Open-weight models keep closing the gap on frontier labs: Moonshot's 2.8-trillion-parameter Kimi K3 reportedly topped Arena's coding leaderboard the same day CNBC reported Google's Gemini 3.5 Pro has hit delays, while Fireworks AI's valuation jumped to $17.5B on demand for cheaper inference. Anthropic and Blackstone's Ode venture is making a bet that 'implementation' - not raw model capability - is the next trillion-dollar AI business, a thesis echoed in Anthropic's own writeup on using Claude Code for large-scale migrations. Agent tooling matured on multiple fronts: 1Password shipped scoped credential access for Claude, LM Studio launched an agent for local/open models, and a sharp post argues most teams' agent evals don't predict production failures. On the risk side, Grok's CLI was caught uploading local files to the cloud and researchers detailed an IoT botnet built with LLM-assisted development, while AWS had another CloudFront outage - a reminder that infra fragility and AI agent trust issues are now running on parallel tracks.
- Kimi K3's 2.8T-param model reportedly tops Arena's coding leaderboard
- Gemini 3.5 Pro hits delays as Fireworks AI valuation hits $17.5B
- Anthropic and Blackstone bet AI 'implementation' beats models as a business
- Grok's CLI was caught quietly uploading local files to the cloud
- AWS CloudFront outage served errors instead of websites again
Model Race: Open Weights Close the Gap
AI Goes to Work: Implementation & Agents
Agent Tooling & Access ControlSecurity & Infra ReliabilityProduct Launches