Hacker News front page, 38 upvotes and 36 comments—AI 2027 has been getting serious traction in AI safety circles lately. The author, Daniel Kokotajlo, comes from the AI safety research community, with team members including Scott Alexander and Thomas Larsen. They split 2025-2027 into two narrative paths: Race (racing ahead) and Slowdown (hitting the brakes).
I read through it once, then went back to examine the key assumptions. Overall impression: the framework is elegant, milestones are pinned to specific months, but a few core assumptions deserve scrutiny.
What AI 2027 Is Actually Predicting
This is a scenario forecasting site, not a paper. Its format is unusual—future AI progress written as a novelized script, one milestone per month.
Core plot:
- 2025: Models approach human-expert level
- 2026: AI Agents take over increasing shares of R&D work
- 2027: AI begins automating AI research itself, triggering an intelligence explosion
The two branches differ in outcome: under Race, humanity doesn't have time to build governance frameworks and AI capabilities spiral out of control; under Slowdown, international coordination succeeds and R&D gets paused for two years.
Source: https://ai-2027.com
Where the Core Numbers Come From
The most interesting part of this forecast is that it's not purely qualitative narrative. The team offers specific numbers: annual training compute growth curves (extrapolated from 2019-2024 scaling data), the share of R&D work taken over by "AIA" (rising from 5% to 90%), probability estimates for alignment failure (~20-30%), and month-level positioning of key milestones.
These numbers shift the conversation from "will AI go out of control?" to "is this specific curve slope realistic?" That's why it resonates in rationalist communities—it breaks vague fears into discussable variables.
The problem: the compute curve is model extrapolation (scaling law extension), while "AI automating AI research" is a back-of-envelope guess. The former has five years of data behind it; the latter has no precedent.
Will an Intelligence Explosion Actually Happen?
This is AI 2027's boldest assumption: AI automates its own research, capabilities explode, things spiral out of control.
I've run similar experiments—using Claude with scaffolding to let the model rewrite its own prompts, tune hyperparameters, and run benchmarks. The real bottleneck isn't compute. It's three things:
- Data feedback loops: AI generating its own training data causes rapid overfitting to a narrow distribution
- Experiment design: Scientific breakthroughs often come from counter-intuitive hypotheses; LLMs have no strong drive to attempt things outside their reward signal
- Evaluation bias: AI optimizes existing metrics, but paper-level progress requires inventing new metrics
AI 2027 says little about these engineering realities. It assumes "AI Agents multiplied by compute equals research output," when the actual equation is far more complex.
The Human Decision Window Isn't That Narrow
The forecast pins 2027 as the critical node. But historically, there's typically a 10-20 year buffer between a technical breakthrough and large-scale deployment.
A few comparisons:
- Nuclear tech: Fission in 1945, large-scale civilian power generation in the 1960s
- GMOs: First transgenic plant in 1983, commercial cultivation in 1996
- mRNA vaccines: The technology matured over decades; COVID forced scaled deployment
The pattern: technical capability does not equal deployment capability. Between them sit regulation, interest-group politics, and public acceptance. AI 2027 conflates the two, assuming a capability breakthrough produces societal impact within six months.
This isn't optimism—the decision window may indeed be wider than the scenario suggests, but it serves a different purpose: not for "slamming the brakes," but for "building guardrails."
Why You Should Still Read It
Despite the critiques above, I think this forecast is worth your time. Not because its predictions are accurate—no scenario predicting 2027 can be accurate.
The reason is that it makes vague fears concrete as traceable milestones. When you see something like "Q3 2026: AI Agents take over 50% of R&D," your brain immediately generates concrete counter-arguments: which companies? How would you verify? Where's the data?
That's its real value. The safety governance community doesn't need more accurate predictions—that's impossible. What they need are scenario scripts that can be discussed, refuted, and revised. AI 2027 is the most systematic version currently available.
Editorial Reflections
AI 2027's real contribution is making disagreements explicit—it forces you to pick a side: do you believe the Race path's loss-of-control risk is larger, or do you believe the Slowdown path's international coordination can work? This "binary choice" framework mirrors how climate modeling and epidemic forecasting operate: divergent predictions come first, then empirical evidence gradually narrows the range.
The biggest risk of contrarian forecasts is that they push decision-makers to over-allocate resources to the worst-case scenario. If there's a 70% probability of a gradual path (AI slowly permeating various industries), but all AI policy is designed around a 5% extreme-disaster scenario, you waste the window for addressing more likely problems.
A contrarian observation: narrative-driven forecasts like AI 2027 influence researcher mindsets far more than they influence policy. They shape the priority ordering of the next generation of AI safety researchers—not necessarily a bad thing, but worth watching. Prediction tools themselves alter the trajectory of what they predict. It's an old problem in forecasting.
I can't confirm whether AI 2027's specific numbers are right, but I'm confident this format—explicit disagreement plus traceable milestones—will become the standard for AI governance discussions over the next few years.