🧠 5 Papers, Explained Simply

A Definition of AGI — a scorecard you can actually use

This paper says: stop arguing about fuzzy AGI definitions. Instead, measure whether an AI matches a well-educated adult across 10 abilities (things like reasoning, reading/writing, math, working memory, long-term memory storage & retrieval, visual & auditory processing, and speed). It borrows from human psychometrics (the CHC model) so we can report an overall AGI Score plus a “profile” that shows strengths and weaknesses.

Why it matters: Research and product teams can set targets per ability (e.g., “improve working memory this quarter”) and avoid being misled by one high score.

DeepSeek-OCR — “optical compression” for fast, high-res reading

The idea: images can compress text efficiently. A slim vision DeepEncoder squeezes a high-res page into a small number of “vision tokens,” then a modest 3B MoE decoder reads it. With <10× text-to-vision token ratios, the model reports ~97% OCR precision; even at 20×, it’s still around ~60%. On OmniDocBench, it beats heavier models using far fewer tokens. They even generate ~200k+ pages/day on a single A100-40G.

Less Is More: Recursive Reasoning with Tiny Networks (TRM)

A tiny recursion model (~27M params) does step-by-step reasoning on puzzle-style tasks (Sudoku, mazes, ARC-AGI), trained on roughly ~1k examples. Reported results include solid scores on ARC-AGI-1 and non-trivial scores on ARC-AGI-2—showing that smart structure + recursion can beat raw size on some reasoning tasks.

Agent Learning via Early Experience — give agents a “childhood”

Before full RL, let the agent gather some low-stakes early experience and learn from it. Two flavors: (1) implicit world-modeling (it learns how the environment behaves) and (2) self-reflection (it critiques and improves its own attempts). Across many environments, this improves success and out-of-domain generalization. On tasks with verifiable rewards (like WebShop or multi-turn function calling like BFCLv3), adding a standard RL step on top gives big additional gains.

DeepAnalyze: Agentic LLMs for Autonomous Data Science

DeepAnalyze-8B is trained to plan, use tools, and write reports across end-to-end data workflows. Its hybrid reward mixes rule-based checks (format, correctness) with LLM-as-a-judge scoring (process quality and report quality) and uses a curriculum (easy→hard). The promise: more robust, auditable data-science agents that care about both the answer and the process.

🧪 Tools to Try (now)

📈 Market & Talent

Nvidia/OpenAI mega-deal chatter: Reporting says Nvidia may lease up to ~5M chips to OpenAI over time (framed as roughly $350B of exposure). If accurate, it shows how fast compute demand is scaling—and the financial risk behind frontier models. Market take: NVDA dipped slightly on the rumor.

Meta hire: Meta reportedly hired Tim Brooks (ex-OpenAI/DeepMind) into Superintelligence Labs—another sign they’re leaning into video-centric world models and longer-horizon reasoning.

🎙 Listening Pick: Andrej Karpathy × Dwarkesh

Watch on YouTube. Themes to notice: tiny, hackable systems (see NanoChat), agent loops that don’t “run away,” and why evals should measure process—not just answers.

🛠 Cheat-Sheet: Borrow These Ideas

• Make an “AGI Scorecard” for your product: pick 6–10 abilities that matter (tool use, memory, vision, latency) and track them monthly.
• Prototype with tiny models + recursion: it’s cheaper to find bottlenecks before you scale.
• Give agents a short “childhood”: a week of world-modeling or self-reflection data, then RL. You’ll see steadier multi-turn behavior.
• Use hybrid rewards: score both outputs and process quality to avoid “shortcut” agents.
• Try a coder-in-the-browser: evaluate Claude Code on the web and Atlas’s agent mode in supervised flows.

You’re reading researchaudio. If a friend sent this, subscribe here.