|
Getting your Trinity Audio player ready...
|
In December, two of the world’s most respected AI researchers, Edward Hughes and Louis Kirsch, walked away from Google to chase a single, unusually specific goal: building an AI system capable of building a better version of itself. Their new London startup, Inherent, now centers its entire operation around a prototype called Faraday, named after the 19th-century physicist, which absorbs enormous volumes of data from the company’s own researchers, emails, meeting transcripts, instant messages, ongoing conversations with Faraday itself, and uses that material to train an improved successor version of itself.
Inherent is not alone in pursuing this idea. Jeff Clune, a computer scientist and co-founder of Vancouver-based startup Recursive Superintelligence, is chasing a nearly identical goal under a name that leaves little ambiguity about the company’s ambitions. Both firms are working toward what computer scientists call recursive self-improvement, a scenario in which an AI system becomes capable of designing, training and refining its own successors with progressively less human involvement, potentially triggering a feedback loop where capability gains compound faster than researchers can track or control them.
The concept itself is not new. AI researchers and forecasters have discussed recursive self-improvement, sometimes described as an intelligence explosion, for years, but it moved from theoretical speculation toward active engineering priority only recently. According to reporting picked up from The New York Times, the core fear driving concern among AI safety researchers is straightforward: if a system can meaningfully accelerate its own development without a human in the loop checking each step, the resulting progress could compound at a pace that outstrips any organization’s ability to test for safety, alignment or unintended behavior before deploying the next version.
Evidence that today’s AI systems are already contributing meaningfully to AI development, even without full autonomy, has been mounting throughout 2026. Large language models can already write functional code, generate synthetic training data, and help optimize the computer chips they run on, according to IBM’s research coverage of the topic. OpenAI Chief Scientist Jakub Pachocki has described automating scientific discovery as one of the company’s core priorities, with an explicit goal of building automated researchers capable of improving AI systems further, a plan first flagged following GPT-5’s gold medal performance at the 2025 International Math Olympiad. The subject has become significant enough within the field that ICML 2026, one of machine learning’s most prestigious academic conferences, hosted a dedicated workshop on AI systems capable of recursive self-improvement.
Not every recent test of these capabilities has gone the way optimists hoped. Research covered by MIT Technology Review in August tested Anthropic’s Claude Opus 4.8 on genuinely novel machine-learning research questions drawn from papers submitted to NeurIPS 2026, one of the field’s top conferences, using an open-source framework called OpenClaw. The results suggested that fully autonomous, self-directed scientific progress remains considerably further off than the most dramatic forecasts imply, indicating that while AI systems can meaningfully assist with research tasks, genuinely novel scientific reasoning without significant human framing and oversight is still a harder problem than compounding-growth scenarios assume.
That gap between current capability and the feared endpoint is a point both leading AI labs have been careful to draw attention to publicly. According to IBM’s coverage, OpenAI has stated that fully autonomous recursive self-improvement is not happening today, and Anthropic has said its own systems cannot yet autonomously build their successors. Both companies, notably, acknowledge that AI already plays an increasingly large role in coding, running experiments and other parts of the AI development pipeline, even while stopping short of the fully self-directed loop that defines the most concerning version of the scenario. That distinction, between AI assisting human-directed development and AI independently driving its own improvement, is where much of the current debate inside the field actually lives.
Academic researchers have also begun trying to formally measure how close current systems are to genuine recursive self-improvement rather than relying purely on speculation. A recent survey published on arXiv interviewed 25 AI researchers drawn from academic literature, industry labs and AI safety organizations to gather a range of perspectives on self-improvement dynamics, part of a broader effort to move the conversation from forecasting exercises like the widely discussed AI 2027 scenario report toward more grounded, evidence-based assessment of where current systems actually stand relative to the capability threshold that would make an intelligence explosion plausible.
The public discourse around this topic has grown loud enough that search interest in terms like “AI liftoff” has climbed noticeably over the past year, driven by a mix of genuine technical concern, media coverage emphasizing worst-case scenarios, and the broader wave of AI safety anxiety that has swept through the industry following a string of public warnings from current and former researchers at major labs. That said, mainstream AI researchers have generally cautioned that no confirmed event or specific technical breakthrough has directly triggered this surge in attention, and most scientists continue to emphasize that today’s systems do not display the kind of rapid, self-directed improvement depicted in the more dramatic popular narratives.
What makes the current moment different from earlier rounds of AI risk discussion is the growing number of well-credentialed researchers who have left major labs specifically to build toward recursive self-improvement rather than to caution against it, a shift that suggests at least part of the field views the capability as inevitable and worth pursuing deliberately rather than something to be avoided entirely. Whether companies like Inherent and Recursive Superintelligence end up demonstrating a genuine, controllable version of self-improving AI, or whether the technical and safety hurdles prove far more stubborn than their founders currently expect, is likely to become one of the defining technical questions the AI industry grapples with over the next several years. Continuing coverage of how frontier AI research is evolving and what it means for safety and industry direction is available on Business Tech. Additional background on how AI labs are approaching self-improvement research can be found through IBM’s coverage of the topic, and further detail on recent experimental results is available through MIT Technology Review’s reporting.
