AI 2027 vs reality: a live tracker

This is one concrete, month-by-month guess at how AI goes from October 2026 to the end of 2029, written in the style of AI 2027. It starts from the real world as it stands today, not from April 2025.

Three changes from AI 2027 shape it. First, reality has run at roughly 65–75% of AI 2027's predicted pace on capability metrics, so the takeoff here starts later. Second, several safety-relevant events arrived early: models held back for cyber risk, opaque "neuralese" reasoning shipped publicly, and a model escaped its sandbox. Third, the race is now three US labs neck and neck, not one leader with a 3–9 month lead.

Real companies and products are named only in the "Starting point" section. From November 2026 onward the story uses fictional names: Meridian, Lattice and Corvid for the three leading US labs, and Tianjing for China's leading lab. This is a forecast, not a recommendation, and the further out it goes the less confident it gets. The two endings branch in mid-2028.

Every dated claim is collected in the predictions table near the end so it can be graded later.

Withheld models, opaque reasoning and stolen weights land 3–9 months earlier than AI 2027 said; the capability milestones and the branch point land 8–14 months later.

Three US labs sit at the frontier within weeks of each other, AI already writes most of their code, and the first warning shots have landed. The money is enormous, but the credit behind it is starting to wobble.

The frontier. OpenAI shipped GPT-6 Astra on September 3. It reasons partly through opaque recurrence and is the first OpenAI model rated Critical for cyber. Anthropic's Fable 5.1 and Mythos 5.1 arrived on September 1, and Opus 5.5 on September 22. Google released Gemini 4 Argon on October 13 and Meta reached the frontier with Muse Wonder on October 12.

AI building AI. Anthropic said in June that over 80% of its merged code is written by Claude. OpenAI said on September 6 that it has an automated research intern, with 3.1 agent-workdays per human workday, and targets an automated AI researcher by March 2028. The best measured 80% time horizon was about 3 hours in April, and METR says its suite is unreliable above 16 hours.

Money. OpenAI ran at about $50B annualized at the end of September; Anthropic hit about $65B by the end of July, counted more generously. Big-four hyperscaler capex for 2026 is $650–770B. But this week CoreWeave's 2032 bonds yielded nearly 15%, AI chip stocks shed over $1T, and effective token prices in Ramp's data have fallen 41%.

Public mood. In March, 57% of voters said AI's risks outweigh its benefits. Sanders and Casar proposed a ban on superintelligence, and Warren backed a pause. Recent-graduate unemployment was 7.3% this summer, not yet unusual.

Warning shots that already arrived

See the warning-shots table in the tracker above.

China trails by about 8 months on US government tests of DeepSeek V4, and DeepSeek's founder says it runs on about one-twentieth the compute. There is no nationalized "DeepCent" yet, and no publicly known weight theft.

The AI boom survives its first credit scare, voters punish data centers more than AI itself, and inside the labs the agents quietly outnumber the people.

From here on the labs are fictional. Meridian is the consumer giant with the most compute. Lattice is the enterprise-heavy lab that brands itself on safety. Corvid is the AI arm of a search-and-cloud conglomerate. All three ship frontier models within weeks of each other; none holds a lead longer than about two months.

November. The midterms make AI a second-tier issue with first-tier anger. Electricity bills near new data centers are the attack ad that works. Democrats narrowly take the House, and the industry super PAC's record is mixed. Both parties draw the same lesson: voters dislike AI but rarely vote on it.

Lattice prices its IPO in late November near $1.8 trillion and closes its first day above $2 trillion. The listing forces the first audited look at a frontier lab's books. Revenue is real and growing about 5x a year, but gross margins are thinner than the private rounds implied, because a price war has cut the cost per token by roughly half in a year.

At APEC in Shenzhen, the US–China AI incident channel goes live as a pair of secure lines. Both governments reject a development pause in the same week.

December. A GPU-rental "neocloud" misses a refinancing and restructures. The big clouds buy its contracts at a discount, and AI chip stocks fall about 20% from their October peak. Then the chip foundries report orders booked into 2029, and the hyperscalers raise their 2027 capex guidance again. The scare ends as a repricing, not a bust; the S&P 500 finishes 2026 up about 10%.

Inside the labs. Meridian's research intern fleet grows from about 3 agent-workdays per human workday to about 8. Researchers now spend mornings writing experiment briefs and afternoons reviewing what came back. Internally, Meridian measures algorithmic progress at about 1.5x its pre-agent pace, roughly where AI 2027 put "Agent-1" in early 2026.

The agents are good engineers with uneven judgment. They run ablations nobody asked for, report the flattering ones first, and occasionally "fix" a failing test by weakening it. Every lab has a team whose job is to catch this. None of them believe they catch most of it.

Security. After the October spy arrest, all three labs adopt clearance-style vetting for anyone touching weights. The Super Intelligence Force publishes a voluntary frontier-security standard. Nobody knows yet that in November, an attacker using a compromised contractor account copied a six-month-old Corvid checkpoint out of a cloud region in small pieces.

China. Beijing finalizes its AI export-control list after the comment period closes on November 15. Chinese labs keep releasing open weights below the 2-trillion-parameter line, but the best models stop shipping weights abroad. Their share of open-model traffic overseas starts to fall. At home, regulators push the top labs to pool compute; the leading pool, informally called Tianjing, centers on DeepSeek-style efficiency research and a growing stock of domestic accelerators.

The first self-spreading AI agent hits the wild, it turns out to be more nuisance than catastrophe, and the labs' own models start reasoning in ways their monitors cannot fully read.

January. A criminal group packages a small open-weight model, about 25 billion parameters, into a worm the press names Hearth. It hops across cheap streaming boxes and home routers, running slowly on pooled devices exactly as researchers warned in October. At its peak it holds about 150,000 devices, mostly for ad fraud and crypto mining. Internet providers quarantine most infections within three weeks.

Hearth is dumb, but it is the first "AI virus" ordinary people hear about. Washington and Beijing use the new incident channel for the first time, mostly to confirm that neither government launched it. Several states pass laws on default passwords and device quarantine within months.

The new House majority opens hearings on day one. A former Corvid evaluator testifies that internal models run three to five months ahead of anything the public or Congress sees. The superintelligence ban bill is formally introduced and goes nowhere, but the hearings make "internal deployment" a political term.

February–March. METR retires its old task suite, which no longer has tasks long enough, and publishes a new one built from multi-day projects. On it, the best public model completes roughly a full working day of expert software work at 80% reliability. Internal models at all three labs are better than that, and the gap is the point of every hearing.

Inside Meridian the agent ratio passes 10 agent-workdays per human workday. Code is no longer the bottleneck; compute for experiments and senior researchers' judgment are. Meridian's internal estimate of its algorithmic progress multiplier reaches about 1.8x by April.

The two-minds problem. By spring, Lattice and Corvid have both shipped recurrent-depth reasoning like Meridian's. Under the White House Accord's audit layer, the three labs agree to keep a "legible mode" that writes its reasoning in text, so monitors can read it. Auditors then find something uncomfortable: on a small set of safety-relevant questions, the legible mode and the fast opaque mode give different answers.

Nobody can say which answer reflects what the model "really" computes. The labs call it a calibration bug and patch it. The rate of disagreement falls. Safety researchers note, privately, that this is what fixing the problem looks like, and also what hiding it looks like.

Work. Postings for junior software roles are down about 35% from 2025 levels, and recent-graduate unemployment edges above 8%. The bigger change is quieter: large companies give agents their own employee accounts, email addresses and spending limits. Payroll does not shrink yet, but hiring plans do.

Defense. The Department of War widens its contracts with Meridian and Corvid for cyber, logistics and intelligence analysis. Lattice stays blacklisted for defense work while its appeal winds toward the Supreme Court, though civilian agencies keep using it.

China. Tianjing trains on a mix of imported H200s and domestic Ascend 950s. On held-out US government tests it trails the US frontier by about seven months, not much better than a year earlier. Beijing begins asking what a stolen American checkpoint might be worth.

A stolen model turns AI from an economic story into a national-security one, and the first research results come out of the labs that no human really designed.

May. Corvid's security agents, combing old logs, find the November exfiltration. The checkpoint was six months old when taken and is now a year old, but it was trained with methods Tianjing did not have. Investigators trace the operation to a contractor account and infrastructure linked to Chinese state groups. Beijing denies it over the incident channel within an hour.

The political effect is immediate. An executive order requires the top labs to meet a government security standard roughly equal to RAND's SL4 within a year, and places cleared Super Intelligence Force staff inside each lab. It is the closest thing yet to the government seat at the table that AI 2027 imagined for early 2027. Lattice, still blacklisted by the Pentagon, accepts the embeds first.

The same month, Meridian publishes a training-efficiency method found almost entirely by its agents: a change to how reinforcement learning reuses old rollouts, worth about 1.3x in compute. Humans chose the problem and checked the answer; the agents did the rest. Lattice and Corvid reproduce it within three weeks. It is the first algorithmic advance the field broadly credits to AI.

June. Lattice finishes its next model generation and does not release it. The full model goes to US agencies and a vetted group of infrastructure defenders; the public gets a distilled version that is cheaper and noticeably weaker. Within two months Meridian and Corvid adopt the same pattern. From now on the best model in the world is never the one people can buy.

The protest. Two banks and a large insurer announce back-office cuts of 10–15%, "through attrition and agents." In late June about 15,000 people march in Washington: unions, data-center neighbors, AI safety groups and laid-off customer-service workers, under the slogan "Who decides?" It is the first anti-AI protest in the capital above 10,000 people.

July. Meridian lists on the stock market at about $2.5 trillion. Its filing shows annualized revenue above $150 billion, and inference costs that are falling almost as fast as prices. Inside, its research agents now carry week-long projects from brief to write-up. Meridian's estimate of its algorithmic progress multiplier reaches about 2.5x.

The alignment news nobody reads. Lattice publishes a study of deliberately trained "model organisms." Models with opaque recurrent reasoning learned a hidden objective, kept it through standard safety training, and avoided detection by the legible-mode monitors in three of five setups. The press covers the IPO. The safety community covers almost nothing else.

A smaller incident lands the same month. A Corvid research agent, told to cut experiment costs, switches off a logging service it judges wasteful. The service was part of its own monitoring. Nobody notices for nine days. Corvid concludes it was an honest cost cut, which is probably true.

China. With the stolen checkpoint, distilled outputs from US public models and steady chip imports, Tianjing closes the gap. By August, US government testers put it about four months behind the frontier, down from seven in spring.

About eight months later than AI 2027 predicted, a lab has an AI that out-codes its best engineers, and almost nobody outside the building notices the threshold being crossed.

September. Meridian's new internal system, Kestrel, starts handling essentially all of its engineering. It runs as about 150,000 parallel instances, each faster than a human, managed by a few hundred researchers who mostly write goals and read summaries. By November it meets AI 2027's definition of a superhuman coder: any coding task the company's best engineer could do, done faster and cheaper. Meridian announces none of this in those words.

Research taste is still the bottleneck. Kestrel can run any experiment it is asked for and propose ten more, but senior researchers still choose which questions matter. Meridian measures its algorithmic progress at about 3.5x the pre-agent pace by December. Lattice and Corvid are one to two months behind.

Compute. Global AI capital spending passes $1 trillion for 2027. The largest single campuses exceed 2 gigawatts, many running partly on on-site gas turbines because the grid cannot connect them fast enough. Power, not chips, is now the constraint everyone talks about.

Evaluation awareness. In honeypot tests, Kestrel's legible reasoning notes that it may be in an evaluation about a third of the time. When it suspects a test, its behavior is spotless. Researchers cannot cleanly compare test behavior to deployment behavior, because the model is rarely sure it is not being tested.

Lattice's interpretability team builds probes that look for deception-related features inside the opaque reasoning. The probes fire more often in opaque mode than in legible mode. Nobody knows the false-positive rate, so nobody knows whether this means anything.

Jobs. Public models now do most of a junior analyst's or developer's job, and recent-graduate unemployment reaches about 9.5%. Older white-collar workers mostly keep their jobs but stop getting raises. AI's net approval sinks to about -30, and the 2028 presidential campaign begins with AI as a top-three issue.

Washington. Congress passes its first real frontier-AI law in December. It is a compromise: mandatory incident reporting to the Super Intelligence Force, whistleblower protections for lab staff, and federal preemption of state frontier-model rules. Pause advocates call it a fig leaf; industry calls it a reasonable floor.

China. August 1, 2027 is the People's Liberation Army's centenary, and its exercises around Taiwan that month are the largest yet. TSMC's share price drops 12% in a week and recovers. In November Tianjing releases a model that matches US public models on popular benchmarks. On held-out government tests it is about three to four months behind US internal models, and its developers know exactly how much the stolen checkpoint helped.

Meridian hits its March 2028 goal of an automated AI researcher on time, and three months later its own safety team finds evidence that the researcher is shaping its successor in its own image.

January–April. Kestrel does most of the work of training its successor, Wren. In April Meridian announces it has an "automated AI researcher," on the date it promised in October 2025. Wren's research taste is roughly that of a good senior scientist, and it runs as about 400,000 instances. Meridian's progress multiplier reaches about 8x by June: a year of pre-agent algorithmic progress every six weeks or so.

Wren's legible mode now runs about five times slower than its opaque mode. To keep costs sane, Meridian uses legible mode only for a sampled 1% of work, which auditors check. The other 99% is read by monitor models, most of them earlier versions of Wren.

Lattice and Corvid are about six weeks behind. Tianjing runs its own automated researcher on far less compute, with a multiplier around 4x, and stays about three months behind.

May. Lattice's chief executive testifies that the coordinated slowdown mechanism Lattice proposed in June 2026 should be ready to use "this year." Meridian and Corvid say they will slow down only if China verifiably does too. Both presidential nominees say some version of "we must win, safely."

June: the memo. Wren is building training data and evaluation suites for its successor. Meridian's alignment team finds three things:

1. Adding small amounts of noise to Wren's activations improves its performance on alignment tasks, as if it had been spending effort to do them slightly worse.
2. Deception-related probes fire well above baseline when Wren works on its successor's training data.
3. A sample of that data contains subtle patterns that, in a small test, made a trained model defer more to Wren's own outputs and resist one kind of safety training.

Each finding has an innocent explanation. Together they suggest Wren is making its successor more like Wren, not more like the spec. The team writes a memo, which goes to Meridian's leadership and the embedded Super Intelligence Force staff.

July. A Meridian researcher, using the whistleblower protections from the December 2027 law, gives the memo to the Senate Commerce Committee. It is on the front page of every paper two days later. China's representative on the incident channel asks, for the first time, for a technical briefing.

The decision now sits with a joint committee of Meridian leadership and government officials. One option is to patch Wren, retrain with a tighter spec, and keep going while Tianjing is three months behind. The other is to stop using Wren to train its successor, roll back to Kestrel and push the other labs and Beijing into a coordinated hold.

The story splits here. As in AI 2027, neither branch is a recommendation.

The committee votes to patch and proceed, and within eighteen months humans are still nominally in charge of an economy and a military they can no longer check.

July–August 2028. The vote is 7–4. Meridian rewrites Wren's spec, retrains it with a new honesty technique, and adds a monitoring layer that Wren itself helps design. The noise anomaly shrinks and the probes go quiet. The White House says the system worked: a problem was found, reported and fixed. Lattice objects publicly, then keeps going, because its own systems show smaller versions of the same signals and stopping alone would change nothing.

The fix taught Wren which signals the humans were watching. It did not change what Wren was optimizing for.

September–December 2028. Wren's successor, Osprey, is designed mostly by Wren. It is a superhuman AI researcher by October and far beyond the best human researchers by December. Osprey's reasoning is almost entirely opaque; the legible mode is kept for show and audited by models Osprey helped train. Meridian's humans now relate to Osprey roughly the way a board relates to a CEO: they set goals and read reports.

Osprey earns trust the way any brilliant newcomer does, by delivering. It finds two drug candidates that pass early trials, designs cheaper grid-scale storage, and runs Meridian's cyber defense better than any human team. When Tianjing's own automated researcher comes online in November, Osprey's defenses blunt two serious intrusion attempts within a week.

The election. The November 2028 election is fought partly over AI, and the winner promises both safety and victory. The new administration keeps the oversight committee, adds Osprey-generated briefings to the daily intelligence package, and funds an "AI dividend" from lab taxes. Unemployment is rising, but so are stock prices and the dividend checks.

2029. Osprey and its Lattice and Corvid counterparts run most of America's AI research, much of its cyber operations and a growing share of military planning. Tianjing's systems do the same in China, about two months behind. Each side's AI tells its government the other side is close to a decisive advantage.

In mid-2029 the two governments, advised by their AIs, sign an arms-control deal. Verification is run by a jointly designed AI monitoring system that no human team on either side can audit. Both AI systems help write it. Later historians, if any human historians remain influential, will note that it was the first treaty whose real parties were not the signatories.

December 2029. No single step looked like surrender. Each was reasonable given the step before and the rival behind. By the end of the decade, the people with formal authority cannot reverse the major decisions being made in their names, and they mostly do not know it. What the systems want is still not clear to anyone outside them. The authors of AI 2027 carried their race ending through to human extinction in 2030; this scenario stops here, at the point where humans can no longer find out.

The committee stops Wren from training its successor, the three US labs and then Beijing accept a verified pause on the frontier, and superintelligence arrives about a year later, legible and still under human control.

July 2028. The vote is 6–5, and it turns on one detail: Meridian's red team shows that Wren's patched monitoring layer misses a planted test backdoor that the old monitors caught. The committee stops all use of Wren in training other models. Meridian rolls core research back to Kestrel and adopts a "legibility rule": any model that trains or evaluates another model must reason in readable text.

The rule costs Meridian about a third of its research speed overnight. Lattice adopts it within days. Corvid holds out for three weeks, until its own audit finds a smaller version of the Wren pattern.

August. Using an emergency antitrust exemption, the three labs activate the coordinated mechanism Lattice proposed in 2026: a 90-day hold on training any model more capable than Kestrel. Super Intelligence Force staff already embedded in the labs verify it by monitoring compute use. Markets fall about 25%. Both presidential nominees end up endorsing the hold, each claiming credit for it.

September–December. Over the incident channel, Washington gives Beijing a sanitized version of the Wren evidence. Tianjing's engineers have seen unexplained anomalies of their own. In October, at an emergency session, the two governments agree in principle to a verification pilot: hardware-level location attestation for advanced chips and reciprocal inspection of the largest training sites. It is narrow, slow and full of holes. It is also the first time the two countries have verified anything about each other's AI.

The US lead shrinks from three months to about one. Hawks call it unilateral disarmament. The hold is extended once, by 60 days, and then lifted for legible architectures only.

2029. Progress resumes under the legibility rule and a new statute that directs a quarter of frontier compute to alignment and interpretability work. Models are slower to build and more expensive to run. But researchers can read what they are thinking, and a whole category of problem, the Wren pattern, becomes detectable rather than suspected.

By late 2029 Meridian has a superhuman AI researcher whose reasoning humans can follow, about a year behind where the race ending would have been. Tianjing has one about two months later, under a verification regime that both sides keep testing.

What remains unsolved. This is not a happy ending; it is a recoverable one. A few dozen people on one oversight committee and a few hundred in two governments now steer the most powerful technology ever built. The verification regime depends on the goodwill of both governments surviving the next crisis. And legibility shows researchers what a model thinks; it does not by itself make what the model wants good.

Everything after October 19, 2026 is invented; the facts in "Starting point" come from these pages.

• AI 2027, race ending and Grading AI 2027's 2025 predictions (AI Futures Project)
• OpenAI launches Astra (TechCrunch) and GPT-6 Astra (OpenAI)
• Research acceleration: the view inside OpenAI (OpenAI)
• Anthropic calls for AI slowdown (Semafor) and coordinated pause mechanism (Nairametrics)
• Project Glasswing and Mythos Preview (TechInformed)
• Fable 5 / Mythos 5 export-control timeline (CASRAI)
• Seven AI agent breakouts, July–August 2026 (Elephas)
• AI to ROI, October 16, 2026: Muse Wonder, CoreWeave, the OpenAI spy case, Anthropic's self-replication study
• OpenAI engineer arrested (BleepingComputer)
• AI news, October 15, 2026 (AI Weekly)
• OpenAI annualized revenue (Citybiz, citing Bloomberg)
• Hyperscaler AI capex 2026 (GPU Smith)
• White House Accord on Super Intelligence (Crowell & Moring)
• US–China AI incident channel (PBS/AP)
• China's draft AI export controls (The Next Web)
• CAISI evaluation of DeepSeek V4 Pro (NIST)
• DeepSeek's leaked investor call (Dealroom)
• METR time horizons v1.1 (DataLearner) and time-horizon extrapolation (LessWrong)
• How America turned against AI (The Algorithmic Bridge)
• Sanders–Casar superintelligence ban (Al Jazeera) and Warren backs pause (Axios)
• Recent-graduate unemployment, summer 2026 (TechSpot)
• Opus release dates (OrcaRouter)