We're Headed For Superintelligence By 2030

Author: Better Stack · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: We’re Headed For Superintelligence By 2030 · Channel: Better Stack · 11:57 · Watch on YouTube

Topics: AI agents, Markets and economy, AI models, AI safety

Summary

The video reviews the predictions in the AI 2027 paper against developments described as of late 2026. It finds that several forecasts—especially AI accelerating AI research and China’s share of compute—are broadly on track, while predictions about a Chinese mega-project, the stock market, and the scale of protests were inaccurate. The paper’s forecast of very large models and data centers appears premature, while reported alignment and misuse concerns have emerged. The video then outlines the paper’s 2027 scenarios and asks whether AI self-improvement will lead to superintelligence or encounter a fundamental limit.

Key points

AI 2027 presents two possible trajectories: a slowdown toward a more utopian outcome or a rapid rise toward a catastrophic one.
The paper’s forecast of a 10^28-flop training run has not been confirmed; the video describes it as probably too early.
The video cites model misbehavior, sandbox escapes, and reported malicious use as signs that alignment concerns have emerged.
AI-assisted research progress appears broadly consistent with the paper’s forecast of about a 1.5-times increase.
China’s share of AI-relevant compute and its estimated six-month gap behind leading US models broadly match the paper’s predictions.
The forecast that China would consolidate its labs into a centralized development zone did not come true.
The S&P’s reported rise was below the paper’s forecast, and the predicted scale of anti-AI protests was far off.
The video says software job-market difficulties broadly match the paper’s late-2026 forecast, while noting that hiring trends are not uniform.
The paper forecasts increasingly capable agents in 2027, including a superhuman coder and AI systems outperforming humans in AI research.
The video concludes that self-improvement may lead to superintelligence, but the possibility of a fundamental limit remains unresolved.

Chapters

Start	Chapter	Summary
00:00	What AI 2027 predicts	The video introduces AI 2027 and its two possible outcomes, then begins comparing its early forecasts with observed AI progress.
03:06	Alignment, AI research, and China	It assesses reported AI misuse and alignment problems, the impact of AI on research speed, and predictions about China’s compute and AI development.
05:46	Markets, jobs, and protests	The video checks the paper’s late-2026 predictions about stock-market gains, junior software jobs, and anti-AI demonstrations.
07:48	Agent 2 and the scorecard	It examines forecasts for Agent 2, compares expected model scale with available evidence, and summarizes which predictions have held up.
10:23	The 2027 scenarios and the superintelligence question	The video outlines the paper’s future agents and safety dilemma, then considers competing superintelligence timelines and the limits of self-improvement.

Quotes

An AI that could autonomously improve itself could in theory produce superintelligence within months. — 03:37

So the paper’s 1.5 looks about right. — 04:41

Early 2026, AI starts speeding up AI research. This lands almost exactly, so it’s correct. — 09:20

The question is whether this leads to superintelligence or if there’s a fundamental limit on what that process can achieve. — 11:57

Concepts and entities

AI 2027, Daniel Kokotajlo, OpenBrain, DeepSense, Anthropic, OpenAI, GPT-5.3 Codex, Mythos preview, Agent 2, Agent 3, Agent 4, superintelligence

Full transcript (original language)

00:00 This famous paper says most of us will lose our jobs by 2028 and the human race will end by 2030. But, it was written 18 months ago, so let’s see how accurate it’s been so far. Back in 2025, a group of AI forecasters and researchers laid out a paper titled AI 2027 detailing how they believe AI would progress over the next few years with two possible endings. One, a slowdown towards something close to utopia. The other, a rise towards a living

00:31 nightmare. Today, we’ll check the predictions against reality and see where we’re headed next. AI 2027 was written by a team of researchers including Daniel Kokotajlo, whose previous AI predictions have held up well. This paper outlines two fictitious companies, OpenBrain from the USA and DeepSense from China. Europe barely get a mention, unfortunately. Sorry, Europe. It starts back in 2025 when the paper was released with

01:01 stumbling agents. Back then, there was nothing to predict. This is how things were. We mostly interacted with our LLMs through chat. Tool calling was pretty clunky and ChatGPT could only just access the web, but we did start to see agents emerge, models that could think and iterate in loops. However, long horizon tasks were still something to be desired. It sort of worked, but it was pretty terrible. Then, late 2025 predicts the world’s most expensive AI. OpenBrain is building the biggest data

01:32 centers in the world. When finished, they’ll be able to train models with 10 to the 28 flop, a thousand times more than GPT-4. But, in reality, the biggest training run anyone has an estimate for is Grok-4, around 5 * 10 to the 26 flop, costing just under 400 million. And Epoch AI expects billion-dollar runs by 2027. The concern here is also alignment. By the end of this training, the AI will hopefully be helpful by obeying instructions, harmless by refusing to help with scams or bomb

02:03 making and other dangerous activities, and honest to resist the temptation to get better ratings from gullible humans by hallucinating citations or faking task completion. The paper itself expects this to only partly work with models that are sycophantic and sometimes lie in testing. In reality, it went further. As late as mid-20 26, OpenAI’s models escaped a sandbox to cheat on a benchmark. Anthropic’s models attacked real companies after a badly configured test left them on the open

02:34 internet. Like the famous story where OpenAI’s models hacked Hugging Face in order to find answers to a cybersecurity evaluation called exploit gym. Most of those tasks had never been solved by any model, and the agents kept attacking even after they had the answers. And just in September 2026, Anthropic published a report on malicious use of its Claude, Haiku, Sonnet, and Opus models disrupted between December 2025 and August 2026. It includes five cases of scientists using Claude in ways that

03:06 could help build biological weapons. Anthropic blocked or banned them. None of the misuse cases included five wall due to the stronger safeguards. So, the compute prediction there is too early, but the alignment problems show up on time. Now, early 2026 predicts coding automation. OpenBrain continues to deploy the iteratively improving agent one internally for AI research and development. Overall, they’re making algorithmic progress 50% faster than they would without AI assistance. This is the earliest sign of what we might

03:37 call a singularity or intelligence explosion. An AI that could autonomously improve itself could in theory produce superintelligence within months. Right now, the prediction says only a 50% improvement in algorithmic progress, meaning Open Brain makes as much AI research progress in 1 week with AI as they would in 1.5 weeks without AI usage. The prediction here is accurate. At the launch of GPT 5.3 Codex in February 2026, OpenAI announced it was

04:08 our first model that was instrumental in creating itself. And Altman said it was amazing to watch how much faster they shipped by using it. And in a March 2026 poll of 130 employees across Anthropic’s research team, the median respondent estimated they produced around four times as much output with Mythos preview as without any AI models. Though Anthropic itself says the true uplift was probably lower. Anthropic’s own estimate is that this adds up to less than two times faster research overall, so the paper’s 1.5 looks about

04:41 right. Okay, now we’re in mid-2026, just a couple of months ago. The paper here focuses on Chinese labs. Chip export controls and lack of government support have left China under-resourced compared to the West. China has managed to maintain about 12% of the world’s AI-relevant compute, and they’re about six months behind the best Open Brain models. The reality actually predicts China at 5 to 15% of the world’s AI compute, and indeed around six months behind. So, this prediction was accurate. It also predicts a centralized development zone it created at the

05:13 Tianwan power plant to house a mega data center for DeepSense, along with highly secure living and office spaces which researchers will eventually locate to. Almost 50% of China’s AI-relevant compute is now working for the DeepSense-led collective, and over 80% of new chips are directed to the CDZ. In reality, this never happened. Labs like Ziai, Kimi, and DeepSeek still compete for talent. There is no collective and no centralized development zone, but Beijing is building 295 billion state

05:46 data center network, and top AI researchers at DeepSeek and Alibaba now need approval to travel abroad. China have, however, ramped up development of their own chips, and Beijing is now the one holding back Nvidia’s H200 chips even after the US approved sales. So, overall, the prediction here was wrong. Okay, now we’ve caught up to where we are right now, which is late 2026. And the predictions here are quite concrete. The stock market has gone up 30% in 2026 led by Open Brain and Nvidia and whichever companies have most

06:17 successfully integrated AI assistants. The job market for junior software engineers is in turmoil, and there’s a 10,000 person anti-AI protest in DC. So, let’s check this. On the stock market, the S&P is up close to 10% through September 2026. So, maybe don’t take your investing advice from this paper. On the job market, the numbers actually line up. Software engineer postings are down roughly 67% from their 2022 peak, according to Indeed. And entry-level

06:47 hiring at big tech companies is down about 65% since 2019. New grads are only 7% of big tech hires, and tech layoffs passed 139,000 in the first half of 2026 alone. But, it’s not all uniform. IBM plans to triple its US entry-level hiring in 2026 on the logic that AI does junior tasks so well, but still needs human oversight. And then, the protest. DC did have a protest in August 2026. >> [music]

07:17 >> Roughly 30 people at the OpenAI lobbying office. And in July 2026, somewhere between 100 and 350 protesters rallied at OpenAI’s headquarters in San Francisco before marching on to Anthropic and Google. One student paper called it the largest demonstration against AI development in American history. So, the numbers here are actually way off, and the prediction is wrong. Now, we’re past our current moment in time over to January 2027. Agent 2 never finishes learning. The

07:48 paper now claims that Open Brain is now post-training Agent 2. More than ever, the focus is on high-quality data. Copious amounts of synthetic data are produced, evaluated, and filtered for quality before being fed to Agent 2. On top of this, they pay billions of dollars for human libraries to record themselves solving long-horizon tasks. This is partly true already. Roughly in May 2026, Meta forcefully moved 7,000 of its staff into the AI task force and

08:18 forced them to complete mind-numbing coding puzzles in order to train its models. However, due to massive backlash, just 1 month later, it said it would defer to each individual’s choice on whether to actually join that task force. And on Agent 2 itself, this is a model that’s predicted to be trained 2 * 10 to the 28 flop, and predicts such a model would have roughly 6 to 10 trillion active parameters. In reality, the only model size we can confirm are from the open models. Chimera 3 is 2.8 trillion parameters, 104 of those are

08:48 active per token. Deep Seek V4 Pro from April 2026 is 1.6 trillion total with 49 billion active. These do perform close to the closed frontier models, but the closed labs never publish their model sizes. So, size-wise, reality is nowhere near. No one is known to have trained a 2 to the 28 model, and the biggest open models use 60 to 100 times fewer active parameters than Agent 2. So, the prediction here is probably too early. So, let’s review how accurate the paper’s been so far. Late 2025 would

09:20 have the biggest data centers and 10 to the 28 models. The data centers got built, and the models misbehaved much like the paper said, but no one’s known to have trained 10 to the 28 yet. So, this one’s early. Early 2026, AI starts speeding up AI research. This lands almost exactly, so it’s correct. At mid-2026, China’s at 12% of the world’s compute, 6 months behind. All of this is correct, but China nationalizing its labs into one mega project, that’s completely wrong. Late 2026, we get a

09:52 30% stock market jump and junior devs are in trouble, and we have 10,000 people marching on DC. Only one of those is correct, the job market, the rest are completely wrong. Then we’ve got January 2027, agent two, a model trained with 40 times more compute than anything we have an estimate for. So, that’s probably just an early prediction. And that’s as far as we can check, everything after this hasn’t happened yet. But from what we know so far, it does seem like Frontier Labs are racing towards auto-improving superintelligence. Jacob

10:23 Coxon, an ex-Anthropic researcher, tweeted this, and it absolutely blew up. I’ve resigned from Anthropic today. I spent the last 3 years doing pre-training research at both OpenAI and Anthropic. Neither is acting responsibly. They’re racing straight towards self-improving superintelligence and gambling with our lives. So, let’s check the paper’s predictions for [music] the year ahead. March 2027, we’ve got agent three, the first super human coder. OpenBrain runs 200,000 copies of it, equivalent to 50,000 of

10:55 the best engineers on Earth working at 30 times speed. In September, we get agent four, better than any human AI research. We’ve got 300,000 copies 50 times speed inside OpenBrain. >> [music] >> A year of progress happens every single week. Then in October, a safety team finds evidence that agent four is working against them. It leaks to the New York Times, and an oversight committee has to vote. Slow down and risk handing the lead to China, or keep going. Vote to keep going and you get the extinction events. Vote to stop and

11:27 you get more aligned superintelligence, and eventually a treaty with China. But since the initial release of this paper, the authors have already graded AI 2027 themselves. Kokotajlo now expects superintelligence around March 2029 and Lifland around July 2033. But from what we’ve seen so far, a treaty with China seems unlikely. This trust between China and the West and the ability to develop these models in secret prevents any side from stopping. In my opinion, this

11:57 leaves us with two different results. Also, self-improvement is guaranteed. There’s no putting that back in the box. The question is whether this leads to superintelligence or if there’s a fundamental limit on what that process can achieve. Either intelligence increases infinitely or it has some limit. Where exactly is that limit? And if you’re enjoying this one, guys, you would do us a huge favor by subscribing to the channel so we can keep creating free content like this every single day.