AI Engineering Signal #83
Kimi K3 open-weight model releases, benchmarks at Claude Opus 4.6 class, priced at Sonnet tier.
Signals
Thinking Machines Inkling: 975B Apache 2.0 multimodal open model
permissive-license multimodal baseline shifts; audit vendor lock-in assumptions before next procurement cycle.
Latent Space
Schema Harness scores ~99% on ARC-AGI-3 public benchmark
ARC-AGI-3 is no longer a reliable deployment gate; replace it in your eval suite now.
Web
Ring-Zero scales zero RL training to one trillion parameters
trillion-parameter RL is documented; revisit compute budget assumptions for reasoning model training.
ArXiv
Android lock screen bug lets Gemini send SMS without PIN
agentic assistants with messaging access are authentication bypass vectors; audit permission scopes on deployed agents.
Web
EEG confirms brain encodes two speech streams simultaneously
dual-stream parallel processing in biological intelligence has empirical support; relevant to multi-agent attention routing design.
Web
The Take
Open-weight frontier models are now matching closed-API flagship performance at Sonnet-tier pricing, and the gap is closing faster than most production cost models assumed. The Gemini SMS bypass is a live reminder that permission design is already lagging capability in shipped products.
Subscribe
Related Signals