The Bulletin.
Notes from the team.
Kimi K3.1 Pro Looks Strongest Where Agents Do Real Work
Kimi K3.1 Pro looks most compelling for coding agents, long-horizon tasks, and 3D work, based on reported benchmark gains and a 1-million-token context window. Developers can try it through Kimi Code or the API, but should treat vendor benchmarks as evidence to test against their own workloads, not a guarantee of results.
Small Decision Models May Deliver More
A 1,085-decision comparison reports that the best open model scored below hosted Jev, but also that the much smaller Kev 4B tied Clef-Flash 9B closely. Those results challenge the assumption that parameter count alone predicts decision quality and give teams a reason to test compact models against their own workloads.
AI Finds Vulnerabilities Faster Than Teams Can Fix Them
AI-assisted vulnerability discovery is now producing validated findings faster than engineering teams can triage, patch, and deploy them. The gap is clearest in open source and AI security, where disclosure and remediation data show substantial backlogs, though noisy detection means discovery volume is not the same as urgent risk.
Jev 1.14 Makes Visual Decisions Possible
Jev 1.14 enables image-based decisions, object localization with bounding boxes, and opt-in 262k-token context, but TypeSafe labels these capabilities experimental rather than production ready. The practical path is to use the stable typed decision model for routing and classification while isolating visual and long-context workflows behind explicit evaluation gates.
Anthropic Replaced Git Because Its Workload Changed
Anthropic built Scion because its AI-heavy monorepo exceeded the operating limits of Git-based infrastructure, not because Git stopped being useful. Most teams should keep Git and measure their own bottlenecks before considering a custom source-control system.
Claude Can Reverse Engineer Software, Until Reality Bites
Claude can accelerate software reverse engineering, but it cannot turn behavioral imitation into a reliable product without human validation, testing, and substantial engineering work. The Adobe-inspired ArtCraft case shows the gap: Claude helped produce early-alpha look-alikes, not a finished replacement for Adobe’s software.
The “AI” Label Is Becoming a Product Risk
AI is simultaneously used as a research field, a product category, a marketing term, and a source of social claims about work, education, and inequality. This post argues that overloaded language creates expectations no implementation can reliably meet.