”
Instead, Valmiz is like having a thousand highly qualiﬁed experts to work on a task to provide input to the human operator at the center of the operation. These abilities enhance decision-making with powerful, precise, and accurate information.
ASTN Group plans to showcase the capabilities of Valmiz to the defense community and consumers during the last quarter of 2023.
ASTN Group · https://valmiz.com