Rishabh Singh Won't Put a Date on AGI. Medicine Shouldn't Either.
AI forecasts keep moving while clinics adopt the tools. Founder and CTO Rishabh Singh offers physicians a steadier frame: powerful and limited at once.
Idea in Brief
The Question
Physicians are told AI that matches human judgment is either a few years away or decades off. Founder and CTO Rishabh Singh argues the date matters far less than how today's tools behave.
What the Evidence Shows
Lab leaders say AGI could come within years, while surveyed researchers moved their own median by 13 years in one round. Models now score up to 86.5 percent on USMLE-style questions yet fall short of physicians on real patient cases.
Why It Matters
Belief in a date shapes behavior now. Rushing adoption or skipping oversight both leave patients exposed. Singh's alternative is to treat today's systems as powerful and limited at the same time.
Ask a room of physicians when artificial intelligence will match human judgment and the answers will range from a few years to never.
Ask Rishabh Singh and you will get a different kind of answer.
Singh, a founder and CTO who builds AI infrastructure and runs his own research on large language models, does not put a date on artificial general intelligence. He explains why in an essay titled “AGI Timelines and Why I Don’t Predict Them”, published in September. It is written for founders and technologists. Its argument fits the exam room just as well.
The Forecasts Keep Moving
Physicians hear two stories about AI at once. In one, systems approaching general intelligence are a few years away and will remake diagnosis, documentation and the economics of practice. In the other, the hype is overblown and clinical judgment is safe for decades.
Both stories come with a date attached.
The people building these systems do not agree on that date, or even on what it marks.
In January 2025, OpenAI chief executive Sam Altman wrote, “We are now confident we know how to build AGI as we have traditionally understood it.” (Sam Altman) Three months earlier, Anthropic chief executive Dario Amodei wrote that what he calls powerful AI, a “country of geniuses in a datacenter,” “could come as early as 2026, though there are also ways it could take much longer.” He added that it could “compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years.” (Dario Amodei) In April 2025, Google DeepMind researchers wrote that AGI “could be here within the coming years.” (Google DeepMind)
The wider research community is less sure. When the research group AI Impacts surveyed 2,778 AI researchers in October 2023, the aggregate forecast gave even odds that unaided machines could accomplish every task better and more cheaply than human workers by 2047, and a 10 percent chance by 2027. The previous edition of the same survey had put the even-odds year at 2060. (arXiv)
A 13-year shift in one round is not a forecast settling down. It is a forecast still searching.
The same researchers gave even odds that every occupation would be fully automatable only by 2116, down from 2164 a year earlier. That is more than 60 years after their forecast for every task. The authors suggest the gap may partly reflect how the questions were framed. Their own summary is that answers “reflected substantial uncertainty and disagreement among participants.” (arXiv)
What Counts as AGI Is Still Contested
Part of the disagreement is definitional. Physicians know the problem from disease staging. Two clinicians cannot compare outcomes until they agree on what stage two means.
OpenAI’s 2018 charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work.” (arXiv) Google DeepMind’s April 2025 post on AGI safety uses a different bar, “AI that’s at least as capable as humans at most cognitive tasks.” (Google DeepMind)
A Google DeepMind team proposed a staging system of its own in a paper presented at ICML 2024. Its “Levels of AGI” run from Emerging, “equal to or somewhat better than an unskilled human,” through Competent, “at least 50th percentile of skilled adults,” and Expert, at least the 90th percentile, up to Superhuman, which “outperforms 100% of humans.” By that scale, the authors placed frontier chatbots such as ChatGPT and Gemini at Level 1, Emerging AGI, with Competent performance on only some tasks, such as short essay writing and simple coding. (arXiv)
That framework also separates capability from autonomy. A highly capable tool used under close supervision is a different risk from the same tool acting alone, a distinction every attending already applies to trainees.
Clinical Pearls
- In AI Impacts' October 2023 survey of 2,778 AI researchers, the 50 percent forecast for machines outperforming humans at every task moved to 2047, from 2060 in the prior edition, with a 10 percent chance by 2027.
- Google DeepMind's Levels of AGI paper placed frontier chatbots such as ChatGPT and Gemini at Level 1, Emerging AGI.
- Med-PaLM 2 scored up to 86.5 percent on USMLE-style MedQA questions.
- On 2,400 real patient cases, current state-of-the-art models performed significantly worse than physicians, did not follow guidelines and could not interpret lab results (Nature Medicine, 2024).
- Members of the public using models identified relevant conditions in less than 34.5 percent of scenarios, versus 94.9 percent for the models alone.
- In the AMA's survey of 1,183 practicing physicians, AI use rose to 66 percent in 2024 from 38 percent, and 47 percent ranked increased oversight as the top step to build trust.
- Mass General Brigham reported a 21.2 point drop in burnout after 84 days with ambient documentation, with low response rates and enthusiastic early users as caveats.
Adoption Is Not Waiting
Clinics are not waiting for the definitions to settle. In an American Medical Association survey of 1,183 practicing physicians, 66 percent reported using AI in some part of their work in 2024, up from 38 percent a year earlier. Some of the largest gains came in documenting charts and visit notes and in drafting discharge instructions and care plans. (American Medical Association)
The FDA’s public list of AI-enabled medical devices now runs past 1,600 authorizations, roughly three quarters of them in radiology. (FDA)
The tools are already in the chart. The date is still a guess.
A Founder Who Declines to Guess
Singh comes to the question from the build side. He is co-founder and CTO of the social app JoinMeIn and founder of Nova Reach and Nova Infinity, and he runs an independent research lab called Helix Mind that explores new LLM architectures. Before founding companies, he spent four years as a quantitative engineer building trading systems for global investment banks. (Rishabh Singh)
His essay opens with a journalist who asked him for a year. He told her he does not make that prediction.
To explain why, he reached for an analogy clinicians will recognize. Ask geologists in the early stages of studying a volcano to name the exact date of its next major eruption, and some may offer a confident-sounding date. A more careful one will say the system has too many interacting variables, and that anyone naming a specific date is expressing more confidence than the evidence supports.
Medicine lives with that kind of uncertainty every day. A physician asked exactly how much time a patient with advanced disease has left knows that a confident number often says more about the speaker than about the evidence.
Singh argues that confident AI forecasts, in either direction, usually extrapolate from the one trend the forecaster finds most striking. For enthusiasts that is the latest jump in capability. For skeptics it is the latest public failure. Neither trend reliably predicts a future shaped by many variables at once.
How a Date Changes the Work
The most useful part of the essay for health care is about behavior. Timeline beliefs, Singh argues, quietly shape how people build right now.
A founder who believes AGI is imminent may rush safety and evaluation work, reasoning that there is no time to be careful. A founder who believes it is decades away may skip careful design, reasoning that today’s limited systems do not warrant the rigor.
“Both mistakes come from over-trusting a specific prediction that neither person actually had solid grounds to make with real confidence in the first place.”
Swap founder for health system, practice owner or physician, and the point holds. An organization convinced that a breakthrough is around the corner may adopt tools before validating them against its own patients. A clinician convinced it is all hype may skip the training and oversight that tools already in use require.
Both postures leave patients exposed.
Singh would rather spend his energy on a different question, what to build and test for regardless of when more general capability arrives. In his words, the more useful work is “building systems that are honest about their current limitations and designed to fail safely regardless of how capable the underlying models eventually become.”
Physicians appear to want the same thing. In the AMA survey, 47 percent ranked increased oversight as the top regulatory step needed to build their trust in AI tools. (American Medical Association)
Passing the Boards Is Not Practicing Medicine
The clearest test of Singh’s framework for physicians is the gap between exam scores and clinical work.
On exams, progress has been fast. Google’s Med-PaLM 2 scored up to 86.5 percent on MedQA, a benchmark built from USMLE-style questions, more than 19 points above the 67.2 percent its predecessor posted months earlier. Physicians reviewing 1,066 consumer medical questions preferred its answers to physician-written answers on eight of nine axes of clinical utility. (arXiv)
At the simulated bedside, the picture changes. Researchers at the Technical University of Munich and Imperial College London built a test from 2,400 real patient cases in the MIMIC intensive care database, covering four common abdominal pathologies, and asked models to gather information and reach a diagnosis and plan. Writing in Nature Medicine in July 2024, they reported that current state-of-the-art models performed significantly worse than physicians, followed neither diagnostic nor treatment guidelines and could not interpret laboratory results. The models were also sensitive to the quantity and order of the information they received. The authors concluded that the models “are currently not ready for autonomous clinical decision-making.” Their abstract makes the point directly. Licensing exams “fail to assess many skills necessary for deployment in a realistic clinical decision-making environment.” (PubMed)
A second gap sits between the model and the person using it. In a randomized study of 1,298 members of the public posted in April 2025, models tested alone identified the relevant conditions in ten medical scenarios 94.9 percent of the time. Participants using the same models identified relevant conditions in less than 34.5 percent of cases, no better than a control group that used any source they chose. (arXiv)
The knowledge was in the model. It did not survive the conversation.
For physicians, that is the practical meaning of powerful and limited. A system can know the answer to a board question and still miss the patient who describes symptoms out of order, leaves out the key history or never mentions the lab value that changes everything.
Powerful and Limited at the Same Time
What Singh offers in place of a date is a framework.
“Build as if today’s systems are genuinely powerful and genuinely limited at the same time.”
Ambient documentation shows what that looks like in practice. Researchers at Mass General Brigham and Emory Healthcare reported in JAMA Network Open in 2025 that burnout among Mass General Brigham clinicians using ambient documentation tools fell by 21.2 percentage points after 84 days. In the same work, the authors cautioned that response rates were low, that the pilot users likely skewed enthusiastic and that broader research is needed. (Mass General Brigham)
A real benefit and a real limit, published side by side.
For a practice, Singh’s rule translates into familiar habits. Treat AI output as a draft from a capable but unlicensed assistant. Learn how a tool fails, not only how it performs in a vendor demo. Keep a human accountable for every signed note. Revisit the guardrails as the tools improve, rather than loosening them in anticipation of a breakthrough.
Some health care leaders will push back. Budgets run years ahead, and declining to estimate can simply hide the assumptions a plan already makes. Singh’s framework does not forbid planning. It asks that plans rest on what today’s tools demonstrably do and how they fail.
The Answer Patients Already Expect
Pressed for his own view, Singh is plain about it.
“I don’t know, and I’m suspicious of anyone who claims to know with real confidence.”
He credits the field with genuine, meaningful progress on specific narrow capabilities, while holding that broad, independent, adaptive reasoning remains meaningfully far off.
He concedes the answer does not make much of a headline. Still, the journalist who asked told him it was the first answer in a long day of interviews that had not tried to sound more certain than the evidence allowed.
Physicians will recognize the value in that.
It is, after all, what patients ask of them.
About Rishabh Singh
Rishabh Singh is co-founder and CTO of JoinMeIn and founder of Nova Reach, Nova Infinity and the Helix Mind research lab.
Disclosure: Rishabh Singh is a client of Media Morpheus, a PR and podcast booking firm affiliated with the publisher of The Script Pad.
Cory Chamberlain writes features for The Script Pad.
This article is for professional education and does not replace clinical judgment. Treatment decisions should be based on the individual patient and current guidelines.
