GLM-5.3 hits 4% control-flow hijack rate, Thorsten Ball asks if model pacing is dead
Register Spill (Thorsten Ball) · rss · 2026-10-04
Thorsten Ball's Joy & Curiosity #102 roundup:
- Gemini 4 Argon autonomously optimized Google's fleet: Argon agents analyzed profiling telemetry and applied memory optimizations, freeing over 300 TiB with 500 TiB–1 PiB in projected savings.
- Safety landscape shift: Anthropic reports GLM-5.3 achieves full control-flow hijacks in 4% of 100 benchmark trials — near Claude Mythos Preview's 6%, while Opus 4.6 and GLM-5.2 scored zero. Simple techniques disabled GLM-5.3's safeguards 64–100% of the time, leading Ball to conclude that any "pacing" strategy is now moot.
- Efficiency rumors: circulating cfo.ai evals claim GPT-6.1 Sol is 3x as efficient as Opus 5.5 and 2.5x vs Sonnet 5.5 — $2.72 vs $7.60 for the same task and score.
- Also: Daniel Lemire on stagnation since 1970; 1,597 Linux kernel vulnerabilities disclosed at once; Idaho's Zero-Based Regulation cut 49.1% of its regulatory code; Katzenberg on AI and creativity — "tools are never the point; what endures is taste and imagination."
More from AGI Musings
- AI-powered forensic metascience: auditing 7 million papers, 500K files scanned, 100+ issues found — gleech · 2026-11-20
- Long Now Labs' first awardees include a thermodynamic clock for LLMs, $25k each — begusgasper · 2026-11-20
- Long Now launches Epistemic Cycles Lab led by Dr. Nina Beguš to map information ruptures — begusgasper · 2026-11-20
- Talia Ringer starts an email thread on open-model advocacy in a 'post-Jensen' world — dinabass · 2026-11-20
- Andriy Burkov's True Positive Weekly #186: what to know when agents do the work — burkov · 2026-11-20
- Brand drift in the agent era: give agents the reasoning behind each rule — round · 2026-11-20