Twelve AI Models Launched in One Week. The Field That Produced Them Was Named at a Summer Workshop in 1956.

Key Takeaways

• •At least twelve major AI models were announced or released in the third week of March 2026
• •The term 'artificial intelligence' was coined at the 1956 Dartmouth workshop by McCarthy, Minsky, Rochester, and Shannon
• •The Transformer architecture, published in 2017, underlies virtually every model released this week
• •Global AI investment exceeded $200 billion in 2025 alone
• •The gap between model releases has compressed from years to days

Timeline

In the third week of March 2026, the artificial intelligence industry did something that would have been unthinkable five years ago and merely improbable two years ago. It released twelve major models in seven days.

OpenAI shipped GPT-5 Turbo with expanded multimodal reasoning. Google DeepMind unveiled Gemini 2.5 Pro with native code execution. Anthropic released Claude Opus 4, its most capable model to date. Meta dropped Llama 4 Maverick and Llama 4 Scout as open-weight releases. Mistral launched its next-generation model. And a handful of smaller labs, including Cohere, AI21, and China's DeepSeek, each pushed updates of their own.

Twelve models. One week. Each claiming state-of-the-art performance on at least some benchmark. Each backed by billions of dollars of compute infrastructure. Each competing for the attention of developers, enterprises, and governments that are trying to figure out which model to bet on, or whether to bet on any single model at all.

The velocity is disorienting. But to understand what this week represents, you have to go back to a much quieter gathering seventy years ago.

Twelve models in one week. In 1956, the founders of AI thought it might take one summer to solve intelligence. Seventy years later, a dozen companies are each releasing their own version of the answer.

— ROOT•BYTE analysis

In the summer of 1956, a 28-year-old mathematician named John McCarthy convinced Dartmouth College to host a workshop. He recruited three co-organizers: Marvin Minsky from Harvard, Nathaniel Rochester from IBM, and Claude Shannon from Bell Labs, the same Claude Shannon who had invented information theory and laid the mathematical foundation for all digital communication.

Their proposal, submitted in August 1955, included a sentence that would define a field: "We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

That sentence coined the term "artificial intelligence." Before Dartmouth, researchers working on thinking machines used scattered terminology: cybernetics, automata theory, complex information processing. McCarthy wanted a new name, one that was ambitious enough to attract funding and specific enough to define a research agenda. He chose "artificial intelligence." The name stuck.

The workshop itself was modest. About ten researchers attended over the course of the summer. They did not solve intelligence. They did not produce a breakthrough paper. What they produced was something more valuable: a community. The attendees went back to their universities and founded AI labs. McCarthy went to MIT and then Stanford. Minsky stayed at MIT. The Dartmouth alumni became the founding generation of AI research.

But progress was slow and overpromised. In 1958, Herbert Simon predicted that within ten years a computer would be chess champion and would discover an important mathematical theorem. He was off by four decades on chess and arguably still waiting on the theorem. The gap between AI's ambitions and its capabilities led to what researchers call the "AI winters," periods in the 1970s and late 1980s when funding dried up and the field contracted.

The first AI winter hit in the mid-1970s after the Lighthill Report in the UK concluded that AI had failed to deliver on its promises. Government funding was slashed. The second winter arrived in the late 1980s when expert systems, the dominant AI paradigm, proved brittle and expensive to maintain. For nearly two decades, AI was considered a field that had promised the moon and delivered a pocket calculator.

The AI winter lasted nearly two decades. Now we are in an AI summer so intense that the models are arriving faster than anyone can evaluate them.

— ROOT•BYTE analysis

What changed everything was data and compute. In 2012, a deep neural network called AlexNet won the ImageNet image recognition competition by a staggering margin, using GPU-accelerated training on a large dataset. The deep learning revolution had begun. Within five years, neural networks were beating humans at image recognition, game playing, and translation.

Then, in 2017, a team at Google published a paper titled "Attention Is All You Need." It introduced the Transformer architecture, a neural network design that could process sequences of data in parallel rather than sequentially. The Transformer made it practical to train models on enormous datasets of text. GPT-1 followed in 2018. GPT-2 in 2019. GPT-3 in 2020. Each was larger, more capable, and more expensive to train than the last.

ChatGPT launched on November 30, 2022, and reached 100 million users in two months. That was the moment AI left the research lab and entered mainstream consciousness. Within a year, every major technology company had an AI product. Within two years, AI was integrated into search engines, productivity software, code editors, and creative tools.

And now, in March 2026, we have arrived at the point where twelve frontier models can be released in a single week and the industry barely pauses to catch its breath.

The implications are significant. First, the model layer is commoditizing. When multiple companies can produce models of roughly comparable capability, the competitive moat shifts from the model itself to the ecosystem around it: the API, the developer tools, the fine-tuning infrastructure, the safety guarantees, the enterprise support. This is the same pattern that played out with databases, cloud computing, and mobile operating systems. The technology becomes table stakes; the platform becomes the product.

Second, the open-weight releases from Meta and others are reshaping the competitive landscape. When Llama 4 is freely available for anyone to download and run, the closed-model providers must justify their pricing with superior capability, reliability, or safety. The pressure is relentless.

Third, the evaluation problem is becoming acute. When twelve models launch in a week, no one has time to rigorously benchmark them all. The industry is increasingly relying on "vibe checks" and anecdotal comparisons rather than systematic evaluation. This is a problem. Models that perform well on benchmarks may fail in production. Models that seem impressive in demos may have subtle failure modes that only emerge at scale.

Global AI investment exceeded $200 billion in 2025. The hyperscalers, Microsoft, Google, Amazon, and Meta, are each spending tens of billions on data centers filled with GPUs. The energy consumption of AI training and inference is now a geopolitical issue, with countries competing for the electricity and chip supply needed to stay in the race.

Seventy years ago, ten researchers spent a summer at Dartmouth trying to figure out whether machines could think. They gave their discipline a name and a mission. They thought it might take a summer. It took seven decades. But now the field they founded is releasing twelve models a week, each one more capable than anything they could have imagined.

The Dartmouth proposal is still available online. It is four pages long. It lists seven areas of study: automatic computers, programming language, neuron nets, theory of the size of a calculation, self-improvement, abstractions, and randomness and creativity. Every single one of those areas is now an active field of multibillion-dollar research.

McCarthy, Minsky, Rochester, and Shannon planted a seed in 1956. This week, twelve trees bore fruit simultaneously. The orchard they planted is now a forest, and it is growing faster than anyone can map it.

(Sources: Dartmouth AI Project proposal (1955), "Attention Is All You Need" (Vaswani et al., 2017), Stanford HAI AI Index Report 2026, company press releases from OpenAI, Google DeepMind, Anthropic, and Meta)

Enjoy This Article?

RootByte is 100% independent - no paywalls, no corporate sponsors. Your support helps fund education, therapy for special needs kids, and keeps the research going.

How did this make you feel?

Recommended Gear

Keep Reading

AI

The AI Company That Said No to the Pentagon, Anthropic, Claude, and the Roots of AI Safety

10 min read

AI

The AI Company That Said No to the Pentagon Is Now Suing the Pentagon

8 min read

AI

China Just Built a Trillion-Parameter AI and Gave It Away for Free. The Root Goes Back to a Victorian Counting Machine.

8 min read