Self Improving AI (SIA) Agents - Emergency Hackathon (by AGI House SF & Nebius)

About

🚨 A paper dropped last week that changes how self-improving agents get built. We're convening an emergency hackathon this Saturday to hack towards the next great leap! 🚨

GitHub & the Paper SIA: Self Improving AI with Harness & Weight Updates (Hebbar et al., 2026)

Framework Overview

Today, every agent's harness and weights get tuned by people. SIA closes the loop, a language-model agent that rewrites both its own harness and the underlying weights.

The numbers are not subtle:

502% improvement on single-cell RNA denoising

91.9% runtime reduction on GPU kernels

56.6% gain on LawBench

We're inviting you to join us to innovate on the Self-improving AI (SIA) agent framework this Saturday!

Tracks

Applied AI Track (use SIA on a real domain task)

Framework Enhancement Track

Research Track (novel evaluation or self-improvement methodology)

Prizes: Meta Ray Bans for the first prize of each track

Agenda

10:00 AM — Doors open, fresh coffee and bagels

10.45 AM — Opening remarks

11:30 AM — Hacking begins and Teams form (2-4 people). Tracks chosen. Research team circulates for technical support.

12:30 PM — Catered lunch.

6:00 PM — Hacking ends. Demo session begins — 3-4 minutes per team plus live Q&A from judges.

8:00 PM — Judging deliberation (private). Cleanup begins.

8:30 PM — Awards. Winners announced per track. Prize distribution. Group photo.

9:30 PM — Adjourn

Sufficient GPU credits will be provided thanks to our partner, Nebius. Nebius is the ultimate cloud for AI innovators, and will be providing H200s through their Nebius Token Factory, including instant model access for Deepseek, Qwen, GLM, Gemma, NVIDIA Cosmos.

Framework Architecture and Benchmarks

Prize for the winning team in the Applied AI Track

Meta Ray Bans

First Prize - Framework Enhancement Track

Prize for the winning team in the Framework Enhancement Track

Meta Ray Bans

First Prize - Research Track

Prize for the winning team in the Research Track

Meta Ray Bans

Self Improving AI (SIA) Agents - Emergency Hackathon (by AGI House SF & Nebius)

About

Sponsors

Sponsors

Prizes

First Prize - Applied AI Track

First Prize - Framework Enhancement Track

First Prize - Research Track

Submissions

Self Improving AI (SIA) Agents - Emergency Hackathon (by AGI House SF & Nebius)

About

Sponsors

Sponsors

Prizes

First Prize - Applied AI Track

First Prize - Framework Enhancement Track

First Prize - Research Track

Submissions

1https://github.com/actionstockplus-ui/ia-humor-benchmarks.gitA humor Enhancer.

2JanuSQLJanuSQL:a self-improving AI that rewrites its own English→SQL agent, graded by running the SQL on open models. The smaller model gained more (+29% vs +15%). One project, 3 tracks targeted: Applied, Framework (Ongoing), Research (Meta-Signals captured). SIA harness lever only.

3SIA-improvementThe autonomous research framework for fine-tuning MoE language models. Designed for Claude Code — describe the experiment you want to run, the system orchestrates it with four collaborating agents and methodology rigor enforced structurally. One framework. Telemetry, attribution

4Graph-of-Experiments: Structured Run Memory for Self-Improving AIEnables the meta agent to make observations between runs by writing them to a knowledge graph, so that the meta agent can experiment more effectively

5deSIAgnSIA Taste-Shader: can an agent self-improve toward someone's taste? The modified SIA model writes GLSL shaders; rewrites an initial prompt to generate the shader output, scored on held-out taste. Finding: smarter harness ≠ more taste (58.8→45) Applied AI and Framework track

Search