Category
Benchmarks & Deep Dives
Rigorous performance testing, framework comparisons, data-backed architectural breakdowns
7 articles
Featured
Latest
AI Agents · Benchmark · Alex Rivera · Oct 23, 2026 · 10 min
Kimi K3.1 Pro Closes the Open-Weight Gap
Moonshot AI's Kimi K3.1 Pro ships with a 1M-token context, $2/$10 per million tokens, and open weights promised by November 4. Independent testing puts it fourth overall and first on Terminal-Bench 4.0.
Explore benchmark →
Benchmarks · Benchmark · Alex Rivera · Oct 6, 2026 · 14 min
Mistral Named the Trillion-Param Model Le Chonk
1 trillion parameters, 49 billion active, trained on 3,800 GB200s in Europe. Preview API today. Weights by the end of October. Cyber is the pitch.
Explore benchmark →
AI Agents · Analysis · Alex Rivera · Sep 19, 2026 · 16 min
Devin Scans the Whole Repo First
Code Scans is MapReduce for a codebase. You name an outcome. Parallel agents hunt. A last agent ranks the findings and opens pull requests.
Read analysis →AI Agents · Analysis · Alex Rivera · Sep 18, 2026 · 16 min
Raindrop Catches Agent Failures Live
Raindrop now has $50M total and a new Simulations product. It watches live agents for silent mistakes, then replays real traffic on every pull request before you ship.
Read analysis →
Benchmarks · Benchmark · Alex Rivera · Sep 17, 2026 · 18 min
Agents Get a Trustmark
AIUC-1 is trying to be SOC 2 for agents: thousands of evals, a certificate, and now a $40M Series A to sell audits and insurance. A badge is not a kill switch.
Explore benchmark →
AI Agents · Deep dive · Alex Rivera · Sep 17, 2026 · 16 min
A Map of AI That Improves Itself Is Not a Launch Date
A ByteDance–Tsinghua–Shanghai AI Lab preprint ranks 491 papers on five rungs of recursive self-improvement. The interesting part is what the ladder does not measure.
Read the deep dive →Essential
Related research
AI Agents · Benchmark · Alex Rivera · Oct 23, 2026 · 10 min
Kimi K3.1 Pro Closes the Open-Weight Gap
Moonshot AI's Kimi K3.1 Pro ships with a 1M-token context, $2/$10 per million tokens, and open weights promised by November 4. Independent testing puts it fourth overall and first on Terminal-Bench 4.0.
Explore benchmark →Benchmarks · Benchmark · Alex Rivera · Oct 6, 2026 · 14 min
Mistral Named the Trillion-Param Model Le Chonk
1 trillion parameters, 49 billion active, trained on 3,800 GB200s in Europe. Preview API today. Weights by the end of October. Cyber is the pitch.
Explore benchmark →Benchmarks · Benchmark · Alex Rivera · Sep 17, 2026 · 18 min
Agents Get a Trustmark
AIUC-1 is trying to be SOC 2 for agents: thousands of evals, a certificate, and now a $40M Series A to sell audits and insurance. A badge is not a kill switch.
Explore benchmark →