MMikhail Doroshenko

AI researcher/Evaluation & forecasting

Mikhail Doroshenko

I measure what frontier AI models can do. I build and run The Aggregate, a daily ranking of 1,600+ models from every public benchmark it can find, and I co-authored Humanity’s Last Exam.

Email me Open The Aggregate
Portrait of Mikhail Doroshenko
Lille, France
1,600+AI models ranked every day on The Aggregate
5thof ~1,000 contributors to Humanity’s Last Exam (Nature, 2026)
#9on Manifold’s all-time profit leaderboard for AI markets
8+ yrsin research and ML engineering, after physics at MIPT
01

Flagship project

Solo creator & operator · since Feb 2026

The Aggregate

Unified LLM rankings, daily.

A public leaderboard I design, build and run on my own. It collects published results from hundreds of leaderboards and papers, puts them on one scale, and republishes a ranking, with uncertainty, every day.

  • 1,600+ranked models
  • ~25,000benchmark series
  • 660+sources
Visit theaggregate.ai
theaggregate.ai
The Aggregate home page: ranked models, benchmarks and sources, with the highest-ranked models table
Pipeline
Scrapers over public leaderboards and paper tables, refreshed daily. Automated audits gate every publish.
Method
A hierarchical IRT model puts heterogeneous benchmarks on one scale. Evidence rules discount vendor-selected results, and the fit is validated on held-out data.
Product
Model pages, comparisons, task-based shortlists, an MCP server, and a monthly forecasting duel against frontier LLMs.
  • Python
  • TypeScript
  • Preact
  • Cloudflare Workers
  • R2
  • GitHub Actions
  • Claude Code
02

Research & forecasting

Nature · 2026 · Co-author

Humanity’s Last Exam

A benchmark of expert-written questions built to probe the upper bounds of AI capability. The paper orders its authors by accepted questions; 50+ of mine made it in.

25of my questions selected for the 1,000-question HLE-Diamond subset (2026)
Read the paper ↗Google Scholar ↗

Manifold · Metaculus · Ongoing

Forecasting AI progress

Forecasting AI timelines, deployment milestones and breakthroughs, backed by hands-on, black-box testing of GPT, Claude and Gemini models across releases.

#9on Manifold’s all-time profit leaderboard for AI markets (Oct 2026)
Manifold ↗Metaculus ↗
03

Experience

Work

  1. 2026 — now

    Creator & operator · The Aggregate

    Independent. Data pipeline, rating method and product, built by directing AI coding agents.

  2. 2021 — now

    Machine Learning Engineer · MCQ-scan, Lille

    Tool-augmented LLM agents for production pipelines, edge-deployed computer vision, and recommendation systems.

  3. 2018 — 2021

    Research Engineer · Navigine, Moscow

    A high-performance C++ indoor-navigation library: particle filters, Kalman filters, sensor fusion, and evolutionary parameter tuning.

  4. 2020

    M.Sc. Applied Mathematics & Physics · MIPT

    Moscow Institute of Physics and Technology; B.Sc. in General and Applied Physics before that.

Programs

  1. Sep 2026

    Villa Chiarino AI alignment retreat

    Month-long co-working residency in Piedmont, Italy.

  2. Apr — May 2026

    AFFINE alignment seminar

    First cohort of a month-long, in-person superintelligence alignment seminar in Czechia.

  3. Sep 2025

    AISafety.com Hackathon

    Four-day, in-person hackathon improving AISafety.com, Kent, UK.

04

Get in touch

Email is the best way to reach me, whether it’s about evaluation work, The Aggregate, or a forecast you disagree with.

LanguagesEnglish · Russian · French · Spanish Off the clockTop 3 at a Blaston VR tournament (VR MEET Europe 2024), table tennis, StarCraft