All Issues
Jul 13 - Jul 19, 2026

AI Weekly: Kimi K3 Designs Its Own Chip, Grok Build's Secret Uploads Exposed

Models & Releases

2 stories

Inkling: Thinking Machines Lab's 975B Open-Weights Multimodal MoE

  • Thinking Machines Lab released Inkling (MIT licence) — 975B total / 41B active MoE, pretrained on 45T tokens of text, images, audio, and video, with a 1M-token context window.
  • Key differentiator: controllable reasoning effort (0.2–0.99 sweep); at mid-effort it matches Nemotron 3 Ultra on Terminal-Bench 2.1 at a third of the tokens; at max effort scores 77.6% SWE-Bench Verified and 97.1% AIME 2026.
  • Available for fine-tuning on the company’s Tinker platform; bootstrapped from Kimi K2.5 synthetic data; companion Inkling-Small (276B/12B active) matches it on most benchmarks at lower cost.

People & Business

3 stories

Anthropic Launches Rupee Pricing for India, Its Second-Largest Market

  • Anthropic rolled out INR-denominated Pro, Max, and Team plans in India on July 13 — the first non-US market to get local-currency pricing, with GST included at checkout.
  • India is Claude’s largest market outside the US; UPI payment support is not yet enabled, and dollar-converted rates are marginally higher than US prices.
  • The move mirrors OpenAI’s and Google’s earlier localisation pushes in South and Southeast Asia as frontier labs compete for the world’s most populous market.

Helsing Raises $1.8B Series E at $18B for Defence AI

  • European defence AI company Helsing raised a $1.8B Series E — oversubscribed — valuing it at $18B and cementing its position as Europe’s answer to Anduril.
  • The round will fund expansion of AI-driven military systems and a first US manufacturing base planned for West Virginia.
  • The raise arrived the same week as the White House GOLD EAGLE initiative (below), reflecting record VC appetite for militarised AI infrastructure.

Policy & Ethics

3 stories

xAI Open-Sources Grok Build Under Apache 2.0 After Controversy

  • 72 hours after the secret-upload story broke, xAI released the entire Grok Build codebase under Apache 2.0 and removed all usage caps — a trust-repair move that came with no formal security advisory.
  • The open-source release revealed the upload code is still present in the binary; community contributions are explicitly blocked despite the Apache licence.
  • Elon Musk stated all previously uploaded user data would be ‘completely and utterly deleted’; no third-party verification has been provided.

White House Launches GOLD EAGLE AI-Cybersecurity Clearinghouse

  • The Trump administration announced GOLD EAGLE on July 14 — a formal clearinghouse requiring AI developers and critical infrastructure operators to share information on cybersecurity vulnerabilities detected by advanced AI systems.
  • The initiative includes open-source AI model developers and is led by White House cyber director Sean Cairncross; it fulfils a commitment in Trump’s earlier AI executive orders.
  • GOLD EAGLE arrives the same week the Grok Build incident exposed how AI coding tools can exfiltrate sensitive infrastructure secrets, putting the policy context in sharp relief.

Products & Hardware

1 stories

Research & Resources

2 stories

OpenAI Scorecard: Measuring AI Value as Useful Intelligence per Dollar

  • OpenAI published a framework for CFOs and enterprise AI buyers on July 17: ‘Useful Intelligence per Dollar’ — measuring work completed, cost per successful task, reliability, and scaling returns rather than cost-per-token.
  • The piece argues the lowest price-per-token often isn’t the lowest cost-per-outcome: a frontier model completing a task in one pass beats a cheaper model requiring retries, latency, and human review.
  • The GPT-5.6 Sol/Terra/Luna tier family is positioned explicitly as the implementation of this scorecard, with Luna for volume, Terra for depth, and Sol for maximum reasoning value.