ChatGPT releases text watermarks, GPT-6 speed 50%
Few AI manufacturers believe that "quantity for quality" is one of them. On October 5th, after 12 days of a series of press conferences, the head of OpenAI Codex, Thibault Sottiaux, announced another 28-day renewal: For each of the next days, the team either issues an update that has a clear value to most Cordex or ChatGPTwork users or executes a full-scale replacement.
The event came to Day 1, OpenAI with three changes in one breath: the default production rate of the GPT-6 series was increased by about 50 per cent; the invisible text-printing, called TextGrain, appeared in response to the EU ' s AI Act ' s requirement that content machines be identifiable; and ChatGPT advertising introduced more visual advertising. Speed up, compliance, commercialization, three lines.
GPT-6 скоростта по подразбиране се увеличава 50%: от 30 жетона до 50 жетона в секунда
According to Tibo, by subscription to GPT-6 Astra and GPT-6.1 Sol, the default generation rate will increase by about 50 per cent, from 30 token per second to 50 token per second, without the user having to adjust any settings.
This boost covers not only OpenAI's own subscription product, but also Sign in with ChatGPT's partner tools, including OpenCode, Pi, Ap and Devin. Users use the ChatGPT account to use subscription capacity in these tools, and within the same range of adjustments, changes become effective within two hours of publication.
For work requiring continuous interaction with the model — cross-reading, modifying, testing, correcting, searching and generating information multiple times — the cumulative experience varies considerably after a one-time waiting period.
TextGrain Invisible Text Watermark: Leave Statistical "Fingerprint" for AI text
The EU ' s AI Act requires the generator AI provider to allow the generated text to be identified in a machine-readable manner. OpenAI offers a technology called textGrain: It does not mark the text corner, nor does it contain hidden characters. Instead, it imposes an invisible statistical signal when modeling words. The detection tool determines whether a text contains OpenAI watermarks by searching for such signals.
Launching the rhythm is in three steps: as of that date, global API clients can initiate text watermarks for selected models (API default shut-down); in the coming weeks, the eligible ChatGPT and Cordex text output in all European Union packages will be added to the invisible watermarks; it will not be set as a global default at the start. OpenAI also plans to open this technology to others.
Официално се признава, че водните знаци осигуряват следи, но те са много ограничени.
Under the 1% target misreporting test, the detection tool is able to identify approximately 80% of the 200-token watermark text, increasing the length to about 95% of 400 token; and the detection performance is significantly lower for small expressions such as mathematics. The editing also weakens the signal: in the 400-token text test, the 10% word is replaced by synonyms, the detection rate is reduced from about 92% to 66%; after the 25% replacement, only 17% is left.
OpenAI ' s official assessment states that textGrain is even more than other programs, including Google Syndium, but it also clearly lists the boundaries of watermarks: it detects watermarks that do not indicate how much judgement, editing and creativity humans have invested in, nor on which to determine the authenticity of copyright, responsibility or content; it does not relate to specific users, accounts, hints or conversations; and, in turn, it does not prove that the text was done independently. Since misreporting and underreporting still exist, testing tools are temporarily open to approved researchers and professional bodies.
In terms of performance, OpenAI did not observe meaningful differences in the inclusion of watermarks in its baseline tests for assessing Astra: Artificial Analysis Inteligence Index was 49.57 to 49.76 points, GPQA Diamond was 94.44 to 93.94 per cent and Terminal-Bench 4.0 to 56.06 per cent.
Also on the same day: a visual advertisement to generate scenes, a weekly business of 1.2 billion people.
OpenAI, the same day, ChatGPT reached 1.2 billion people a week. The new form of visual advertising will begin with testing in the United States later this month, showing how the selection is presented in the context of ChatGPT's production of pictures, and how to use them. Advertisements are clearly marked, separated from images that are being generated and do not affect ChatGPT responses.
OpenAI has also expanded its advertising attribution tool to connect conversion data with platforms such as Highouch, Telium and LiveRamp. Early data from several groups of collaborators: Weightwatchers on ChatGPT Ads had 15.3 per cent lower share-takers than their combined-paying search benchmark; 67 per cent of the health brand Dose's incremental purchases came from new customers; and 93 per cent of Portland Leather's visitors from ChatGPT Ads were new visitors.
Малко наблюдение.
The three changes in Day 1 coincided with the three current lines of OpenAI: making paid users feel that subscriptions are more valuable, explaining to regulators where AI's content comes from, and turning the growing size of users into a business that advertisers are willing to pay. Beyond the model, the AI platform is getting closer and closer to what we know about the Internet. 28 Days of Renewed Commitment has only just begun, and it is worth looking forward to what will follow.