From Chatbot to Coworker: Nine Years That Rewired AI

Elon Musk and Sam Altman launch OpenAI as a nonprofit AI lab, pledging to counterbalance Google's DeepMind and share its research openly.^[1]

Why it mattered The countermove to Big Tech AI was itself a startup — one that would later shed both 'nonprofit' and 'open.'

Eight Google researchers publish the Transformer, a neural network architecture built entirely on attention that drops the recurrence and convolution earlier language models relied on.^[2]

Why it mattered It replaced the sequential building blocks of language models with a parallelizable attention mechanism that scaled cleanly with data and compute.

OpenAI introduces GPT-3, a 175-billion-parameter language model that performs many tasks from a few examples in the prompt, with no task-specific fine-tuning.^[3]

Why it mattered It demonstrated that simply scaling a Transformer produced few-shot abilities that smaller models lacked.

OpenAI releases ChatGPT as a free research preview built around conversational back-and-forth.^[4]

Why it mattered AI moved from lab demos and API calls into an everyday product anyone could talk to in plain language.

Ten days after launch, OpenAI's CEO publicly warns that ChatGPT is impressive but unreliable for anything that matters.^[5]

Why it mattered The company that shipped the product was also the one tapping the brakes on trust.

Meta publishes LLaMA, a family of foundation models in 7B, 13B, 33B, and 65B parameter sizes, under research-focused access.^[6]

Why it mattered A major lab put a capable foundation-model family into researcher hands rather than only behind a product API.

OpenAI releases GPT-4, a markedly more capable model that passes professional exams and accepts image input.^[7]

Why it mattered The performance jump made AI feel less like autocomplete and more like a general reasoner.

Anthropic publicly launches Claude after a closed alpha with partners including Notion, Quora, and DuckDuckGo.^[8]

Why it mattered A credible competitor entered the frontier race the same day GPT-4 shipped.

The Biden administration signs Executive Order 14110 on safe, secure, and trustworthy AI, directing federal agencies across safety, security, labor, civil-rights, and innovation priorities.^{[9] [10]}

Why it mattered The U.S. government moved from observer to rule-setter through an executive order.

Google introduces Gemini 1.0 in Ultra, Pro, and Nano sizes, describing it as built from the ground up to be multimodal.^[11]

Why it mattered Google DeepMind put a flagship multimodal model family into the frontier race.

The New York Times files a copyright suit alleging its articles were used without permission to train large language models.^[12]

Why it mattered The training-data question moved from think pieces into federal court.

Anthropic releases the Claude 3 family: Haiku, Sonnet, and Opus.^[13]

Why it mattered Anthropic claimed Claude 3 set new industry benchmarks across a range of cognitive tasks.

OpenAI announces GPT-4o, a flagship model that can reason across audio, vision, and text in real time.^[14]

Why it mattered OpenAI moved multimodal interaction closer to the core model rather than a pipeline of separate models.

The EU's risk-tiered AI regulation is published in the Official Journal, entering into force in August 2024 with obligations phasing in over the following years.^[15]

Why it mattered The world's first comprehensive AI law set binding rules tied to risk category.

OpenAI releases o1, a model that spends inference-time compute working through problems step by step before answering.^[16]

Why it mattered Progress shifted from bigger training runs toward thinking longer at inference.

In his 'Reflections' post, OpenAI's CEO predicts AI agents will start doing real work inside companies in 2025 — the exact opposite of his 2022 caution.^[17]

Why it mattered The same leader who said 'don't rely on it' now framed AI as a member of the workforce.

OpenAI and SoftBank announce Stargate, a new company intending to invest $500 billion over four years in U.S. AI infrastructure for OpenAI, with Oracle and MGX among initial equity funders.^[18]

Why it mattered The race became a capital-and-infrastructure contest measured in hundreds of billions of dollars.

Reuters reported that DeepSeek's low-cost model sparked a global AI sell-off, with Nvidia falling just under 17% and losing about $593 billion in market value on Jan. 27, 2025.^{[19] [20]}

Why it mattered A low-cost Chinese challenger rattled the assumption that frontier AI valuations would keep rising with compute spending.

Anthropic announces Claude 3.7 Sonnet and Claude Code, a command-line tool for agentic coding in limited research preview.^[21]

Why it mattered Anthropic presented Claude Code as a tool that can search and read code, edit files, run tests, commit and push to GitHub, and use command-line tools.

Anthropic releases Claude Opus 4 and Sonnet 4, positioning Opus as a heavyweight model built to run long, autonomous coding and agent tasks.^[22]

Why it mattered Anthropic split its line into a heavyweight Opus tier for sustained, multi-step work and a faster Sonnet tier for everyday use.

Anthropic's CEO tells Axios that AI could eliminate up to half of all entry-level white-collar jobs and push unemployment to 10-20% within one to five years, and urges the industry to stop sugar-coating it.^[23]

Why it mattered A frontier-lab CEO put a hard number on the job losses his own products might cause.

OpenAI introduces GPT-5 as a unified system that can answer quickly or route harder problems to deeper reasoning.^[24]

Why it mattered OpenAI moved quick-response and deeper-reasoning behavior into one routed system.

OpenAI ships an opt-in setting that limits ChatGPT's outbound requests to blunt prompt-injection attacks that try to exfiltrate a user's data.^[25]

Why it mattered A frontier lab shipped a user-facing security control aimed squarely at prompt injection, the defining attack on AI agents.

An Anthropic Institute report argues AI is increasingly used to build AI, and urges the field to be ready to slow down if recursive self-improvement takes hold.^[26]

Why it mattered A frontier lab put recursive self-improvement — AI accelerating AI research — on the record as a near-term concern, not science fiction.