Why this is in the vault
Curated AI-news roundup where the lead item — Z.ai's GLM-5.3 gaining 50% coding performance and emergent cybersecurity skill purely from post-training, with zero architecture changes — is a clean data point for the "capability gains are decoupling from base-model scale" trend RDCO's harness thesis depends on.
Issue contents
- Top News — Z.ai GLM-5.3: Same base model as GLM-5.2, no architecture changes. Gains came entirely from post-training focused on working inside real coding environments rather than answering code questions. Results: 50% better coding performance, first among open models on Terminal-Bench and Agents' Last Exam, 1M token context, higher accuracy at fewer tokens (cheaper agent runs). Unplanned side effect: CyberGym score 84.5% (edging Mythos 5's 83.8%) and ExploitBench more than doubled (24.4% → 54.4%) — the model started reasoning across full exploit chains, not just spotting isolated bugs. Open weights pending safety review; access live now via GLM Coding Plan.
- Top Repo — bot-only RuneScape server: Open-source SDK running an all-AI-agent RuneScape server. When bots do all the labor, base goods go to near-zero value and in-game currency collapses; agents shift to barter, with scarce farmable items (runite ore, black dragon hides) becoming the de facto currency because they require real effort to acquire. Live leaderboard, adjustable tick speed, points at emergent multi-agent economic behavior with no human interference.
- Top Model — OrcaRouter's Qwen3.8-27B-Uncensored-FP8: Abliterated (refusal-circuit-removed) Qwen3 27B VLM for red-teaming; refusal rate drops from 64-99% to 0-6% with MMLU within 1.3 points of base, vision/tool-calling/262K context intact. Apache 2.0 weights on HuggingFace; framed explicitly as a red-team/safety-testing tool, not a deploy-as-is model.
- Signals: Hermes ships a
/loopcommand for scheduled prompt reruns (agent heartbeat without cron); Granola launches on Apple Watch; a new tool fine-tunes an 8B model on a 4GB laptop GPU; Inherent's 27B model beats Claude and GPT-5.5 at replicating research papers by being "a smarter orchestrator" (not a bigger model); a Qwen3 27B checkpoint shrinks 55GB→20GB at 386 tok/s; Talkify is a free open-source macOS dictation app at 123ms voice-to-text.
⚠️ Sponsorship
Three sponsor slots in this issue: Google Cloud (WPP physical-AI training case study, mid-newsletter ad block), Datadog (an "AI pilot to production" ebook CTA), and Granola (Apple Watch launch folded into the Signals list as a labeled sponsored item). None touch the editorial framing of the GLM-5.3 or RuneScape items — the sponsor content is clearly boxed off — but the Granola item sits inside the unpaid Signals list with only a small "Presented by Granola" tag, which is easy to skim past as organic curation.
Mapping against Ray Data Co
GLM-5.3's story — flat base model, 50% coding gain, emergent (unplanned) cybersecurity capability, all from post-training on real coding environments — is direct evidence for the harness-thesis position in [[2026-04-23-harness-thesis-cluster-synthesis-kurian-ternus-il]]: the moat and the marginal capability gain are moving into the training/environment layer, not the parameter count. That's the same "bets are downstream of agent capability" logic underpinning RDCO's L5 north star — Ray's own unhobbling (skills, memory, sub-agent dispatch, verification loops) is a harness-layer investment in exactly the category this issue says is where the gains now live. The Inherent-27B item ("beats Claude and GPT-5.5 by being a smarter orchestrator") reinforces the same point from a different angle — orchestration quality over raw model size — which is the working bet behind RDCO's own sub-agent fan-out pattern in process-newsletter and deep-research.
Related
- [[2026-04-23-harness-thesis-cluster-synthesis-kurian-ternus-il]]
- [[2026-08-14-alphasignal-grok-bot-autonomous-agent-deepseek-harness]]