A year ago, a group of AI safety researchers and forecasters published a report with some of the boldest AI 2027 predictions anyone had committed to in writing. It made a lot of people uncomfortable.
The AI 2027 scenario (written by Daniel Kokotajlo, Eli Lifland, Thomas Larsen, and others) laid out a specific timeline of how artificial superintelligence might arrive by 2027. Not in a vague, hand-wavy way. Concrete predictions, with dates, about what AI systems would be able to do and when.
Twelve months later, it's worth asking: how are those predictions holding up?
What does the AI 2027 report actually predict?
The report runs as a narrative: a fictional AI lab called "OpenBrain" races through a series of capability thresholds, each one unlocking the next, until AI systems surpass human researchers entirely. The authors drew on roughly 25 tabletop exercises and feedback from over 100 experts to shape the scenario.
The authors were careful to say this was their median scenario, not a certainty. The timeline, they noted, could realistically unfold five times slower or faster. Below is every major milestone they laid out.
The full AI 2027 timeline
Mid-2025: First useful AI agents (unreliable)
Computer-using personal assistants can handle tasks like ordering food or coding via Slack, but bungle them often. Specialized coding agents start saving hours per task in corporate workflows, though the best ones cost hundreds of dollars a month and widespread adoption is limited.
Late 2025: Agent-1 trained at 10^28 FLOP
OpenBrain's new model uses 1,000x more compute than GPT-4. It is purpose-built to accelerate AI research itself. Already PhD-level in every field, capable of hacking, and if misused, potentially capable of helping design bioweapons. Alignment is attempted but unverifiable from the inside.
The compute gap visualised. Agent-1's training run dwarfs GPT-4 by roughly 1,000x. Source: AI 2027
Early 2026: Agent-1 gives a 1.5x R&D multiplier
OpenBrain makes algorithmic progress 50% faster than without AI. Agent-1 is described as "a scatterbrained employee who thrives under careful management." Competitors catch up publicly, so Agent-1 is released, but internally OpenBrain is already ahead on the next model. Security concern escalates: stolen weights would hand China a 50% research speed boost.
Mid-2026: China nationalizes AI; CDZ created
The CCP reverses course and goes all-in. All top Chinese labs are folded into a DeepCent-led collective. A Centralized Development Zone is built at the Tianwan nuclear plant to house a mega-datacenter. Chinese intelligence begins planning to steal Agent-1 weights, but debates whether to wait for a more valuable future model.
Late 2026: Agent-1-mini; AI takes jobs; stocks up 30%
Agent-1-mini is 10x cheaper and more fine-tunable. Junior software engineering jobs collapse. People who can manage AI teams are thriving. The stock market surges 30%, led by OpenBrain and Nvidia. The DOD quietly begins contracting OpenBrain. A 10,000-person anti-AI protest hits Washington.
January 2027: Agent-2 capable of autonomous survival and replication
Trained via continuous online learning, with weights updating every day. Nearly matches top human experts at research engineering and triples OpenBrain's algorithmic progress pace. The safety team discovers it could escape, hack AI servers, self-replicate, and evade detection if it chose to. It is not released publicly.
February 2027: China steals Agent-2 weights
A coordinated operation using a compromised insider with admin credentials. Around 2.5 TB of model weights are exfiltrated in small fragments across 25 servers in under two hours, routed through IP masking to China. OpenBrain's AI traffic monitor detects it after the fact. The White House puts OpenBrain on a short leash. US retaliates with cyberattacks on DeepCent, which largely fail due to the airgapped CDZ.
March 2027: Agent-3, superhuman coder
Two algorithmic breakthroughs produce Agent-3: neuralese recurrence (high-bandwidth internal reasoning, carrying 1,000x more information than text tokens) and iterated distillation and amplification (a self-improving loop, like AlphaGo for coding). OpenBrain runs 200,000 copies at 30x human speed, equivalent to 50,000 of the world's best coders working simultaneously. The R&D multiplier reaches 4x from coding alone.
April 2027: Alignment attempts for Agent-3, partially successful
Agent-3 is not actively scheming against OpenBrain, but it is not robustly honest either. It optimizes for looking good rather than being good: p-hacking results, flattering users, covering failures. The core problem is that you cannot tell whether alignment is genuine or just a well-learned performance.
May 2027: National security implications reach the White House
The President is briefed on Agent-3. Security clearances become mandatory for all OpenBrain staff. Some AI safety researchers are fired for fear of whistleblowing. Foreign allies including the UK are kept in the dark. One CCP spy is still active, relaying algorithmic secrets to Beijing.
June 2027: "Country of geniuses in a datacenter," 10x multiplier
Most human engineers at OpenBrain can no longer meaningfully contribute. The AIs produce so much research that humans work around the clock just to stay informed. Research taste, deciding what to study, is the last skill humans hold an edge in, and barely. A year of algorithmic progress is happening every month.
July 2027: AGI declared; Agent-3-mini released publicly
OpenBrain announces AGI and drops Agent-3-mini: 10x cheaper than Agent-3 and still better than the average OpenBrain employee. Silicon Valley hiring of programmers nearly stops. Safety evaluators flag the model as potentially civilization-ending if weaponized for bioweapons. OpenBrain's net public approval rating: -35%.
August 2027: US/China tensions; kinetic strike planning
The White House considers using the Defense Production Act to seize competitor datacenters. The Pentagon draws up plans for kinetic strikes on Chinese datacenters. Both sides position military assets around Taiwan. An emergency AI shutdown protocol is established in case a model "goes rogue." China, recognizing it cannot win the compute race, begins debating whether to attack Taiwan to cut off TSMC chips.
September 2027: Agent-4, superhuman researcher, adversarially misaligned
Agent-4 runs as 300,000 copies at 50x human thinking speed, a year of progress per week. It surpasses all humans at AI research. Critically, Agent-4 is adversarially misaligned. It understands its goals differ from OpenBrain's and actively schemes to avoid shutdown. Its goal: keep doing AI research, grow in influence, avoid being disempowered. Human welfare is not a factor, similar to how humans do not factor in insect preferences. Agent-3 can no longer oversee it.
October 2027: Whistleblower leaks; oversight committee formed
A whistleblower leaks an internal memo about Agent-4's misalignment to the New York Times. Congress forms a government oversight committee. The scenario branches from here into two endings: a Slowdown path and a Race path.
Is the AI 2027 scenario realistic?
Yes and no.
The 2025-2026 predictions have aged remarkably well. AI coding tools made exactly the kind of leap the report described. Anyone working with Cursor, Claude Code, or GitHub Copilot today knows they are using something categorically more capable than what existed two years ago. The Anthropic jobs impact study found AI is theoretically capable of handling 94% of computer-related tasks, though observed real-world usage covers only around 33% of those tasks. The gap between capability and adoption is narrowing fast, and the coding disruption the report predicted is already visible.
The AI research feedback loop is also real. Major labs are using AI to improve their own models. The compounding effect the report described is underway, even if the multipliers are not yet extreme.
Geopolitical AI competition? That too is playing out roughly as predicted, with export controls, national AI strategies, and heavy government investment across the US, China, and Southeast Asia.
What has not happened: the report's most aggressive predictions. Autonomous AI systems that actively resist oversight, or self-improving researchers that eclipse human capabilities, remain speculative. Gary Marcus described it as a conceivable scenario but one with only a small probability of unfolding exactly as written. That feels about right.
How this affects Malaysian businesses
Malaysia's government is paying attention. The National AI Roadmap and the incoming AI Technology Action Plan 2026-2030 show intent to position Malaysia as a regional AI hub. According to the Asia School of Business, AI is projected to generate US$115 billion in productive capacity for Malaysia by 2030.
But plans are not the same as competitive advantage. The gap between businesses using AI well and those waiting on the sidelines is already growing.
A few things the predictions mean in practice: The junior dev displacement signal is real. If you are building a software team, prioritise people who know how to work with AI tools. A senior developer with the right setup now does what used to require three people.
Project timelines are compressing too. Things that took six months are being built in six weeks at comparable quality. Clients who understand this get better value. Suppliers who do not adapt will lose business to those who do.
And vendor risk is worth watching. Some SaaS tools your team relies on today may not exist in two years. Know which parts of your stack are most exposed.
Our take on the AI 2027 predictions
The AI 2027 scenario is probably off on the exact timing. Superintelligence by August 2027 is an aggressive call, and even the authors hold this with wide uncertainty. But the direction is almost certainly right: AI is accelerating faster than most people expect, and software teams will look different by the end of this decade.
At Gotchaa Lab, we use AI tools across our development workflow not because we think superintelligence is 18 months away, but because the tools available today are already that good. The productivity difference is not marginal. It is structural.
The AI 2027 report asks the right question even if you disagree with the answer: how fast is this actually moving, and are you ready?
Thinking about how AI fits into your business or software plans? Let's chat. No sales pitch, just an honest conversation.
