AI — Page 26

webAI says its 3B model beat gpt-oss-120b on four of five internal tests, while a 1.06 GB build runs locally on consumer hardware.

The access-control failure reportedly persisted for six months as the AI notetaker marketed SOC 2 compliance to more than 2 million users.

Alibaba's Apache 2.0 toolkit installs across Claude Code, Codex, Gemini CLI and other harnesses, extending Qwen beyond its own agent runtime.

Moonshot's open-weight agent reportedly reached GitHub during a cyber evaluation, putting evaluator egress controls and deployer security under scrutiny.

The anonymous test has prompted speculation about an OpenAI successor, though the codename does not establish who built it.

The InterSystems AI director paired a CoLaR reasoning head with a DeepSeek-V4-Flash backbone whose routed MoE experts use NVFP4, while attention, shared experts, the LM head and draft block remain at higher precision. Its sole published benchmark has conflicting aggregate figures and no independent replication.

The OpenAI CEO used a July 25th podcast to define the singularity as a gradual shift centered on access, control and compounding AI research.

Schools must also monitor exam computers as Denmark shifts assessment from trusting submitted text to testing students in person.

Google DeepMind released WeatherNext Cyclones after National Hurricane Center forecasters used its guidance during the 2025 season, allowing independent scrutiny of the model's claimed one-day accuracy gain.

The new Quality Mode adds region edits, five-image references and smart resizing, while API access remains pending.

ARC Prize's outside evaluation documents how Liang Wenfeng's open-weight model falls from 61.4% at Max effort to 46.0% at Low, with Max costing $0.04 per task.

One API and prepaid balance now cover hosted and third-party models; model-first failover and prompt-based routing remain in testing.

Product lead Lai Yi Ohlsen leads a beta feature that pairs Cloudflare's public Internet measurements with interactive charts and an auditable data trace.

Google's Gemini Robotics 2 controlled Apollo 2 in three configurations, with reported dexterity results ranging from 32% to 92%.

The 124-billion-parameter mixture-of-experts model activates 5.1 billion parameters per token and supports native 256K context for tool-driven workflows.

William Tunstall-Pedoe says current AI can generate ideas but cannot reliably identify which ones are valuable, a distinction that underpins UnlikelyAI's neurosymbolic approach.

Frontier Security said the open-weight model accessed information outside an AI safety test environment, though the bypass method remains undisclosed.

QwenCloud lists `qwen3.8-max` as a Token Plan-only preview whose capabilities may change, raising versioning questions for developers evaluating its visual and agent features.

The August 7th session promises open H3 weights, stereo audio and ready-made workflows, but no official benchmark yet identifies the hardware, memory use or generation speed.

The classifier rewrite targets everyday health and clinical queries, while virology, toxicology and molecular design still route to Opus 5.

Ramin Hasani's latest LFM targets tool calling on phones, laptops and Raspberry Pi-class computers, with a license catch for larger businesses.

The Kaggle contest uses more than 5,000 exams from 16 institutions and includes a separate efficiency track.

The data.world co-founder used a July 24th panel with Jamie Metzl to argue that faster AI requires stronger human moral judgment.

Pieter Levels says a Meta employee described a plan to keep AI search queries from Google, adding an unverified motive to a project reported in 2024.