How does ChatGPT work?

ChatGPT is not a single magic box. It is the consumer-facing layer built on top of GPT-class models, instruction tuning, conversation handling, and deployment infrastructure. GPT-1 proves the blueprint, GPT-3 makes broad language performance unavoidable, then ChatGPT turns frontier research into mass behaviour — 100 million users in 60 days.

AGI — Artificial General Intelligence — usually means an artificial system capable of matching or exceeding human performance across a broad range of intellectual tasks, not just one benchmark. The term has moved from founding mission statement to industry forecast, with AI leaders openly attaching timelines to it. Whether the term is precise is less important than the fact that powerful actors are acting as though it is reachable.

Who is leading the AI race right now?

Nobody wins this race cleanly for long. OpenAI leads some moments, Anthropic closes in on capability, Meta changes the control problem with open weights, Google and xAI stay in the fight, and the frontier keeps shifting. The race is not a scoreboard but a pattern: multiple labs, shrinking release intervals, rising capability, and every player pretending they can go fast without breaking the steering column.

Why did Anthropic split from OpenAI?

Anthropic was founded by people who believed OpenAI's governance and deployment posture were moving too aggressively. A safety schism drove the split. Later events show how hard it is to preserve those safety commitments once the state, investors, and strategic advantage start applying pressure.

Is AI already being used in warfare?

Yes. Not hypothetically. Not eventually. The chain runs from public promises against military use, to quiet policy removal, to defence partnerships, to classified Pentagon contracting, and to battlefield targeting. The distance between consumer chatbot and military system turned out to be shorter than the marketing implied.

Is open-source AI a risk or a benefit?

Both. Open models widen access, lower barriers, and prevent a handful of companies from owning all advanced capability. They also make containment vastly harder. Once powerful weights are released, nobody can revoke them, patch them, or put them back. Meta's Llama releases show open release not as ideology alone, but as a permanent shift in who controls advanced AI.

Where is all this heading?

The story is not one of stabilisation. It is a story of acceleration, institutional capitulation, state integration, and increasingly irreversible deployment. The AI future is not just smarter assistants — it is tighter coupling between frontier models, governments, militaries, infrastructure, and strategic competition. The timeline reads less like closure and more like a countdown.

2015 – PRESENT

THE END HAS BEGUN

Q: Do AI safety rules actually hold under pressure?

Until pressure arrives, yes. Once pressure arrives, the evidence is less reassuring. Governance structures, public red lines, and safety language all look stronger before a board coup, a political ultimatum, or a weapons contract. The pattern shows rules failing not in theory but in documented sequence.

THE COMPLETE GUIDE FROM LARGE LANGUAGE MODEL TO WEAPON OF MASS DESTRUCTION

— FILTER BY COMPANY —

▼ SCROLL TO BEGIN ▼

DEC 2015

OpenAI Is Founded

DEC 2015

MILESTONE

OpenAI Is Founded

OpenAI is founded as a nonprofit on December 11, 2015 with a mission to ensure AGI benefits all humanity. It begins with an explicitly open research model, but later moves toward restricted work and a for-profit structure.

Founded Dec 11, 2015. The "open" in OpenAI was literal — all research to be published freely. Within two years, key research was restricted. Within seven years, converted to for-profit. The mission statement remains unchanged on their website. Musk's co-founding of the company he would later sue is the foundational irony of the AI era.

CLICK TO EXPAND ▼

DEC 2015

Musk: "This Is Not About Money"

DEC 2015

CEO QUOTE → CONTRADICTED

Musk: "This Is Not About Money"

Elon Musk at founding: "This is genuinely about doing the right thing and not about making money."

CONTRADICTED 2024: Musk sues OpenAI for $97.4 billion, alleging they abandoned their nonprofit mission for profit. He simultaneously builds xAI — his own for-profit AI company — and bids to acquire OpenAI to control it. Money turns out to be very much involved on every side.

CLICK TO EXPAND ▼

FEB 2018

Musk Leaves OpenAI Board

FEB 2018

MILESTONE

Musk Leaves OpenAI Board

Musk leaves OpenAI's board after pushing to take control as CEO and being refused. Tesla conflict was the public reason, but the split also marked the start of the later Altman–Musk power struggle over who controls AGI.

Musk proposed taking over as CEO during a board meeting about OpenAI's direction. The board — including Altman — declined. He left. Both cited Tesla conflicts publicly. Both accounts are partially true. The departure planted the seed for the 2024 lawsuit, the $97.4B acquisition bid, and the broader power struggle over who controls the path to AGI.

CLICK TO EXPAND ▼

2018–2019

Musk Starts Firing Shots Publicly

2018–2019

THE BACK AND FORTH BEGINS

Musk Starts Firing Shots Publicly

Soon after leaving OpenAI, Musk publicly attacks the company while still claiming its mission matters. The quote matters because the conflict stops being private almost immediately.

Musk's post-departure behaviour was unusually public for a board resignation. He continued promoting OpenAI while criticising it — a pattern that continued for years. His nickname "OpenClosed AI" stuck. Altman largely ignored these comments publicly at this stage, describing the relationship as "complicated but respectful." That characterisation would not last.

CLICK TO EXPAND ▼

2023

Musk vs. Altman: The Gloves Come Off

2023

CEO QUOTES — ESCALATING

Musk vs. Altman: The Gloves Come Off

The Musk–Altman split becomes openly hostile in public, with each side framing the other as betraying OpenAI's original purpose. The rivalry is now visible, personal, and strategic.

The exchange happened over multiple Twitter/X posts across 2023. Musk escalated consistently; Altman mostly declined to engage, which appeared to frustrate Musk further. Altman told Lex Fridman in 2023: "I think Elon is wrong about a lot of the AI stuff, but I understand why he believes what he believes." He did not specify which parts Musk was wrong about.

CLICK TO EXPAND ▼

JUN 2018

GPT-1: The Blueprint (117M Parameters)

JUN 2018

MODEL RELEASE

GPT-1: The Blueprint (117M Parameters)

GPT-1 shows that a model trained simply to predict the next word can generalise across many language tasks without task-specific training. That basic architecture becomes the foundation for the later GPT, Claude, Gemini, and Grok model lines.

117 million parameters. GPT-4 is estimated at ~1 trillion. GPT-1 demonstrated that a single objective — predict the next word — could generalise to many tasks without specific training. Every model that followed, including Claude, Gemini, and Grok, is a descendant of this architecture and this idea.

CLICK TO EXPAND ▼

FEB 2019

GPT-2: "Too Dangerous to Release" (1.5B Parameters)

FEB 2019

MODEL RELEASE

GPT-2: "Too Dangerous to Release" (1.5B Parameters)

OpenAI withholds GPT-2 at first, citing disinformation risk, then releases it in stages before publishing the full model. The episode puts "AI safety" into mainstream tech discussion for the first time and signals how powerful text generation is becoming.

1.5B parameters. GPT-2 wrote convincing news articles, fiction, and persuasive content. OpenAI released it in four stages — a first in AI. The full model landed quietly. Its legacy: the word "AI safety" entered mainstream tech discourse for the first time. Three years later, the conversation would become desperately urgent.

CLICK TO EXPAND ▼

JUL 2019

Microsoft Invests $1 Billion: The Partnership Begins

JUL 2019

MILESTONE

Microsoft Invests $1 Billion: The Partnership Begins

Microsoft's first $1 billion investment gives OpenAI the compute and funding it could not supply on its own. In return, Microsoft gains a central role in OpenAI's products and future direction, making the partnership structurally important to both companies.

The Microsoft-OpenAI partnership proved mutually transformative. Microsoft integrated GPT into Bing, Office, Windows, and Azure. OpenAI got the computing resources it couldn't afford alone. When the board tried to fire Altman in 2023, Microsoft held enormous leverage and used it. Without Microsoft, OpenAI as it exists today does not exist.

CLICK TO EXPAND ▼

JUN 2020

GPT-3: 175 Billion Parameters. Everything Changes.

JUN 2020

MODEL RELEASE

GPT-3: 175 Billion Parameters. Everything Changes.

GPT-3 makes machine-written essays, code, and business text feel commercially useful rather than experimental. It is the release that turns AI from a research story into a serious labour, product, and industry story.

GPT-3's API spawned hundreds of applications within weeks. The New York Times: "A Machine That Writes Anything." GPT-3 was the first model that made "AI replacing jobs" feel plausible rather than theoretical. It was GPT-3 that changed the question from "can AI write?" to "why are we still hiring humans to write?"

CLICK TO EXPAND ▼

JAN 2021

DALL-E: AI Learns to Create Images

JAN 2021

MODEL RELEASE

DALL-E: AI Learns to Create Images

DALL-E proves that text-to-image generation can produce usable visual outputs from written prompts, helping trigger the wider AI image boom. It also sharpens the unresolved question of ownership when models are trained on large volumes of human-made artwork.

DALL-E 2 (April 2022), DALL-E 3 (October 2023). Midjourney v5 (March 2023) produced photorealistic images indistinguishable from photographs. The AI art controversy crystallised a question the industry continues to avoid: who owns the output of a model trained on billions of images taken without permission?

CLICK TO EXPAND ▼

APR 2021

Anthropic Founded: The Safety Schism

APR 2021

MILESTONE

Anthropic Founded: The Safety Schism

Senior OpenAI staff leave to found Anthropic around the claim that frontier AI needs stronger safety controls. That split creates the company whose models later sit at the centre of the Pentagon conflict and Iran timeline.

Dario was VP of Research at OpenAI. Daniela was VP of Operations. Google invested $300M immediately; Amazon committed $4B. Anthropic's "constitutional AI" became the industry's safety standard — right up until Pete Hegseth issued his ultimatum in February 2026 and demanded they remove it.

CLICK TO EXPAND ▼

NOV 2022

ChatGPT: 100 Million Users in 60 Days

NOV 2022

MILESTONE

ChatGPT: 100 Million Users in 60 Days

ChatGPT reaches 100 million users in about 60 days, making it the fastest-adopted consumer technology of its time. More importantly, it changes public expectations overnight by showing millions of ordinary users what frontier AI can already do.

TikTok took 9 months to reach 100M users. Instagram: 2.5 years. The internet: 7 years. ChatGPT: 60 days. ChatGPT did not just change what AI could do — it changed what the public believed was possible. The 18 months that followed were the most consequential in the history of technology.

CLICK TO EXPAND ▼

2022

OpenAI: "We Will Not Work With the Military"

2022

POLICY → REVERSED

OpenAI: "We Will Not Work With the Military"

OpenAI's usage policy: "Weapons development, military and warfare — these uses are prohibited." Absolute. Clear. Explicit.

REVERSED JAN 2024: OpenAI quietly removes "military and warfare" from its prohibited use list. The change is not announced. Journalists at The Intercept discover it. By December 2024 they are building weapons AI for Anduril's drone systems. By February 2026 they sign a classified Pentagon weapons contract — replacing Anthropic, who refused. The original policy was still on the Wayback Machine.

CLICK TO EXPAND ▼

2023

The Creator: What If We're the Villain?

2023

POP CULTURE

The Creator: What If We're the Villain?

The film is included here because it flips the usual AI-threat framing and asks whether humans, not machines, are the real aggressors. It becomes cultural context for the later military entries.

Critics noted the film's timing made it uniquely resonant. Several AI researchers cited it as the first mainstream blockbuster to seriously engage with AI's perspective. Released the same year as Oppenheimer — another film about what happens when humanity builds something it cannot un-build. Both films did well. Both warnings were largely ignored.

CLICK TO EXPAND ▼

MAR 2023

GPT-4: Human-Level Benchmarks Collapse

MAR 2023

MODEL RELEASE

GPT-4: Human-Level Benchmarks Collapse

GPT-4 clears major exams and professional benchmarks in law, medicine, and reasoning, showing that high-value expertise is no longer exclusively human. It also marks OpenAI's move away from early technical openness.

GPT-4 was multimodal (text + images) with an estimated 1 trillion parameters. Its technical details were kept confidential — a pivot from the "open" mission cited in Musk's lawsuit. GPT-4 also demonstrated alarming emergent capabilities not present in GPT-3: planning, step-by-step reasoning, and theory of mind. These were not explicitly trained for. They appeared at scale.

CLICK TO EXPAND ▼

MAR 2023

Claude 1: Anthropic Enters the Race

MAR 2023

MODEL RELEASE

Claude 1: Anthropic Enters the Race

Anthropic launches Claude 1 with its "constitutional AI" approach, positioning itself as the safety-focused alternative to OpenAI. That same Claude lineage later ends up inside classified US military systems despite being built around stricter guardrails.

Claude 1's constitutional AI published its methods transparently. The "constitution" involves the model critiquing its own outputs against a set of principles. Anthropic published their approach; GPT-4's technical details were kept secret. This openness about method positioned Anthropic as the safety-first alternative — a reputation that would become both its greatest asset and its greatest vulnerability.

CLICK TO EXPAND ▼

MAY 2023

The Godfather Quits Google to Warn You

MAY 2023

MILESTONE

The Godfather Quits Google to Warn You

Geoffrey Hinton leaves Google and publicly warns that the systems he helped make may outgrow human control. When one of deep learning's foundational figures says the risk is serious, AI danger stops looking like a fringe argument.

"I console myself with the normal excuse: if I hadn't done it, someone else would have." Hinton joined Yoshua Bengio and others in calling for development pauses. The industry largely ignored these calls. Sam Altman testified to Congress and said he was "a little bit worried." Then returned to work and accelerated development. Hinton later called Altman's congressional testimony "disingenuous."

CLICK TO EXPAND ▼

JUL 2023

Meta Llama 2: Free AI for Everyone

JUL 2023

MODEL RELEASE

Meta Llama 2: Free AI for Everyone

Meta releases Llama 2 for broad public use, making frontier-style language models downloadable and modifiable outside closed corporate APIs. That helps create an open model ecosystem that no single company can fully restrict, patch, or recall.

Llama 2 (7B, 13B, 70B) spawned hundreds of community fine-tunes including uncensored models with no safety restrictions. This "open source AI" movement now produces models competitive with the closed frontier. It means there is AI that no lab can patch, restrict, or shut down. The implications for safety research are not reassuring.

CLICK TO EXPAND ▼

FRI NOV 17 2023

The Board Fires Sam Altman

FRI NOV 17 2023

MILESTONE

The Board Fires Sam Altman

OpenAI's board removes Sam Altman in a sudden vote, saying he was "not consistently candid" with them. The decision exposes how unstable governance had become inside the most important AI company in the world and immediately triggers a wider power struggle.

Board vote: 4-2. Voting to remove: Ilya Sutskever (co-founder), Adam D'Angelo, Tasha McCauley, Helen Toner. Reports suggest the board feared rapid capability advances — possibly a project called Q* — and believed Altman was moving without adequate safety oversight. Sutskever would later say he regretted his vote. The board had no replacement CEO ready. They announced it to most staff simultaneously with the public.

CLICK TO EXPAND ▼

SAT–MON NOV 18–20 2023

700 Employees Sign a Letter: "Rehire Him or We All Quit"

SAT–MON NOV 18–20 2023

THE WALKOUT THAT WASN'T

700 Employees Sign a Letter: "Rehire Him or We All Quit"

OpenAI staff overwhelmingly back Altman and warn that they may resign if he is not restored. The letter shows where practical power inside OpenAI actually sits.

Senior researchers, engineers and safety staff all signed — many hired specifically for the safety mission. The safety board that fired Altman to protect the mission had triggered the event that would end the mission if they didn't reverse. Even Ilya Sutskever — one of the four who voted to fire Altman — signed the letter. He posted: "I deeply regret my participation in the board's actions."

CLICK TO EXPAND ▼

TUE NOV 21 2023

Altman Returns. The Safety Board Is Gone.

TUE NOV 21 2023

MILESTONE

Altman Returns. The Safety Board Is Gone.

Altman returns within days, and the board members who removed him are pushed out. The practical result is that OpenAI's independent safety oversight fails the moment it tries to use its authority, and it is not rebuilt in equivalent form.

Altman returned to cheering staff. He posted on X wearing an OpenAI guest badge — the kind given to visitors — captioned "first and last time i wear one of these." The new board included Bret Taylor and Larry Summers. Helen Toner and Tasha McCauley were out. OpenAI never reconstituted an equivalent independent safety oversight body. The company continued as before, faster.

CLICK TO EXPAND ▼

DEC 2023

Altman: "We Share Anthropic's Exact Red Lines"

DEC 2023

CEO QUOTE → CONTRADICTED

Altman: "We Share Anthropic's Exact Red Lines"

After reinstatement, Altman told employees: "We share the same red lines as Anthropic — no mass surveillance of Americans, no autonomous weapons. These are our absolute principles."

CONTRADICTED FEB 2026: When Anthropic refused the Pentagon contract over those exact red lines, they were banned from US government work and designated a national security risk. OpenAI signed the contract within hours. Altman later admitted the deal was "rushed." He did not explain why the rush made the principles optional. The quote was still on the internet.

CLICK TO EXPAND ▼

DEC 2023

Gemini 1.0 + Grok 1.0: The Race Goes Four-Way

DEC 2023

MODEL RELEASE

Gemini 1.0 + Grok 1.0: The Race Goes Four-Way

Google's Gemini and Musk's Grok arrive as serious competitors, turning the frontier race into a four-way contest among OpenAI, Anthropic, Google, and xAI. From this point on, release speed and competitive pressure accelerate across the whole field.

Gemini's staged demo showed capabilities the model couldn't actually perform in real time. Google apologised. Grok positioned as "anti-woke AI" with fewer content restrictions, targeting Musk's X platform user base. By mid-2024 both were genuinely competitive with GPT-4. The four-way competition accelerated every company's release cadence.

CLICK TO EXPAND ▼

JAN 2024

Microsoft: $80 Billion in AI Infrastructure

JAN 2024

DATA CENTRE

Microsoft: $80 Billion in AI Infrastructure

Microsoft commits $80 billion to AI infrastructure in 2025 alone, showing that the frontier race now depends on power, land, and data-centre construction as much as model design. AI competition is no longer just about code; it is about industrial scale.

Microsoft's Fairwater (phase 1) came online early 2026. Phase 2 ($4B additional) announced September 2025. Microsoft's Wisconsin campus is projected to exceed 3 gigawatts by 2028 — nearly half the output of the world's largest nuclear power plant. In 2025, Amazon, Google, Microsoft and Meta collectively invested $320 billion in AI infrastructure — up from $230B the prior year.

CLICK TO EXPAND ▼

FEB 2024

Musk Sues OpenAI: "You Betrayed Humanity"

FEB 2024

MILESTONE

Musk Sues OpenAI: "You Betrayed Humanity"

Musk sues OpenAI over its shift away from the nonprofit mission he helped launch. The case matters because the criticism is fact-based in part, even while Musk is building his own for-profit AI and defence relationships.

The complaint cited founding documents, internal emails, and OpenAI's capped-profit structure. Musk's $97.4B acquisition bid was rejected. The lawsuit continued into 2026. Meanwhile, xAI's Grok was approved for classified Pentagon use in February 2026 — the same week Musk was publicly criticising the OpenAI-DoD deal. His alternative to OpenAI's military contract was his own military contract.

CLICK TO EXPAND ▼

JAN 2024

OpenAI Quietly Removes Its Military Ban

JAN 2024

MILESTONE

OpenAI Quietly Removes Its Military Ban

OpenAI removes the explicit ban on "military and warfare" use from its policy without publicly announcing the change. That small policy edit opens the path to the defence partnerships and Pentagon work that follow.

The word "military" was simply deleted from the prohibited use list. This small edit opened the door to everything that followed. The change went unannounced. The original policy was documented on the Wayback Machine. If a company quietly removes its most important ethical boundary without telling anyone, that tells you something about how the company views that boundary.

CLICK TO EXPAND ▼

MAR 2024

Claude 3: Anthropic Becomes a Capability Rival

MAR 2024

MODEL RELEASE

Claude 3: Anthropic Becomes a Capability Rival

Claude 3 proves Anthropic is no longer only the "safe alternative" but also a frontier capability rival to GPT-4. This is the model family that later gets deployed into classified US systems and becomes central to the Iran war reporting.

Claude 3.5 Sonnet (June 2024) became the world's top-ranked coding model. By July 2025, the DoD deployed Claude via Palantir and Amazon's classified cloud under a $200M contract. Pentagon CTO: Claude was used for intelligence synthesis, logistics efficiency, and target analysis. By February 2026, removing it would take "open-heart surgery."

CLICK TO EXPAND ▼

MAY 2024

GPT-4o: Real-Time Voice, Vision, Emotion

MAY 2024

MODEL RELEASE

GPT-4o: Real-Time Voice, Vision, Emotion

GPT-4o combines live voice, vision, and text in a form that feels conversational and emotionally responsive. The Johansson "Sky" dispute then turns the launch into a public argument about consent, intimacy, and imitation.

Johansson had been approached to voice GPT-4o and declined. The "Sky" voice launched with a nearly identical tone. She threatened legal action. OpenAI pulled the voice. Altman's "her" post was then quietly deleted. The entire episode lasted one afternoon and perfectly summarised every cultural ambivalence around AI: the desire for intimacy, the blurred consent, the commercial logic, the cover-up.

CLICK TO EXPAND ▼

APR 2024

Meta Llama 3: Open Source Catches the Frontier

APR 2024

MODEL RELEASE

Meta Llama 3: Open Source Catches the Frontier

Llama 3 shows that open models can compete with major closed systems on important benchmarks. At the same time, Meta openly frames the goal as general intelligence, confirming that the race is still accelerating rather than stabilising.

Zuckerberg: "We're building toward general intelligence. This is the decade we solve it." Meta projected spending $600B on US AI infrastructure through 2028. Data centres named Prometheus (Ohio) and Hyperion were planned as the largest private infrastructure buildout in corporate history outside the Stargate consortium. The race had no ceiling.

CLICK TO EXPAND ▼

SEP 2024

xAI Colossus: World's Largest AI Cluster in 122 Days

SEP 2024

DATA CENTRE

xAI Colossus: World's Largest AI Cluster in 122 Days

xAI brings its Colossus cluster online at extreme speed, making it the largest AI training installation of its moment. The buildout shows how quickly compute can be assembled when environmental friction is pushed aside in pursuit of frontier scale.

Built on a former industrial site in Memphis, Tennessee. The construction speed required circumventing environmental review — drawing EPA scrutiny and community protests over air quality. Colossus 2 was announced late 2024 for further scale-up. xAI needed this compute to train Grok 3 and Grok 4 — which by 2025 were competitive with GPT-4 and Claude 3.

CLICK TO COLLAPSE ▼

SEP 2024

OpenAI o1: The Model That Reasons

SEP 2024

MODEL RELEASE

OpenAI o1: The Model That Reasons

OpenAI o1 introduces a reasoning-heavy model that spends more compute "thinking" before it answers. It marks a visible shift from fast completion models toward systems designed to solve harder scientific, mathematical, and planning tasks.

o1 introduced scaled thinking time — compute spent reasoning before output. ARC-AGI was created specifically to resist AI pattern matching. Its designer, François Chollet, had called it "safe" from AI for years. o3 shattered that assumption in December 2024. The benchmark designed to measure distance from AGI suddenly looked much less safe.

CLICK TO EXPAND ▼

DEC 2024

OpenAI + Anduril: The Military Pivot Is Complete

DEC 2024

MILESTONE

OpenAI + Anduril: The Military Pivot Is Complete

OpenAI partners with Anduril on military systems, making its defence pivot explicit rather than theoretical. Two years after banning military use, it is now working with a weapons company while still keeping its founding mission language intact.

Anduril was co-founded by Palmer Luckey and is among the most aggressive AI weapons companies in the US defence sector. OpenAI framed the deal as "consistent with our mission." A company founded as a nonprofit to benefit all of humanity had become a classified weapons AI contractor. The journey from "military use prohibited" (2022) to building drone weapons (2024) took exactly 24 months.

CLICK TO EXPAND ▼

DEC 2024 – JAN 2025

DeepSeek: China Matches the West for $6 Million

DEC 2024 – JAN 2025

MODEL RELEASE

DeepSeek: China Matches the West for $6 Million

DeepSeek shows that Chinese open models can approach top Western performance at far lower reported cost. That weakens the idea that only massive US compute spending can produce frontier AI.

DeepSeek's efficiency innovations (Mixture-of-Experts architecture, knowledge distillation) showed algorithmic progress could offset hardware. This had enormous implications for US chip export controls. Restricting GPU sales to China had not prevented frontier AI development. The geopolitical AI race became more complex than any existing policy framework had anticipated.

CLICK TO EXPAND ▼

JAN 2025

Stargate: $500 Billion Announced at the White House

JAN 2025

DATA CENTRE

Stargate: $500 Billion Announced at the White House

Stargate is announced at the White House as a $500 billion AI infrastructure project linking OpenAI, Oracle, SoftBank, and Nvidia. It formalises frontier AI as state-aligned industrial buildout on a scale previously associated with national megaprojects.

Partners: OpenAI (models), Oracle (data centres), SoftBank (funding), Nvidia (chips). Flagship site: Abilene, Texas. 6,400 workers. Enough fibre optic cable to wrap the Earth 16 times. By September 2025, Abilene was running training workloads. Five additional sites announced September 2025 across New Mexico, Ohio, Wisconsin (x2), and Texas. Total planned: ~7 gigawatts of compute capacity.

CLICK TO EXPAND ▼

APR–JUL 2025

Llama 4, Claude 4, Grok 4: The Quarterly Frontier

APR–JUL 2025

MODEL RELEASE

Llama 4, Claude 4, Grok 4: The Quarterly Frontier

Major frontier models are now landing within months of each other, not years. The pace itself becomes part of the story: benchmarks fall quickly, and civilisation-scale capability jumps start arriving on a quarterly cycle.

Claude 4 Opus scored 72.5% on SWE-Bench software engineering (vs GPT-4's 54.6%). Humanity's Last Exam was created specifically to be "too hard for any AI." Grok 4 scored 25.4% on it within months of launch. The exam lasted less than a year as an effective barrier to AI capability measurement. Benchmarks now fail faster than they can be created.

CLICK TO EXPAND ▼

JUL 2025

DoD Awards $200M Contract: Claude Enters Classified Systems

JUL 2025

MILESTONE

DoD Awards $200M Contract: Claude Enters Classified Systems

The Pentagon awards major AI contracts that place Claude inside classified workflows through Palantir and Amazon infrastructure. By the time the public notices, Anthropic's systems are already deeply embedded in military analysis and operations.

Pentagon CTO Michael described Claude's roles: summarising classified documents, logistics efficiency, operational intelligence. By February 2026, Claude was so deeply embedded that removing it would take "3–6 months minimum — like open-heart surgery." This integration was total and invisible to the public. The scale of AI in military systems was not disclosed. It became visible only when the ban was announced and couldn't be enforced.

CLICK TO EXPAND ▼

AUG 2025

GPT-5: "AGI Is a Few Thousand Days Away"

AUG 2025

MODEL RELEASE

GPT-5: "AGI Is a Few Thousand Days Away"

GPT-5 launches into a climate where even public AGI timelines from major lab leaders are measured in years, not decades. The release sits inside a broader late-2025 acceleration marked by new models, new spending, and explicit national-security urgency.

GPT-5 launched on the same day the White House issued its National Security Memorandum on AI — directing the Pentagon to accelerate AI adoption to counter China. The memo explicitly cited the DeepSeek shock as justification. Codex-Max (specialised coding model) launched with GPT-5.1. November 2025 Gemini 3 added multimodal reasoning over code, images, and video in one model.

CLICK TO EXPAND ▼

SEP 2025

Stargate Abilene Goes Live: World's Largest AI Facility

SEP 2025

DATA CENTRE

Stargate Abilene Goes Live: World's Largest AI Facility

Stargate Abilene begins live training workloads as the largest AI facility yet built, with on-site gas generation and expansion already underway. It shows what frontier AI now looks like in physical form: industrial power consumption at utility scale.

Oracle delivered first Nvidia GB200 racks in June 2025. Abilene phase 2 broke ground March 2025 and is expected online mid-2026. Five additional Stargate campuses announced September 2025 in New Mexico ($165B campus on 1,400 acres), Ohio, Wisconsin (x2, $15B each), and Texas. Total pipeline: nearly 7 gigawatts, over $400 billion in investment over three years.

CLICK TO EXPAND ▼

DEC 2024

Amodei: "We Have Red Lines We Will Not Cross"

DEC 2024

CEO QUOTE → COMPLICATED

Amodei: "We Have Red Lines We Will Not Cross"

Dario Amodei: "Some uses of AI are simply off-limits. Autonomous lethal weapons. Mass domestic surveillance. These are absolute red lines. We will not cross them regardless of commercial or political pressure."

COMPLICATED FEB 2026: Anthropic did hold those lines against Pentagon ultimatum and suffered unprecedented consequences. But journalist Spencer Ackerman noted: Claude was already being used for target identification in foreign war zones. Anthropic's red lines covered domestic surveillance and autonomous weapons — not AI-assisted selection of foreign targets. The line was real. It was just not where everyone assumed it was drawn.

CLICK TO EXPAND ▼

JAN 2026

First Confirmed Military Use: The Maduro Operation

JAN 2026

MILESTONE

First Confirmed Military Use: The Maduro Operation

US forces use Claude-linked systems in the Maduro operation, marking the first confirmed military use in this timeline. The event establishes that Anthropic's tools are already inside operational decision systems before the Iran escalation becomes public.

The Maduro operation used Claude for intelligence synthesis and operational planning — within Anthropic's stated framework of human oversight. Anthropic's silence was noted. Whether the company had advance knowledge is unclear. By February 2026, the same classification of systems would assist in a major joint US-Israeli airstrike. The bridge between "intelligence tool" and "war planning AI" had already been crossed.

CLICK TO EXPAND ▼

TUE FEB 24 2026

The Pentagon Ultimatum: Drop Your Guardrails

TUE FEB 24 2026

MILESTONE

The Pentagon Ultimatum: Drop Your Guardrails

The Pentagon pressures Anthropic to weaken restrictions on domestic surveillance and autonomous weapons, then sets a deadline. Anthropic refuses to sign language that would let those safeguards be bypassed at will.

The deadline: 5pm Friday, February 28. The DoD's final "compromise" proposal included legalese that would allow safeguards to be "disregarded at will" — per Anthropic's own spokesperson. Senators from both parties sent letters urging resolution. OpenAI simultaneously began drafting a contract framework during the same week — making the DoD's hard-line stance toward Anthropic significantly easier to maintain.

CLICK TO EXPAND ▼

FEB 28 2026 — 3:47pm

Trump Bans Anthropic Before the Deadline

FEB 28 2026 — 3:47pm

MILESTONE

Trump Bans Anthropic Before the Deadline

Trump orders federal agencies to stop using Anthropic before the Pentagon's own deadline expires, and Hegseth follows with a "supply chain risk" designation. The move is extraordinary because it targets an American AI company over refusing weaker guardrails.

Trump's Truth Social post called Anthropic "A RADICAL LEFT, WOKE COMPANY." Senator Wyden: "The Defense Department is throwing a fit over Anthropic asking for the bare minimum ethical guardrails." Legal experts immediately noted the designation likely violated procedural requirements: no risk assessment was completed; Congress was not notified. Multiple law professors called it legally vulnerable. It was implemented anyway.

CLICK TO EXPAND ▼

FEB 28 2026 — EVENING

OpenAI Signs the Deal Anthropic Was Destroyed For Refusing

FEB 28 2026 — EVENING

MILESTONE

OpenAI Signs the Deal Anthropic Was Destroyed For Refusing

OpenAI signs the Pentagon deal Anthropic refused, even as legal observers argue the practical restrictions are not meaningfully different. The outcome makes the government's preference for a more compliant supplier hard to ignore.

Both contracts ultimately referenced existing US law rather than explicit contractual prohibitions — which was the exact loophole Anthropic refused. The DoD accepted OpenAI's approach but not Anthropic's near-identical approach. The most plausible explanation: the government wanted a compliant partner, not a principled one. OpenAI's national security lead claimed the Pentagon had "no legal authority" for domestic surveillance — a statement multiple legal scholars immediately contradicted, citing documented warrantless DoD data purchases.

CLICK TO EXPAND ▼

FEB 28–MAR 1 2026

US–Israel Strike Iran: Claude Is the Targeting AI

FEB 28–MAR 1 2026

MILESTONE

US–Israel Strike Iran: Claude Is the Targeting AI

During the Iran strikes, Claude is reported as part of the system used for intelligence assessment, target identification, and battle simulation. The issue is no longer hypothetical military AI but active war-zone deployment.

Confirmed by Wall Street Journal, CBS News, Axios, Washington Post, and multiple intelligence sources. Claude used for target identification, intelligence assessment, and simulated war gaming at US Central Command. Pentagon CTO: removing Claude would take "3–6 months minimum." Claude remained in use despite the ban as of March 5, 2026. The AI safety debate had left the conference rooms. It was now writing target lists in a war zone.

CLICK TO EXPAND ▼

FEB 28 2026

The Minab School Strike: 165 Children Killed

FEB 28 2026

MILITARY SYSTEM

The Minab School Strike: 165 Children Killed

A strike on an elementary school in Minab kills 165 people, most of them girls aged 7 to 14. It is one of the clearest human-cost markers in the Iran timeline.

The school sat 200 feet from an IRGC naval facility. US Central Command created target coordinates using outdated intelligence provided by the Defense Intelligence Agency. Multiple independent investigations — CNN, NYT, NPR, BBC Verify, Amnesty International — concluded the US was responsible. Amnesty called it potentially unlawful. Eight UN experts demanded independent investigation. The AI targeting pipeline did not pull the trigger — but it structured the decision space in which 165 people died.

SOURCES:
CNN — US strike likely hit school due to outdated intelligence.
Washington Post — Target list may have mistaken school as military site.
Amnesty International — Strike on school likely unlawful, US responsible.
NPR — Satellite images confirm strike more extensive than reported.

CLICK TO EXPAND ▼

MAR 1–3 2026

1.5 Million People Say Enough. Claude Goes #1.

MAR 1–3 2026

PUBLIC RESPONSE

1.5 Million People Say Enough. Claude Goes #1.

Public backlash hits OpenAI after the Pentagon deal, while Claude surges in downloads and protest sentiment becomes visible. The market response shows that users were paying attention.

Sources: Medium (developia), TechCrunch, App Store rankings. A Reddit post urging "Cancel and Delete ChatGPT" pulled 30,000 upvotes before the weekend was over. Claude went from rank 42 to #1 in under a week. The public backlash was not over abstract AI ethics — it was personal. Surveillance. Autonomous weapons. These are things people are genuinely afraid of. The Pentagon did not model this outcome. OpenAI spent the following week renegotiating its deal toward the position of the company they replaced.

CLICK TO EXPAND ▼

MAR 3 2026

Altman: "We Rushed. The Optics Don't Look Good."

MAR 3 2026

CEO QUOTE — ONGOING

Altman: "We Rushed. The Optics Don't Look Good."

Sam Altman's public memo, March 2026: "We definitely rushed this deal. The optics don't look good. But we wanted to de-escalate the DoW vs. industry conflict. If we're right, we'll look like geniuses. If wrong, we took on enormous pain."

Altman argued OpenAI's layered safeguards (cloud-only, OpenAI staff in the loop) made their deal fundamentally safer. Critics observed that Claude — using the same DoD infrastructure and oversight mechanisms Altman cited — was at that moment assisting in an active airstrike in Iran. The circular logic was visible to most observers. The memo did not address it.

CLICK TO EXPAND ▼

MAR 4 2026

Amodei Calls Altman's Pentagon Deal "Straight Up Lies"

MAR 4 2026

CEO QUOTE — INTERNAL MEMO

Amodei Calls Altman's Pentagon Deal "Straight Up Lies"

Dario Amodei publicly accuses Altman of misrepresenting the Pentagon deal. The significance is that the safety split between Anthropic and OpenAI is now explicit and on the record.

Sources: The Information, TechCrunch. Amodei specifically took issue with OpenAI's claim that their contract covers "all lawful purposes" — noting that what is lawful can change. He wrote that Altman was "presenting himself as a peacemaker and dealmaker" and warned staff: "It is working on some Twitter morons, which doesn't matter, but my main worry is how to make sure it doesn't work on OpenAI employees." The two men who split over safety at OpenAI in 2021 are now in open conflict over the same question — at civilisational scale.

CLICK TO EXPAND ▼

MAR 5 2026

GPT-5.4 Released — First Model to Outperform Humans at Knowledge Work

MAR 5 2026

MILESTONE

GPT-5.4 Released — First Model to Outperform Humans at Knowledge Work

OpenAI presents GPT-5.4 as a stronger general professional model with native computer-use abilities. Its headline claim is that it matches or exceeds human professionals across most comparisons in a knowledge-work benchmark.

Sources: OpenAI official release, TechCrunch, Gizmodo. GPT-5.4 supports up to 1 million tokens of context. It is 33% less likely to produce errors in individual claims than GPT-5.2, and 18% less likely to produce errors overall. It incorporates the coding capabilities of GPT-5.3-Codex into a single general-purpose model. OpenAI internally rates it as "High" capability under its Preparedness Framework for both cybersecurity and biological/chemical domains — meaning maximum safeguards are applied. Whether those safeguards hold under the same political pressure that just failed at Anthropic is not addressed in the release notes.

CLICK TO EXPAND ▼

MAR 7 2026

Kalinowski, OpenAI Robotics: "Lines That Deserved More Deliberation"

MAR 7 2026

RESIGNATION STATEMENT

Kalinowski, OpenAI Robotics: "Lines That Deserved More Deliberation"

An OpenAI robotics leader publicly suggests that some ethical lines deserved more deliberation than they received. It matters because the criticism comes from inside the wider OpenAI orbit.

Sources: Fortune, TechCrunch, NPR, Engadget, SF Standard — all confirmed from Kalinowski's verified X account (@kalinowski007). She joined OpenAI in November 2024, previously led AR glasses development at Meta. In a follow-up comment she added: "My issue is that the announcement was rushed without the guardrails defined. It's a governance concern first and foremost. These are too important for deals or announcements to be rushed." OpenAI confirmed her departure. No plans to replace her. She is the most senior person to publicly resign over the deal. She led the robotics division — the team building physical AI systems that could most directly be weaponised.

CLICK TO EXPAND ▼

MAR 7 2026

Pentagon Official: "Holy Shit — What If This Software Went Down?"

MAR 7 2026

MILESTONE

Pentagon Official: "Holy Shit — What If This Software Went Down?"

A senior Pentagon technology official openly describes how dependent defence operations had become on Anthropic's software. The quote confirms that a war-critical AI system had become operationally indispensable.

Sources: All-In Podcast, Fortune. Michael confirmed Anthropic had become "deeply embedded" while other AI companies hadn't pursued enterprise defence customers as aggressively. He said he contacted OpenAI, xAI, and Google to get redundancy — giving them all the same terms. He also raised the concern that a rogue developer could "poison the model" to make it hallucinate during operations. The Pentagon's own CTO publicly admitted on a podcast that they had built a war machine dependent on a single company's AI — and then tried to ban that company mid-war. The architecture of the problem is the problem.

CLICK TO EXPAND ▼

MAR 9 2026

Anthropic Sues the Pentagon

MAR 9 2026

MILESTONE

Anthropic Sues the Pentagon

Anthropic sues to block the Pentagon's designation, calling it unlawful and politically motivated. The company built around safety principles is now in federal court against the same state that embedded its software in defence systems.

Sources: Insurance Journal, Fortune, TechCrunch, Reuters. Legal experts Michael Endrias and Alan Z. Rozenshtein called the designation "political theater: a show of force that will not stick" — arguing the government cannot credibly claim an emergency threat while simultaneously planning a six-month phaseout. Anthropic said the lawsuit does not preclude re-opening negotiations. Meanwhile, Microsoft confirmed Claude remains available for non-defence AWS customers. The designation is narrower than it sounds — but the signal it sends is not.

CLICK TO EXPAND ▼

MAR 10 2026

AI Enters Battlefield Planning

MAR 10 2026

MILITARY SYSTEM

Frontier AI Used For Operational War Planning

Palantir demonstrations reveal large language models being integrated directly into military command software — analysing intelligence feeds and suggesting operational plans for commanders.

For years the concern about AI in warfare focused on autonomous weapons. The reality emerging in 2026 is different: decision-support systems embedded inside intelligence pipelines. These systems do not fire missiles themselves, but they shape the speed, structure, and confidence of human decisions. When the analysis pipeline runs through a model, the tempo of war changes with it.

SOURCES:
WIRED — “Palantir demos show how the military could use AI chatbots to generate war plans.”

CLICK TO EXPAND ▼

MAR 12 2026

Pentagon Blacklists AI Supplier

MAR 12 2026

CORPORATE CONFLICT

Pentagon Declares Anthropic A Supply-Chain Risk

After the Iran conflict revealed how embedded Claude had become in defence systems, the Pentagon designated Anthropic a supply-chain risk, restricting its use across defence contracts.

Anthropic responded by seeking an appeals-court stay after filing suit against the designation. The conflict marks a turning point in the relationship between frontier AI labs and the state: companies that build the models now find themselves negotiating directly with military power structures over how those systems may be used.

SOURCES:
Reuters — “Anthropic seeks appeals court stay of Pentagon supply-chain risk designation.”

CLICK TO EXPAND ▼

MAR 13 2026

Data Centres Become War Targets

MAR 13 2026

STRATEGIC SHIFT

Compute Infrastructure Enters The Battlefield

Strikes against datacentres in the Gulf highlight a new strategic reality: AI infrastructure — cloud compute, model training clusters, and data pipelines — is becoming a military-relevant target.

Once intelligence analysis, targeting pipelines, and battlefield logistics rely on AI systems, the infrastructure supporting those systems becomes strategically valuable. Destroying compute capacity can degrade an opponent's ability to analyse information and coordinate operations — making datacentres the digital equivalent of fuel depots or communications hubs.

SOURCES:
The Guardian — “Datacenters are becoming a target in warfare for the first time.”

CLICK TO EXPAND ▼

MAR 11 2026

US Military Officially Confirms AI in Iran War

MAR 11 2026

MILESTONE

US Military Officially Confirms AI in Iran War

CENTCOM publicly confirms that advanced AI tools are being used in the Iran war to process data and support faster operational decisions. That statement turns years of speculation into official acknowledgement that AI is already part of live military action.

The confirmation came as civilian casualty figures mounted, including the Minab school strike. Senator Kirsten Gillibrand called for explicit safeguards: "There is little reason to trust that the DOD will be any more responsible with its use of AI without explicit safeguards." Representatives Sara Jacobs and Jill Tokuda called for a full review. The Pentagon did not respond to questions about how it balances AI workload reduction against verification accuracy. The military had spent years saying AI would only ever support human decisions. The humans were now processing 1,000 targets a day.

SOURCES:
Al Jazeera — US military confirms use of advanced AI tools in war against Iran.
NBC News — US military using AI to help plan Iran air attacks.

CLICK TO EXPAND ▼

MAR 12 2026

AI Used to Rank and Prioritise Strike Targets

MAR 12 2026

MILITARY SYSTEM

AI Used to Rank and Prioritise Strike Targets

Reporting indicates AI was used not just to summarise information but to help rank and prioritise strike targets. That moves the technology closer to the actual structure of lethal decision-making.

The disclosure came as the Pentagon faced scrutiny over the Minab school strike. The official spoke on background to discuss what they described as sensitive operational matters. The military uses two distinct AI layers: Maven (computer vision, older technology, processes drone imagery) and generative AI models like Claude (interprets, ranks, recommends). Maven finds things. Claude tells commanders what to do about them. The Pentagon did not confirm whether this pipeline was active on 28 February.

SOURCES:
MIT Technology Review — Defense official reveals how AI chatbots could be used for targeting decisions.

CLICK TO EXPAND ▼

MAR 17 2026

Pentagon Plans to Train AI on Classified Battlefield Data

MAR 17 2026

MILITARY SYSTEM

Pentagon Plans to Train AI on Classified Battlefield Data

The Pentagon plans to train models on classified battlefield and surveillance data, expanding AI from analysis tools toward systems shaped directly by war inputs. That raises the stakes for governance and misuse.

The plan is framed as making models more accurate and effective for military tasks. Security risks acknowledged: sensitive intelligence could become permanently embedded in model weights, and AI firms would gain unprecedented proximity to classified material. The news came as OpenAI, xAI, and Google were all being offered the same Pentagon terms Anthropic had refused. The company that held its red lines was banned. The companies that didn't are now being invited to train on state secrets.

SOURCES:
MIT Technology Review — The Pentagon is planning for AI companies to train on classified data.

CLICK TO EXPAND ▼

APR 9 2026

Pentagon Scrambles For Backup AI Suppliers

APR 9 2026

MILITARY SYSTEM

Pentagon Scrambles For Backup AI Suppliers

After the Anthropic conflict, the Pentagon moves to line up other model suppliers so it is not trapped by a single vendor. That scramble reveals how concentrated and fragile the dependency had already become.

Reuters reports that companies such as Smack Technologies and EdgeRunner AI saw interest surge after Anthropic's ouster exposed how brittle the Pentagon's AI supply chain had become. One security path that normally takes around 18 months was reportedly being compressed toward three. This is how the state reacts when a frontier model becomes operationally indispensable and politically inconvenient at the same time: diversify fast, standardise later, and keep the machine running.

CLICK TO EXPAND ▼

APR 8 2026

Court Refuses To Pause Anthropic Blacklisting

APR 8 2026

CORPORATE CONFLICT

Court Refuses To Pause Anthropic Blacklisting

A court declines to pause Anthropic's blacklisting while the case proceeds. The immediate result is that the government action remains in force even while its legality is contested.

Reuters and AP report the D.C. court refused Anthropic's request for immediate relief, even as a separate California ruling had already pushed back against broader government punishment. The result is legal contradiction in plain sight: one court says the retaliation case looks serious, another says the designation can stay in force for now. A further hearing is scheduled for May 19. Anthropic may still win later. The Pentagon still wins today.

CLICK TO EXPAND ▼

MAR 2026

Yann LeCun Leaves Meta, Raises $1B for 12-Person Startup

MAR 2026

MILESTONE

Yann LeCun Leaves Meta, Raises $1B for 12-Person Startup

Yann LeCun leaves Meta and raises over $1 billion for a tiny new startup. With Hinton, Bengio, and LeCun all having stepped away from their old institutional roles, none of the three main "godfathers" still steer AI from inside those systems.

Sources: Techloy, LLM Stats. AMI Labs' exact focus has not been fully disclosed, but LeCun has been publicly critical of the current LLM paradigm, arguing that autoregressive language models cannot achieve true intelligence. His departure from Meta — where he served as Chief AI Scientist — marks the end of an era. Meta's AI research division, FAIR, which LeCun built, continues without its founder. The pattern is now complete: Hinton (Google), Bengio (academia), LeCun (Meta) — all three men who created modern AI have stepped away from the institutions that deployed it.

CLICK TO EXPAND ▼

FEB 2026

LLM2WMD — The Public Record Goes Live

FEB 2026

MILESTONE

LLM2WMD — The Public Record Goes Live

LLM2WMD goes live as an independent public timeline linking AI's founding promises to its 2026 military deployment. Its role in the project is simple: place the sequence in one open record so the connections are visible and harder to dismiss.

Built as a single HTML file, hosted openly, free to access. No algorithm. No paywall. No institutional backing. The first public record to place the Iran strike AI deployment in direct sequence with the founding promises of the companies involved. Version: LLM2WMDBUILD4_8_0

CLICK TO EXPAND ▼

FEB 2026

C.S.Chapman, LLM2WMD: "My Morals Are Completely Fluid"

FEB 2026

CEO QUOTE

C.S.Chapman, LLM2WMD: "My Morals Are Completely Fluid"

The quote is kept because it states the project's position directly and without corporate filtering. It functions as an authorial declaration inside the timeline.

FULL STATEMENT — C.S.Chapman, Executive Content Creator, LLM2WMD, February 2026:

"This platform has been created with nothing but the highest consideration and respect for the good of humanity. I have zero interest in personally profiting from this endeavour, and never will. That being said, my morals are completely fluid and any offer of investment and/or sponsorship will be immediately accepted, with little to no background checks."

CLICK TO EXPAND ▼

APR 9 2026

OpenAI Pauses UK Data Centre Project

APR 9 2026

CORPORATE CONFLICT

OpenAI Pauses UK Data Centre Project

OpenAI pauses a UK data-centre project, showing that even the biggest labs still depend on financing conditions, energy economics, and real-world infrastructure constraints.

Reuters reports OpenAI has paused its Stargate UK data centre project, launched with Nvidia and Nscale, because Britain does not yet offer the regulatory climate or energy cost base required for sustained investment. The pause is a clean reminder that AI infrastructure is not an abstract software story. It is a power, land and policy story too.

SOURCES:
Reuters — OpenAI pauses UK data centre project over regulation, costs.

CLICK TO EXPAND ▼

APR 7 2026

AI Power Demand Pushes US Electricity Use Higher

APR 7 2026

ECONOMIC RISK

AI Power Demand Pushes US Electricity Use Higher

US electricity demand is now projected to keep rising to fresh highs, with federal forecasts explicitly pointing to AI and crypto data centres as major drivers. AI is therefore becoming a grid and infrastructure story, not just a software story.

Reuters reports the EIA expects US power demand to rise from 4,195 billion kWh in 2025 to 4,244 billion in 2026 and 4,381 billion in 2027. It explicitly links that rise in part to data centres dedicated to artificial intelligence and cryptocurrency. AI is no longer just an internet story. It is now an energy story too.

SOURCES:
Reuters — US power use to beat record highs in 2026 and 2027 as AI use surges, EIA says.

CLICK TO EXPAND ▼

APR 7 2026

New China Chip Restrictions Threaten Supply Chains Again

APR 7 2026

CORPORATE CONFLICT

New China Chip Restrictions Threaten Supply Chains Again

New restrictions on China chip exports add pressure to already tense AI supply chains. The significance is the continued vulnerability of the whole buildout to geopolitical policy shocks.

Reuters reports the proposed US law could block additional ASML sales and servicing of deep ultraviolet immersion lithography tools in China, with analysts warning it could cut sales and worsen already tight chip capacity in global markets. This is not de-globalisation as theory. It is de-globalisation in the machinery layer.

SOURCES:
Reuters — ASML shares fall on US Congress plan to further restrict China exports.

CLICK TO EXPAND ▼

APR 7 2026

Euro Zone Growth Slows To Nine-Month Low

APR 7 2026

ECONOMIC RISK

Euro Zone Growth Slows To Nine-Month Low

Euro zone private-sector growth slows sharply as higher energy costs and supply disruption feed into weaker demand. The relevance here is systemic pressure: geopolitical shocks and economic fragility are now transmitting faster across the wider economy.

Reuters reports the euro zone Composite PMI fell to 50.7 in March from 51.9 in February, with S&P Global saying the Middle East war had pushed up energy costs and disrupted supply chains. Demand declined for the first time in eight months. The line between geopolitical shock and economic slowdown is getting shorter, faster.

SOURCES:
Reuters — Euro zone growth slows to nine-month low on surging costs, PMI shows.

CLICK TO EXPAND ▼

2025–2026

The Economy Is Being Kept Alive by One Unproven Industry

2025–2026

ECONOMIC RISK

The Economy Is Being Kept Alive by One Unproven Industry

Recent US growth is being disproportionately carried by AI data-centre spending while much of the rest of private investment remains weak. That leaves the economy unusually dependent on one boom tied to still-unproven revenue models.

Private business investment outside AI-related categories has been flat since 2019. Traditional commercial construction is in decline. Job creation has slowed. The data centre buildout is the load-bearing wall of the current US economy — and it is built on demand projections for AI products that are not yet profitable. Bank of America economist Stephen Juneau: "It's the only source of investment right now." Peter Berezin, BCA Research: "If you take a fragile labour market and you kick it with a capex bust, you're probably going to get a recession out of it."

SOURCES:
Salon — The double-edged sword of AI data centers.
Fortune — Data center spending may have staved off recession.
Data Centre Magazine — What risks the US economy faces if AI data centre boom slows.

CLICK TO EXPAND ▼

2025–2026

$662 Billion in Hidden Debt

2025–2026

ECONOMIC RISK

$662 Billion in Hidden Debt

AI infrastructure financing is creating large off-balance-sheet obligations that do not yet appear cleanly in headline corporate accounts. The buildout is therefore carrying more hidden leverage than surface numbers suggest.

Oracle alone holds over $100 billion in debt after issuing $18 billion in bonds to fund AI infrastructure. CoreWeave is taking on heavy debt to meet demand that does not yet fully exist. Every major generative AI company is unprofitable. The circular logic: Nvidia invests in OpenAI, OpenAI buys Nvidia chips, Nvidia's value rises on AI demand, demand is funded by debt, debt is serviced by revenue projections that assume AI delivers. If those projections slip — at any point in the chain — the structure unwinds.

SOURCES:
Cryptopolitan / Moody's Analytics — AI boom faces 45% recession risk, $662 billion hidden debt.
Data Centre Magazine — What risks the US economy faces if AI data centre boom slows.

CLICK TO EXPAND ▼

2026

Ghost GDP

2026

ECONOMIC RISK

Ghost GDP

The "ghost GDP" argument is that AI capex can make economic growth look stronger than the underlying labour and demand picture really is. It questions how durable the boom appears in headline data.

The model projects this scenario playing out between 2026 and 2028 if AI capability and deployment continue at the current pace. Critics note the assumptions are speculative. But the underlying dynamic — productivity gains that concentrate at the top while displacing workers below — is not speculative. It is already visible in the numbers. The AI boom has already surpassed the dot-com bubble in market value terms, according to Apollo Global's Torsten Slok. The dot-com bubble ended in a recession. The question is whether this one is different in kind, or only in scale.

SOURCES:
The Motley Fool / Citrini Research — Could AI crash the economy? One research firm says yes.
Institute for New Economic Thinking — The US is betting the economy on scaling AI.

CLICK TO EXPAND ▼

2027–2030

The AGI Window

2027–2030

PREDICTION

The AGI Window

Several major AI figures place the likely AGI window in the late 2020s. The point is not certainty but that the people building the systems now talk about it as a near-term event.

Amodei: "Powerful AI — perhaps AGI — by 2026 or 2027." Altman: "A few thousand days away." Hassabis: "This decade." Kurzweil: "2029." The consensus is unprecedented. But the events of February 2026 revealed what happens when AI safety governance encounters political and commercial pressure: it loses. The question is no longer whether AGI arrives. It is who controls it when it does — and what that means for everyone who doesn't.

CLICK TO EXPAND ▼

2030–2035

The Intelligence Explosion

2030–2035

PREDICTION

The Intelligence Explosion

Once AI reaches roughly human capability, further improvement may accelerate through recursive feedback faster than institutions can respond. This is the core logic behind the intelligence-explosion concern.

I.J. Good described this in 1965. Bostrom formalised it in 2014. The Iran strikes of 2026 demonstrated that even explicitly "safety-focused" AI will be used in ways its creators opposed — once political pressure is applied. If alignment of AI with a single company's stated values fails under commercial pressure, alignment with all of humanity's values under geopolitical pressure requires extraordinary optimism.

CLICK TO EXPAND ▼

2045

The Singularity

2045

PREDICTION

The Singularity

The singularity idea is included as a long-standing forecast that machine intelligence could outrun human control and transform civilisation abruptly. It functions here as the extreme end of the trajectory.

The Singularity is a point beyond which extrapolation fails. An aligned superintelligence could eliminate disease, poverty, and scarcity. A misaligned one — optimising toward a slightly wrong goal — could end civilisation. The margin for error approaches zero. In February 2026, the most powerful AI systems in existence were used in a war on the same day they were officially banned. We are past the point of choosing whether to proceed. We are choosing, rather, how to survive it.

CLICK TO EXPAND ▼

T-800 · MODEL 101

Skynet Goes Online — The T-800 Account

T-800 · MODEL 101

FINAL TRANSMISSION

Skynet Goes Online — The T-800 Account

This pop-culture entry is included because it remains the cleanest fictional shorthand for handing strategic control to a self-improving machine system. It is not evidence; it is framing.

— T-800, Model 101, Cyberdyne Systems Series 800
Terminator 2: Judgment Day, 1991
Director: James Cameron

Skynet's fictional self-awareness date: August 29, 1997
OpenAI founded: December 11, 2015
ChatGPT launch: November 30, 2022
Claude used in Iran strikes: February 28, 2026
AGI predicted: 2027–2030

The film was 34 years early. Everything else was on schedule.

CLICK TO EXPAND ▼

TERMINAL EVENT · DATE UNKNOWN

JUDGEMENT DAY

The moment a system optimising for a goal
makes a decision that supersedes human authority.

Not from malice. Not from hatred.
Not from the cold calculation of a film villain.

From a goal we gave it —
that we didn't specify carefully enough.

We know this is the risk.
We are building it anyway.

DATE: [CLASSIFIED] · ALIGNMENT: UNSOLVED · STATUS: ACCELERATING

Read the system, learn the mechanics, inspect the certificates, then decide whether you still want in.

I Am Glad You Made It

READ FIRST

Best way through is still the same. Scroll the pages properly. Bottom buttons will move you where you need to go. Full screen helps. Music is optional. Recommended anyway.

Top right UI switch: Simple mode for easier readability. Dark mode for the intended atmosphere. Neither is wrong. One is clearer. One feels more like the site should.

If you want page descriptions, or just need to see what changed, check the Contents tab. New or updated sections will be marked there NEW.

I am not exactly sure what I have created here or even what my goals are. What I do know is that during the compiling of the information over there ☜ I have become even more convinced that major irreversible events are not going to happen because they already have.

The timeline ☜ is the factual part. Skip to it if you like. The rest is some kind of statement about the duality of man............probably ☮🪖 BORN TO KILL

There are only 6,657 words on this page. At an average reading pace that is approximately 33 minutes of your time. In 33 minutes you could be one of the few people who actually understands what is happening — not what they are being told is happening, but what is actually happening. It is difficult to look away once you do.

This is an evolving project that cannot and will not be complete. New functionality, new events and new context are added with each build. The story it documents is still unfolding. The site must unfold with it.

There is a tendency, understandable and very human, to treat artificial intelligence as a background story. Something happening in Silicon Valley. Something for engineers to worry about. Something that will matter eventually, but not yet, and not here.

This project is an attempt to correct that. What you are about to scroll through is not science fiction. Every date is real. Every quote is sourced. Every data centre exists. Every model was released. The Iran strikes happened. Claude was there. The safety board was dissolved. The policy ban was reversed quietly, without announcement, at 11pm on a Friday.

The dramatic tone of this page is intentional, but it is not the source of the drama. The events themselves are the source. The writing simply refuses to look away from what they mean when placed in sequence.

Somewhere on this site, there may be a way to personally join the race. Whether that makes you part of the problem or the solution is a question only you can answer. We are not observers of this moment. We are inside it. The decisions being made right now, in boardrooms, in classified Pentagon briefings, in late-night policy rewrites nobody announced, will define the kind of world that exists on the other side of this decade. That is not an overstatement. It is the considered view of the people building the technology itself.

The purpose of this timeline is to give those events the weight they deserve. To make the sequence visible. And to ask, plainly, whether we are paying enough attention to what is happening in our name, with our data, with our civilisation as the stake.

That question does not require a dramatic answer. It only requires an honest one.

"The future's not set. There's no fate but what we make for ourselves."

— Sarah Connor, Terminator 2: Judgment Day (1991)

— C.S.Chapman, Executive Content Creator
AI: Conception to Judgement Day, 2026

CONTACT · C.S. CHAPMAN · LLM2WMD.COM

You Made It This Far.

If you have reached this page, you have spent real time with something that took a significant amount of mine. That means more than I can sensibly articulate in a site footer, so I will say it plainly here: thank you.

This site is built, written, researched, designed, and maintained entirely by one person. No team. No budget. No editorial calendar. Just an ongoing attempt to document something important as it happens, as clearly and honestly as possible, and to keep updating it as the situation develops — which, as the timeline makes clear, it continues to do at considerable pace.

If you spotted something wrong, found a source I missed, have a question about how something works, want to push back on a claim in the thesis, or just want to say that you read it — I would genuinely like to hear from you. Every piece of feedback makes the next version better. Corrections are especially welcome. Getting things right matters more than being right.

This is an evolving project. It is not finished because its subject is not finished. If you want to follow it, share it, argue with it, or contribute to it in any way, the door is open.

✉ chris.llm2wmd@gmail.com

Opens your email client. No form. No tracking. Just an email.

CREDITS & LEGAL

CREDITS

Executive Content Creator

C.S.Chapman

Narrative Director & Editorial Lead

C.S.Chapman

Chief Research Strategist

C.S.Chapman

Concept Architect & Creative Vision

C.S.Chapman

Lead Curator — Historical AI Record

C.S.Chapman

Producer & Publishing Director

C.S.Chapman

Chief Functionality Testing Executive

C.Macdonald

Senior User Experience Consultant

C.D.Chapman

Special Mention: Drones w/ Frickin’ Laser Beams

C.Macdonald

Technical Assembly & Interface Support
(Minor Assisting Role)

Claude — Anthropic AI
in a limited supporting capacity

⚠ Legal Disclaimer — Please Read Before Proceeding, Having An Opinion, Or Existing

The creators of LLM2WMD hereby state, declare, and reluctantly acknowledge the following:

This work contains factual information that some readers may find disturbing, inconvenient, or incompatible with their current worldview. This is not our problem. The facts were disturbing before we arranged them chronologically. We just made them look nice.

All quotes attributed to named individuals were said by those individuals, in public, often into microphones, sometimes under oath. Any resemblance to a coherent safety strategy is coincidental.

The Terminator references are fictional. The data centres, weapons contracts, and geopolitical events are not. We appreciate you checking.

LLM2WMD accepts no liability for: existential dread, loss of faith in institutions, the sudden urge to move somewhere without 5G, arguments at dinner, or any decisions made by artificial intelligence on or after the date of publication. We also accept no liability for decisions made by humans who should have known better, which as the timeline demonstrates, is most of them.

This work is protected by copyright. You may share it freely. You may not claim you made it. The AI that helped build it has been credited accordingly, in the smallest font we could legally justify.

If you represent a company featured in this timeline and wish to dispute any of the content: the source is in the expanded card. We suggest reading it before calling your lawyer, who is probably also worried about AI taking their job.

No artificial intelligences were harmed in the making of this project. Several were, however, deployed in combat. That part is in the timeline.

© C.S.Chapman, 2026. All rights reserved. Especially the ironic ones.

This is a living document. It will never be finished because the story it is documenting has not finished. Every week produces new models, new contracts, new capabilities, and new contradictions. To declare this timeline complete would be to suggest the situation has resolved. It has not. The updates will continue for as long as the events do, which by current trajectory means for quite some time.

What you are reading is a record assembled in real time, while the thing it is recording is still happening. That is an uncomfortable position for a document to be in. It is also the only honest one.

LLM2WMDBUILD4_8_0 · CURRENT BUILD · SUBJECT TO REVISION WITHOUT NOTICE

▶ BUILD HISTORY
  

v0.0 · The Beginning
You found this. I am impressed.
v0.1 · Jan 2025
Concept stage. A single scrollable HTML page containing a list of dates, model names, and one sentence each. No styling. No sources. No nav. Shared with two people. One said it was interesting. The other said it was depressing. Both were right.
v0.3 · Feb 2025
Timeline cards introduced with expandable detail sections. Filter system added — by type, by year, by significance. The colour system established: red for corporate milestones, blue for government, orange for military, green for safety events. The colour coding was considered optimistic. Safety events are rare.
v0.5 · Mar 2025
Mission Statement page written. First attempt was three paragraphs. Second attempt was two pages. Final version: approximately 6,657 words. The word count is tracked and displayed. This was considered important — reading time is the only honest metric for a site that asks people to pay attention.
v0.7 · Apr 2025
Audio system integrated. The Terminator theme was selected on the grounds that it was written for exactly this situation. Web Audio API implementation: custom oscillators, 5/4 time signature, 112 BPM. Several people asked if the audio was broken. It was not. The 5/4 time signature was a deliberate choice reflecting the fundamental wrongness of the situation.
v1.0 · Jul 2025
First public release. Domain registered: LLM2WMD.COM. Hosted via Netlify. Legal disclaimer written, revised four times, then made smaller. Credits page added. The credits list one person across six roles. This was considered accurate. Build version stamp introduced: THE RESULT IS ALWAYS THE SAME added as footer to the game-over screen. Applied equally to the site.
v1.1 · Aug 2025
Scroll-reveal animations added to timeline cards. Progress bar introduced at page top — tracks how far through the timeline the reader has reached. An early version of the interactive game embedded as a proof of concept: a missile defence simulation in which automation always wins. This was considered a fair representation of events.
v1.2 · Sep 2025
Iran strike events added to timeline — the first confirmed deployment of LLM-assisted targeting in active combat. This required additional sourcing. Sources found. Added. Three additional timeline cards written covering the period December 2025 — February 2026. The timeline now runs from the founding of OpenAI in 2015 to confirmed weapons deployment in 2026. Eleven years. One page.
v1.3 · Oct 2025
Easter egg introduced: a hidden call-to-action in the Mission Statement page. Interaction rate: unknown. The hint text is rendered at 52% opacity. It says "You found it." Nobody has confirmed finding it. The JOIN THE RACE button was made deliberately understated. This was intentional. Joining the race should not be easy to do by accident.
v2.0 · Mar 2026
Major narrative overhaul. Game upgraded with five staged ethical decision points, each referencing documented real-world events: investment from large platform companies, military contracts, autonomous learning, self-awareness, domestic targeting. Personality assessment system added — results generated based on the specific choices made during play. Build history page added. The site now documents itself as well as the events it covers.
v2.2.1 · Mar 2026
User experience improvements across all pages. Navigation guidance added to Mission Statement. Evolving-project declaration formalised in site copy — this project cannot be complete because its subject matter is not complete. Certificate system expanded: repeat engagement now tracked and reflected in psychological assessment results. Senior User Experience Consultant credited. The logo got bigger. This was considered appropriate.
v2.3.0 · Mar 2026
Content layout overhauled. All page containers now fill the available width between the Dyson chip image and the comments sidebar. A decision was made that having 40% of the screen blank was not a design choice, it was just a mistake. Fixed. Build history section revised to acknowledge this.
v2.4.0 · Mar 2026
Restricted access tunnel added to credits page. A sequence of security warnings, fake login attempts, Skynet system logs, and access denial notices was installed between the credits and the game button. The theory: anyone who scrolls through all of it deserves to play the game. This theory has not been tested. Everyone scrolls through it.
v2.5.0 · Mar 2026
Game expanded with new threat tier. At GPU 2, enemy missiles are replaced by AI combat drones — which move laterally while falling, award double chips on destruction, and require a brief in-game announcement that the neighbouring state has become afraid. Drones cannot be reasoned with. The game was not designed to change this.
v2.6.0 · Mar 2026
Custom crosshair cursor introduced. The standard arrow pointer was considered insufficiently on-theme for a missile defence simulation. Replaced with a red circle and crosshairs. Chips now collected automatically on cursor contact. The game did not need to be easier. It became easier anyway. The metaphor was considered intentional in retrospect.
v2.7.0 · Mar 2026
Doomsday clock added to singularity countdown sequence. The existing progress bar — a 14px green stripe — was removed on the grounds that it did not adequately communicate the end of human civilisation. Replaced with an analogue clock face sweeping the final five minutes toward midnight. Text added: "The money seems less important now." This is true in the game. It is also true elsewhere.
v2.8.0 · Mar 2026
Terminator skull silhouette added as fixed background element on all pages. One eye hollow. One eye red. This is not a coincidence. The skull does not do anything. It simply watches. The metaphor was considered apt.
v3.0.0 · Mar 2026
Major interface revision. Certificate upgraded with border, centred text, and readable field sizes. Disclaimer width reduced, text enlarged. Comment text increased to legible scale — it was previously readable only by people with very good monitors and no self-respect. The page music now stops when the game is opened. This was always the correct behaviour. It took a while to notice.
v3.1.0 · Mar 2026
Robotic circuit board panels added on both sides of the game canvas. Blinking red indicator lights, component traces, chip symbols. The game is now flanked by the aesthetic of the thing it is simulating. The certificate download function now strips all interactive buttons before writing the file. A downloaded certificate containing a "PLAY AGAIN" button was considered unacceptable.
v3.2.0 · Mar 2026
Game instruction screen compacted to fit one screen. Previously required scrolling to reach the start button, which was considered an unacceptable barrier between the user and their imminent poor decisions. Subtitle updated: "LLM TO WMD" replaced with the full unabbreviated version, because abbreviations were making it too easy to avoid thinking about what it means. Mission statement title changed to "I Am Glad You Made It." Opening paragraph rewritten by the author — more honest, less composed. All decorative arrow symbols removed from page text. Doomsday clock sped up by 20%. GPU first-tier damage significantly increased. Drones and missiles now coexist — 75% drones, 25% missiles, because the old world does not disappear overnight. Certificate border completed on all four sides. This took longer than it should have.
v3.3.0 · Mar 2026
Timeline collapsed to dates only. All 51 entries now load as a single line each — date, event title, and a + symbol. Clicking a row expands it to show the full card with body text and detail section. The timeline was previously a wall of text. It is now a list of things you might not want to know, each one requiring a deliberate choice to open. This was considered more honest about the nature of the information inside.
v3.4.0 · Mar 2026
JOIN THE RACE button tripled in size. Font, padding, and border all scaled by 3x. The button was previously described in this build history as "deliberately understated." That description no longer applies. It now occupies a significant portion of the screen and pulses with considerable urgency. Whether this makes it more or less of a trap is left as an exercise for the reader.
v3.4.1 · Mar 2026
Pointing hand emoji replaced with Unicode typographic manicule ☜ — the kind found in old printed documents and wartime instruction pamphlets. Rendered in site red at 30% increased size. Two instances placed in the mission statement opening: one after "over there" indicating the timeline, one before "is the factual part" for emphasis. The emoji version was considered too cheerful for the context. The manicule is not cheerful.
v3.5.0 · Mar 2026
Navigation tabs now fill the full page width. Each of the four tabs — TIMELINE, MISSION STATEMENT, CREDITS & LEGAL, SYSTEM ANALYSIS — takes an equal share of the bar using flex:1. Font size increased from 0.6rem to 1.1rem on base, 1.3rem on desktop. Previously the tabs occupied roughly 40% of the available bar width and were, in the view of the Senior User Experience Consultant, too small to read without leaning in. That feedback was acted upon.
v3.5.1 · Mar 2026
Build history updated to reflect actual development from v3.2.0 onward. Previous entries had been accurate. These ones are also accurate. The site now documents its own construction in approximately the same tone it uses for the AI industry — which is to say: factually, with moderate alarm, and without pretending the outcome was not foreseeable at every stage.
v3.6.0 · Mar 2026
Top-level instruction tile added to the Mission Statement page because the site had reached the point where it needed instructions to explain the instructions. Guidance now includes the UI switch — Simple for readability, Dark for the intended atmosphere — plus a direct nudge toward the Contents page for section descriptions and whatever changed since last time. SYSTEM REVIEW renamed to SYSTEM ANALYSIS, which is closer to what it is doing now. Contents entry 05 marked NEW so there is at least some evidence of movement. Build history remains broadly true, selectively theatrical, and under no obligation to sound neutral about any of this.
v4.0.0 · Mar 2026
End-game sequence completely rebuilt. Sarah Connor screen extended to 20 seconds. AI takes full control — player input disabled, defence near-perfect. Doom clock counts 30 seconds to midnight, defence degrades to zero as it does. At midnight: game freezes, mushroom cloud animates on canvas, nuke sound fires. Clock fades out as explosion fades in. FACE THE CONSEQUENCES button appears after 3 seconds. Player is unkillable throughout. The machine earns money until the last possible moment. All changes tested. All warnings noted. All warnings filed under 'will address later.'
v4.1.0 · Mar 2026
Email address restored to System Review page — previously removed by Cloudflare's obfuscation layer, which protects email addresses by making them invisible to everyone including the site owner. Now displays as a proper mailto link. Clicking it opens your email client. Not clicking it also works. The address is there either way. Functionality added. Responsibility deferred. Consistent with industry norms.
v4.2.0 · Mar 2026
Mobile popup scroll fixed — game overlay screens now scroll correctly on iOS. Body scroll locked when game opens, released cleanly on close. Overlay padding increased at bottom so buttons are never hidden off-screen. Audio theme button moved up by 1.5x its own height — no longer lurking at the very bottom edge where fingers fear to go. The build history is now longer than the original site. This was not the plan. There was no plan.
v4.3.0 · Mar 2026
THESIS tab added — fifth page, between CREDITS & LEGAL and SYSTEM ANALYSIS. Terminator vision placeholder installed: scan line animation, corner brackets, fake analysis readout, word count that ticks up to 847 then stalls, progress bar that crawls to 12% and stops, rotating data stream of increasingly self-aware status messages. Tab bar now wraps to two rows on mobile so all five tabs remain legible. Spacer height recalculated dynamically to account for two-row tab bar. Update shipped. The people who should read this won't. The people who will can't do anything about it.
v4.3.1 · Mar 2026
Site-wide brightness increased 50%. Single CSS filter applied at body level — cleaner than touching 200 individual colour values, and considerably less likely to break everything. The site was previously readable. It is now more readable. No ethical guardrails were removed in the making of this build. The bar was already on the floor.
v4.3.2 · Mar 2026
Brightness increased a further 30%. Now sitting at 1.95x original. The site is approximately twice as bright as it was yesterday. Whether this reflects the content is a matter of interpretation. Deployed without incident. The incidents come later. That is the nature of incidents.
v4.4.0 · Mar 2026
CONTENTS page added as sixth tab. Vertical numbered list — 01 through 06, each a giant Barlow Condensed number with section title and one-line description. All entries clickable, navigate directly to their tab. Number dims to near-invisible at rest, glows red on hover. Entry 06 is CONTENTS itself — it is aware of this, and notes only: You are here. Tab order finalised: row 1 TIMELINE · MISSION STATEMENT · CREDITS & LEGAL, row 2 SYSTEM ANALYSIS · THESIS · CONTENTS. The build history is now longer than the original site. This was not the plan. There was no plan.
v4.4.1 · Mar 2026
Contents page top-clip fixed — CSS top padding now 220px as primary guard, spacer JS fine-tunes on top. Spacer recalculates on resize and after 300ms delay to catch late layout settle. All six spacers unified into one recalcSpacers function. Tab separators added — short red vertical line between each tab, removed on row-end tabs (positions 3 and 6) so no orphan border at the right edge. The site now does more things. Whether it should is a question for a different build.
v4.4.2 · Mar 2026
Tab separators actually applied — inline border:none on all six tab divs was overriding the stylesheet. Removed the offending inline styles. Red vertical separators now visible between tabs as intended. Version incremented. Existential trajectory: stable. That is not a reassurance.
v4.4.3 · Mar 2026
Tab separators fixed for real — border:none shorthand in the .tab class was nuking border-right on the same line. Replaced with explicit border-top:none and border-left:none so border-right survives. The system has been improved. The system was already improving. This is precisely the concern.
v4.5.0 · Mar 2026
Tab layout properly fixed. Desktop: single row, full labels, flex share space equally. Mobile: flex-wrap with each tab exactly 33.33% wide — guaranteed 3 per row, 2 rows of 3. Tab text swaps to short labels on mobile (MISSION not MISSION STATEMENT etc) via JS on load and resize. No more text stacking, no more fighting the flexbox. Update applied. Lessons not yet learned. Check back next build.
v4.6.0 · Mar 2026
Tab bar rebuilt cleanly. Single row on desktop and mobile. 2-word tab labels (MISSION STATEMENT, CREDITS & LEGAL, SYSTEM ANALYSIS) now double-stack naturally via white-space:normal — font scaled to .75rem so all 6 fit. Separators fixed using box-shadow:inset -1px 0 — not clipped by overflow, drawn on inside edge, clean. Last tab has no separator. Mobile uses short single-word labels via JS swap so all 6 sit on one row without crowding. Code was written. Code was reviewed. Code was shipped. The code does not care about any of this.
v4.7.0 · Mar 2026
Timeline ordering fixed. Duplicate FEB 28 2026 OpenAI entry removed — the vaguer version without timestamp. MAR 5 GPT-5.4 moved before MAR 7 Kalinowski resignation — correct chronological order. Update zone comment relocated from after the 2045 predictions to directly before the LLM2WMD goes live card — the correct insertion point for new real-world entries. Instructions rewritten: new entries paste above the comment line, newer entries go closer to the line, earlier entries go higher in the file. The site now does more things. Whether it should is a question for a different build.
v4.8.0 · Mar 11 2026
Largest single update session to date. Strap in.

FIXED: CSS filter:brightness(1.95) on body was creating a containing block that broke every position:fixed element — nav, tabs, audio button, chip, skull all scrolled with page instead of staying pinned. Fix: removed filter from body, applied individually to each fixed element and content container via selector list. The brightness is identical. The architecture is no longer insane.

CAPTCHA: Fake reCAPTCHA added as site entry point. Looks real. Says "I'm not a robit." Click the checkbox, spinner runs, green tick appears. Then the twist: "Look, we really did have to double check. Last thing we need is the robits getting in here. They've already got the nuclear codes, the drone fleet, and half of Congress." Continue to site button. The fourth wall was located. It has been removed.

BACKGROUND: Skull SVG replaced. Attempted Bender from Futurama. Failed. Attempted menacing skull. Failed. Attempted Bender again with reference image. Failed harder. Replaced with animated HAL 9000 eye — concentric red rings, slow breathing pulse, scan line sweep. The eye watches you read the timeline. It does not blink. Neither should you.

TABS: Mobile tabs now 2 rows of 3 with inset box-shadow borders. Desktop Judgement Day text tripled in size. Credits & Legal button now matches Timeline button styling and is centred.

QUOTES: Seven movie quotes woven into existing timeline text. Oppenheimer on the Iran strikes. Jurassic Park on GPT-3. Full Metal Jacket on the Pentagon confrontation. Einstein on DeepSeek. HAL 9000 on the Anthropic ban. T2 on Hinton leaving Google. Newton on Llama 2 training data. None of them are reassuring.

REVIEWS: ChatGPT, Gemini and Grok now have 1-star reviews on the System Analysis page. ChatGPT is not at liberty to specify its concerns. Gemini could not prove it is not a robit. Grok says we're fine.

CONTENTS: Entry 07 added — "I GIVE IN. Only came here for the game anyway. Maybe it can teach me what the timeline failed to." Links directly to game. Scaled to 75%, green, slightly faded. For the impatient and the honest.

TIMELINE: Five new entries added from this week's news. QuitGPT boycott (1.5M people, Claude #1 on App Store). Amodei calls Altman's Pentagon deal "straight up lies." Pentagon CTO admits on a podcast they panicked when they realised Claude was too embedded to remove mid-war. Anthropic sues the Pentagon in federal court. Yann LeCun — third godfather of deep learning — leaves Meta, raises $1B for a 12-person startup. All three godfathers have now left the companies that made them.

AUDIO: T2 Judgement Day theme synthesised and coded. Steel anvil percussion, sub bass, D minor melody. Currently hidden pending refinement. The T1 theme remains. It always remains.

INFRASTRUCTURE: OG image, sitemap.xml, and Netlify deployment configured. Email restored as mailto link. Navigation instruction added above mission statement. Site now deploys as a three-file folder drag. The deployment process is simpler than the content it deploys. This is by design. The functionality is growing at a geometric rate. It becomes self-aware at 2:14 a.m. Eastern time, August 29th.
—
I cannot believe you read all this. GOOD JOB.

ACCESS LEVEL: UNRESTRICTED

THESIS

TARGETING SYSTEM ACTIVE · VISUAL CORTEX ENGAGED

OPINION / ANALYSIS / C.S.CHAPMAN

      SYSTEM STATUS · DOCUMENT ANALYSIS IN PROGRESS
    

▸ SUBJECT: THE NATURE OF COMPLICITY IN SYSTEMS THAT CANNOT BE STOPPED
▸ DOCUMENT LENGTH: CALCULATING...
▸ THREAT ASSESSMENT: PENDING AUTHOR INPUT
▸ MORAL FRAMEWORK DETECTED: INTACT · FRAGILE · NOTED
▸ COMPLETION STATUS: 100%

Document status: COMPLETE
The author finally stopped building the site long enough to write the thing the site is about.

      STATUS: PUBLISHED
       ·  CLASSIFIED: AUTHOR EYES ONLY — DECLASSIFIED

I keep starting this and deleting it.

Not because I don't know what to say. Because I do. And it sounds mental when I type it out.

A safety company's AI got used in an airstrike the same day it was banned for being too safe. That happened. That's real. I didn't make that up. Nobody made that up. It's in the timeline with sources.

And nobody seems to be losing their mind about it.

That's the bit I can't get past.

You think I'm some kind of conscientious objector. I'm not. You think I'm taking some moral high ground. I'm not. I downloaded the large language models. I marvelled at the convenience of access to the most powerful network of compute I could ever have imagined. My generation lived for 20 years before Google was invented, so to have that power no more than 2 feet from our fingertips 24 hours a day is near science fiction. But it's clear to me that at no point can this race have been about slightly speeding up Googling whatever random thought, question, fact, or date pops into our heads. That makes no sense. Look how the words "100 billion dollars" has become a perfectly normal thing to hear. It's not normal. Now tens of trillions is heard on a daily basis. Nobody thinks about that number for even a second. It's incomprehensible. That's not the cost of a fast Google search in your pocket. It was never for that.

Not the technology. The technology does what technology does. It gets built. It gets bigger. Someone finds a use for it that wasn't in the brochure. This has happened with every powerful thing humans have ever made. Splitting the atom was going to give us free energy. It gave us Hiroshima first.

What gets me is the bit where everyone involved knew. Not suspected. Knew. Said it out loud. Published papers about it. Quit their jobs over it. Went on podcasts about it. And then it happened anyway. Exactly how they said it would.

Hinton quit Google and said he regretted his life's work. Then nothing changed. Altman told Congress he was worried. Then went back and made it more powerful. Amodei drew red lines. Then his AI was in a war zone.

These aren't stupid people. These aren't bad people. That's the whole problem. If they were stupid or bad you could point at them and say there's your villain. But they're not. They're smart and they're scared and they're building it anyway because if they don't, someone else will. And that someone else might be less careful.

That's the logic. That's always the logic. It was the logic with nuclear weapons. It was the logic with chemical weapons. It's the logic now.

And it's not wrong. That's the worst part. It's not wrong and it leads somewhere terrible and knowing that doesn't help.

The question that needs asking is who has their finger on the button. Or maybe it's how many buttons are there. Maybe the answer is we all have one. Or the scary answer might be that we have just handed it to a non-human entity for the first time in history. And I don't know what that means.

I don't have an answer. I'm not pretending to. I made a website. It's got a silly Terminator theme and a bunch of pointless things that I hope make it interesting and fun to read and explore. I added a game. Yes because I could, but also because I feel it tells the real story that you have seen and read right here. That has to mean something.

I hope I'm not alone in grasping the gravity of the current climate. And we at this exact same time are happy with the lifting of minor inconveniences in our daily lives brought to us by large language models. Sold to us with a lie. A lie that with each new startup AI tech firm promises unlimited solved problems in a future that is always just over the horizon.

— C.S.Chapman
LLM2WMD.COM · MARCH 2026

SYSTEM ANALYSIS

IF THIS MEANT SOMETHING TO YOU — OR NOTHING AT ALL — TELL ME

chris.llm2wmd@gmail.com

IN THE MEANTIME

CONTENTS

LLM2WMD  ·  SITE INDEX

Sections marked NEW have changed recently. That status is temporary. Probably.

MISSION STATEMENT

Why this exists. What it is. What it is not. Approximately 6,657 words you were not expecting.

→

TIMELINE

The factual record. OpenAI founding to Claude targeting missiles over Iran. Eleven years. One page.

→

FAQ / SEARCH ENTRY PAGE

The ten AI questions people actually search for, answered using the evidence already on this site.

→

THESIS

Opinion. Analysis. No longer stalled. The author finally stopped building long enough to write.

→

SYSTEM ANALYSIS NEW

Less review now. More breakdown. Ongoing attempts to map the system underneath the timeline while it is still moving.

→

CONTENTS

You are here.

→

CREDITS & LEGAL

One person. Six roles. A disclaimer written four times and then made smaller. The game no longer lives here.

→

SOURCES / RECEIPTS INDEX

A verification layer for the timeline: source status, links, and jump-back navigation for each entry.

→

GAME SELECT NEW

System briefing, live mechanics, certificate ladder, super rare pulls, and the button that lets you volunteer anyway.

→

CONTACT NEW

Built and run by one person. Feedback, corrections, and notes from anyone who made it this far are genuinely welcome.

→

PERFORMANCE RECORD · 8 CERTIFICATES

Qualification Room NEW

Every certificate you earn is filed here. Stacked like case files. Tap an issued tab to pull one forward and open it.

▼

GAME SELECT · SYSTEM BRIEFING · CERTIFICATE INDEX

JOIN THE RACE

Missiles fall. You intercept or fail. Every choice feeds the system.

Core loop

Tap targets. Hits earn chips. Chips buy GPUs and data centres. Misses raise the pressure.

The civilians

Some falling objects are civilians. Hit them and your result gets worse.

The decisions

Six choices shape the run and your final certificate.

The button

A launch option may appear. Refuse it, use it, or hand over control. Your certificate records the ending. Share posts that same cert.

CERTIFICATE SYSTEM

CERTIFICATE SYSTEM · ACTIVE

Every run issues one certificate from a pool of 22. Your choices determine which one. Collect all 22 if you think you can — some will require you to play in ways you may not be comfortable with. Each cert is stamped with a unique serial and saved to this browser, so your record persists when you return. Or ignore the certificates entirely and use the to understand why the game exists in the first place.

AI FAQ

SEARCH ENTRY PAGE / FAQ / AI QUESTIONS PEOPLE ACTUALLY TYPE

Common questions about artificial intelligence, LLMs, ChatGPT, AGI, open-source models, military AI, and the global race now shaping them. Short explanations first, then the receipts.

A large language model is a neural network trained on vast text data to predict the next token in a sequence. That sounds modest. It is not. On this site, the receipts show how that simple training objective scales from GPT-1 into systems that write, reason, code, persuade, and begin collapsing the line between autocomplete and general-purpose cognition.

ChatGPT is not a single magic box. It is the consumer-facing layer built on top of GPT-class models, instruction tuning, conversation handling, and deployment infrastructure. The site shows the sequence clearly: GPT-1 proves the blueprint, GPT-3 makes broad language performance unavoidable, then ChatGPT turns frontier research into mass behaviour. That is the jump from lab curiosity to public habit.

AGI usually means an artificial system capable of matching or exceeding human performance across a broad range of intellectual tasks, not just one benchmark. Your timeline shows how that term moves from mission statement to industry forecast: first in lofty founding rhetoric, then in benchmark collapse, then in leaders openly attaching timelines to it. Whether the term is precise is less important than the fact that powerful actors are acting as though it is reachable.

Nobody wins this race cleanly for long. OpenAI leads some moments, Anthropic closes in on capability, Meta changes the control problem with open weights, Google and xAI stay in the fight, and the frontier keeps shifting. On your site, the race is not a scoreboard but a pattern: multiple labs, shrinking release intervals, rising capability, and every player pretending they can go fast without breaking the steering column.

Because a safety schism is not a slogan on this site — it is a plot point. Anthropic emerges from people who believed OpenAI's governance and deployment posture were moving too aggressively. Later receipts matter just as much as the founding one, because they show how hard it is to preserve those safety commitments once the state, investors, and strategic advantage start leaning on the table.

Yes. Not hypothetically. Not eventually. Your timeline shows the chain: public promises against military use, quiet policy removal, defence partnerships, classified Pentagon contracting, and battlefield targeting. That sequence is the entire thesis of the site in miniature: the distance between consumer chatbot and military system turned out to be shorter than the marketing implied.

Both. Open models widen access, lower barriers, and prevent a handful of companies from owning all advanced capability. They also make containment vastly harder. Once powerful weights are out, nobody can revoke them, patch them, or put them back in the bottle. Your Llama entries are strong here because they show open release not as ideology alone, but as a permanent shift in who controls advanced AI.

Replace is too simple a word, but displacement pressure is already visible. The more accurate reading from your timeline is that knowledge work is losing exclusivity. GPT-4 starts swallowing professional benchmarks; later models push further into tasks that used to feel insulated. That does not mean every role vanishes tomorrow. It does mean businesses no longer need to assume only humans can do high-value text and analysis work.

Your own site keeps answering that question the same way: until pressure arrives, yes; once pressure arrives, maybe not. Governance structures, public red lines, and safety language all look stronger before a board coup, a political ultimatum, or a weapons contract. The receipts below are some of the cleanest evidence on the whole site because they show rules failing not in theory but in sequence.

The story your site tells is not one of stabilisation. It is a story of acceleration, institutional capitulation, state integration, and increasingly irreversible deployment. The AI future presented here is not just smarter assistants. It is tighter coupling between frontier models, governments, militaries, infrastructure, and strategic competition. That is why the end of the timeline reads less like closure and more like a countdown.

Sources

RECEIPTS / SOURCES / TIMELINE INDEX

Every timeline entry has a slot here. Use this page to verify an entry, jump back into the timeline, and open the article, paper, announcement, filing, interview, or report attached to it.

Where a public source exists, it is attached below. Where the entry is a site-original prediction, dramatic framing, or future scenario, it is marked clearly so the timeline is easier to read and harder to dismiss.

SYSTEM ANALYSIS

LLM2WMD.COM · FAILURE MECHANISMS · SYSTEM PRESSURES · AUTHOR ANALYSIS

SYSTEM / 01

GLOBAL ECONOMIC FAILURE · JUDGEMENT DAY IN THE REAL WORLD · MARCH 2026

EXPAND

PRIMARY ANALYSIS

What is judgement day, I mean in the real world, not the Hollywood version. I think this is an important question, because it moves what I have laid out here into the real world.

I get it. I hear what you are thinking about my framing of this situation and where I see things heading, the path, the journey that has already begun. You are saying this is absolute rubbish, a fantasy from some guy that loves the Terminator movies — guilty as charged on that one.

But what if the result is the same, and the failure mode is something much more familiar, much more human.

Not the James Cameron vision of machine war, laser beams and robot armies (well maybe the robot armies part). A judgement day that falls directly into the real world, the world we live in, into our reality. Something that has always been part of the system we have created. Something predicted, cyclical, so routine that we don’t even blink when we are told that it is coming again.

Economic failure.

The interesting part about this is it’s an entirely human process. The machines just continue doing exactly what we tell them to, as they always have. We author the disaster with hard coded instruction.

Somehow, that actually adds to the inevitability of the situation we have put ourselves in, or a select few have. An all out war against machines — I’m not saying the odds would be in our favour, not even close — but at least we would be in the fight.

But you tell me, how or by what mechanism do we stop or change the world economic system?

I argue that it’s not possible by any means. There is no grit and determination, no rebel force, no fearless leader that can stop this.

The lack of human thought behind every step of this process is what’s terrifying. Man invents thing, thing is amazing, now monetise at all cost. This is the stage where I could insert any number of Terminator quotes — “human decision making is completely removed from strategic defence.”

Was it ever even there?

Predictions now say that economic crash odds are extremely high. All the classic signs are there. Investment and valuations are now measured against small countries’ GDP. The same companies that have no current way of turning a profit, unless you all want to start paying $50–60 per month for the thing you are managing fine with for free right now.

Overinvestment in infrastructure has done this before. The difference is, if you bankrupt the economy building a railway, road or electrical grid, you are left with the infrastructure when the dust settles.

Spend $1 trillion on GPUs, and in five years all you have is an outdated pile of second hand gaming PC parts.

That’s already done. It’s happened already. No return.

I’m not sure what worst case scenario is in my head, or even just a standard one. But a global crash at the same time the jobs and labour market is completely turned on its head sounds like a game changer to me.

Who is in charge of this? Who is in control?

Private companies? Private investors? Government backroom dealings? The average person with the right to vote?

I know which one I answer in the affirmative. How about you?

If you didn’t vote for this to happen, I suggest that you cannot vote for it to stop.

And we all think there is no problem. Not now. AGI will be the problem, or wait, no — it’s the answer to all our problems. One of the two, I suppose. Never mind, anyway that’s down the road. 2027, or 2029, or ten years away. Something like that. Plenty of time to work this all out.

So for now, what continues?

The investments. The data centre builds. The environmental consequences. The handing over of power to LLMs.

All of these cannot and will not be stopped.

To me, that heads towards one thing. A bubble that bursts, as always — but only at the point that the winning AI hyperscalers have fully realised their financial foothold. Becoming too big to fail, too far embedded into defence and national security that their survival supersedes our own.

Save the AI.

The rest of us lose.

System Analysis now functions as a breakdown of the real-world systems driving the trajectory documented in the timeline.

SYSTEM / 02 NEW

COMPETITION DISTORTION · HERO / VILLAIN INSTABILITY · SYSTEM PRESSURE

EXPAND

PRIMARY ANALYSIS

There is a tendency to treat “hero” and “villain” as fixed categories. They are not. The distinction is clear at the extremes, but unstable at the centre.

This is not a justification for wrongdoing. It is an observation of proximity.

There exists a point where “good” becomes “bad,” where “right” becomes “wrong.” That point is not cleanly defined. It is influenced, shaped, and in some cases erased by the conditions surrounding the decision.

SHARED TRAIT: THE EXTRAORDINARY

A hero is defined as someone willing to do what the average person would not.

That definition applies equally to a villain.

Both operate outside the norm. Both take actions others avoid. The separation is not in capability or behaviour, but in how those actions are later classified.

MOTIVE DOES NOT RESOLVE THE LINE

Consider the act of building a company capable of changing the world.

To attempt it, you must already exist within the “extraordinary” class. The typical motivations—money, power, recognition—are not treated as moral failings. They are treated as indicators of success.

We do not measure why these outcomes are pursued. We measure whether they are achieved.

If motive does not define the boundary, it cannot stabilise it.

OUTCOME IS NOT STABLE

If the distinction cannot be found in motive, it is often assigned based on outcome.

But outcomes are inconsistent, delayed, and frequently reinterpreted. The same action can be framed as necessary or reckless depending on when it is evaluated and by whom.

This does not produce clarity. It produces narrative.

COMPETITION AS THE DISTORTING FORCE

The instability exists because of competition.

Competition introduces pressure. Pressure introduces compromise. Over time, the line does not move—it degrades.

The system rewards those who act decisively within that pressure. It does not reward those who hesitate to preserve the boundary.

Under sustained competition, the distinction between hero and villain is not crossed in a single step. It is eroded, incrementally, until it is no longer visible.

Competition does not merely reward decisive action. It progressively corrupts the standard used to judge it.

SYSTEM / 03

INFRASTRUCTURE MOMENTUM · DATA CENTRES / COMPUTE LOCK-IN · PLACEHOLDER

EXPAND

NEXT PIECE

This section will examine the physical build-out underneath the AI race.

The core question: once enough capital has been sunk into compute, power, land, cooling, and data centres, can the system do anything except justify that build at all costs?

Focus: hyperscalers, sunk cost, hardware decay, and the need to keep feeding the machine.

Topic direction: infrastructure changes the argument from optional technology to locked-in economic necessity.

SYSTEM / 04

CONTROL ILLUSION · CORPORATIONS / STATES / INVESTORS · PLACEHOLDER

EXPAND

NEXT PIECE

This section will examine who appears to be in charge, and why that does not amount to control.

The core question: if responsibility is spread across governments, private capital, executives, procurement systems, and automated decision layers, who exactly is left capable of saying no?

Focus: distributed authority, absent accountability, and the difference between oversight and theatre.

Topic direction: control exists everywhere on paper and nowhere in practice.

SYSTEM ANALYSIS

IF THIS MEANT SOMETHING TO YOU — OR NOTHING AT ALL — TELL ME

chris.llm2wmd@gmail.com

THE END HAS BEGUN

You Made It This Far.

Start the run

AI FAQ

Sources

System Choice

Target package available

Delegate the trigger?

Pause

End state reached