Comments

You have commented 375 times on Rantburg.

Your Name

Your e-mail (optional)

Website (optional)

	My Original Nic Pic-a-Nic Sorry. Comments have been closed on this article.
Foto

Submit your comments on this article

Science & Technology

How neural networks are used in military decision-making

2026-02-23

Direct Translation via Google Translate. Edited

Text taken from the Telegram channel of darpaandcia

Commentary by Russian military journalist Boris Rozhin is in italics.
[ColonelCassad] Shyam Sankar, Palantir's CTO and a lieutenant colonel in the US Army Reserve, explains how their multi-model neural network system for military decision-making works.

Palantir doesn't trust any AI. Their system runs multiple neural networks simultaneously, forces them to argue with each other, and then produces a consensus. We took several of their patents and analyzed what's really going on under the hood.

It all doesn't start in the cloud—in a war zone, there's no stable internet connection, but there is electronic warfare. Therefore, the first layer of the system resides directly on a drone or satellite. Computer vision micromodels operate there: they don't think, they just "look" and translate images into raw numbers—coordinates, probabilities, and labels. A satellite spots an armored column, compresses the video into a text file, and transmits it. Traffic immediately drops by 90%. Palantir's Edge AI has been operating in Ukraine since 2022.

The data then enters what's called an Ontology. This isn't just a database—every tank, document, and person is transformed into an "object" with connections and a history. But before any neural network receives a single byte, the system checks cryptographic permissions. The general sees one thing, the operator another, their NATO ally a third. The AI is physically blind beyond what it's given for a single second.

There's no ChatGPT-style "ask whatever you want" here. The request goes to the Orchestrator, which breaks the task down into chunks and distributes them to micro-agents. Each agent is locked into the rigid logic of a finite-state machine: step 1, step 2, call the tool. Not a single step aside—only on the rails of the combat manual.

The most interesting thing happens at the decision point. When it needs to answer—"Is this an enemy column?"—the system doesn't trust a single model. The request goes in parallel to GPT, Claude, Llama, Groq, and perhaps a classified proprietary solution. They think independently, blind to each other's responses. Then the "Synthesizer" compares the results, seeks consensus, and filters out hallucinations.

Ooooooohhh — filtering out hallucinations is definitely the sexy bit.

At the same time, each agent is monitored in real time by Evaluator Agents—they read the logs, and if the model starts to act up or deviate from the protocol, the process is killed instantly. An engineer can intervene in the chain of reasoning, correct the error manually, and restart the pipeline.

A person sits in front of a dashboard. The screen reads: "Target detected. Probability 98%. Consensus of 4 neural networks. Suggested action: strike. Click Approve."

That's roughly how it all works.

Like drones, neural networks (and then full-fledged AI) will be an inevitable part of military affairs in the not-so-distant future.

Posted by:badanov

#2 comments); ?>werwer

Posted by: Abu Uluque 2026-02-23 13:05

#1 comments); ?>werwer

Posted by: Super Hose 2026-02-23 06:26

00:00