# Why 2031 Might Be the Last Year Humans Do AI Research > Source: > Published: 2026-08-20 11:46:25+00:00 For the past decade, the rate of progress in artificial intelligence has been constrained by a single critical bottleneck: the human brain. Every architectural breakthrough, hyperparameter tweak, data-filtering pipeline, and distributed training setup has been conceived, debugged, and executed by human computer scientists. Even as AI models have grown exponentially in raw compute and capability, human researchers at labs like OpenAI, Anthropic, and Google DeepMind have remained the central intelligence steering the wheel. That status quo is rapidly approaching an expiration date. In a debate on the *Dwarkesh Podcast*, Ryan Greenblatt, Chief Scientist at Redwood Research, outlined a concrete timeline for a profound historical transition: **by 2030 or 2031, AI research and development will be fully automated by AI systems.** Once artificial intelligences match the top human experts in machine learning R&D, a feedback loop known as **Recursive Self-Improvement (RSI)** will trigger. In this scenario, human researchers will no longer write the code or design the architectures. Instead, superintelligent AI systems will automate their own upgrades, compressing four to five years of historical AI progress into a single calendar year(leading directly to Artificial Superintelligence (ASI) by 2033). Here is the technical case for why 2031 may mark the final year that human beings conduct frontier AI research. To understand why AI research will be automated long before many other human professions, one must look at the specific nature of Machine Learning (ML) engineering. Unlike fields that require slow, messy real-world feedback (such as trial law, geopolitical negotiation, or clinical medicine), AI R&D is uniquely **verifiable, containerizable, and amenable to empirical hill-climbing.** ``` ┌────────────────────────────────────────────────────────┐│ The RSI Feedback Loop │├────────────────────────────────────────────────────────┤│ ││ ┌───────────────┐ ┌────────────────────┐ ││ │ AI Researcher │ ────────> │ Builds Smarter │ ││ │ Model │ │ Next-Gen AI │ ││ └───────────────┘ └────────────────────┘ ││ ▲ │ ││ │ │ ││ └───────────────────────────────┘ ││ Feeds Back into R&D ││ │└────────────────────────────────────────────────────────┘ ``` As Greenblatt highlights, frontier AI companies are explicitly prioritizing making their own models exceptional at AI research. The domain possesses structural qualities that make it an ideal playground for Reinforcement Learning (RL): AIs can be aggressively trained via RL on thousands of small-scale, containerized ML environments. For example, a model can be assigned to run hyperparameter tuning, optimizer tweaks, or architectural experiments on small compute clusters (e.g., eight H100 GPUs) running mini-models similar to Andrej Karpathy’s nanoGPT speedruns. The model’s objective is simple and verifiable: lower the target loss function as quickly as possible. Unlike pure mathematics or theoretical physics, which often require deep, non-obvious abstractions to make a single leap, progress in machine learning is heavily additive and multiplicative. Innovations (ranging from context window extensions and better data curation to novel RL objective functions) generally stack cleanly on top of one another without interfering. In math, it is difficult to know if a researcher is halfway to proving a conjecture. In machine learning, if the goal is to cut training loss or achieve a benchmark 2x faster, progress is visibly measurable at every step along the training curve. Because machine learning is fundamentally an empirical discipline driven by trial and error within software containers, AIs can run millions of experiments in parallel, acquiring intuition and “taste” at a pace thousands of times faster than any human researcher. How exactly will an AI lab transition from human researchers to a fully automated AI research team? The process will not happen overnight in a single giant jump, but through a structured, multi-tier training pipeline. Imagine a near-future model (let’s call it **GPT-7.5)**. To turn GPT-7.5 into a world-class machine learning researcher capable of designing **GPT-9**, the lab will put it through three distinct layers of training: ``` ┌─────────────────────────────────────────────────────────┐│ AI Researcher Training Stack │├─────────────────────────────────────────────────────────┤│ 1. Small-Scale Pre-Training (GPT-2 Scale) ││ • Hyperparameter search, architecture tweaks ││ • Mini-model speedruns on small GPU clusters │├─────────────────────────────────────────────────────────┤│ 2. Mid-Scale Fine-Tuning (GPT-6 Scale) ││ • Post-training & mid-training recipe optimization ││ • De-risking medium-scale experiments │├─────────────────────────────────────────────────────────┤│ 3. Real-World Production R&D ││ • On-policy & off-policy RL on critical research path││ • Large-scale bug hunting & hyperparameter tuning │└─────────────────────────────────────────────────────────┘ ``` GPT-7.5 is placed in thousands of isolated environments where it is tasked with pre-training small, GPT-2-sized models across diverse modalities (image classification, video generation, game-playing RL models, and text LLMs). The model is rewarded purely on how efficiently its algorithmic changes improve the target metrics. The model is given access to mid-tier compute resources (e.g., fine-tuning or post-training runs on a model like GPT-6). Here, it learns how to tweak post-training pipelines, design reasoning environments, and de-risk experimental setups that closely resemble frontier architectures. Finally, as GPT-7.5 operates on the critical path to building GPT-8, its real-world research wins are converted into RL rewards. If GPT-7.5 proposes a post-training method that successfully de-risks a major frontier run, that rollout is heavily reinforced. Furthermore, training AIs to detect subtle bugs in complex, distributed infrastructure will be among the easiest tasks to automate. Greenblatt points to a well-known industry rumor regarding legendary engineer Noam Shazeer when he re-joined Google DeepMind (GDM): “There’s a rumor that right after Noam Shazeer joined GDM… they had a new really good training run, and the reason why is that Noam Shazeer just looked at their codebase and found a bunch of bugs, because he just knew where to look.” Training AIs to replicate this “Shazeer intuition”(scanning massive distributed codebases, pointing out subtle training bugs, and designing small-scale de-risking experiments to prove the bug exists) is a highly verifiable, containerizable task. An AI trained on millions of bug-injection environments will far surpass human ability at keeping massive, multi-gigawatt training runs bug-free. When an AI lab successfully automates its AI R&D, what does the speed of progress actually look like? Greenblatt’s median forecast is that automating AI research will yield **four to five years of historical AI progress compressed into a single calendar year.** To understand the sheer magnitude of this leap, consider the historical progress of AI over the past six years. Around 2020, OpenAI released **GPT-3**. Six years later, the industry moved to frontier models like **Mythos 5** and Anthropic’s **Fable 5**. MetricGPT-3 (circa 2020)Mythos 5 / Modern Frontier (circa 2026)**Training Compute**~3e23 FLOPs~3+ Orders of Magnitude Higher (~1e27 FLOPs)** Capabilities**Basic text completion; poor zero-shot reasoning; weak coding.Advanced agentic execution; complex multi-hour coding; long-horizon reasoning. Under Recursive Self-Improvement, the jump from a GPT-3-level baseline to a Mythos-level baseline will no longer take six years (it will happen in **12 months**, using the *same relative compute budget)*. How can an AI researcher model achieve five years of progress in one year without waiting for massive new hardware builds? By bridging compute gaps through **hyper-accelerated algorithmic progress.** Historically, AI progress has been driven by a combination of compute scaling and algorithmic efficiency. Greenblatt estimates that to achieve five years of progress in a single year, automated AI researchers will need to generate roughly **eight years’ worth of algorithmic progress** in 12 months. By inventing vastly superior data-filtering techniques, discovering more efficient transformer architectures, optimizing hyperparameters, and developing hyper-dense reasoning architectures, automated AI researchers can train models that match giant, compute-heavy systems while using a fraction of the hardware. We are already seeing early signs of this dynamic today. Token prices for frontier models have remained remarkably low ($30 to $50 per million tokens) despite massive leaps in intelligence. This suggests that labs are deliberately shifting toward training smaller, algorithmically denser models that allow for faster iteration, rather than relying solely on giant, brute-force compute runs. A primary objection to the 2031 automation timeline(frequently raised by skeptics like Dwarkesh Patel) is the **data bottleneck**. Patel argues that modern AI progress relies on a deca-billion-dollar human data industry. Frontiers models are good at coding, law, and infrastructure because armies of human experts write specialized SFT traces and RL environments. If human expert data is required to make AIs smarter, doesn’t human data generation inherently bottleneck the speed of recursive self-improvement? Greenblatt dismantles this objection with two key counterarguments: Frontier AI labs currently spend overwhelmingly more on compute than on human expert data. The spending split between compute and data is roughly **10:1 or 20:1**. If human expert data were the primary limiting factor of progress, labs would be directing far higher percentages of their capital toward data labeling rather than gigawatt data centers. The massive leaps in pre-training data quality in recent years (such as the transition from raw web scrapes like OpenWebText to highly refined datasets like FineWeb) were not caused by hiring millions of humans to type custom text. They were achieved through **scientific breakthroughs in data curation, automated filtering, and schleppy engineering.** ``` [Raw Web Scrapes] ──> (AI Filtering & Curation Science) ──> [High-Density Datasets] ``` An automated AI researcher does not need humans to write new data. It uses its intelligence to better understand what data matters, filter out garbage, design synthetic RL environments, and orchestrate automated evaluations. Even if one accepts that AI R&D can be automated in software, can an AI really master the real world without human-like physical experience? What about complex, non-verifiable domains like semiconductor manufacturing, hardware engineering, or geopolitical maneuvering? The key insight is that an AI need not be a master of political maneuvering to radically transform the physical world. It only needs to master **hardware R&D, chip design, and robotics.** ``` ┌─────────────────────────────────────────────────────────┐│ Industrial Explosion │├─────────────────────────────────────────────────────────┤│ 1. AI automates software AI R&D ││ │ ││ ▼ ││ 2. AI automates chip design & semiconductor R&D ││ │ ││ ▼ ││ 3. AI designs next-gen automated fabs & robotics ││ │ ││ ▼ ││ 4. Physical explosion of compute & robotic labor │└─────────────────────────────────────────────────────────┘ ``` Once AI researchers achieve superhuman capabilities in AI software, they will immediately apply those capabilities to **hardware co-design**: This creates an **industrial explosion**. Even if an AI model is initially unversed in the subtle nuances of Texas politics or international diplomacy, its ability to manufacture millions of autonomous robots, build gigawatt data centers, and design novel physical hardware will completely alter the global economy. It is the 21st-century equivalent of entering an 18th-century political debate with a fleet of steamships and telegraph lines: physical and technological dominance renders local friction secondary. If Ryan Greenblatt’s median forecasts hold, the timeline for human technological supremacy looks like this: When we look back on the history of science, the period between 1950 and 2030 will be viewed as a brief, transitional interregnum where biological human brains designed primitive computing systems. By 2031, the rate of technological discovery will far outpace the biological clock of human neurons. The researchers of the future will not be computer science PhDs working in Silicon Valley (they will be vast clusters of silicon intelligence running millions of experiments per second, designing minds that humans can no longer fully comprehend). [Why 2031 Might Be the Last Year Humans Do AI Research](https://blog.devgenius.io/why-2031-might-be-the-last-year-humans-do-ai-research-fb84a8bdaaa8) was originally published in [Dev Genius](https://blog.devgenius.io) on Medium, where people are continuing the conversation by highlighting and responding to this story.