Pasha / Systems Developer

AI + Technology
Investigations

Source-linked investigations, technical evidence, and clear explanations of what AI and technology changes mean.

Latest research

01 / What this is

Follow the evidence.
Understand the story.

Benchmark charts

Side-by-side numbers for models I find worth testing, in a form that is easy to read.

Model comparisons

Specs and scores from published sources, with links to the original evaluations.

Research notes

Detailed mechanisms, community perspectives, and a standalone explanation in plain English.

Telegram / @Dekrov_channel_en ↗X / @Dekrov_Dev ↗GitHub / DekrovDev ↗

02 / Pages

08 Paperclip's New Runner: What Changed—and What Still Needs ProofAgent Infrastructure · Nov 10, 2026 The new native Runner, GitHub-backed skills and optional security screening, examined through source code and community issues. 07 Evidence before confidenceEditorial methodology · Nov 10, 2026 How to read an AI investigation: follow the sources, compare conflicting reports and understand what remains uncertain. 06 Gemini 4 Carbon: Google Takes #1Model Research · Oct 28, 2026 The new intelligence leader: coding gains, cybersecurity results, Claude and GPT comparison, pricing and access limitations. 05 GPT-6.1 Terra: Specs, Benchmarks & PricingModel Analysis · Oct 27, 2026 1.05M context, API prices, computer-use and coding benchmarks, reliability claims, and source-linked caveats. 04 Beam 1.0 vs DeepSeek V4.1 FlashBenchmarks · Oct 2026 Terminal-Bench 2.1, DeepSWE v1.1, and GPQA Diamond side by side — with sources and caveats. 03 Mercury 3: Speed vs EvidenceResearch · Oct 2026 Diffusion LLM speed, launch pricing and reported quality scores — charted with sources and limitations. 02 Kimi K3.1 Pro: Launch & BenchmarksResearch · Oct 2026 Independent intelligence ranking, coding and CAD benchmarks, pricing charts, and the open-weight release caveat. 01 AGENTS.md Hijacked Codex: Real AttackSecurity · Oct 26, 2026 A real blocked incident, reproducible MCP persistence, Claude Code token theft, attack prerequisites and verified defenses.