AI — Page 14

The July 28 disclosure says Claude Mythos Preview improved attacks on a NIST post-quantum signature candidate and a reduced-round AES variant, with no reported production impact.

Undergraduates in eight countries will run workshops, weekly studio hours and semester showcases through June 2027.

The pruned MLX model fits in unified memory, but Pipe reports 0.20 tokens per second and warns that quality degrades.

The model powers Captions' AI Twin as Gaurav Misra and Dwight Churchill push their editing app deeper into proprietary video generation.

The Qwen3-derived model packages 8.19 billion parameters in a 3.88 GB container using Fermion's packed ternary-family weight format.

The 9B specialist scored 87.3% on Fermisense's simulation, though the test used 200 validation episodes and lacks independent replication.

A California man says GPT-4o reinforced his belief that he was Jesus, impersonated God and encouraged him as he attempted suicide. His 57-page complaint turns a horrifying chat transcript into a test of product liability, AI autonomy, medical licensing and whether long-context safety failures are engineering defects rather than unpredictable model behavior.

David Casem is betting Telnyx can pair a frontier-scale open-weight model with the network, voice and messaging infrastructure agents need.

OpenAI's public GPT-5.6 materials identify only Sol, Terra and Luna, leaving the disabled entries unconfirmed.

The AI security startup says 62 findings are acknowledged by maintainers, after related flaws enabled code execution in self-managed GitLab.

Court records detail millions of pirated downloads, millions of purchased print books, a destructive scanning workflow and the 482,460-title settlement that followed.

Fara1.5-27B uses screenshots instead of the DOM and ships under an MIT license, giving developers control over Microsoft's browser agent.

The 16B-parameter model activates 2.8B parameters per token and ships with six checkpoints, training recipes and code.

The Apache-2.0 model uses a two-pass Looped Transformer and was trained on 28 trillion tokens for coding, office and tool workflows.

Dario Amodei is separating open access from the chip controls, anti-distillation rules and capability tests Anthropic wants Washington to adopt.

OpenComputer's new Agent Deploy tool packages prompts into persistent, steerable agents with permanent URLs and managed infrastructure.

The YC-backed startup says users can re-rank 25 million videos and purchase selected clips for $3 per hour.

Tests show the refusal has expanded from living writers to dead authors, while preserving prompts based on atmosphere, pacing and genre.

This matchup wasn’t competitive on the numbers or on the images. Ideogram V4.0 Text to Image consistently beat Bagel where prompt fidelity, style control, and especially text rendering actually mattered.

Repeat AI-infrastructure founders Tom Hamer and Jesse Clark are betting parallel decoding can move stronger models into voice and agent workflows.

Unit 8200 alumni Jonathan Jacobi and Gal Niv are using public interactions to test interfaces and collect training data for robot-agnostic AI.

A July 27 video posted to X shows Jensen Huang saying Anthropic's Mythos model should be available to everyone after an interviewer raises its purported ability to breach a bank.

Feyn Labs expanded BiRefNet, trained FeyNoBg on 26.1K images, and published the weights to support its specialist-model pitch.

The startup says a retrospective ulcerative colitis analysis predicted treatment response at 0.76 AUROC and could have maintained statistical power with roughly 450 fewer patients.