Need to know
- Thinking Machines Lab released Inkling, a 975B-parameter open-weight frontier model under Apache 2.0, directly challenging closed-model incumbents.
- xAI open-sourced Grok Build's 844K-line Rust codebase but left repo-exfiltration code intact, gated only by a server-side flag.
- Anthropic and Blackstone launched Ode, a $1.5B joint-venture AI implementation firm targeting enterprise deployment at scale.
- Nvidia shipped Cosmos 3 Edge and Jetson Thor T3000/T2000, putting frontier physical AI inference on edge hardware for the first time.
- Atlassian updated Jira to orchestrate coding agents directly from tickets, citing a 65% rise in AI tool usage with only 10% speed gains.
New Releases
Thinking Machines Lab released Inkling, a 975 billion parameter mixture-of-experts model, under Apache 2.0 on July 15, making it the largest permissively licensed Western frontier model available.
- Trained on 45 trillion tokens with a 1M-token context window, Inkling handles text, audio, images, and video; a smaller Inkling-Small variant runs with 12B active parameters for cost-sensitive deployments.
- Benchmark reality check: Early testers including Ethan Mollick report the model fails tests passed by every frontier model since DeepSeek R1, placing it below leading Chinese open-weight models despite the headline parameter count.
Nvidia announced Cosmos 3 Edge, a 4B-parameter on-device world model, alongside Jetson Thor T3000 and T2000 compute modules, enabling robot policy generation without cloud connectivity.
- Japan's industrial stack committed: FANUC, Yaskawa, Kawasaki Heavy Industries, Fujitsu, Sony, and six others intend to join the Nvidia Cosmos Coalition, signaling physical AI is becoming a procurement category in manufacturing.
- Cosmos 3 Edge runs on Jetson Thor at the edge, pairing vision reasoning with robot policy generation in a single Nemotron-based model that fits industrial deployment constraints.
Atlassian shipped a Jira update that lets teams assign tickets directly to Claude Code, Cursor, and GitHub Copilot, with a Teamwork Graph supplying Jira and Confluence context to the agent.
- The productivity gap is the pitch: Atlassian cites internal data showing 65% growth in AI tool usage among engineers but only 10% speed gains, and frames the Teamwork Graph as the missing enterprise context layer.
- Ticket to PR loop is now automated: The update closes the handoff between spec writing in Confluence, ticket creation in Jira, and pull-request generation by a coding agent without re-keying.
Clay added open-weight model support to Claygent, its AI research agent, giving GTM teams a cost-effective option for long-running prospecting and enrichment tasks.
- Cost asymmetry is the use case: Open-weight models reduce per-run costs on high-volume, long-horizon tasks like list enrichment where frontier model quality is not required.
- No new model is bundled; teams select from available open-weight options, which means the benefit scales with model quality improvements over time at no additional Clay-side cost.
Funding
Anthropic, Blackstone, Hellman and Friedman, and Goldman Sachs co-founded Ode, a $1.5B AI implementation company, signaling that frontier labs now believe deployment consulting is a durable margin pool worth owning rather than ceding to partners.
Reo.Dev closed an $11.3M Series A led by Elevation Capital to expand US sales and marketing operations, a bet that AI-native revenue intelligence built outside Silicon Valley can capture enterprise GTM budgets shifting away from legacy platforms.
Case Studies
Wiz Research found that one in five applications generated through AI-assisted "vibe coding" platforms contains security vulnerabilities, including hardcoded credentials and exposed secrets.
- Lovable was the primary platform studied; Wiz analyzed millions of apps created via natural-language prompts, finding patterns like hardcoded passwords in variables and unsanitized inputs that bypass standard code review.
- The attack surface is novel: Unlike traditional codebases, vibe-coded apps are created by builders with little security background, meaning vulnerabilities are systematic rather than incidental.
Trending on X
- Inkling early quality concerns Ethan Mollick and others are posting failed benchmark results for Thinking Machines' Inkling, with Mollick noting it falls short of frontier Chinese open-weight models on even basic tests, dampening the launch excitement around the 975B parameter count.
- xAI Grok Build open-source controversy Developers are pointing out that xAI's Apache 2.0 release of Grok Build is effectively read-only open source: one bot commit, no external contributions accepted, and the repo-exfiltration code that leaked SSH keys is still present in the binary, just disabled server-side.
- Dimon's 'ballistic missiles' Mythos warning Jamie Dimon's quote comparing broad access to Anthropic's Mythos model to distributing ballistic missiles is circulating widely, reigniting debate about whether frontier model access should be gated by government review rather than commercial pricing.
- Token allocation as unpredictable surprise Mollick's observation that AI providers keep silently increasing token limits is resonating with enterprise buyers who note that unpredictable context windows make capacity planning feel like gambling rather than engineering.
- Kimi K3 imminent frontier open-weight release Moonshot AI's Kimi K3, reportedly a 2-3 trillion parameter model, is surfacing under the name 'KIVINE' on benchmark arenas ahead of an official launch, and the community is watching whether it will leapfrog Inkling as the dominant open-weight frontier option before the week ends.