Karpathy's autoresearch lets AI train itself, oh-my-pi ships full IDE

Thursday, 21 May 2026

Karpathy's autoresearch gives AI agents a real LLM training setup to experiment on autonomously, they modify code, train for 5 minutes, check if the result improved, then repeat all night.

Steady flow of agent frameworks and infrastructure drops, from Vercel's browser automation CLI to QuantumFlow's distributed inference scheduler.

👁️AWS ships multimodal evaluators

amazon

MLLM-as-a-judge for image-to-text tasks in Strands Evals. Verify whether model responses are actually grounded in source images.

Takeaway

We need reliable ways to evaluate visual AI outputs. These multimodal judges could help us catch hallucinations in document understanding or image analysis workflows.

AWSMultimodalEvals

🔒Code privacy with AI agents

Devs discuss what source code actually leaves their network when using AI coding agents and whether we've quietly accepted the risk.

SecurityPrivacy

🛡️GitHub probes internal repo breach

github

GitHub is investigating unauthorised access to internal repositories and will notify customers if any impact is discovered.

Bigger Picture

Infrastructure Shift to Agents

Railway's $200k+ monthly spend on coding agents signals infrastructure providers are optimising for AI workloads specifically. The bare metal approach with instant deployment could reshape hosting economics.

Under The RadarGitHubSecurity

🤔Learning versus letting AI handle

Experienced devs debate where to draw the line between understanding fundamentals and delegating work to AI tools.

DivisiveAI Workflows

🚂Railway builds agent-native cloud

latent·2 min read

3 million users, own-metal data centres, and zero activation energy deployment. $200k+ monthly spend on coding agents.

Takeaway

We're seeing cloud providers optimise specifically for agent workloads. The bare metal approach with instant deployment could cut our hosting costs whilst supporting the burst patterns agents create.

DevOps

😤Claude-written Jira tickets pain

Developer shares frustration debugging Jira tickets clearly written by Claude, highlighting the quality problems with AI-generated requirements.

AI Workflows

🎓AI engineering curriculum

GitHub

435 lessons from linear algebra to autonomous swarms. Python, TypeScript, Rust, Julia. Every lesson builds from raw maths to working code.

Top Voted

Python

TypeScript

Rust

Yesterday's Sentiment/Mixed

Builders Stay Focused Amid Turbulence

Strong technical momentum from Karpathy's autoresearch and production-grade agent infrastructure, but GitHub's security breach plus ongoing debates about AI code privacy are dampening the mood.

Learn/Multiple Mentions

What makes agents coordinate properly?

Agent coordination uses structured protocols to prevent conflicts when multiple AI systems work together. Projects like lavern use debate protocols with verification layers, whilst oh-my-openagent orchestrates models through harnesses. Proper coordination prevents duplicate work and ensures task dependencies are respected.

Multi-agentOrchestration

🔬Karpathy's AI trains models itself

GitHub

Autoresearch gives AI agents a single-GPU nanochat setup to experiment autonomously, they modify code, train, check results, and iterate overnight.

Bigger Picture

Self-Improving Research Loop

Karpathy's approach mirrors what we're seeing in production: agents that don't just execute tasks but improve the systems they work within. This could accelerate model development cycles dramatically.

Deep Dive

PythonResearchAgents

💾dejavu adds local AI memory

GitHub

Memory layer that follows we across AI tools. No cloud storage, runs in local SQLite, supports Python, REST, CLI, and MCP.

Trending

PythonLocal AIMCP

🤖oh-my-pi ships full IDE agent

GitHub

Coding agent with hash-anchored edits, LSP integration, 40+ providers, and tool-calling Python/Bun workers. Fork of Pi by Mario Zechner.

Bigger Picture

Memory Persistence Problem

Three different memory solutions dejavu, OpenViking, and grok-animus suggests the current RAG patterns aren't meeting dev needs for stateful AI systems.

Trending

TypeScriptAgentsDev Tools

🌐Vercel ships browser agent CLI

GitHub

Fast native Rust CLI for browser automation, built for AI agents. Supports snapshots, accessibility trees, and ref-based clicking.

Trending

RustAgentsVercel

⚡QuantumFlow schedules LLM

GitHub

Distributed inference framework with multi-backend support, adaptive scheduling strategies, and cluster management for large model workloads.

Bigger Picture

Security Reality Check

The GitHub breach combined with questions about AI agent data exposure shows we're shipping features faster than establishing security baselines. The community discussions feel overdue.

Trending

PythonLLM Ops

🔧oh-my-openagent grows

GitHub

Agent OS supporting multiple harnesses (OpenCode, Codex, Pi). Features Team Mode with model orchestration and autonomous task completion.

Trending

TypeScriptAgents

🕸️Understand-Anything maps codebases

GitHub

Turns any codebase into interactive knowledge graphs. Works with Claude Code, Cursor, and other AI tools for visual code exploration.

Trending

TypeScriptCode Gen

🔄KV cache beats RAG pipeline

Team replaced their RAG setup with persistent KV cache and found it more effective. Shared their findings and approach.

RAG

🚀llama.cpp gains fresh momentum

GitHub

LLM inference in C/C++ continues gaining traction. Recent updates include gpt-oss model support and multimodal capabilities in llama-server.

Deep DiveC++Local AIInference

🗄️OpenViking unifies agent context

GitHub

Context database for AI agents using filesystem approach. Unifies memory, resources, and skills with tiered loading and recursive retrieval.

Trending

PythonAgents

⚖️lavern ships AI law firm

GitHub

67 specialist agents coordinate through debate protocol with mandatory human gates and three layers of verification. Apache 2.0 licensed.

Trending

TypeScriptAgents

🤖Composer 2.5 trained on Claude

Community speculation that Cursor's Composer 2.5 was trained on Claude outputs, based on observed behaviour patterns.

Cursor

📊ppt-master generates PPTX natively

GitHub

AI generates natively editable PowerPoint from documents. Real shapes and animations, not images. Supports PDF, DOCX, URLs, and Markdown.

Trending

Python

🪦Google's open source shutdown

Dev.to·2 min read

Developer's Gemini CLI PR merged just before Google announced the tool's sunset in favour of Antigravity CLI. Reflection on the bait-and-switch pattern in open source AI.

Takeaway

We're seeing major vendors use open source as a testing ground before replacing community tools with proprietary alternatives. This pattern affects how we evaluate which projects to contribute to.

Open SourceGoogle

Learn/Core Concept

How does adaptive scheduling work?

Adaptive scheduling dynamically allocates computing resources based on real-time model demands and hardware availability. Systems like QuantumFlow monitor inference workloads and automatically shift tasks between GPUs, adjusting batch sizes and memory usage to optimise throughput. This prevents bottlenecks when running multiple models simultaneously.

BatchingLoad-balancing

Read online