Extending test-time compute can collapse LLM accuracy. Here is the forensic math on Goodhart gaming, attention drift, and why latent recurrent depth wins.
Benchmarks & Hype Checks
89 Articles in this Category
Independent evaluations of synthetic benchmark claims: ARC-AGI-3, SWE-bench Verified, HumanEval, and live latency audits.
Disclosures broken by developer Lyra confirm Anthropic is closed-beta testing Claude Opus 5.5 (‘claude-wafer-eap’), cutting pricing to $4/$20 per M tokens.
Step 5 Preview redefines the AI Pareto frontier. With a 600B/27B sparse MoE and 1M context, it matches Kimi K3 Max (AA Index 44) at ~$0.71 task cost.
Quick Answer · Featured Snippet Target What is PrismML Ternary Bonsai 2 27B? It is a 1.76-bit…
GPT-6 Astra cracked an unsolved 1918 German ADFGVX cipher using historical monograph search and key transposition in 170 chars. Here is the technical audit.
Executive Summary & Position #0 Answer Did Google “benchmax” Gemini 4 Pro? Yes, the empirical and telemetry…
Qwen3.8-Omni-Flash beats Gemini 3.8 Flash on WildClawBench (71.0 vs 58.9) and AliMeeting (89.7 vs 37.1) at 4.2x lower video cost. Read our full teardown.
Forensic audit of stealth/union-alpha: 74% DeepSWE score, MoA gateway architecture, Austrian Compunect GmbH trail, and developer outputs from X.
Google Dream-RSI cuts code search calls by 161.5× and hits 2,350ms SOTA on Lasso using offline replay simulators—leaving LLM weights 100% frozen.
Google DeepMind is stealth-testing Gemini 4 Pro under gemini-3.8-flash in LMArena. Forensic audit of the 3D voxel pagoda, SVG pelican, and TPU speed.