After nearly ten years at META, TONY TUNG joined PRESAGE as co-founder and scientific director. On the FrenchWeb set, he explains how the startup wants to simulate the effects of a software change before its deployment and what it remains to demonstrate to cloud operators.
A configuration change can relieve one service and shift the problem to another. For teams operating cloud infrastructure, the challenge is knowing where the pressure will fall before making the change. It is this decision that PRESAGE is interested in, which is developing a model intended to anticipate deterioration and to simulate the effects of an intervention on the infrastructure.
TONY TUNG solved this problem using computer vision. During a trip to France, while he was still working at META, he met BENJAMIN REY, who presented his project to him. REY and ARTHUR CHEVALIER launch PRESAGE, then TUNG joins them to take scientific direction.
TONY TUNG, from computer vision to cloud infrastructures
- TELECOM PARIS. PhD and specialization in computer vision.
- TRIMBLE. He worked in France before leaving for Japan.
- UNIVERSITY OF KYOTO. Teacher-researcher in computer vision.
- RAKUTEN. Creation of a computer vision team within the Japanese e-commerce group.
- META. Almost ten years in Silicon Valley. He works on vision applied to virtual and augmented reality, then on multimodal generative models for AI agents at META SUPERINTELLIGENCE LABS.
- PRESAGE. Return to France. He joins BENJAMIN REY and ARTHUR CHEVALIER as co-founder and scientific director, with the objective of forming the research team in Paris.
Journey retraced from the interview given to FrenchWeb.
Simulate the consequences of an intervention
At PRESAGE, TONYTUNG transposes an idea from world models. A model learns how an environment evolves and seeks to predict its next state. In the case of Cloud-1this environment is an IT infrastructure. The model must learn how its components react to each other and then estimate the consequences of an action. Adding servers can, for example, reduce a queue while sending more requests to a database already close to saturation.
Operators already have tools to detect drift and limit the risk of deployment. DATADOG offers forecasting and anomaly detection functions. AWS allows you to cause controlled failures to observe the reaction of an infrastructure. GOOGLE documents incremental deployments that expose a change to a fraction of actual traffic. PRESAGE seeks to intervene upstream, when teams evaluate the consequences of a modification.
TUNG points out a difficulty which also arises with the agents responsible for correcting incidents. An automated intervention can itself cause a surge in load or move a bottleneck. In the architecture he describes, language models mainly serve as an interface with operators. Cloud-1 must estimate the effects of the intervention on the system.
Operators hold the data necessary for proof
To train its models and compare their predictions to the behavior of the cloud, the startup must access data that operators rarely share. TONYTUNG explains that some partners provide anonymized data and computing resources, like SCALEWAY
With a team of five people, PRESAGE focuses its efforts on research and targets large cloud operators and data center operators. An initial funding of 2.3 million was used to constitute the team and a new round is in preparation.
In the longer term, TONY TUNG envisions data centers capable of supporting part of their own operations. The next step will be an ordinary decision for a cloud operator. When a team is about to modify its infrastructure, will Cloud-1 be able to spot an effect that it had not anticipated, early enough and with enough precision for it to review its deployment?
FRENCHWEB PARTNER · CANALCHAT GRANDIALOGUE
Make your speaking engagements an interactive event
Are you preparing a show, a meeting with your clients or an internal event? Canalchat Grandialogue brings together live video production, live streaming and interactivity to bring together managers and employees, brands and customers.
A single point of contact to combine the realization of your event and the participation of your audiences.

