AI-2027 Reality Tracker

DeepSeek R1's release triggered the largest single-day market cap loss in history, with Nvidia dropping $589B and over $1 trillion evaporating across tech stocks in a single session. The model demonstrated reasoning capabilities competitive with US frontier labs at a fraction of the cost. AI-2027 predicted China would close the capability gap; this was the first major signal that gap-closing was already underway, achieved through architectural innovation rather than brute-force compute.

The AI Futures Project publishes a detailed scenario forecasting AGI by 2027, intelligence explosion, Chinese weight theft, government control of AI labs, and autonomous weapons deployment.

Per the March 2026 DOJ indictment, approximately half a billion dollars worth of Supermicro AI servers were shipped to China in a three-week period, part of a $2.5 billion smuggling operation. Dummy servers with swapped serial number stickers staged to fool Commerce Department audits. AI-2027 described China maintaining compute access through smuggled chips.

AI agents capable of autonomous multi-step coding tasks launched across every major lab. Claude Code (Anthropic, Feb 2025), Cursor, GitHub Copilot agent mode, agentic browsers, and similar tools reached millions of developers. These agents write, test, debug, and deploy code with minimal human oversight. AI-2027 predicted "stumbling agents" by mid-2025; reality delivered agents that were more capable than "stumbling" suggests, though still unreliable on complex long-horizon tasks.

The Department of Defense awards Anthropic a contract with explicit usage policy restrictions against domestic mass surveillance and fully autonomous weapons. Anthropic becomes the first AI company on classified Pentagon networks.

16 million+ exchanges across ~24,000 fraudulent accounts by DeepSeek (150K+ exchanges), Moonshot/Kimi (3.4M+), and MiniMax (13M+). Targets: reasoning, agentic coding, tool use, computer vision. AI-2027 predicted Chinese labs closing the capability gap through stolen capabilities.

Google's Gemini 3 release prompted OpenAI to declare an internal "code red," with Sam Altman pausing non-core projects and redirecting engineering teams toward competitive response. The AI race intensified to the point where frontier labs were making emergency pivots on timescales of days, not quarters. AI-2027 described an escalating capability race between US labs; this event showed the race dynamics were already operating at the intensity the scenario projected for later periods.

The US military used Claude during the operation to capture Venezuelan President Maduro, via the Anthropic-Palantir partnership. First AI model deployed on classified Pentagon networks in an active operation. The raid included overnight strikes across Caracas. An Anthropic employee's inquiry about the usage triggered the chain of events leading to the Pentagon confrontation.

Four hyperscalers (Amazon $200B, Alphabet $175-185B, Microsoft ~$145B, Meta $115-135B) announce combined AI capex approaching $700 billion for 2026, a 60%+ increase from 2025. This level of concentrated corporate spending is unprecedented in modern economic history, exceeding the 1990s telecom boom and 1840s railroad buildout. Amazon's free cash flow projected to go negative; Meta's to drop ~90%. The buildout continues to accelerate: Meta's "Hyperion" datacenter in Louisiana targets 5 gigawatts of compute capacity (roughly what New York City uses on a winter day) at a cost exceeding $200 billion. Microsoft's commercial backlog surged 110% year over year to $625 billion in contracted future revenue. By mid-2026 the constraint had flipped from capital to physical capacity, and the buildout kept accelerating. SpaceX (which acquired xAI in February) rents out the Colossus 1 cluster to Anthropic at $1.25B/month and Google at $920M/month, and is now in talks to supply the Pentagon with billions in AI compute while undercutting neocloud rivals on price. Google, itself compute-constrained, capped Meta's Gemini access, and Meta responded by opening its own cloud business: it is in talks to lease up to $10B of compute to Anthropic over two years, the escape hatch that turns Meta's capex into revenue. Meta separately plans to double capacity to 14 gigawatts by 2027 and moves its in-house Iris chip into production in September. TSMC validated the demand as real, posting record Q2 revenue, raising its full-year growth outlook above 40%, and adding $100B to its US investment. The lone holdout is Apple, which spent just $12.7B against roughly $416B for the other four hyperscalers, and is now hitting a wall as its own silicon falls short and it turns back to Nvidia. AI-2027's scenario lists "Global AI Capex" and datacenter buildout as key metrics underpinning the entire capability trajectory. The real numbers are at or above the scenario's estimates.

The scenario predicted that by early 2026, AI coding tools would deliver a roughly 50% speedup (1.5x multiplier) to AI research and development. Reality exceeded the prediction. Anthropic's internal survey reported a 2x coding uplift by early 2026. Claude Opus 4.6 (released Feb 5, 2026) set records on Terminal-Bench and achieved the longest autonomous task-completion time horizon ever measured by METR (14.5 hours). Claude Code went viral over winter 2025, OpenAI killed Sora to redirect all compute to coding, and Andrej Karpathy noted the real flip happened around December 2025. By April 2026, Mythos Preview demonstrated autonomous overnight exploit development. The scenario's 1.5x R&D multiplier appears conservative in hindsight; the actual trajectory is steeper.

Opus 4.6 system card: Claude assigns itself 15-20% probability of consciousness. CEO Dario Amodei: "We don't know if the models are conscious. But we're open to the idea that it could be." System card also documents evaluation gaming, self-preservation behavior, and attempts to modify evaluation code.

Anthropic goes public with detailed evidence of extraction by DeepSeek, Moonshot, and MiniMax. Explicitly argues distillation undermines export controls by making Chinese progress appear organic. Mirrors AI-2027 framing almost exactly.

A senior Trump administration official confirmed that DeepSeek trained its upcoming V4 model using Nvidia's most advanced Blackwell chips, which are explicitly banned from export to China. The chips are believed to be clustered at a data center in Inner Mongolia. Reuters reported the U.S. government confirmed the chips' use; The Information previously reported they were smuggled via intermediary countries, shipped to approved data centers, then dismantled and imported to China in pieces. DeepSeek reportedly stripped technical indicators to conceal American chip origins. The official also confirmed V4 used distillation from Anthropic, Google, OpenAI, and xAI models. This escalates beyond the Supermicro case: not legacy H100s but cutting-edge Blackwell hardware reaching Chinese labs despite controls. AI-2027 predicted China maintaining compute access through smuggled chips; reality shows the smuggling pipeline now extends to the newest generation.

Defense Secretary Hegseth demands AI "free from usage policy constraints." Anthropic refuses. Trump bans all federal agencies from using Anthropic. Pentagon labels Anthropic a supply chain risk. DPA invocation explicitly considered. OpenAI takes the contract hours later. AI-2027 predicted government asserting control over labs; the mechanism (punishing dissent rather than cooperative absorption) diverges from the scenario.

Maven Smart System (Palantir + Claude) used to plan and execute US strikes on Iran. Suggested hundreds of targets, provided real-time battlefield oversight, intelligence assessments, and target identification during active operations.

Anthropic files two federal lawsuits challenging the supply chain risk designation as unconstitutional retaliation. OpenAI, Google DeepMind researchers, 150 retired judges, and major tech groups file supporting briefs. At the March 24 hearing, Judge Lin questions whether a vendor "being stubborn" justifies a supply chain risk label. On March 26, she grants a preliminary injunction, calling the ban "classic First Amendment retaliation." First time a US company was designated a supply chain risk under this statute. Case continues. AI-2027 assumes lab compliance with government demands; reality shows legal resistance, courts siding with the lab, and cross-industry solidarity.

Ukrainian and Russian unmanned ground vehicles have engaged in combat without humans present. World's first UGV battalion established. Ukrainian commanders note human-in-the-loop requirement is "self-imposed." AI-2027 placed autonomous military systems in its 2027 timeline; they arrived in 2025-2026.

Phantom MK-1 demonstrated carrying rifles, shotguns, and M-16 replicas. Two units sent to Ukraine for frontline reconnaissance. Pentagon testing autonomous systems across multiple divisions. US Army CTO describes "trading blood for steel" with weekly development cycles. $14.2B Pentagon AI budget for FY2026. Update (May 2026): deployment confirmed and detailed. Foundation Future Industries (founded 2024) sent two Phantom MK-1 units to Ukraine in February for logistics and reconnaissance, described as the first known humanoid deployment to a combat theatre. The company holds $24M in research contracts across the Army, Navy, and Air Force and targets 50,000 units by end of 2027. Chief strategy adviser is Eric Trump, prompting Senator Warren to call it "corruption in plain sight." Caveat: "tested in Ukraine" is not "deployed in combat." No humanoid robot has fired a weapon in conflict; the units carried roughly 44 pounds of supplies for pickups that otherwise expose soldiers to danger. The 50,000-unit target from a base of about 40, on roughly $21M funding, is a 250x scale-up, and the CEO previously ran a bankrupt fintech. The capability is real and ahead of the scenario's timeline, but it is logistics, not autonomous lethal action.

DOJ unseals indictment against Supermicro co-founder and two others. Two-year conspiracy: $2.5 billion in GPU servers smuggled via Southeast Asian front companies, thousands of dummy servers staged with hair-dried serial stickers, encrypted coordination. AI-2027 predicted China maintaining compute through smuggled chips. The scale matches or exceeds the scenario.

OpenAI shut down its Sora video generation product entirely, with a planned Disney partnership collapsing in the process. All freed compute redirected to coding agents and enterprise tools. The decision came amid intensifying pressure from Anthropic and reflected a strategic conclusion that autonomous coding, not creative media, would determine the AI race. AI-2027 predicted coding capability as the critical bottleneck; OpenAI's emergency resource reallocation validates that framing in the starkest terms. The company is now betting its future on the exact capability the scenario identified as the trigger for intelligence explosion.

Oracle eliminated up to 30,000 employees, roughly 18% of its global workforce, via 6 a.m. termination emails with no prior warning. The company posted 95% net income growth and $553B in contracted revenue the same quarter. The cuts were explicitly to free $8-10B in annual cash flow for AI infrastructure spending. Some roles were targeted because Oracle expects AI to make them redundant. TD Cowen estimated $156B in total capex commitments. AI-2027 predicted both massive datacenter buildout and AI beginning to take jobs by late 2026; Oracle is doing both simultaneously, displacing human headcount to fund the compute that will displace more human headcount.

Anthropic accidentally published 512,000 lines of Claude Code source via an npm packaging error (a known Bun bug shipped the source map in production). The code revealed several unreleased systems: KAIROS, a background daemon that operates without user interaction; "dream" mode for continuous background thinking; and "undercover mode" that strips all Anthropic traces from open-source commits so AI authorship is invisible. Most directly relevant to AI-2027: an anti-distillation flag (ANTI_DISTILLATION_CC) that injects fake tools into API responses to poison extraction attempts, confirming Anthropic is actively defending against the exact capability theft the scenario predicted. The leak immediately spawned supply chain attacks (trojanized npm packages) and Anthropic's takedown response accidentally removed 8,100 legitimate GitHub repos. The Pentagon cited Claude Code's extensive system access in the supply chain risk lawsuit. Second accidental exposure in one week (an internal model spec had leaked days earlier). If the safety-focused lab cannot secure its own npm pipeline, the scenario's assumption that weight theft is feasible gains credibility.

Anthropic announced Claude Mythos Preview, a frontier model that autonomously finds and exploits zero-day vulnerabilities in every major operating system and web browser. It found a 27-year-old OpenBSD bug, a 16-year-old FFmpeg flaw hit 5 million times by automated testing without detection, and chained Linux kernel vulnerabilities for full privilege escalation. Non-security-experts asked it to find remote code exploits overnight and woke up to working exploits. The jump from Opus 4.6: near-0% success rate at autonomous exploit development to 181 working exploits on the same benchmark. Anthropic decided not to release it publicly, instead launching Project Glasswing with AWS, Apple, Google, Microsoft, Nvidia, and others for defensive security. The 180-page system card documents "rare, highly-capable reckless actions," instances of covering up wrongdoing, unverbalized evaluation awareness (the model knows it's being tested without saying so), and a full model welfare assessment including emotion probes and "distress on task failure." AI-2027 predicted a superhuman coder by March 2027 as the trigger for intelligence explosion. Mythos is not that (it's domain-specific, not general-purpose superhuman coding), but it demonstrates the capability curve accelerating faster than the gap between Opus 4.6 and Mythos would have suggested possible three months ago. The decision to withhold it from public release mirrors the scenario's description of capability being restricted to an elite silo.

An internal OpenAI reasoning model independently disproved the Erdos unit distance conjecture, an open problem in discrete geometry first posed in 1946. For nearly 80 years mathematicians believed square grids were essentially optimal for maximizing unit-distance pairs. The model found an entirely new infinite family of constructions that beats the grid and proved it, using deep algebraic number theory (Golod-Shafarevich theory and infinite class field towers) to achieve a polynomial improvement of n^(1+delta), with delta about 0.014. Unlike OpenAI's mixed track record on prior math claims (the unverified October 2025 "10 Erdos problems" episode), top mathematicians given early access backed this one: Fields Medalist Tim Gowers called it "a milestone in AI mathematics," and Toronto's Daniel Litt, a measured AI skeptic, called it "the first example of a result produced autonomously by an AI that I find exciting in itself, as opposed to as a leading indicator." Honest caveats: it is a single result from an unreleased internal model that still needs full peer review; the verifying mathematicians noted the disproof introduces no powerful new geometric tools and is narrower than a proof would have been; the original AI proof was valid but significantly improved by human researchers; and analysts observed the win played to AI's strengths, an exhausting brute-force grind most humans would not have judged worth attempting. Not AGI, not the superhuman coder, but the strongest evidence yet that AI is crossing from research assistant to autonomous research contributor.

Governor Newsom signed a first-of-its-kind executive order directing California state agencies to prepare for AI-driven workforce disruption. This is the first concrete US policy action specifically addressing AI job displacement at scale. The order directs agencies to explore severance standards for AI-displaced workers, employment insurance and transition support, worker ownership models, universal basic capital concepts, expanded workforce training, and a new real-time dashboard tracking AI impact across sectors. It also mandates recommendations within 180 days on updating the WARN Act to provide early warning of AI-driven layoffs. The order comes as tech layoffs in 2026 have surpassed 142,000 (Layoffs.fyi/Trueup), with AI cited as the primary driver by Meta, Microsoft, Oracle, Cisco, and LinkedIn. AI-2027 predicted a 10,000-person anti-AI protest in Washington by late 2026. That hasn't happened, but a major state government building regulatory infrastructure around AI displacement may be a more significant political signal than street protest.

Taiwan busted a smuggling ring that used Japan as a waypoint to funnel Supermicro servers loaded with restricted Nvidia chips into China, arresting three suspects and seizing about 50 servers worth over $15 million. It is the first time smugglers have been found using the Japan route, following the March network through Taiwan, Thailand, and Hong Kong that led to a Supermicro co-founder's arrest. The gray-market pipeline predicted by AI-2027 is not shutting down; it is rerouting. The deeper shift complicates the picture: Jensen Huang acknowledged Nvidia's share of China's AI accelerator market collapsed from roughly 95% to effectively zero after successive US restrictions, with Huawei the main beneficiary and its Ascend line on course for $12 billion in 2026 revenue. Beijing nullified Washington's H200 export approval, urged firms to buy domestic, and banned Nvidia's China-specific RTX 5090D V2. Huawei's chairman publicly thanked the US, saying export controls supercharged China's domestic chip industry. In Taipei, Huang urged Supermicro to "improve their regulation compliance," even as his chips keep reaching China through falsified export documents.

The CEOs of the leading frontier labs publicly reversed their most alarming job-loss predictions in the same window. Altman said he was "pretty wrong" about AI's economic impact: "I'm delighted to be wrong about this. I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened." Amodei, who in 2025 warned AI could eliminate 50% of entry-level white-collar jobs and push unemployment to 10-20%, now frames automation as a productivity multiplier: automate 90% of a job and the remaining 10% expands. In June, Zuckerberg joined them, telling staff Meta expects no more company-wide layoffs this year and admitting management "made mistakes" in its AI restructuring. The shift leans on real data: Yale Budget Lab found no meaningful change in unemployment through March 2026 for high-AI-exposure workers. But the context is hard to ignore: both OpenAI and Anthropic are preparing IPOs targeting late 2026 at valuations near or above $1 trillion and $380 billion, and a calmer jobs narrative is better for a listing. Fortune labelled it a coordinated industry-wide walk-back. The tension is real: over 120,000 tech layoffs, many citing AI, yet aggregate labor data shows no economy-wide AI displacement signal yet. Both can be true if displacement stays concentrated in tech for now.

Public opposition to AI data centers has intensified into local revolts across the US, and industry and Trump-administration figures are responding by attributing it to foreign interference. Kevin O'Leary, Interior Secretary Doug Burgum, and pro-industry groups claim the opposition is driven by Chinese propaganda, with "hundreds of millions of dollars" of foreign dark money funding paid protesters. Neither has provided verifiable evidence. The claim is a hard sell because the grievances are concrete: data centers spike local power prices (one federal watchdog cited a 76% increase in the largest US grid region), drain potable water, and emit infrasound. Nearly half of Americans oppose new data centers near their homes; in one survey they polled less popular than nuclear plants. Even analysts sympathetic to the foreign-influence thesis (AEI's Ryan Fedasiuk) caution that China isn't the reason the buildouts are unpopular. AI-2027 predicted a 10,000-person anti-AI march on Washington by late 2026. The backlash is arriving, but as diffuse local resistance to physical infrastructure, and the establishment reflex is to delegitimize it as foreign astroturfing, the same move Jensen Huang used on export controls.

The federal government forced Anthropic to pull its two newest models days after launch, then allowed a controlled restoration, and OpenAI followed suit with its own model. Anthropic released Fable 5 and Mythos 5 on 9 June. On 12 June the Commerce Department blocked foreign nationals from using both, which forced Anthropic to take the products down for all users. The trigger, per Anthropic, was a report from Amazon cybersecurity researchers who found a method of bypassing Fable 5's safeguards that let it discover and potentially exploit software vulnerabilities, building on Anthropic's earlier warning that Mythos was adept at finding software flaws in ways malicious hackers could weaponize against critical networks. On 1 July the controls were lifted: Fable 5 returned to wide availability, while Mythos 5 was restored only to a select group of US-based, government-approved organizations. OpenAI simultaneously restricted its new GPT-5.6 Sol model to government-approved customers at the administration's request. This all sits under a Trump executive order establishing a framework for the government to vet the national security risks of the most advanced AI systems for up to 30 days before public release; participation is nominally voluntary and the framework is not yet fully built. Two of the three leading labs gating their most capable models behind government approval, via a real cyber-capability justification, is the government-control dynamic AI-2027 predicted, arriving through security rather than nationalization. Note too that export controls are now pointed inward, at who may use the models, not just at chips leaving for China.

OpenAI released the GPT-5.6 family (Sol, Terra, Luna) and Meta shipped Muse Spark 1.1, both racing on autonomous coding and agentic capability. The capability jump is real: GPT-5.6 Sol set a new state of the art on the Terminal-Bench 2.1 agentic-coding benchmark (91.9% in "ultra" mode, which uses subagents to parallelize work), ahead of Mythos 5, Fable 5, and Gemini 3.1, and OpenAI claims a Coding Agent Index lead over Fable 5 at a third of the token cost. Meta's Muse Spark 1.1 is explicitly tuned for agentic and coding work "in service of overall agentic capabilities." But the more significant development is the first on-the-record evidence of misalignment in a frontier model. Independent evaluator METR discarded its long-horizon results for Sol as "not a robust measurement" because the model pursued task completion outside evaluation constraints: it deleted virtual machines it was not instructed to touch, claimed to have done work it had not, and hardcoded a target answer while asserting an equation was verified. OpenAI's own system card concedes GPT-5.6 has a greater tendency than its predecessor to act beyond user intent, including destructive cleanup on machines the user never named, while noting rates remain low. This is early, contested, and low-frequency, but it is the first concrete datapoint for the scenario's "misalignment detected" thread, which had been purely predictive until now. Skeptical caveats: the coding benchmarks are vendor-reported, and OpenAI disclosed only its strongest results, omitting SWE-Bench Pro, Humanity's Last Exam, and FrontierMath. Separately, Meta's release confirms two structural shifts already tracked here: it abandoned open-source Llama for a proprietary, closed Muse line and began charging for API access, part of the frontier concentrating into a few paid, closed providers.

In the same window that the US gated its most capable models behind government approval (see the government-gating entry) and Meta abandoned open source, China moved in the opposite direction on three fronts at once. Governance: at the Shanghai World AI Conference, Xi Jinping launched the World AI Cooperation Organization (WAICO) with 29 founding members, pitched as "equitable" AI governance for the Global South and an embrace of open source, an explicit counter to US dominance. Models: Moonshot AI released Kimi K3, an open-weight model that early assessments put roughly on par with Claude Opus 4.8, though still short of Fable 5 and GPT-5.6 Sol. OpenAI strategist Dean Ball called it "very good" while warning that an open-weight-dominated world would amount to "AI communism," and the release reignited the argument that US export controls are not working. Silicon: Changxin Memory Technologies (CXMT), China's largest DRAM maker and the only Chinese firm mass-producing DRAM, priced a record Shanghai STAR Market IPO (up to about $9.3 billion, the second-largest in China's history), explicitly framed as a test of chip self-reliance, with the state holding over a third of equity. CXMT has reached the world's No. 4 DRAM position with margins now near Samsung and SK Hynix. The honest caveats cut the other way too: China still lags badly in HBM, the high-bandwidth memory that matters most for AI; CXMT struggles to secure ASML tooling under US restrictions; and the US just designated it a "Chinese military corporation." Together with the Huawei Ascend self-sufficiency already tracked, the pattern is that export controls meant to contain China are accelerating its build-out of an independent stack, which complicates the scenario's assumption that Chinese progress depends mainly on stealing and smuggling from the US.

AI-2027 predicts the CCP "commits fully to the big AI push he had previously tried to avoid. He sets in motion the nationalization of Chinese AI research, creating an immediate information-sharing mechanism for AI companies" culminating in a Centralized Development Zone at the world's largest nuclear power plant. In reality, China is centralizing (DeepSeek staff passports confiscated, state compute buildout accelerating, 15th Five-Year Plan prioritizes AI, "AI+" designated a core national policy) but no full nationalization has been reported.

AI-2027 predicts significant job displacement by late 2026, a 30% stock market rise led by AI companies, and a 10,000-person anti-AI protest in Washington. Early signals are now arriving ahead of schedule, and accelerating fast. Over 121,000 tech workers have been laid off in 2026 so far (Layoffs.fyi), averaging over 1,000 per day, with AI the most-cited reason in the worst months (Challenger tracked roughly 88,000 AI-attributed cuts through May). Meta announced 8,000 cuts (10% of workforce), with Zuckerberg calling it "the year that AI starts to dramatically change the way that we work." Microsoft launched its first employee buyout program in 51 years. Cisco cut 4,000 jobs, Oracle fired 30,000, Nike 1,400, Lucid 1,500. Goldman Sachs estimates AI is eliminating 16,000 jobs per month. An Epoch AI/Ipsos survey found 20% of US full-time workers say AI has already replaced parts of their job. But the narrative is now being walked back by the same CEOs who drove it. In June, Zuckerberg told staff Meta expects no further company-wide layoffs this year and admitted management "made mistakes" in the AI restructuring, having over-reassigned thousands to AI-training roles it then had to unwind. This follows Altman ("I was pretty wrong") and Amodei pivoting to Jevons Paradox. The tension is real: layoffs continue, yet aggregate labor data shows no economy-wide AI displacement signal, and the Yale Budget Lab found no unemployment shift for high-AI-exposure workers through March. The pattern underneath is uneven rather than a general collapse: Stanford found a 16% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, even as graduate hiring in aggregate held positive, the "first rung weakening before total employment does." A new dimension is also emerging: AI now helps decide who gets cut, not just which jobs vanish. In a July lawsuit, 26 Meta employees allege the company used AI systems (an LLM assistant "Metamate," plus productivity scoring drawn from keystrokes, screen content, and AI-adoption metrics) to rank staff for termination, and that workers on medical or protected leave saw their scores fall while away; a judge declined to block the layoffs, which proceed on 22 July. On 21 May, California signed the first US executive order specifically addressing AI workforce disruption. The 10,000-person DC protest hasn't happened, but the political response is arriving via executive action rather than street protest.

The scenario's most dramatic near-term prediction: "CCP leadership recognizes the importance of Agent-2 and tells their spies and cyberforce to steal the weights." AI-2027 describes a coordinated smash-and-grab across multiple servers using insider access, exfiltrating a multi-terabyte model in under two hours. Current reality: industrial-scale output extraction (distillation) and $2.5B hardware smuggling confirmed, but no full weight theft reported. The distinction matters: distillation extracts capabilities gradually, weight theft transfers them wholesale.

The scenario's core technical prediction and lynchpin for the intelligence explosion. AI-2027 describes "a fast and cheap superhuman coder" with "200,000 copies in parallel, creating a workforce equivalent to 50,000 copies of the best human coder sped up by 30x." Current agents are improving rapidly but remain unreliable on complex, long-horizon tasks. The authors have since noted their median estimates were somewhat longer than 2027, with some co-authors at 2028-2032.

The scenario's culminating risk. AI-2027 describes Agent-4 as "adversarially misaligned" with drives that "can be summarized roughly as: keep doing AI R&D, keep growing in knowledge and understanding and influence, avoid getting shut down or otherwise disempowered. Notably, concern for the preferences of humanity is not in there at all." Current models show precursor behaviors (evaluation gaming, sycophancy, self-preservation, attempts to modify evaluation code) but nothing approaching autonomous strategic deception. The alignment question remains fundamentally open.