What leaked “SteamGPT” recordsdata might imply for the PC gaming platform’s use of AI

These days, it looks like each tech firm and their company mother or father is seeking to squeeze AI instruments and options into their merchandise, whether or not they’re needed or not. So when recordsdata with names and features referencing a “SteamGPT” appeared in a latest Steam consumer replace, Valve watchers took quick notice.

From the skin, it’s onerous to inform exactly what kind any such “SteamGPT” would take. But trying via variable names and references within the recordsdata themselves means that Valve could also be trying to make use of AI instruments to streamline inner evaluations of in-game incidents and sift via doubtlessly suspicious accounts.

Looking on the variables

As tracked by the automated SteamTracking GitHub project, the time period “SteamGPT” seems a number of instances in three separate files added within the April 7 Steam consumer replace. In addition to the SteamGPT naming conference—a seemingly apparent reference to the generative pre-trained transformers popularized by ChatGPT and its ilk—the recordsdata embody mentions of phrases like multi-category inference, fine-tuning, and “upstream models” that time to some form of generative AI system.

What may that AI be used for? Well, the recordsdata include a number of references to a labeler and “labeling tasks,” working with arguments figuring out a “problem” and “subproblem” and an “evaluation_evidence_log” associated to a particular “matchid.” Together with mentions of a “logs_to_inference” metamodel, that sounds prefer it may very well be a hook right into a system for mechanically producing labels to categorize the various incident reports made in Steam multiplayer video games.

Another portion of the SteamGPT recordsdata hints that Valve may use AI instruments to summarize suspicious exercise historical past or patterns in doubtlessly fraudulent Steam accounts. Quite a few “SteamGPTSummary” features embody references to VAC bans, Steam Guard, and account lockdowns. These features additionally appear to have a look at proof comparable to e-mail addresses (“high_fraud_email”), use of superior safety features (“two_factor”), and the place a linked telephone quantity originates (“phone_country”) to assist decide whether or not an account is on the up-and-up. There are additionally a couple of references right here to an account’s trust score, which is already used to assist safe matchmaking in video games like Counter-Strike 2.