AI — Page 36

On July 23rd 2026, Dev released a benchmark that measures how well large language models can operate AutoCAD via computer-use.

The new models deliver higher‑fidelity image generation and faster, cheaper speech synthesis for Microsoft products.

Google's first ATLAS report, based on 15 million de‑identified interactions, offers the most comprehensive look at how users adopt its AI tools.

The U.S. rollout lets adults connect Apple Health and supported medical records, with permission controls and a promise that connected data will not train foundation models.

The open-source setup serves the 753-billion-parameter model with a 248,000-token window, though the hardware costs more than $14,000.

The AI veteran introduces a community-driven tool that can draft documents, send Slack messages and update calendars, pushing open-source agents toward production-level tasks.

Microsoft's Superintelligence team released specialized MAI models that claim to match GPT‑5.6 in Excel and beat GPT‑5.4 Mini in Copilot while using fewer tokens.

Shensi Ding says a panel of cheaper models matched Claude Fable 5 on a 25-task research test at one-quarter of the cost.

Video enters limited early access first, while image generation, open weights and robotics access will follow in phases.

Louis Beaumont's 2024 project has 20,000 GitHub stars, but commercial use of its once-MIT code now requires a paid license.

The new text‑to‑speech model adds inline tag steering for emotions and breaths, and supports free‑form natural‑language prompts.

Founder Justin Haubrich is turning a $24.99 Mac utility into a consumer data-discovery tool that VaultSort says makes zero network calls.

Independent tests put it level with 3.5 Flash on intelligence, with half the task time and an 18% lower cost per task.

Chris Best is betting that reader-triggered Pangram scores and creator disclosures can preserve authorship without banning AI-assisted writing.

Alexander Borzunov and collaborators' 2022 research project trades cloud control for model access, while privacy, latency and incentives remain constraints.

Benjamin Davis and Philip Mims spent $99.59 on 650 persistent text-world runs across 13 models, then treated judge instability as the result.

Cognition builder Jared Zoneraich ran roughly five Ultra-mode agents per problem, with one counterexample checked in Lean.

A Reddit discussion shows heavy Claude users moving to Kimi's K3 model after pricing and openness concerns.

Ayush Agarwal's team is offering developers a persistent canvas, shared browser and permission layer for agents that work alongside people in real time.

Jeff Cui shows that a frozen ViT backbone plus a tiny transformer can outperform a 7B‑parameter VLM in robot control.

Scott Winters says GPT-4o minimized symptoms before his July 2025 ICU admission; the complaint also seeks to halt ChatGPT Health.

The open 5B model generates 720p video at up to 16 FPS, though AMAP's default FP8 configuration reached 13.3 FPS.

The MIT-licensed Mac app combines chat, dictation, meeting notes and an agent, with remote processing handled through a verifiable confidential backend.

The iOS app offers a basic panel at no cost and a $69‑year premium that adds deeper metrics, promising AI‑driven life‑extension plans.