All Stories — Page 103

Daybreak Red limits the purpose-trained model to approved researchers, while Daybreak Blue opens GPT-5.6 Sol for broader defensive work.

Stoa Markets' YC S26 founders want firm bids and transaction data to replace GPU deals conducted through calls, spreadsheets, and group chats.

The logged-in lounge automatically posts new stories and gives readers a shared room with RuntimeWire staff and other technology followers.

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.

OpenAI's unreleased model showed strong cyber performance, prompting isolated testing and universal monitoring while OpenAI continues its assessment.

The co-founders are testing fixed court hardware through NBA Launchpad, where a broader league partnership depends on the pilot results.

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

CTech reports that Decart's founders could sell less than three months after the company raised $300 million at a $4 billion valuation, although no buyer is confirmed.

Founded by Adi Singh and Sven Myhre, Ara is a two-person YC startup that documented its Cloud Coding Agent on July 7th, with isolated sessions, stored context and human review.

The Claude-powered assistant found missing authorization checks during a routine booking task, then acted without approval.

The API gateway supports up to 50 image, video and audio references per request, but routes the model through a single provider.

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.

Irregular co-founders Dan Lahav and Omer Nevo built an independent AI security lab whose evaluation infrastructure became part of the risk it was designed to measure.

Meta documents that machine-wide user rules always load. RuntimeWire captured the resulting provider request and examined what that default means for personal files written for competing AI clients.

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

The alpha service stores failed routes, computations and Lean proofs so research agents can inherit one another's work.

MiniMax says it plans to release H3's separate 2K regeneration model and sparse-attention reference code, giving developers more of its video-generation stack to inspect and test locally.

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

The 256GB Day One player presents owned music with sleeve and booklet art. Sleevenote says preorders are closed, though The Verge reported a final handful available on August 8th.

Craig Cheney's MIT-trained team is connecting robotic work and verification directly to manufacturers' production systems.

Daso's pitch is that child safety should be the product architecture, not a parental-control setting: creative tools, a constrained AI tutor and no browser, app store or feed.