The Rise of Open Weights: A Long Scroll Story

The Rise of Open Weights — and the fall of commercial AI?

The Rise of Open Weights

Scott Logic · The story of open weights

The Rise of Open Weights

And the fall of commercial AI?

Scroll to begin

Artificial Intelligence (AI) has dominated the technology and mainstream media headlines for years. Names like OpenAI and ChatGPT have become household terms. But the dominant position of these tech giants, known as frontier labs, led by OpenAI and Anthropic, who have grown their revenues faster than almost any tech company before them, is now under threat from so-called open weights models — predominantly emanating from China.

This is the story of open weights AI models: what they are, where they come from, and why they may disrupt the leading frontier labs before those companies have a chance to fully capitalise on the technology they spent so much money developing.

But to understand how we got here—and this is a story with more twists, reversals and ironies than you might expect—we need to start our journey a few years earlier, in the run-up to GPT-3 and ChatGPT, when this technology first began making headlines.

Section 2 · A seven-year build-up

2015–2022 — The road to ChatGPT

ChatGPT looked like an overnight success. In truth it was seven years in the making — walk the timeline, one milestone at a time.

2015

2016–17

2018

2019

2020

2021

2022

2015

A non-profit is born

OpenAI was founded in December 2015 by Elon Musk, Sam Altman and others, out of a worry that recent advances in Artificial Intelligence might lead to what is known as Artificial General Intelligence (or AGI). Google had recently acquired DeepMind, a leading research lab that was considered the favourite for developing AGI first. The OpenAI founders considered that AGI shouldn’t be owned by a single profit-driven company, creating the company as a non-profit lab, backed by ~$1 billion in pledges, chartered to benefit humanity and share its research openly.

2016–2017

A research lab, exploring

The early days of OpenAI involved research across a wide range of fronts, including Gym their reinforcement-learning research, where they taught a simulated robot to backflip from human feedback. They also created bots that competed in online computer games.

2018

The first GPT

In mid 2018, OpenAI published GPT-1, their first "Generative Pre-Trained Transformer" introduced in the paper "Improving Language Understanding by Generative Pre-Training". More commonly known as a Large Language Model (LLM), it had ~117 million parameters and demonstrated a now-foundational recipe: pre-train a model on a large text corpus, then fine-tune it on specific tasks. Modest by today's standards, it proved the approach that GPT-2, GPT-3 and ChatGPT would scale up. The GPT-1 architecture was based directly on Google's Transformer and its self-attention mechanism, which was published the year before in the landmark paper "Attention Is All You Need". Ironically the core invention that underpins OpenAI's flagship GPT line of models was published by Google - the very company whose AI dominance had spurred OpenAI's founding.

2019

Too dangerous to release?

While GPT-1 was an impressive model in its own right, OpenAI continued to research a wide range of AI applications, and as a result their compute capacity was spread thin. A key discovery that ultimately caused the company to divert their attention and compute towards developing GPT models was the scaling law - the empirical finding that a language model's performance improves predictably as you scale up model size (parameters), the training data and compute (for training). With this finding, OpenAI scaled up across all three dimensions, GPT-1 had 117 million parameters, GPT-2 had up to 1.5 billion. GPT-2 was so fluent that OpenAI initially held back the full model, citing misuse fears, before releasing it in stages. Notably as a non-profit they couldn't raise the cash needed to fund the power and compute for ever-larger training runs. As a result it shifted to a “capped-profit” structure and took a $1bn investment from Microsoft.

2020

Scale changes everything

OpenAI released GPT-3 with 175 billion parameters and a public API. A single model that could now write, summarise, translate and code with no task-specific training. The training run for this model was estimated to cost several million dollars. The new capped profit structure allowed OpenAI to start monetising the AI models they were producing, with the argument that the pursuit of AGI requires enormous capital and compute, hence the need to create commercially successful products. In tandem, OpenAI was becoming more closed. GPT-1 was fully open from the outset (paper, model weights and code), whereas GPT-2 was released in stages over a 9 month period. Whereas GPT-3 was entirely closed, with none of the technical details released. Furthermore Microsoft secured an exclusive license to the underlying model.

2021

From lab to product

In the year that followed, OpenAI released a succession of commercial products, including DALL·E that turned text into images and Codex which powered GitHub's Copilot, an AI assistant that helps software engineers write code. Copilot was their first genuinely mainstream commercial success, that gradually became a standard part of the developers toolkit over the next couple of years.

2022

The moment

Another important development happened the year after, with OpenAI's safety and alignment team developing InstructGPT which trained language models to follow instructions based on human feedback. Prior to that, models would often emit offensive or dangerous material from their now vast training dataset. While the technique was safety focussed, a side-effect of their approach was that it actually made talking with GPT models more pleasant and the output fundamentally more useful. With this final advancement, OpenAI wrapped their GPT-3.5 in a simple chat interface, and tentatively launched it as a "research preview" later that year. They initially expected a few thousand users and allocated server capacity for around 100,000 users. It reached ~1 million users in five days and ~100 million within two months and become the fastest-growing consumer app of its time.

The upshot

In just seven years, OpenAI had become almost the opposite of what it set out to be. Founded as a non-profit, it had gradually closed its doors with each successive model and product release, with its inner workings undisclosed and ultimately licensed to Microsoft. The overall mission - the pursuit of safe AGI to benefit humanity - hadn't changed, but the means had changed beyond recognition. But finally, almost by accident, the runaway success of ChatGPT suddenly turned OpenAI into a household name and lit the fuse of the AI 'arms race' that followed. The year ended with Google CEO Sundar Pichai declaring an internal "code red", having realised a rival had beaten it to market with its own foundational technology; AI was now a strategic priority.

Section 3 · The turn

2023 — “We have no moat”

While the Code Red signalled that Google leadership thought OpenAI was winning the race, a highly prescient leaked memo pointed out that they were perhaps looking at the wrong race altogether.

On 4 May 2023, a Google engineer’s internal memo was leaked to the analyst newsletter SemiAnalysis, carrying the punchy title "We Have No Moat, And Neither Does OpenAI".

A moat, in investor speak, is a durable competitive advantage that protects a company from competitors. While OpenAI, Google and now Anthropic (spun out from OpenAI in 2021) were pouring money and resources into scaling their AI efforts, the open source community were rapidly gaining ground, based on the surprise leak of LLaMA, a model developed by Meta.

The memo generated interest at the time of writing, but soon faded from view. OpenAI had just launched GPT-4, venture capital was flooding into frontier labs, and training runs were already costing hundreds of millions of dollars. The idea that a loose community of open-weight developers could keep pace looked quite unlikely. Three years later, that memo reads less like an opinion and more like a prediction.

Section 4 · What 'open' really means

Open weights, open source and the four freedoms

This story is about open weights AI, and LLaMA's somewhat unintentional release was something of a catalyst. But before telling that particular story, let's take a look at what open weights and open source are - the subtle and sometimes controversial differences.

Open source has a very long history in the field of computing, and describes the principle of publishing the source code of applications so that others can review, learn from and modify the code, as well as simply run it themselves. Put simply if an application or utility is open source it means that anyone is free to re-produce and recreate it. Open source is about freedom.

The majority of modern software is open source, giving consumers transparency and freedom from vendor lock-in. Somewhat surprisingly, it is entirely possible to make a commercial success of a business that releases all its software as open source and a great many do.

The GPT series of models are all based on a technique known as deep learning. The models themselves are a massive layered network of nodes and connections with associated parameters, resembling the structure of a human brain. The training process involves showing the model a vast quantity of data (e.g. written text), and training it to replicate the patterns it observes. This involves a vast amount of time and computation as the model gradually adjusts the parameters that form the connection between the nodes across the various layers. Fundamentally the model learns how to perform tasks by observing and mimicking its training data, with this learning 'baked in' to the model weights.

Once model training is complete, the training dataset is no longer needed. The model has learnt the underlying patterns, and trained to exhibit specific behaviours. As a result, using an AI model, through what is known as inference, only needs the model architecture and weights. Incidentally, the inference cost (i.e the cost of using an AI model) is many orders of magnitude less than the training cost.

So how do we apply open source to AI models?

Open source software is built around four key freedoms - the freedom to use, study, modify and share. Access to the model architecture and weights gives you the freedom to use and share, but without having access to the training dataset, it isn't possible to study or easily modify the model.

As of 2026, the vast majority of 'open' AI models share their architecture and weights, but not the details of how they were trained, and as such, they are not open source. As a result, they are described as 'open weights' models to make the distinction clear.

Even though they are not open source, there is still a tremendous amount of value in models being released as open weights. This allows anyone to host and run the model themselves - something which is becoming increasingly important due to geopolitics.

With the open source back story complete, let's return to LLaMA and the 'we have no moat' memo.

Section 5 · The model that got out

2023 — The LLaMA leak

Meta has had a long history of AI research, with their lab led by Turing Award winner Yann LeCun. Their team had a deep-rooted open-science culture, genuinely open sourcing their work. Furthermore, Meta didn't intend to monetise their AI developments, their core business has historically been built on advertising and engagement.

In February 2023 Meta released LLaMA 1 (Large Language Model Meta AI) as a research release, with weights available to academics. Notably the smaller 13B parameter model beat the larger 175B parameter GPT-3 on many benchmarks - although OpenAI would release the far more powerful GPT‑4 just weeks later.

[truncated for AI cost control]