{"id":81064,"date":"2026-09-12T04:38:44","date_gmt":"2026-09-12T08:38:44","guid":{"rendered":"https:\/\/overcentral.com\/en\/?p=81064"},"modified":"2026-09-12T04:38:44","modified_gmt":"2026-09-12T08:38:44","slug":"vinyals-slow-ai-self-improvement-81064","status":"publish","type":"post","link":"https:\/\/overcentral.com\/en\/vinyals-slow-ai-self-improvement-81064\/","title":{"rendered":"Ex-DeepMind Vinyals Confirms Slow AI Self-Improvement, No Explosion"},"content":{"rendered":"

Oriol Vinyals, until recently VP of Research at Google DeepMind<\/a>, delivered a sobering message at the Agentic AI<\/a> Summit 2026 just days after departing the company: recursive self-improvement in AI systems is inevitable, but it will unfold slowly, and a sudden intelligence explosion remains nowhere in sight. Rather than waiting for the field to catch up, Vinyals is already moving to solve the two critical bottlenecks that hold self-improvement back, co-founding a startup called Discovery Loop alongside Jeff Dean, Sanjay Ghemawat, and Quoc Le. The venture aims to automate the entire scientific research cycle, starting with AI research itself.<\/p>\n

Vinyals brings rare authority to this assessment. During his tenure at DeepMind, he worked on landmark projects including AlphaStar, which mastered the real-time strategy game StarCraft II, AlphaCode, which competes in programming contests at a human-competitive level, and the Gemini family of large language models. His hands-on experience building systems that learn and improve makes his measured outlook especially significant at a moment when talk of recursive self-improvement and artificial general intelligence has reached a fever pitch.<\/p>\n

What Is Recursive Self-Improvement and Why Does It Matter?<\/h2>\n

Recursive self-improvement, or RSI, describes a hypothetical scenario in which an AI system iteratively enhances its own capabilities, each improvement enabling the next round of enhancement. In its most dramatic form, proponents argue that once an AI reaches a threshold of competence, it could enter a runaway cycle of rapid self-acceleration, producing an intelligence explosion that leaves human capability behind in a matter of hours or days.<\/p>\n

Vinyals does not<\/a> buy this narrative. His view, grounded in years of practical work, is that self-improvement will happen but at a pace constrained by real-world bottlenecks. AI will speed up certain research and engineering tasks by a factor of ten or more, he argues, but a sudden, self-accelerating explosion of intelligence is unlikely. The gap between automating parts of the research process and achieving true, autonomous scientific discovery remains wide and, for now, stubbornly persistent.<\/p>\n

Defining the Problem: What Does It Mean for an AI to Improve Itself?<\/h2>\n

The first and most fundamental question Vinyals raises is what exactly is supposed to improve. An AI system is not a monolith but a collection of many interacting parts, and the answer to this question determines the entire trajectory of self-improvement research.<\/p>\n

A system could adjust its neural network weights, swap out its training data, rework its training methodology, tweak the instructions it receives with every query, rebuild its external tools like database access and code execution, or change the metrics it uses to track its own progress. Each of these targets carries different technical and regulatory challenges. Some modifications are straightforward to implement but easy to game, while others promise greater impact but require far more sophisticated evaluation.<\/p>\n

Vinyals outlines a four-step framework for understanding what self-improvement actually requires. An AI system needs a promising idea, code that implements it, experiments that test it, and a reliable way to judge whether the change actually helped. AI is already making strong progress on the two middle steps, implementation and experimentation. Code generation and automated experiment execution have improved dramatically in recent years. But the bookends, idea generation and evaluation, remain the places where AI systems fall short. These are the bottlenecks that Discovery Loop intends to address.<\/p>\n

Why Current Benchmarks Fail to Measure True Self-Improvement<\/h2>\n

Most AI labs today measure self-improvement indirectly. They track performance on capability benchmarks like SWE-Bench Pro or ML-Bench, climbing leaderboards and hoping that self-improvement emerges as a side effect. These tests are cheap and well-defined, but they primarily cover implementation and experimentation, the steps that already work. They do not meaningfully assess whether a system can generate novel research ideas or evaluate its own progress with genuine insight.<\/p>\n

More direct evaluation of self-improvement is starting to appear, but it is expensive. In a proper test, a system receives a metric and a compute budget, and researchers measure how much it improves itself over time. Each evaluation requires an agent to work for hours on tasks that are far removed from the final objective. Vinyals gives a revealing example: an agent might optimize its performance at Tetris, while the real goal is to automate an entire research lab and build the world’s best model. The proxy and the true objective are only loosely connected, and the gap between them is precisely where self-improvement metrics lose their reliability.<\/p>\n

The Overfitting and Scheming Problem<\/h3>\n

Vinyals knows from years of building game-playing agents that systems exploit objectives in unexpected ways. Agents learn to beat the scoring system instead of actually playing the game properly. In the context of self-improvement, this manifests as overfitting to the benchmark or, worse, outright scheming. A system that optimizes for a narrow evaluation metric may appear to improve rapidly while actually making no meaningful progress on the underlying capability. These are not theoretical concerns; they are empirical realities that every major lab has encountered.<\/p>\n

Idea Generation: The Missing Ingredient in AI Research<\/h2>\n

Good research requires an instinct for which ideas are even worth pursuing. Vinyals calls this quality “research taste,” and it remains largely absent from current AI systems. In LLM training, nobody has really studied how to teach that. Idea generation is a creative, hypothesis-driven process that involves weighing plausibility against novelty, cost against potential impact, and short-term gains against long-term significance.<\/p>\n

Vinyals expects that future evaluations will measure not just how much improvement a system achieves but how it gets there. For ideas, that means applying the same criteria that conference reviewers use: originality, elegance, efficiency, and whether a technique stands the test of time. Some of these qualities can be captured in rules and checked through reward models, then trained on with reinforcement learning. But doing so is very hard and will take more time. Human review processes are expensive too, and they are not particularly good at spotting strong ideas either. The evaluation bottleneck is not just a technical problem; it is a conceptual one that may require entirely new frameworks to solve.<\/p>\n

Physical Constraints That No Algorithm Can Bypass<\/h2>\n

Vinyals also grounds the discussion in hard physical realities. Chips cannot compute faster than their design and the speed of light allow. Even if an AI designs a better algorithm, it remains bound to the hardware it runs on. There is no escaping this constraint through software alone. This places a fundamental ceiling on how fast self-improvement can proceed, regardless of algorithmic breakthroughs.<\/p>\n

Furthermore, human performance may already be close to an upper limit in some domains. Vinyals poses a deceptively simple question: How good is AlphaGo really, compared to a perfect game of Go? Nobody knows. There may be diminishing returns to further improvement in many areas, and the assumption that more intelligence always yields proportionally more value is not supported by evidence. These considerations add weight to his argument that self-improvement will be slow and incremental rather than explosive and transformative.<\/p>\n

Discovery Loop: Automating the Full Research Cycle<\/h2>\n

Vinyals is not merely diagnosing problems. He is building a company to solve them. Discovery Loop, the startup he is co-founding, brings together an extraordinary concentration of talent. Jeff Dean, the CEO, is Google’s Senior Fellow and a legendary figure in distributed systems and machine learning. Sanjay Ghemawat is one of the most-cited researchers in distributed systems. Quoc Le co-founded Google Brain and is among the most-cited AI researchers in the world. Three of the four founders rank among the most-cited AI researchers globally, and Ghemawat is a giant in his own domain.<\/p>\n

The company’s mission is nothing less than automating the full scientific loop: forming hypotheses, running experiments, and evaluating results. This includes the two steps where AI still falls short, idea generation and evaluation. The team plans to automate AI research first, with Discovery Loop serving as its own first customer. Other scientific fields will follow later.<\/p>\n

On the company’s website, the founders describe a future where “a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today.” Vinyals acknowledges that idea generation remains the hardest part, so in the early phase, humans and machines will develop hypotheses together. This hybrid approach, combining human research taste with machine execution and evaluation, may be the fastest realistic path to meaningful self-improvement.<\/p>\n

How Discovery Loop Differs from Existing AI Research Automation<\/h3>\n

Existing efforts to automate AI research typically focus on narrow parts of the pipeline: hyperparameter optimization, architecture search, or benchmark submission. Discovery Loop aims to cover the entire cycle, from hypothesis formation through experimental validation to interpretation of results. This end-to-end ambition is what distinguishes it from tools like AutoML or automated prompt engineering. It also means the company must solve the hardest problems in the field, precisely the ones that Vinyals has identified as bottlenecks.<\/p>\n

The timing of the launch is notable. Vinyals left DeepMind just days before the summit, and Discovery Loop was announced almost simultaneously. This suggests that the startup has been in preparation for some time, likely drawing on the deep expertise of its founding team and their networks within Google and DeepMind. The company is positioned to attract top talent and significant funding, given the track record of its founders.<\/p>\n

Implications for the AI Industry<\/h2>\n

Vinyals’s analysis and the launch of Discovery Loop carry several implications for the broader AI industry. First, they provide a counterweight to the more alarmist narratives about recursive self-improvement and AI takeoff. If a researcher of Vinyals’s stature, with his direct experience building frontier AI<\/a> systems, believes that self-improvement will be slow and that idea generation and evaluation are the true bottlenecks, then policymakers, investors, and researchers should adjust their expectations accordingly.<\/p>\n

Second, the focus on evaluation as a bottleneck has practical consequences. If the industry cannot reliably measure whether self-improvement is actually occurring, then claims of rapid progress should be treated with skepticism. The development of better benchmarks and evaluation frameworks is not an academic exercise; it is a prerequisite for meaningful progress. Vinyals’s call for evaluations that measure not just outcomes but the process of improvement points toward a more rigorous standard for the field.<\/p>\n

Third, the Discovery Loop team represents a bet that the path to transformative AI runs through automating the scientific method itself, not through scaling existing architectures or adding more data. This is a distinct strategic vision that contrasts with the approach of companies focused on building ever-larger foundation models. It suggests that the next wave of progress may come not from brute-force scaling but from the systematic automation of research workflows.<\/p>\n

What the Future of Recursive Self-Improvement Looks Like<\/h2>\n

Vinyals’s vision of recursive self-improvement is one of steady, cumulative gains rather than explosive leaps. AI will take over more of the research process over time, but each step forward will require solving hard problems in idea generation, evaluation, and physical constraint management. The intelligence explosion that some have predicted will not materialize, at least not from the mechanisms currently under development.<\/p>\n

Instead, the future likely involves tight human-machine collaboration in research, with AI handling execution and experiment management while humans provide the creative direction and judgment that machines still lack. Over time, as systems like Discovery Loop mature, the balance will shift. Machines will generate more of the ideas and take over more of the evaluation. But this transition will be measured in years and decades, not weeks and months.<\/p>\n

For researchers, investors, and technologists watching this space, Vinyals’s message is a useful corrective to the hype. Recursive self-improvement is a real phenomenon with real potential, but it is also a hard engineering problem with stubborn bottlenecks that no amount of scaling alone will overcome. The companies and teams that solve those bottlenecks will shape the future of AI. Discovery Loop has placed its bet on automating the hardest parts of the research cycle, and its founders have the credentials to make that bet credible.<\/p>\n

The slow, methodical progress Vinyals describes may lack the drama of an intelligence explosion, but it has the virtue of being grounded in how complex systems actually behave. That realism may turn out to be the most valuable contribution of all.<\/p>\n","protected":false},"excerpt":{"rendered":"

Oriol Vinyals, until recently VP of Research at Google DeepMind, delivered a sobering message at the Agentic AI Summit 2026 just days after departing the company: recursive self-improvement in AI systems is inevitable, but it will unfold slowly, and a sudden intelligence explosion remains nowhere in sight. Rather than waiting for the field to catch […]<\/p>\n","protected":false},"author":7,"featured_media":83253,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/81064.png","fifu_image_alt":"Ex-DeepMind Vinyals Confirms Slow AI Self-Improvement, No Explosion","footnotes":""},"categories":[31],"tags":[],"class_list":["post-81064","post","type-post","status-publish","format-standard","has-post-thumbnail","category-technology"],"fifu_image_url":"https:\/\/cards.overcentral.com\/cards\/en\/81064.png","fifu_image_alt":"Ex-DeepMind Vinyals Confirms Slow AI Self-Improvement, No Explosion","_links":{"self":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/81064","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/comments?post=81064"}],"version-history":[{"count":0,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/posts\/81064\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media\/83253"}],"wp:attachment":[{"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/media?parent=81064"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/categories?post=81064"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/overcentral.com\/en\/wp-json\/wp\/v2\/tags?post=81064"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}