The previous article established the slope: capability doubling on a measured trend,
forecast medians between 2030 and 2035, error bars stated. This one walks that slope
outward into the world it touches – six fields, each sorted into three honest tiers.
No field gets hype. No field gets dismissal. Each gets a verdict you can check.
// TL;DR – what you'll take away
- Capability moves fastest where feedback is cheap, digital, and verifiable – and slowest where it is physical and expensive. That gradient predicts each field's pace better than any launch announcement.
- Coding leads, hacking inherits coding's curve with opposite incentives, biology pioneered the safeguards, robotics remains the honest wall, politics arrives as atmosphere, and forecasting is quietly joining the fields it measures.
- Every claim is tiered: exists now (checkable today), emerging (in pilots and papers), or science fiction (no demonstrated path yet) – a device borrowed from the AI-2027 scenario's capability radar.
The AI-2027 scenario tracks AI capability across a radar of fields – hacking,
bioweapons, coding, robotics, politics, forecasting – and sorts every claim into
three buckets: currently exists, emerging tech, science fiction.
It is the most honest device in AI communication, because it forces a writer to say
which tier each claim sits in. I am borrowing it for this entire article.
● exists now – checkable today
◐ emerging – pilots, papers, scenarios
✦ science fiction – no demonstrated path
// The organising principle
Capability advances fastest where feedback is cheap, digital, and verifiable –
code compiles or it does not – and slowest where feedback is physical, expensive,
and slow. Every field below sits somewhere on that gradient, and its position
predicts its pace better than any lab announcement.
field 01 · coding
● exists now
Agentic coding on production codebases. Over 80% of Anthropic's own production code is Claude-authored; Stripe reports months of migration work compressed into days; METR measures autonomous task horizons in the double-digit hours. This is operational reality, not demo.
◐ emerging
Multi-day autonomous tasks – beyond the 16-hour ceiling current measurement suites can verify. Coordinated agent fleets (the 5–10 parallel agents from Article 03's startup story). The AI-2027 scenario projects a "fast and cheap superhuman coder" by spring 2027 – a scenario claim, running so far at ~65% of predicted pace.
✦ science fiction
End-to-end product engineering with no human gate anywhere in the loop – and the closed loop itself: AI systems autonomously building their successors. Anthropic's own roadmap treats that third scenario as possible but undemonstrated.
// what to watch · METR time horizons quarterly; the first credibly verified week-long autonomous engineering task
field 02 · hacking
● exists now
Vulnerability triage and code auditing at useful quality, capture-the-flag performance at strong human level, convincing phishing and social-engineering content on demand – and the same models working the blue side: detection engineering, log analysis, patch suggestion.
◐ emerging
Autonomous exploit chains against known vulnerability classes, and AI-versus-AI security operations. The AI-2027 scenario's 2026 frame: agents "only a little worse than the best human hackers, but thousands of copies can be run in parallel, searching for and exploiting weaknesses faster than defenders can respond." A scenario – but one the security industry is already building against.
✦ science fiction
Reliable autonomous discovery and weaponisation of novel zero-days at scale, across hardened targets, faster than any patch cycle, without attribution. Pieces exist in research settings; the assembled capability does not.
// what to watch · the offence/defence balance: whether AI-assisted patching and detection keep pace – Article 14 in this series goes deep here
field 03 · robotics
● exists now
Warehouse and logistics automation at commercial scale. Driverless taxis carrying paying passengers in a handful of cities. Vision-language-action models doing genuinely impressive manipulation – in labs and constrained demos.
◐ emerging
Humanoid pilots in factories and general manipulation in semi-structured environments. Improving sim-to-real transfer, and foundation models giving robots the language understanding they always lacked. Notably: the AI-2027 scenario itself keeps robotics late in its timeline – even the aggressive forecast respects the physical wall.
✦ science fiction
The household robot that handles an unfamiliar home, and the proverbial robot plumber – the long-standing benchmark for why the physical world is hard. Nothing deployed today generalises across messy, unstructured physical environments.
// what to watch · humanoid pilot-to-production conversions; the cost curve of physical interaction data collection
field 04 · politics
● exists now
Synthetic media appearing in real election cycles. AI-generated arguments rated as persuasive as human-written ones in controlled studies. Microtargeting infrastructure, automated content farms, and moderation tooling on the defensive side.
◐ emerging
Personalised persuasion at scale – the AI-2027 evaluations flag persuasion capability explicitly. AI-drafted policy and lobbying material. And the risk Dario Amodei ranks among his five most serious: state-level integration of AI into surveillance and propaganda.
✦ science fiction
Reliable, undetected steering of democratic outcomes – and the tail risk Amodei names directly: AI-enabled permanent authoritarian lock-in. Neither is demonstrated; both are why provenance standards and transparency legislation matter before the capability matures.
// what to watch · deepfake-incident-to-response time in major elections; adoption of content provenance standards
field 05 · biology
● exists now
Expert-level biology knowledge in frontier models, and real research uplift: Fable 5 outperformed dedicated protein language models on gene therapy design tasks. Equally real: the safeguard layer – biology and chemistry classifiers, gated trusted-access programmes for vetted researchers, gene synthesis screening. The defence infrastructure exists because the capability does.
◐ emerging
Meaningful uplift for non-experts – Amodei's stated concern that models are "approaching (or may already have reached)" dangerous end-to-end knowledge. The AI-2027 scenario's mid-2027 evaluation imagines exactly the assessment safety teams run today. This is why model-weight security has become a national-security topic, not a developer detail.
✦ science fiction
Novel pathogen design beyond known templates, and mirror-life organisms – which Amodei places one to a few decades out, while noting AI could compress that. The catastrophic scenarios remain undemonstrated; the policy work to keep them that way is happening now.
// what to watch · whether gated-access and classifier regimes hold as open-weight models advance – the field's entire safety model depends on it
field 06 · forecasting
● exists now
AI matching average human forecasters on some geopolitical question sets, producing calibrated probability estimates, and serving as forecasting infrastructure: base-rate retrieval, scenario enumeration, evidence aggregation across more sources than any human team reads.
◐ emerging
Hybrid human-AI forecasting that beats both alone. AI entries climbing forecasting leaderboards against ranked human forecasters – the AI Futures survey infrastructure from Article 04 is exactly where this shows up first. In the AI-2027 scenario, agents do strategic planning internally; in reality, forecasting teams already draft with models.
✦ science fiction
Superhuman strategic foresight – an AI that reliably outpredicts the best human teams on long-horizon outcomes. If that ever exists, every other field on this page changes simultaneously, which is precisely why the claim deserves its tier.
// what to watch · AI versus ranked humans on public forecasting leaderboards; whether the next AI-2027 self-grading cycle uses AI on both sides
honest uncertainty
// What I am genuinely unsure about
- Whether the gradient holds under a data breakthrough. A cheap way to collect physical-interaction data at scale would re-order robotics overnight – gradients predict pace, not surprises.
- How much weight to give scenario claims. I have labelled every AI-2027 projection as scenario, but the honest truth from Article 04 cuts both ways: reality ran slower than their script – and still ran.
- My own vantage point. I see coding clearest because I live there; my tiers for biology and politics lean harder on Amodei, the AI-2027 team, and the safety literature than on first-hand judgment.
// I believe this
The question "what can AI do?" has no honest single answer – it has six different
answers moving at six different speeds, set by how cheaply each field can tell
the machine it was wrong. Sort every claim you hear into exists, emerging, or
science fiction, and most of the noise resolves on its own.
This completes The Futures arc: Article 04 gave you the slope and the error bars,
this article gave you the map of where it lands. From here the series returns to
the engine room – the nine SDLC phases where AI adoption actually breaks (Article 06),
and the eight forces that decide whether your team compounds or fragments (Articles
07–10). The fields above will keep moving. The practices below are how you stay
standing while they do.
// field references last reviewed · 2026