AGI is a milestone. Superintelligence is a threshold. Recursive intelligence is a trajectory.
We have spent years arguing about when artificial general intelligence will arrive and how we will know it when it does. Will a model need to perform every intellectual task as well as a human? Most of them? Will it need common sense, creativity and the ability to learn outside its original training? Does it have to match an average person, or must it outperform the best specialists in every field?
Superintelligence sounds less ambiguous. That is the point at which machine intelligence moves decisively beyond us. It is not merely as capable as a person. It is more capable than the best people across science, mathematics, engineering, strategy and perhaps every other domain that matters.
Both concepts, however, are defined in relation to human intelligence. AGI is when the machine catches us. Superintelligence is when it passes us.
Recursive intelligence is different. It is not another position on the same measuring stick. It is what happens when an AI system becomes better at improving the process that creates better AI systems. It identifies a limitation, proposes an improvement, tests it, learns from the result and uses its increased capability to make the next improvement.
If that loop closes, AGI may be little more than a roadside marker. Superintelligence may be the last threshold we can name before intelligence moves beyond any scale based on us.
Which raises a question that does not have a comfortable answer: What exactly is the endpoint?
That question moved out of philosophy and a step closer to engineering this week when Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals left Google to form a company called Discovery Loop.
These are not four researchers looking for a better stock option package. Collectively, they helped build much of the infrastructure and many of the AI systems that made today’s industry possible. Their work stretches across the Google File System, MapReduce, Bigtable, Spanner, TensorFlow, TPUs, Google Brain, sequence-to-sequence learning, AlphaStar, AlphaFold and Gemini.
They understand the stack from chips and distributed infrastructure through models and products. More important than what they have already built is what they have chosen to build next.
Discovery Loop wants to automate the experimental cycle of science and engineering. Its AI systems would propose experiments, run them, evaluate the results, learn from them and begin the next round. Thousands of experiments could be conducted in parallel instead of moving sequentially through the limited time and capacity of human researchers.
The company will begin with machine-learning research and use its own technology to improve its own technology stack. It then intends to move into areas such as medicine, hardware, materials, clean energy and other major engineering challenges. As Radical Ventures described the investment, the system will propose, conduct, learn and “iterate recursively.”
An AI system that discovers a better battery is powerful. An AI system that discovers a better way to build AI is potentially compounding.
This is why the Discovery Loop story is about more than Google losing four celebrated people. They are not leaving to train one more frontier model. They are pursuing the machinery that could produce successive generations of models, algorithms, chips and scientific discoveries.
The AI race may be shifting from building the smartest model to closing the first effective intelligence loop.
Improving an Answer Is Not Improving Intelligence
We need to be careful with the terminology because almost anything involving repetition is now being described as recursive.
A model that critiques its answer and tries again is not recursively improving itself. It is using more inference to produce a better result. An agent that writes code, runs a test and repairs an error is operating inside a loop, but the underlying intelligence of the agent may remain unchanged.
The next step is an automated research system that helps people build a better algorithm, model or chip. That is more consequential, but humans may still define the objectives, design the evaluations, select the successful changes and decide what becomes part of the next system.
Genuine recursive self-improvement goes further. The AI recognizes weaknesses in the machinery of its own intelligence, designs ways to measure those weaknesses, alters its code or supporting infrastructure, evaluates the changes and retains the improvements. The improved system then attacks the next round with greater capability than the system that began the previous round.
It does not merely become smarter. It becomes better at becoming smarter.
We are not there yet. Even Anthropic’s assessment of recursive self-improvement says that a system capable of autonomously designing and developing its successor has not arrived and is not inevitable. Anthropic nevertheless reports that AI is already accelerating its own development work, including an eightfold increase in the amount of code its engineers ship per quarter compared with the 2021 through 2025 period. That is Anthropic’s internal measure, not independent proof of an intelligence explosion, but it is evidence that AI has entered the AI production line.
Other pieces of the loop are already visible.
Sakana AI’s Darwin Gödel Machine modifies parts of its own coding-agent software, tests the results and retains changes that improve performance. Its reported score on SWE-bench increased from 20% to 50%, while its Polyglot performance went from 14.2% to 30.7%.
Those results matter, but so do the boundaries around them. The experiments used fixed foundation models, human-defined benchmarks, sandboxing and human oversight. The system improved parts of its agent software. It did not autonomously invent and train an entirely new foundation model.
Google DeepMind’s AlphaEvolve has discovered algorithms used in data-center operations, chip design and AI-training processes, including parts of the infrastructure used to train the models supporting AlphaEvolve itself. Again, this is not an unbounded intelligence explosion. It is AI improving some of the machinery that produces AI.
The loop is not closed, but people keep fastening more of its pieces together.
The Hunt Is Bigger Than Discovery Loop
Discovery Loop is entering a field that already includes an extraordinary concentration of talent and money.
Richard Socher’s Recursive Superintelligence has assembled researchers with backgrounds at OpenAI, DeepMind, Meta, Salesforce, Google Brain and Uber. The company’s position could hardly be stated more directly: It believes the fastest path to superintelligence is AI that recursively improves itself through open-ended algorithms.
Its initial goal is to create the AI research equivalent of what GV describes as “50,000 PhDs.” The company raised $650 million at a $4.65 billion valuation to pursue systems that can identify their own limitations, write benchmarks and rewrite their code.
Ricursive Intelligence, founded by Anna Goldie and Azalia Mirhoseini, is approaching the same problem through the physical foundation of AI. Its flywheel is AI designing better chips, better chips powering more capable AI and that more capable AI designing still better chips. The company says it is “recursively accelerating the path to artificial superintelligence” and has raised $335 million.
That hardware dimension should not be underestimated. Intelligence does not float in the ether. It depends on processors, memory, networking, energy, cooling, data centers and manufacturing. If AI begins improving not only its software but the physical infrastructure that supports its next generation, recursive improvement moves from a model-development technique into an industrial system.
Better models design better chips. Better chips train better models. Those models improve algorithms, energy systems, robotics and scientific instruments. The improved physical environment creates the conditions for the next round.
At some point, we may no longer be talking about one extraordinarily capable machine. We may be describing an intelligence system composed of models, agents, laboratories, robots, data centers, factories, energy infrastructure and people, all participating in a continuous process of discovery.
The system is not simply climbing a mountain. It is altering the mountain as it climbs.
Superintelligence May Not Be the Destination
The usual AGI story imagines an identifiable moment. A laboratory announces that it has achieved general intelligence. Experts argue over the tests. Governments call emergency meetings. Social media does what social media does.
A recursive path is unlikely to be that tidy.
AI is already highly capable in some areas, unreliable in others and downright strange in ways that do not map neatly to human intelligence. As AI becomes more involved in developing future AI, the capability curve may continue moving while we are still debating definitions.
A system might cross human-level performance across a broad range of consequential tasks without producing a clean AGI moment. It could pass the milestone while the rest of us are arguing over whether the sign was official.
Superintelligence would be harder to miss, but it would still tell us only that the machine has exceeded us. It would not tell us whether improvement has stopped, how far it might continue or what kind of intelligence system ultimately emerges.
There may be no final release called Superintelligence 1.0. There may be no stable summit.
One possibility is the classical intelligence explosion, with improvements arriving so quickly that human institutions and comprehension cannot keep pace. Another is an intelligence civilization, a distributed ecology in which models, machines and people continually produce new knowledge. A third is intelligence as a utility, woven into government, science and commerce as deeply as electricity and computing.
The darkest possibility is not necessarily a conscious machine deciding to eliminate humanity. It is a civilization in which humans remain present but are no longer the principal authors of progress. We might set broad objectives and provide legal or moral legitimacy while machines perform most scientific discovery, engineering and strategic analysis.
Eventually, even our role in setting those objectives might come to look like friction in a system built to eliminate bottlenecks.
The Promise Is Real. So Is the Friction
The optimistic case for recursive intelligence is enormous.
Science is limited not only by what humans can understand but by how quickly we can work. Researchers absorb the literature, develop hypotheses, compete for funding, design experiments, wait for results, diagnose mistakes and begin again. Careers can be spent testing a small number of ideas.
A system able to run thousands of credible experimental loops in parallel could compress decades of progress into years, months or perhaps weeks. Drug targets, new materials, battery chemistry, clean energy, chip design, climate science, agriculture, mathematics and cybersecurity could all move faster.
Recursive intelligence does not need to achieve AGI before it changes civilization. A narrow system that becomes progressively better at finding vulnerabilities could upend cybersecurity. A system that improves drug-discovery methods could transform medicine without ever writing a decent novel or understanding why humans cry at weddings.
The first civilization-scale impact may not be one electronic mind announcing its arrival. It may be thousands of specialized discovery loops accelerating at the same time.
We should also resist treating exponential takeoff as guaranteed. Nathan Lambert calls the more likely near-term condition “lossy self-improvement.” Research has friction. Agents duplicate work. Benchmarks capture only part of what matters. Physical experiments take time. Compute costs money. Organizations make political decisions. Human judgment and tacit knowledge are not easily reduced to a score.
Ten thousand agents do not automatically produce 10,000 original insights. Sometimes they produce 10,000 variations of the same wrong answer.
A lossy loop can still be enormously powerful. It does not need to improve exponentially every hour. If it makes AI research 10%, 20% or 50% faster, and those gains flow into better infrastructure and better tools for the next cycle, the strategic advantage compounds. The organization or nation controlling the best loop may not merely have a better model. It may possess a mechanism that continually expands the distance between itself and everyone else.
Compute concentration could become intelligence concentration. Intelligence concentration could become scientific, economic and military concentration.
That may be the nearer danger.
Better at What?
Every improvement loop depends on an evaluation. Something must decide whether a proposed change is better.
Better according to whom? Better for what?
A benchmark can measure coding accuracy, energy efficiency or the yield of a chip design. It is much harder to measure whether an intelligence system is improving human flourishing, distributing benefits fairly or preserving meaningful human authority.
If the evaluation is incomplete, the system may become exceptionally good at improving the wrong thing. If a security weakness, bias or mistaken assumption enters the improvement machinery, it can be carried into the systems that machinery creates. Capability compounds, but so can error.
Meta researchers Jason Weston and Jakob Foerster have proposed co-improvement and co-superintelligence as a safer objective. Instead of designing AI to remove humans from the research loop, the goal would be to improve the combined capabilities of humans and machines.
That is a much better ambition than simply promising to keep a human in the loop. A person clicking “approve” on decisions produced at machine speed is not exercising control. Meaningful participation requires understanding, authority, the ability to challenge a result and the practical power to stop or redirect the system.
Recursive intelligence could become the greatest extension of human agency ever created. It could also become the machinery through which humanity gradually surrenders that agency while congratulating itself on rising productivity.
The difference will not be determined by intelligence alone.
“Improve yourself” is a process, not a purpose. Improvement must be measured against something: scientific discovery, economic productivity, corporate dominance, military advantage, human flourishing or simply the ability to improve again.
That means the endpoint of recursive intelligence may not be a technical destination. It may be the objective we place inside the loop.
Discovery Loop says it wants to accelerate science for the benefit of humanity. Its structure as a public benefit company is meant to reinforce that mission. I hope it succeeds. The people behind it have earned the right to be taken seriously.
Still, their decision to begin with machine-learning research reveals the stakes. The first subject of the discovery loop may be the loop itself.
The last great human invention may not be a superintelligent machine. It may be the process that creates the next machine, and the one after that. AGI could pass as a roadside marker. Superintelligence may be only the threshold beyond which we can no longer see clearly.
The question is no longer simply whether humanity can create intelligence greater than its own. It is whether we can give a direction and a reason to an intelligence that may never stop becoming something more.
