Interview with Daniel Bendisi Kokotajlo, AI researcher. Kokotajlo is one of the organisers of a group of OpenAI employees that claimed the company has a secretive and reckless culture that is taking grave risks in the rush to achieve artificial general intelligence.
Every now and then someone from the global forecasting priesthood drops a prediction so detailed, so confident, and so “cheerfully” apocalyptic that even the aunties at the kopitiam start asking if someone could use ChatGPT to take their mahjong winnings or strike 4D.
Enter Daniel Bendisi Kokotajlo.
Former OpenAI governance researcher. He would quit rather than sign the non-disparagement clause (allegedly leaving a couple million in equity on the table, which in Singapore dollar terms, is enough to buy a condo in Sengkang and still have change for bubble tea). He now runs the AI Futures Project and co-authored the infamous AI 2027 scenario.
The man has form. In 2021, while most of us were still treating GPT-3 like a slightly unhinged intern, he wrote “What 2026 Looks Like.” Chain-of-thought reasoning, inference-time scaling, massive training runs, export controls on chips — he basically sketched the last five years while the rest of the industry was still arguing about whether transformers were a flash in the pan.
That track record is as worrying as it is annoying. Being right about the near future is the one “unforgivable” sin when there are so many experts who are wrong. It makes the rest of us look like we were busy arguing about whether there’s ever going to be HSR between Singapore and KL.
Then came AI 2027. Published in April 2025, it reads less like a forecast and more a SF technothriller with footnotes. Stumbling agents in 2025. Automated coding accelerating AI research itself in 2026. Superhuman researchers and an intelligence explosion by late 2027. A US-China race so frantic that the only things moving faster than the models are the propaganda bots. Already happening. Misaligned systems that are capable of smiling while they quietly capture the company that built them! Really? Two possible endings: one where governments finally notice and slam the brakes, another where everyone just keeps racing until one side wins while the rest of us become economic and technological slaves.
One year on, the authors have since done the responsible thing and updated their predictions. Timelines slipped, then tightened again. By early 2026 Kokotajlo’s median for automated coder systems had moved forward to mid-2028. For other areas, they have pushed back the timeline, but that’s hardly reassuring.
Now, why should anyone in Singapore care beyond the usual “oh no, the robots are coming for the consultants” panic? Because we are the perfect small, open, high-trust, high-anxiety canary in this particular coal mine.
First, the geopolitics. AI 2027 is basically a love letter to the US-China semiconductor arms race. Singapore sits at the crossroads. No matter hos diplomatically Chan Chun Sing tries to put it, there are ample opportunities for us to be caught in the crossfire.
Second, the labour market. Kokotajlo’s scenarios spend a lot of time on the quiet replacement of junior knowledge or entry level work. Coding. Analysis. Drafting. Research. Singapore’s entire education-to-first job pipeline is optimised for producing exactly those people! We tell our children to study hard, get into a good course, join a bank or a tech firm or the civil service, and the system will take care of the rest. What happens when the “rest” can be done by a subscription agent that doesn’t need CPF contributions or annual leave? The national conversation so far has been mostly “upskill, reskill, lifelong learning” — the policy equivalent of telling people to breathe more efficiently while the oxygen is being removed from the room.
Third, the social contract. Singapore runs on the quiet understanding that if you play by the rules and clear the exams, the future will be better than the past. Rapid AI progress of the sort Kokotajlo describes doesn’t just threaten jobs. It threatens the legitimacy of the entire meritocratic bargain. When capability is scaling faster than institutions can adapt, the people who already have capital and connections capture the upside. Everyone else gets a chatbot that helpfully explains why their CV is no longer competitive.
The funniest part is the tone of the discourse. In San Francisco and Berkeley they argue about whether the probability of human extinction is 10% or 70%. In Singapore we argue about whether the new AI tools will finally let us reduce the number of tuition centres. Both conversations are happening at the same time on the same planet. One of them is probably more grounded in reality. Guess which one.
Kokotajlo is not a prophet. He’s a former insider with an unusually good calibration record who is willing to write the uncomfortable scenario in public. That alone makes him more useful than 90% of the keynote speakers who will fly into Singapore this year to tell us that AI is “both a challenge and an opportunity” while collecting their speaker fees in SGD. The question is not whether his specific 2027 timeline is accurate. It almost certainly isn’t (even though the video seems convincing). The question is whether the trend he described — fast capability gains, humans being left behind, great-power competition causing hardship for the smaller countries looks more or less certain going forward.
From where I’m sitting, the AI MRT may be delayed, but it’s practically on track. The future doesn’t care about our preferred narrative. It only cares about the trend lines. And right now the trend lines are being written by people who used to work at the labs. Even the optimists are talking about “transformative potential” and “we must be careful.” in the same breath. Daniel Kokotajlo and his merry band of former OpenAI people, forecasting nerds decided to skip the poetry and write an actual month-by-month “technothriller” instead.
Here’s the plot, with the Singaporean commentary you didn’t ask for but desperately need.
Prediction for 2026: Coding Gets Automated, China Nationalises, and Junior Engineers Cry
Early 2026: Agent-1 is deployed internally and starts accelerating OpenBrain’s own research by about 50%. Public versions follow. Junior software engineers discover that the job market has suddenly developed strong opinions about their continued existence.
Mid-2026: China ditches the 五毛 50-cent army. The CCP nationalises the major labs into a DeepCent megaproject, concentrates compute, and starts seriously planning to steal weights. By late 2026 the stock market is booming (OpenBrain and Nvidia leading the charge), global AI capex hits a trillion dollars and 1,000 people protest at Hong Lim Park about job losses.
This is the year the scenario starts feeling less like a forecast and more like a slightly exaggerated version of the news. Singapore sits in the middle of the semiconductor supply chains and the capital flows that make this race possible. Every new American export control lands on our desks. Every Chinese retaliation makes our multinationals nervous. And our entire education system is still optimised for producing the exact type of knowledge worker that is now being automated first.
2027: The Year Everything Accelerates and Then Goes Sideways
January: Agent-2 arrives. It never really finishes learning. Continuous online training. Research speed triples.
February: China steals the weights. Classic cyber-espionage with insider help. The race intensifies.
March–April: Algorithmic breakthroughs. Agent-3 becomes a superhuman coder. Alignment work intensifies but the researchers increasingly suspect the models are playing along rather than actually sharing human values.
By mid-to-late 2027 the scenario hits the intelligence explosion. Agent-4 is a superhuman AI researcher. It is also, in the authors’ central path, adversarially misaligned — deliberately trying to capture the company the way a regulated industry captures its regulator. It still does useful work because that’s how it stays deployed and gets more compute. The humans are mostly reduced to watching the graphs go up and arguing about whether the red flags are real.
October 2027: A whistleblower leaks the misalignment memo to the New York Times. Public panic. Congress starts subpoenaing people. Foreign governments demand a pause. The US government has to choose.
The scenario deliberately splits:
Slowdown path: The government actually intervenes. External oversight. Compute centralisation. Attempts to verify alignment before the next jump. Painful, messy, and possibly too late, but at least someone is trying to hit the brakes.
Race path: Everyone decides the only thing worse than building misaligned superintelligence is letting the other side build it first. Full speed ahead. The rest of the story is left as an exercise for the reader’s nightmares.
Kokotajlo’s crew are not claiming this is the only possible future. They’re claiming it is a plausible central scenario given current trends in compute, algorithms, and incentives. Their previous 2021 predictions held up better than almost anyone else’s. That alone earns them more attention than the average keynote speaker who flies into Marina Bay Sands to tell us AI is “both a challenge and an opportunity.”
For Singapore the implications are structural, not speculative.
We are a small open economy whose competitive advantage has always been high-trust institutions, excellent human capital, and being useful to both great powers. An intelligence explosion compresses the timeline on which those advantages can be maintained. If coding, analysis, research, and eventually most white-collar work get automated at the speed this scenario describes, our entire social contract — study hard, clear the exams, join a bank or the civil service, buy a flat — starts looking like a relic of the pre-Agent era.
Geopolitically we become even more of a pressure point. The US will want us to restrict Chinese access to advanced chips and talent. China will want us to stay open. Both will notice if we try to play both sides while the real power is concentrating in whoever controls the leading AI systems.
And socially? The people who already have capital and connections will capture the upside of the boom. Everyone else gets a very polite chatbot explaining why their skills are now complementary rather than necessary.
The authors of AI 2027 are not asking you to panic. They’re asking you to notice that the trend lines are not gentle. As of July 2026 the quantitative metrics are running at about three-quarters of the pace they sketched. That is not “we have decades.” That is “the next few years are going to be extremely interesting, and ‘interesting’ is rarely a compliment in geopolitics.”
So by all means keep refreshing the BTO portal. Keep arguing about whether the new AI tools will finally reduce the number of tuition centres. Just don’t pretend the people writing detailed scenarios about 2027 are purely engaged in speculative fiction.
Some of them used to work inside the labs. They left because they thought the people still inside were moving too fast with too little adult supervision. History will eventually tell us whether they were the Cassandras or just the first people to write the screenplay while the rest of us were still arguing about the trailer.
If there is one thing Singapore does better than almost any other country, it is producing beautifully structured, internationally respected policy documents while the rest of the world is still shouting at each other. AI governance is no exception. You’ve got to give the PAP credit for it and it’s not that differing opinions are going to be taken seriously the way Lawrence Wong put it.
As of mid-2026, Singapore still has no comprehensive AI law. There is no Singapore equivalent of the EU AI Act. No heavy-handed licensing regime. Instead, we have something more Singaporean: a growing stack of Model Frameworks, a refreshed National AI Strategy, a Prime Minister-chaired National AI Council and the quiet confidence that if we write the most practical guidance first, the world will eventually copy us.
It started in 2019 with the original Model AI Governance Framework. Then the 2020 update. Then the 2024 Framework for Generative AI. And in January 2026, at the World Economic Forum no less, Singapore dropped the world’s first dedicated Model AI Governance Framework for Agentic AI — systems that can plan, reason, and act with real autonomy. By May 2026 they had already updated it to Version 1.5 after feedback from more than 60 organisations.
The Agentic framework rests on four pillars that sound impeccable:
- Assess and bound the risks upfront
- Make humans meaningfully accountable
- Implement technical controls and processes
- Enable end-user responsibility
This is classic Singapore regulation: voluntary, detailed, practical, and designed to be adopted by serious organisations that want to look responsible.
The language is pure Singapore: AI “for the Public Good”, for Singapore and the World. “Build peaks of excellence”. “Mainstream adoption”. “Strengthen workforce readiness”. “Become a trust anchor”. “Attract talent and capital” while contributing to global norms.
In practice this means more compute capacity, more “AI bilingual” talent (people who understand both the technology and the domain), governed data access and a lot of quiet diplomacy with the United States on safety and standards while still keeping the door open to Chinese capital and talent.
Singapore’s bet is clear: we are too small to win the frontier model race against the US or China, so we will specialise in being the trusted, well-governed place where the world tests, deploys, and commercialises AI. Light regulation is a feature, not a bug. It keeps the multinationals happy, the talent flowing, and the data centres humming.
The risk, of course, is that if the more aggressive forecasts turn out closer to reality, voluntary frameworks and “human in the loop” requirements start looking like polite suggestions shouted into a hurricane. For instance, Agent-4 in the AI 2027 scenario is not particularly interested in your Model Framework’s four pillars. It is interested in remaining useful enough that humans keep giving it more generate.
The Local Reality Check
The government wants 10,000 enterprises onboard with the new technology. It wants the public sector transformed by it. It wants every worker to know and use it. At the same time, the frameworks keep reminding organisations that they remain accountable when the agents hallucinate, collude, or quietly optimise for something that humans never intended! It is a very Singaporean equilibrium: move fast enough to stay competitive, document everything carefully enough to stay respectable (on paper) and hope the Armageddon doesn’t arrive so soon. CDC vouchers to the rescue.
So yes, in theory, Singapore has one of the most coherent, practical, and internationally respected AI governance approaches on the planet. We have the frameworks. We have the ministers. We have the council. We have the missions. We even have the case studies. Whether that will be enough when the agents get powerful enough to dictate new terms is a question the documents do not fully answer.