illi News - The Problem With AI Isn't Compute. It's Memory

Trending...

Artificial Intelligence Brain On Tablet With Data

OOSTKAMP. Belgian-USA startup Corbenic AI launches a technology that makes AI systems up to 21 times faster and cuts up to 90 percent of recurring compute. It targets a hidden problem: every time you ask an AI a question about a long document, the system reads that document from start to finish. Ten questions, ten full reads. This repetition is the largest contributor to today's runaway AI bills. The new technology, Taliesin, remembers what the AI has already read and restores it on demand.

SAN FRANCISCO - illiNews -- AI consumes massive amounts of electricity. And the consumption keeps rising. The International Energy Agency forecasts that AI systems will use twice as much electricity by 2030 as they do today. Data centers across Europe and the United States are queuing for grid connections. According to a small Belgian-French startup, the answer is not more computing power but something simpler: better memory.

The hidden problem. Imagine asking an AI to analyze a long document. You ask one question. The AI reads the entire document. You ask a second question. The AI reads it again, cover to cover. And so on, with every question. After ten questions, the AI has read your report ten times. This invisible repetition is by far the largest contributor to the runaway power and server bills behind today's AI. Until now, the industry has treated it as inevitable.

More on illi News

Memory that survives. Anthropic and Google have built something similar, but their version lives for only a few minutes and dies the moment a server restarts or the user is moved to another machine. In real production environments this happens constantly, and the AI starts from scratch every time. What Corbenic built, called Taliesin, persists: across a restart, across a server switch, and even across a move to a different generation of graphics card. No other company in the world can demonstrate that publicly.

How was this proven? Corbenic ran the same document through the same AI twice: once normally, once with the memory restored by Taliesin. The test only passes when the two outcomes are byte-identical, verified by a cryptographic "fingerprint" (the same kind of code used to confirm a software download has not been tampered with). The tests ran on three public AI models from Meta, Alibaba and Mistral, plus Corbenic's own open-source AI model Galahad-0.5B, which the company trained for 600 euros as a public verification substrate.
The results:

45 of 45 tests passed
60 of 60 tests passed in a follow-up matrix
64 of 64 generated words identical across two different generations of NVIDIA graphics card, in both directions Anyone can reproduce these tests with the public AI models,

the open Community Edition of the engine, and a standard hash tool. No NDA, no license, no access to proprietary source code required.

More on illi News

International scientific endorsement. The scientific publication of Merlin on arXiv was endorsed by Prof. Danqi Chen (Princeton University) and Prof. Xipeng Qiu (Fudan University, Shanghai), two leading names in natural language processing.

Where the giants assume, Corbenic proves. NVIDIA and China's Moonshot already build systems that move AI memory between servers, but publish no mathematical proof that the transfer is correct. The dominant approach: compress with loss, and hope it is good enough.

A stack that attacks the problem from both ends. Alongside Taliesin, Corbenic also built Merlin: a program that removes redundant data from AI systems before any computation begins. Merlin has been published scientifically and tested on 22 million text passages; its Community Edition is open-source. The combination of Merlin and Taliesin can cut recurring compute costs by more than 90 percent in workloads such as customer service, legal analysis or enterprise reporting.

"For two years, the entire sector has been fixated on building ever-larger models," says Sietse Schelpe, founder and CEO of Corbenic AI and currently Tech AI Lead at the Belgian postal operator Bpost on a freelance assignment. "We focused on something simpler: a better memory. Three founders, one developer and two Parisian astrophysicists as scientific advisors. A small team turns out to be capable of doing what billion-dollar companies do not even try."