🆕 Latest from the labs
🧭 Tracking the Path to AGI

The State of AGI Research, Mapped and Explained

Real-time analysis of the labs, researchers, and breakthroughs racing toward artificial general intelligence — investigations, comparative research maps, and founder playbooks, grounded in sourced 2026 reporting, not hype.

AGI Capability Pulse — 2026 Snapshot Tracking weekly
Reasoning & Math 92%
Coding & Tool Use 88%
Long-Horizon Agents 61%
Multimodal Grounding 74%
Continual Learning 34%
Alignment & Interpretability 41%
Directional estimates synthesized from our lab & benchmark coverage — see the Research Frontier Map for sourcing.
Latest from the Academy

New articles and the threads they continue

The newest posts on the site. The dots under each one trace its series: click any dot to read the earlier or later parts.

🧭Four Ways to Build an Agent: Frameworks, Harnesses, Runtimes and Builders7 Nov 2026 · 5 min
Frontier Lab Updates

Latest from the Labs

Every release and announcement from 13 frontier labs over the last five years — models, products, APIs, pricing and research — newest first, each linked to its source.

QwAlibaba Qwen · 🧠Qwen4-MaxAlibaba's new flagship after a limited preview: a sparse mixture-of-experts model reported at 2.4T total and 71B active parameters, native vision and a 1M-token context. API only, no weights; Alibaba says smaller Qwen4 models will be open-sourced later. Alibaba reports about 60% lower cost per task than its predecessor on its own benchmark. Figures from secondary coverage. GDGoogle DeepMind · 🧠Gemini 4 HydrogenThe fast, low-cost member of the Gemini 4 family for everyday coding and agent work. It replaces Gemini 3.8 Flash in Search, the Gemini app and developer tools. Google says it scores above Gemini 4 Argon on the Artificial Analysis Intelligence Index at its default thinking level and runs about 2.5 times faster than 3.8 Flash (company claims). MeMeta AI · 🧠Muse EmberMeta Superintelligence Labs' lightest and most cost-efficient Muse model so far, aimed at running agents. Reported via secondary coverage. AnAnthropic · 🧠Claude Fable 5.2 and Claude Mythos 5.2Anthropic's new top models for hard technical work. Fable 5.2 is generally available; Mythos 5.2 goes only to vetted organisations through Anthropic's life-sciences and cyber verification programmes. Anthropic reports 69.9% on FrontierSWE v2 for Fable 5.2 (Opus 5.5: 62.3%); press coverage also reports lower prices and fewer refusals than Fable 5.1. Benchmark and pricing figures are the company's, as reported. Disclosure: Claude, made by Anthropic, helps write this site. OAOpenAI · 🧠GPT-6 CyberA cybersecurity-focused version of GPT-6 Astra, limited to approved customers in OpenAI's Daybreak Red programme. OpenAI rates its vulnerability-finding and exploit capability as 'critical'; reported benchmark gains (e.g. ExploitGym 48.5% vs 32.1% for GPT-5.6 Cyber) are OpenAI's own figures from secondary coverage. OAOpenAI · 🧠Images 2.6 (GPT Image 2.6)OpenAI's new image model in ChatGPT, Codex and the API, with two API models: GPT-Image-2.6 Flare (speed) and Sunburst (quality). OpenAI says it is 50% faster than Images 2.5; reported additions include animated GIFs, persistent reference images and SVG output.
See all 329 updates, filter by lab →
The Long View

AI Research Progress Through the Years

How AI research got here, year by year: the ideas, papers, systems and turning points that led from early neural networks to today's frontier models and agents.

📅 AI Research Progress · 19 Articles
AI Research Progress Through the Years
A year-by-year record of the papers and ideas that moved AI research forward — explained in plain English with animations, and linked to the sources.
New · Community-voted
What the Community Reads: Hugging Face Paper Themes
Each month's most-upvoted Hugging Face papers grouped by theme, to show which topics rise and fade. Upvotes measure attention, not quality.
📈 Open the full HF Paper Themes page: streams, heatmap, month bubbles and sources →
🚇
AI Idea Flow: How Ideas Flowed and Merged, 1950 to Today
Interactive metro map · 1950 → 2026
📅
AI Research in 2026: The Papers and Ideas Reshaping the Field
40 min · 14 papers, ICLR & ICML 2026
🚀
AI Research in 2025: The Year Reasoning, Agents and Test-Time Compute Broke Out
45 min · 18 papers, ICLR, ICML & NeurIPS 2025
🧭
AI Research in 2024: The Year Context, Agents and World Models Took Shape
45 min · 21 papers, ICLR, ICML & NeurIPS 2024
🏗️
AI Research in 2023: The Year Foundation Models Became Platforms
45 min · 17 papers, NeurIPS, ICCV & more
🌱
AI Research in 2022: The Year Reasoning, Alignment and Generative AI Converged
45 min · 17 papers, NeurIPS 2022 & more
🧱
AI Research in 2021: The Year the Foundation-Model Stack Emerged
40 min · 14 papers, ICML 2021, Nature & more
📐
AI Research in 2020: The Year Scaling Became the AI Strategy
40 min · 12 papers, NeurIPS & ECCV 2020
🏗️
AI Research in 2019: The Year Pretraining Became the Default
45 min · 16 papers, NeurIPS, CVPR, Nature & more
🔓
AI Research in 2018: The Year Pretraining Changed Everything
40 min · 14 papers, NeurIPS, NAACL, Science & more
🌱
AI Research in 2017: The Year the Modern AI Architecture Was Born
40 min · 11 papers, NIPS, ICLR, Nature & more
⚙️
AI Research in 2016: The Year Deep Learning Became a General-Purpose Engine
40 min · 12 papers, CVPR, ICML, Nature & more
🔧
AI Research in 2015: The Year Deep Learning Became Trainable, Attentive and Interactive
40 min · 12 papers, three Test of Time awards
🎭
AI Research in 2014: The Year Deep Learning Learned to Generate, Translate and Remember
40 min · 12 papers, four test-of-time winners
📐
AI Research in 2013: The Year Representation Learning Became the Language of AI
40 min · 12 papers, word2vec, VAE & more
💥
AI Research in 2012: The Year Deep Learning Broke Through
35 min · 8 papers, AlexNet & more
🌅
AI Research in 2011: The Year Before the Breakthrough
40 min · 13 papers, Hogwild!, Watson & more
🧭
AI Research in 2001–2010: The Decade That Set the Stage for Deep Learning
45 min · 18 papers across a decade
📼
AI Research in 1980–2000: The Two Decades That Invented Modern AI's Core Ideas
50 min · 20 papers, two decades
📟
AI Research in 1950–1980: The Three Decades That Founded the Field
50 min · 20 papers, three decades
00 · Follow the Frontier

Where AGI Research Is Heading

Long-form, fully sourced series on recursive self-improvement and the capabilities that still separate today’s models from AGI.

✅ Complete · 7 of 7 Parts + 1 Bonus Report + 3 Research Updates
The Recursive Self-Improvement Series
A seven-part deep dive into whether AI can meaningfully accelerate the process that builds better AI — from the 1965 origin of the "intelligence explosion" concept through the real, independently verified systems (Dream-RSI, ModularRSI, ScienceBuddy, AlphaEvolve) attacking pieces of the problem today, with every claim checked against primary sources before publication.
New · Interactive map
RSI Evolution Map: 524 papers by autonomy level
How AI self-improvement research climbed from executing human-designed procedures (L1) to improving its own improvement mechanism (L5). Built on the CC0 Awesome-RSI catalog.
♻️ Open the full RSI Evolution Map: loop explorer, three views and industry cases →
New · Interactive map
Frontier Labs × RSI Levels
Where Google DeepMind, Anthropic, OpenAI, Meta and the Chinese labs sit on the five levels of self-improvement, with every dot linked to the lab's own post. Claude, made by Anthropic, helps write this site.
🏛️ Open Frontier Labs × RSI Levels: sort, evidence and caveats →
New · Interactive map
Model Release Map: foundation, image, video, speech and world models
Which lab shipped which model and when, how each model line evolved, and how long open-weight models took to catch up with closed ones.
🧭 Open the full Model Release Map: every view, filter and source →
New · Interactive tracker
Frontier Benchmark Tracker: who leads, on what, measured by whom
The latest scores of frontier models across eight kinds of task, with every number labelled by who reported it: the model's own lab, a competitor's launch table or an independent evaluator.
📊 Open the full Benchmark Tracker: lead over time, saturation, model profiles and lab vs independent →
01 · Build the Machine

From Electrons to Frontier Models

Three series, one stack: the physical machinery of AI (power, chips, memory, networks and money), how a frontier model is produced on top of it (followed end to end with Llama 3 405B), and the people who run each layer of the factory's software.

New · Interactive map
The AI Factory: inside view and from above
🏭 Open the full cutaway: every floor, its bottleneck and all the articles that cover it →
New · Interactive timeline
Stack Innovation Map: 20 years of hardware and software leaps
🧬 Open the full Stack Innovation Map: every layer, generation lines and cross-layer chains →
🧱 The Physical Layer
🖥️ AI Compute Stack · 18 Articles
The AI Compute Stack: From Electrons to Tokens
What it physically takes to run AI: power, factories, chips, memory and the economics that tie them together. Power → private energy → factories → vs supercomputers → cooling & water → chips → supply chain → networks → memory → custom silicon → inference → edge devices → token prices → financing → China → frontier sites → economics → the horizon. Each part stands alone; together they explain why AI is now an energy and hardware story as much as a software one.

🗺️ See the whole stack on one page: key numbers, bottlenecks and reading tracks →

🧑‍🔧 The People Layer
🧑‍🔧 Who Runs the AI Factory · 12 Articles
Who Runs the AI Factory: The Software Stack, Role by Role
The AI factory's software stack, one layer at a time, through the people who own it: what they run, the tools they use, what breaks and how they hand off to the next layer. Bring-up & firmware → OS & runtime → networking → storage → scheduling → kernels & compilers → training frameworks → data platform → experiment tooling → inference → observability → security.

🗺️ See every role on one page: stack map, role finder, skills matrix and career paths →

🧩 Agentic AI System Designs
🧩 Agentic AI System Designs · 7 Articles
Agentic AI System Designs: The Agent Systems the Labs Have Built, Taken Apart
Four kinds of agent system (frameworks, harnesses, runtimes, builders), each system dismantled layer by layer: an eight-layer architecture, one task traced through it, failure modes, security, how to build one from scratch and what it costs. Wave 1: the map and five harnesses.

🗺️ Every agent system on one page: by kind, compared, and a chooser →

New · Live architecture
The AI Software Stack, Live: 11 layers, 72 technologies, data in motion
🧬 Open the full live architecture: every layer, its innovations, and six flows from training steps to agent tasks →
Run the Machine · The Maestro

Building an AI Factory From Zero, One Hat at a Time

Everything in section 01, put to work. Gopala, a researcher-founder, takes one self-improving (L4) idea all the way to a frontier-style model and client products at startup scale, wearing every hat in turn: researcher, founder, infrastructure engineer, ML engineer, evaluator, safety lead and operator. Real-time logs, runnable code, budgets and an honest line on what one person cannot do alone.

New · Interactive map
Gopala's Journey: the whole Maestro series on one road
Season 1's eighteen stops, now extended into Season 2 as each phase ships. Press play to watch the journey, or click a stop for its decisions, setbacks and the layers at work.
🗺️ Open the full map of Gopala's Journey →
🎼 The Maestro, Season 2
🎼 The Maestro, Season 2 · 19 Articles
The Maestro, Season 2: Scaling the Factory
Phase 5 (founder to CEO) → Phase 6 (the research engine) → Phase 7 (Maestro-2) → Phase 8 (the product grows up) → Phase 9 (competing at the frontier). 23 parts, published phase by phase; Phase 5 is out now.

🎼 New here? Start with Season 1: 18 parts from an idea to a shipped model →

AI Research Progress 00 · Frontier Recursive Self-Improvement Capability Stack 01 · Build the Machine AI Compute Stack Producing Intelligence Who Runs the AI Factory AI Factory Case Studies AI Factory as a Service Inside the ML Layer Inside the Data Layer Inside the Evaluation Layer Inside the Agent Layer Agentic AI System Designs Run the Machine The Maestro The Maestro, Season 2 02 · How LLMs Are Built Frontier Architecture & Reasoning Harness Engineering Practical RSI Architecture & Learning Topic Explainers Model Lineage Model Case Studies Reference Series 03 · Courses Language Modeling from Scratch Self-Improving AI Agents Course 04 · Researcher Researcher Foundations Engineering Practicum Researcher's Path 05 · Company Founding a Startup Startup Lane Deep Dives Economics & Society
02 · Learn How LLMs Work

How LLMs Are Built: From Raw Text to Your Screen

Fourteen steps across four acts — pre-training, alignment, deployment, and what happens the instant a user hits send — with the objective functions behind each one. Take the interactive tour first, then go deeper: every step links to a full article with runnable code.

Act 1 of 4 · Pre-training
Data Collection
🛠️ Go Deeper · 18 Articles
How LLMs Are Built: One In-Depth Article per Stage
The full lifecycle of a large language model, one in-depth article per stage — the transformer, pretraining, fine-tuning, post-training, inference, scaling with distributed training, the production system around the model, context engineering and memory, retrieval-augmented generation, tool calling, agentic systems, multi-agent systems, agent memory and planning, MCP, production agent architecture and agent evaluation — starting with an interactive map of the complete stack — each with runnable code, real output, and sources checked.
🗺️
From Transformer to Production Agentic AI: The Complete Stack
Interactive · the whole series on one map
⚙️
How a Transformer Works: A Decoder-Only Model, End to End
45 min · interactive + runnable code
🏭
How LLM Pretraining Works: From Raw Text to a Foundation Model
40 min · runnable mini pipeline
🎯
How LLM Fine-Tuning Works: Adapting a Base Model to a Task
40 min · full FT vs LoRA experiment
🧭
How Post-Training Works: SFT, RLHF, DPO and RLAIF
45 min · SFT vs DPO vs RLHF experiment
⚡
How LLM Inference Works: From a Request to a Streamed Response
40 min · runnable mini inference engine
📈
Scaling Laws, GPUs and Distributed Training: How Frontier Models Get Built
40 min · scaling study + all-reduce check
🏗️
Production LLM System Design: Anatomy of a ChatGPT- or Claude-Style Assistant
35 min · runnable mini assistant runtime
🧩
Context Engineering and Memory Architecture for LLMs
30 min · five memory strategies measured
🔎
RAG: Retrieval-Augmented Generation, End to End
35 min · RAG pipeline measured on real docs
🛠️
Tool and Function Calling: How an LLM Goes from Answering to Acting
30 min · runnable tool orchestrator
🤖
Agentic Systems: From Single Tool Calls to Multi-Step Goal Completion
30 min · runnable agent loop
👥
Multi-Agent Systems: End-to-End Architecture
30 min · cross-checking and coordination measured
🗂️
Agent Memory and Planning Patterns: How Agents Remember, Plan, Reflect and Keep Going
30 min · checkpoints, lessons and replanning measured
🔌
MCP and Tool Ecosystems: How Agents Discover, Connect to and Use External Capabilities
30 min · MCP server from scratch, SDK-verified
🏛️
Production Agent Architecture: End-to-End System Design
30 min · platform mechanics measured
📏
Evaluating Agents: Measuring Success, Safety and Reliability
30 min · runnable evaluation harness
🧰
LLM Harness Engineering: Engineering the System Around the Model
20 min · same model, five harnesses
🧬 Frontier Architecture & Reasoning · 4 Articles
Latest in Frontier Model Architecture & Reasoning — 4-Lens Review
The latest in how frontier models are designed and how they reason, as of September 2026, reviewed through four lenses: the research changing what's inside the models, the engineering to train and serve them at scale, the market of labs, open models, chips and clouds, and case studies of real model families.
🧰 Harness Engineering · 10 Articles
Latest in Harness Engineering (Agentic Scaffoldings) — 4-Lens Review
The latest in the systems built around models to turn them into working agents — context and memory, tools, planning loops, verification, guardrails and evaluation — reviewed through four lenses: the research advancing agent scaffolding, the engineering to run it reliably in production, the market of agent frameworks, platforms and protocols, and case studies of real agent harnesses.
🧬 Architecture & Learning Technique · 5 Articles
Architecture & Learning Technique: How the Field Actually Got Here
The history of neural network architectures, the separate thread of learning techniques layered on top, and the live debate over what comes after Transformers.
🔍 Topic Explainers · 19 Articles
Topic Explainers: Standalone Deep Dives
World models, retrieval, agents, safety, reasoning, chips, robotics, multimodality, voice, image/video, pricing, agentic loop engineering, brain–computer interfaces, and fresh launches like TypeSafe's Jev — each explained on its own.
✅ Complete · 8 Labs, Fully Cross-Referenced
The Model Lineage Series
Eight consolidated comparisons tracing every major model from OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek, Meta, Mistral, and Qwen — architecture, data, safety frameworks, and disclosure philosophy, each lineage cross-linked to the other seven for direct side-by-side reading.
03 · Take a Course

Full University Courses, Lecture by Lecture

Guides to Stanford courses on building language models from scratch and on self-improving agents.

🧱 Lecture Series · Part 4 of 19
Language Modeling from Scratch: A Guide to Stanford's CS336
A lecture-by-lecture deep-dive series on Stanford's from-scratch LLM-building course — tokenization, architectures, systems, scaling laws, data, and alignment, with real citations, original code walkthroughs, and hands-on directions tied to the actual public assignment repos. One full article per lecture, no length cap.
🎓 Course Guide · Handbook + 5 Deep Dives (Complete)
Self-Improving AI Agents: A Guide to Stanford's CS329A
A standalone companion to Stanford's course on the mechanics behind self-improving agents — test-time compute, verification, agentic feedback loops, planning as search, RL for reasoning, and long-horizon evaluation. A different branch of "self-improvement" from this site's RSI series, covered on its own.
04 · Become a Researcher

From Foundations to a Frontier-Lab Career

The curriculum, the engineering practice and the career path.

🧭 The Researcher's Path · 3 Articles
The Researcher's Path: Careers, Labs & Field Notes
What it actually takes to work on AGI — papers to read, labs to target, a real day in the life, and a full investigation into one of the field's most secretive labs.
05 · Build a Company

Three Lanes, Three Very Different Bets

Every AGI-adjacent startup — and increasingly, every researcher's career choice — falls into one of three strategic lanes.

💼 Economics & Society · 2 Articles
Economics & Society: What AI Progress Does to the Rest of the World
Two pieces on the productivity paradox — what happens to jobs, growth, and the shape of the economy as AI adoption accelerates.
Go Deeper

Explore Everything

The full archive, or the interactive course that started it all.