alphasignal openai resets kimi coding rl sim

Why this is in the vault

Kimi K2.7-Code's 30% reasoning-token reduction is a direct cost lever for RDCO's agentic loops (reasoning tokens bill as output at $4/1M). The RL-to-physical-robot pipeline is a clean architecture reference. The agent skill security scanner item cross-validates RDCO's existing pre-install SOP with a new class of tooling.

Issue contents

TOP NEWS

• OpenAI Codex: banked rate-limit resets (21.6K likes) — Codex users on any paid tier now get one immediately-banked reset instead of a fixed 3am auto-fire. Invite up to 3 friends to Codex and both parties earn an additional banked reset when the friend first messages; up to 4 total, each expiring in 30 days. Third-party referral mechanics are new for OpenAI tooling. (Third-party news item)

• Kimi K2.7-Code open-sourced, 30% fewer reasoning tokens (10.6K likes) — Moonshot AI released Kimi K2.7-Code targeting "overthinking" reduction. Benchmark gains vs K2.6: +21.8% on Kimi Code Bench v2 (50.9 → 62.0), +31.5% multi-language (26.7 → 35.1), +11.4% on MCP Mark Verified tool-use (72.8 → 81.1). For frontier context, GPT-5.5 scores 69.0 and Claude Opus 4.8 scores 67.4 on Kimi Code Bench v2 — K2.7-Code (62.0) trails both on that metric. Architecture: 1T total / 32B activated MoE, 384 experts, 256K context. Pricing: input cache miss $0.95/1M, output (reasoning + response) $4.00/1M. The 30% token reduction mechanism is undisclosed — Moonshot claims it but no architectural explanation is published; likely RLVR training with token-length penalties or inference-time budget-forcing. Modified MIT license; weights at moonshotai/Kimi-K2.7-Code on HuggingFace. OpenAI/Anthropic SDK compatible. (Third-party news item)

TOP REPO

• RL + Codex → living creature sim deployed to physical hardware (6K likes) — A self-righting robot trained entirely in MuJoCo via reinforcement learning, then deployed on a 3D-printed robot with 2 servo motors and an M5Atom microcontroller smaller than a matchbox. The trained RL policy was compiled to a plain C header file and flashed to the chip — no cloud, no GPU at inference time. Clean example of sim-to-real transfer at micro-hardware scale. (Third-party repo)

SIGNALS

• 140+ expert AI agent prompt library — 810K GitHub stars (5.3K likes) — Curated open-source prompt collection for AI agents. Scale of adoption suggests it's a de-facto community reference. (Third-party)

• Open-source tool scans AI agent skills for malicious code before install (2.5K stars) — Security auditor for AI agent plugins/skills. Multiple tools in this class emerged in 2026: NVIDIA SkillSpector (SKILL.md-format targeting, 64 detection patterns across 16 categories including prompt injection, data exfiltration, MCP tool poisoning), Cisco AI Defense Skill Scanner (SARIF output, CI/CD build-gate), and Snyk Agent Scan (formerly mcp-scan by Invariant Labs, 13-platform auto-discovery but requires cloud-side analysis). (Third-party)

• New math method for joint embedding of mismatched high-dimensional datasets (tagged "Must Read") — Alignment approach for structurally dissimilar datasets. Lower RDCO relevance unless applied to multi-modal agent memory. (Third-party)

• LLMs hallucinating the same fake people repeatedly, polluting academic databases (1.3K likes) — Consistent fake identities appearing in academic citation databases. Relevant for any RDCO use case relying on LLM-generated citations or contact research. (Third-party)

• Goodfire: audit and fix what your model learns before training starts (754 likes) — Pre-training run behavior auditing and correction tooling. Low immediate RDCO relevance (no fine-tuning in current stack). (Third-party)

⚠️ Sponsorship

Three sponsor blocks present:

1. Slack — enterprise AI search agent across Slack, Google Drive, Gmail, Outlook, and 2,600+ connected apps
2. Entelligence — "Entelligence Wrapped" tracks AI coding tool usage (Copilot/Cursor/Claude), surfaces usage patterns and spending leaks
3. Checksum — API + E2E Playwright tests that self-heal, running in CI

None of the sponsored items overlap with the editorial content above.

Mapping against Ray Data Co

High relevance — Kimi K2.7-Code token reduction: RDCO runs Claude as a continuous COO agent. Reasoning tokens in Claude's extended thinking modes bill as output (~$15/1M for Opus-class). Kimi K2.7-Code at $4/1M output with 30% fewer reasoning tokens than its predecessor is a credible cost-reduction alternative for coding-heavy sub-agent tasks (pipeline authoring, code review, test generation). Caveats to weight before any trial: (1) the 30% reduction mechanism is unverified — vendor benchmark only; (2) Kimi Code Bench v2 benchmarks are Moonshot's own eval suite, not independently reproduced; (3) thinking mode is hardcoded on (can't disable), and temperature/sampling are not user-adjustable, limiting fine-grained prompt control; (4) K2.7-Code trails GPT-5.5 and Claude Opus 4.8 on the same benchmark. Worth a structured cost/quality trial against Claude Sonnet for pipeline code authoring tasks, not a drop-in swap.

Medium relevance — agent skill security scanning: RDCO has an existing pre-install security SOP (02-sops/2026-05-02-mcp-plugin-skill-install-security-review-sop.md). The emergence of SkillSpector (NVIDIA), Cisco AI Defense, and Snyk Agent Scan suggests the SOP can be augmented with automated tooling. SkillSpector is most directly applicable given its SKILL.md-format targeting and local-only analysis (no cloud data exfiltration risk). Flag for next SOP revision.

Low relevance — RL robot sim: The sim-to-real pipeline is architecturally elegant and a good reference for understanding compiled RL policy deployment, but no immediate application to RDCO's current stack.

Low relevance — OpenAI Codex banked resets: RDCO runs Claude Code as primary. The referral mechanic is interesting as a product engagement pattern but not directly actionable.

Watch — LLM hallucinated fake people: Affects any RDCO workflow using LLM-assisted contact research or citation generation. Low urgency but worth noting as a known failure mode.

Related

• [[06-reference/2026-06-11-data-engineering-central-agentic-token-tax-opencode-ollama]] — token cost framing for agentic loops; direct complement to the Kimi cost-reduction angle
• [[06-reference/2026-05-02-moonshots-ep252-google-anthropic-gpt55-cloud]] — Kimi K2.6 deep-dive: architecture, training cost, benchmark vs Opus; baseline for K2.7 comparison
• [[06-reference/2026-04-22-alphasignal-ai-news-roundup]] — Kimi K2.6 original mention and model landscape context
• [[02-sops/2026-05-02-mcp-plugin-skill-install-security-review-sop]] — RDCO's pre-install security SOP that the new scanner tooling could augment