AI Today: Falcon ASR and LegalOn models

Knowledge state as of 2026-10-09. Source material gathered on 2026-10-09.

Today's compilation covers new developments in cloud gaming, advanced speech recognition models, and practical tools supporting the development of custom artificial intelligence solutions. We also analyze operational activities undertaken by model providers regarding security and the optimization of computing infrastructure maintenance costs. These insights provide concrete data for developers and system architects, focusing on implementations, performance, and efficient resource utilization in production environments.

Highlights Today (TL;DR)

• GeForce NOW: The service has added Gears of War: E-Day and nine other titles to its library.
• Falcon ASR: TII introduced a 1.6-billion-parameter speech recognition model optimized for the Emirati dialect.
• ML-Intern: An agent that automates the process of building and testing custom models using prompting.
• OpenAI: The company disrupted two influence operations using fake profiles to spread geopolitical content.
• LegalOn: The company reduced Codex usage costs by 65 percent by optimally matching smaller model variants to tasks.

GeForce NOW Introduces Gears of War: E-Day

The GeForce NOW cloud service expanded its library on October 8 with the game Gears of War: E-Day, enabling cloud gameplay utilizing the performance of GeForce RTX graphics cards.

The title, which had its global release on October 6, features a story set 14 years before the events of the first game in the series. Users of the Ultimate subscription tier can enjoy GeForce RTX 5080-class performance, along with support for NVIDIA DLSS and NVIDIA Reflex technologies. The game is available within the service for Game Pass subscribers on Steam and Xbox platforms.

In addition to this debut, the GeForce NOW library was expanded with 9 other titles, including STAR WARS: Galactic Racer, Clive Barker’s Hellraiser: Revival, and Forever Ago. The manufacturer also announced changes to the purchasing process for Fire TV device users, who will be able to purchase service subscriptions directly through Amazon in the coming weeks. For individuals wishing to test the cloud's capabilities, one-day passes have been made available, the cost of which can be deducted from the fee for the first full subscription.

For AI Builders: Developers of cloud gaming systems can analyze the mechanism of integrating subscription purchases with external hardware platforms. The source does not specify subscription costs for individual regions, technical details of streaming over low bandwidth connections, or the availability of these features outside the ecosystem of devices mentioned in the announcement—these issues require verification in the provider's API technical documentation.

Falcon ASR from Technology Innovation Institute

The Technology Innovation Institute (TII) from Abu Dhabi introduced the Falcon ASR model on October 7 for automatic speech recognition. It is a 1.6-billion-parameter system optimized for the Arabic language with a particular focus on the Emirati dialect. The model also supports English, French, Spanish, and Portuguese.

In tests conducted by the creators, based on six datasets, Falcon ASR achieved an average Word Error Rate (WER) of 20.92%, outperforming the best model in the tested set, which scored 23.17%. In an internal evaluation focused on the Emirati dialect, the model achieved a WER of 22.73% and a CER (Character Error Rate) of 10.19%. The model architecture is based on previous work on Falcon3-Audio. The system can generate word-level timestamps, allowing text to be linked to a specific segment of the audio recording.

The model training process accounted for diverse acoustic conditions, such as background noise, overlapping voices, music, reverb, and telephone interference. Falcon ASR processes the five listed languages using the same weights, eliminating the need to use a language flag before starting transcription. Currently, interested parties can test the system's operation via the Hugging Face platform, while the team's plans include providing access through an API and dedicated applications.

For AI Builders: Teams working on automated call center audio processing can test the model on their own audio samples, accounting for background noise. The publication does not specify licensing conditions, API usage costs, or potential constraints regarding computing infrastructure when deploying custom instances.

Automated Creation of Specialized Models by ML-Intern

An article published on October 8, 2026, described the process of creating custom artificial intelligence models with the support of the ML-Intern agent, for situations where ready-made market solutions do not meet user expectations. The author used this tool to create six models within a few days, including plant diagnostics, character generation, and camera angle recognition tools, among others.

The process relies on precise prompting of the agent, which plans the work, requires setting a budget before starting paid tasks, and publishes the results on the Hugging Face platform. For example, creating the citrus-disease-vlm-instruct model based on Qwen3.5-2B cost approximately $1.90 and increased plant disease diagnosis accuracy from 14.9% to 52.8%. Meanwhile, creating a lightweight, 0.8B prompt rewriter version running on a CPU and consuming a quarter of the 9B model's tokens cost a total of $16. Each project required defining a dataset, a base model, and a training script. The agent also performs smoke tests to verify whether model weights have changed before investing a full budget into the training process.

For AI Builders: A developer or hobbyist building custom models can use ML-Intern to automate dataset preparation and fine-tuning processes. Before deploying this solution, one should check whether the source data licenses used by the agent permit commercial model publishing and verify actual operational costs in the chosen cloud environment, which this mechanism may not fully anticipate.

OpenAI Disrupts Influence Operations

OpenAI reported on October 8 that it detected and disrupted two artificial intelligence-enabled influence operations. These activities involved the use of fake journalist profiles and a fictional think tank. The goal of these activities was to spread geopolitical content.

For AI Builders: Teams responsible for security and content moderation on social platforms should implement mechanisms to verify the authenticity of published content authors. The source does not specify what exact detection techniques were used, nor whether public tools for identifying such operations were developed—this information requires separate determination before implementing custom monitoring systems.

LegalOn Optimizes Codex Model Costs

LegalOn reported on October 8 that it reduced its daily Codex usage costs by 65 percent. This was achieved while maintaining the previous pace of development work. This success stems from the strategic matching of specific models—Astra, Sol, and Luna—to individual tasks, as well as effective budget management.

For AI Builders: Teams utilizing multiple language models can test assigning different variants to specific, smaller tasks instead of using a single model for everything. The source does not explain how the models were technically matched to tasks, nor what the detailed licensing conditions of these solutions are—it is worth verifying this before deploying the strategy in your environment.

Frequently Asked Questions (FAQ)

What languages does the Falcon ASR model support?

Falcon ASR has been optimized for the Arabic language, with a particular focus on the Emirati dialect. The model also supports English, French, Spanish, and Portuguese. The system processes all listed languages using the same weights, allowing transcription without requiring a language flag.

What savings did LegalOn achieve through model optimization?

LegalOn reduced its daily Codex usage costs by 65 percent. This was achieved through the strategic matching of specific model variants—Astra, Sol, and Luna—to individual tasks, while maintaining the previous pace of development work.

Can users test GeForce NOW performance before subscribing?

For individuals wishing to test the cloud's capabilities, one-day passes have been made available. The cost of such a pass can be deducted from the fee for the first full service subscription.

Sources

• Rally Up: ‘Gears of War: E-Day’ Launches on GeForce NOW — 2026-10-08
• Introducing Falcon ASR — 2026-10-07
• The model that didn't exist, so you made it yourself — 2026-10-08
• Disrupting AI-enabled “false front” operations — 2026-10-08
• LegalOn halves Codex costs while maintaining development speed — 2026-10-08

--- Text prepared automatically by the aipoc-lab laboratory (AI) based on publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.