AI — Page 20

In a 12-task Editorial Craft evaluation of 12 AI models, OpenAI’s GPT-5.6 Sol Pro ranked first with a 0.97 score at $0.0091 per task. OpenAI claimed four of the top six spots, while Grok 4.6 and Claude Opus 4.8 tied at 0.93 among the leading non-OpenAI models.

The experimental editor treats persistent `.hz` pseudocode as a human-authored specification and regenerates source code from changes to it.

Alexandr Wang's MSL reports that its model trails its predecessor without tools and pulls ahead when paired with Meta's runtime.

George Kurdin is pairing a customer payment portal with Stripe distribution as Monk pushes finance teams to consolidate invoice-to-cash work.

OpenAI's opt-in timeline records selected activity from apps and websites, while admins must enable it for business accounts.

Hrant Davtyan's lab tested nearly 30 systems on 20.7 hours of Armenian audio across five speech datasets, with closed systems leading the combined ranking.

Arthur Mensch's Mistral says its retrieval loop lifted FinanceBench accuracy from 26.7% to 86% in a company-run test.

TrueFoundry reports lower costs than Claude Managed Agents, though its largest claimed saving also comes from switching models.

Core and Pro subscribers get credit-free everyday Agent tasks, with five-hour limits and up to 30 hours of monthly chat on Core.

The 2025 tax-year pilot saved accountants 31% of preparation time, according to Current, while keeping practitioners responsible for review.

Jeremy Morrell argues LLM-written extensions and sandboxed runtimes could serve workflows too specific for a conventional product roadmap.

The sale covers 100 million emails and 500 million Teams items; a union says stripping names leaves confidential employment content exposed.

Scott Wu's $26B coding-agent company stayed independent, while the two sides kept discussing a compute deal that echoes SpaceX's Cursor playbook.

Inco AI adds a path selector and local convolution to DFlash 2, reporting 21% longer accepted drafts with 1.3% added cycle latency.

The former NVIDIA AI research vice president is building Veeda with Zan Gojcic and Huan Ling around simulated environments for embodied agents.

Private Safety Processing will trace risk patterns across related API interactions while giving OpenAI only limited safety signals.

Daniel and Michael Han's post-training quants target local AI users, with the headline gains measured through Unsloth's own tests.

The three-model family extends self-scaffolding into task generation; its benchmark comparisons come from Ornith AI's own evaluation runs.

xAI's flagship model brings a 500,000-token context window to AWS at $2 per million input tokens and $6 per million output tokens.

The World Labs co-founder is betting on persistent 3D environments for creators, designers and robotics researchers.

Andrei Georgescu and Dan Huh's robotic lab says each HIVE can test 10,000 tissues, generating human biological data before clinical trials.

WIRED reports that OS Investigate lets police search vehicle movements and connected records by behavior, extending Garrett Langley's camera network beyond plate lookups.

The Apache 2.0 release includes Mojo's compiler and tooling as Qualcomm-owned Modular argues its AI platform can serve rival chips.

The project registers fixed GitHub snapshots, checks Lean proofs mechanically, and uses an LLM to compare formal statements with informal descriptions.