AI Roundup March 2022: DeepMind's Chinchilla, OpenAI's davinci-002, and Midjourney's quiet beta

TL;DR

March 2022 was a deceptively heavy month. DeepMind dropped the Chinchilla paper and quietly broke the "just make it bigger" consensus that had governed LLM training for years. OpenAI shipped a new instruction-tuned generation of API models (text-davinci-002 and code-davinci-002) that became the workhorses of the GPT-3.5 era. GitHub put Copilot inside Visual Studio, and Midjourney cracked open its Discord beta, planting the seed for the text-to-image explosion that would dominate the rest of the year.

DeepMind's Chinchilla says your model is too big and undertrained

On March 29, DeepMind posted "Training Compute-Optimal Large Language Models" (arXiv:2203.15556), and it landed like a quiet earthquake. The team trained over 400 models, from 70M to 16B parameters, across datasets from 5B to 500B tokens, to answer one question: given a fixed compute budget, how should you split it between model size and training data?

The answer overturned the prevailing wisdom. Model size and training tokens should scale roughly in lockstep: every time you double parameters, you should double the data. The rough rule of thumb that came out of it, around 20 tokens per parameter, became known as "Chinchilla optimal." To prove the point, DeepMind trained Chinchilla, a 70B model, and had it outperform the 280B Gopher using the same compute budget. Most of the giant models built up to that point were, in the paper's framing, massively oversized and massively undertrained.

Why it matters to builders: this is the paper that made small-and-well-trained the smart play. It is the intellectual ancestor of every compact, data-hungry open model you can actually run on your own hardware. If you have ever been grateful that a 7B model punches way above its weight, you have Chinchilla to thank for the math.

OpenAI quietly ships text-davinci-002 and code-davinci-002

On March 15, OpenAI rolled out new versions of GPT-3 and Codex in its API under the names text-davinci-002 and code-davinci-002, with edit and insert capabilities and a training cutoff of June 2021. These were instruction-following models that built on the InstructGPT work OpenAI had announced earlier in the year, and they were noticeably better at actually doing what you asked.

There was no splashy launch event, just a changelog entry and some new model strings. But these became the default workhorses for a huge wave of early LLM products. The edit and insert endpoints in particular gave developers structured ways to do things beyond plain completion, which mattered a lot when you were trying to build real features rather than demos.

If you were building on the OpenAI API in 2022, davinci-002 is probably the model your prototype ran on. It is the bridge between raw GPT-3 and the ChatGPT moment still nine months away.

GitHub Copilot lands in Visual Studio 2022

On March 29, GitHub brought Copilot to the Visual Studio 2022 IDE, expanding its AI pair programmer beyond the VS Code and JetBrains crowd into the heart of the .NET ecosystem. At this point Copilot was still in technical preview, free to use for accepted testers, and would not become a paid subscription product until that summer.

The significance here is reach, not novelty. Putting code completion that good in front of millions of Visual Studio developers normalized AI-assisted coding for a giant, often more conservative slice of the industry. The debates about training data, licensing, and what "your code" even means were already simmering, and they would only get louder. But for working developers, the productivity case was getting hard to argue with.

Midjourney opens its Discord beta

In March, Midjourney quietly launched its closed beta on Discord, with David Holz's team inviting early users to generate images through a bot rather than a traditional web app. It was scrappy, it was weird, and the Discord-native interface was unlike anything else in the space.

This was not the open public launch, that came in July, but March is when the engine started turning. Alongside the diffusion research and other image tools brewing at the time, Midjourney's beta marked the beginning of the text-to-image wave that would define the back half of 2022. For anyone watching closely, it was an early signal that generative imagery was about to go from research curiosity to cultural phenomenon.

Key Takeaways

• Chinchilla changed the rules. The compute-optimal scaling result made smaller, better-trained models the rational choice, and it underwrites basically every capable model you can run locally today.
• davinci-002 was the quiet workhorse. OpenAI's instruction-tuned March models powered a generation of early LLM apps and set the stage for the ChatGPT era.
• AI coding went mainstream-adjacent. Copilot reaching Visual Studio pushed AI assistance into the .NET world and made AI-in-the-editor feel normal for millions more developers.
• Text-to-image started its takeoff. Midjourney's Discord beta was a small spark in March that helped ignite the generative image boom later in 2022.
• The cutoff matters. davinci-002's June 2021 training cutoff is a reminder that even the best models of the moment have a knowledge horizon, something every builder still has to design around.