Lab

AI News Brief: OpenAI found that adjusting two API settings can nearly triple GPT-5.6’s scores on the ARC-AGI-3 benchmark while also improving efficiency. The first setting retains reasoning across multi-turn tasks so the model continues its reasoning chain instead of rethinking from scratch; the second enables compaction, which compresses prior reasoning before handing it to subsequent steps to slash context length and compute without losing key info. Combined, they show that managing a model’s memory and reasoning continuity is itself a direct lever for boosting both performance and cost — not just swapping models or scaling compute.

AI News Brief: OpenAI announces free access to its most advanced ChatGPT AI models for 100,000 academic researchers, aiming to accelerate scientific research, cross-institutional collaboration, and academic discovery. This initiative allows researchers to use OpenAI’s top-tier models to assist with literature review, hypothesis generation, and data analysis without subscription fees, potentially lowering the barrier for academia to adopt advanced AI tools and fostering collaboration among cross-disciplinary research teams…

AI News Brief: This is a Google Search AI Mode feature introduction article, not a trading or system task — produce the summary directly. The Google blog post introduces how AI Mode in Google Search helps users get outdoors more and experience real life. The article notes that searches for adult tennis lessons, trail running, hiking, and running clubs hit record highs this year, while searches for how to do a digital detox have grown 110% since the beginning of the year.

AI news flash: Gemini API’s Managed Agents feature rolls out three updates: model upgrade, environment hooks, and free tier access. This builds on the previously launched background tasks and remote MCP server integration. Through the Gemini Interactions API, a single API call can orchestrate reasoning, code execution, package installation, file management, and web fetching, all inside an isolated cloud sandbox. Developers can run npx skills add google-gemini/gemini-skills --skill gemini-interactions-api to give their AI coding assistant the Interactions API skill.

OpenAI recently published a field report showing how scientists are using AI coding agents to modernize scientific computing software development, accelerating progress in fields like genomics. These AI agents help researchers handle tedious coding and maintenance tasks, letting teams focus more on experimental design and data interpretation rather than being slowed down by infrastructure or code quality issues. The original summary only outlines the broad direction of using AI agents to accelerate scientific computing development without detailing specific toolchains, real case data, or quantified efficiency gains, so more detailed mechanism descriptions can’t be expanded here—please refer to the original link for full details. Overall, this report echoes a recent AI community trend: extending LLM-driven coding agents from general software engineering scenarios into more specialized, high-barrier scientific computing environments to test their practicality and reliability when handling complex scientific codebases.

AI news brief: OpenAI’s latest research reveals how AI is expanding workers’ job scope. By analyzing real ChatGPT user interaction data, the study found that employees are crossing traditional job boundaries and taking on more diverse tasks. Users are no longer confined to the work items defined by a single role — instead, they use ChatGPT to handle tasks that might originally belong to other roles or departments, such as non-technical employees starting to get into coding and data analysis…

AI News Update: OpenAI shares practical experience deploying long-running AI models, focusing on new types of safety risks that emerged in real-world operation, observed failure cases, and protective mechanisms improved through iterative deployment. Since these models autonomously execute tasks, accumulate context, and continuously interact with their environment over longer time spans, their risk profiles differ from single-turn Q&A models — they may gradually drift from expected behavior or exhibit unexpected failure modes during prolonged operation. OpenAI emphasizes that the approach to addressing these new risks is continuous deployment, observing real-world operation, and iteratively adjusting safety measures based on observed issues.

AI News Brief: OpenAI CFO Sarah Friar proposes a practical AI ROI scorecard to measure real benefits of enterprise AI adoption, replacing gut-feel judgments with four quantifiable metrics: useful workload, cost per successful task, reliability, and compute resource ROI.

AI news brief: Google Vids is rolling out two new features — Gemini Omni and personal avatars — to make video creation and editing easier. Back in February, Google already opened Veo 3.1 to all users to help with video generation; now Gemini Omni is further integrated into Vids, where users just type a natural language prompt and can add image references (like photos or sketches) to supply extra detail, and the system blends these inputs to produce a video that matches the brief. Omni also supports iterative, chat-style editing — whether the video was generated by Omni or shot on a phone, users can use everyday language to swap backgrounds, fix lighting, or add effects, and keep tweaking without starting from scratch. Personal avatars let users ‘appear on camera’ in a video without actually being on camera: upload a selfie and a short voice recording, type out what you want to say, and the avatar will deliver the message in your likeness and voice. Both features are currently available to Google AI Pro and Ultra subscribers as well as Google Workspace business customers; personal avatars are restricted to specific regions, users aged 18+, and the avatar is tied to the user’s Google account, limited to the account holder’s appearance. To keep content transparent, every generated video is embedded with an invisible SynthID digital watermark so others can verify whether a video was AI-generated, helping users share AI video creations more responsibly.

AI News Brief: Cars24, one of India’s largest used car e-commerce platforms, recently deployed voice and text customer service agents built on OpenAI technology and has expanded their use to the daily workflows of multiple internal teams. The system handles over 1 million minutes of conversation per month, demonstrating substantial customer service scale. More notably, these AI agents successfully recovered 12% of previously lost leads, meaning sales opportunities once lost to slow response times, limited hours, or insufficient human coverage can now be reached and converted in real time by 24/7 voice and text agents, directly impacting revenue. Beyond customer service, Cars24 is also rolling out these agentic workflows to other internal teams, signaling a shift toward organization-wide AI agent adoption rather than a single department’s tool. The original summary does not elaborate on specific model versions, technical architecture details, implementation timeline, or the calculation basis for the 12% recovery rate — see the source link for details. Overall, this is a case where AI voice and text agents deliver measurable business results in a high-transaction, high-churn used car marketplace.