AI Roundup July 2022: DALL-E 2 Opens Up, Midjourney Goes Public, BLOOM Drops 176B Weights

TL;DR

July 2022 was the month text-to-image went from research demo to something you could actually use. OpenAI opened DALL-E 2 to a million beta users with real commercial rights, Midjourney threw open its public beta, and the BigScience collective dropped BLOOM, a fully open 176B-parameter multilingual LLM. On the science side, DeepMind released predicted structures for nearly every protein known to science. The generative-AI gold rush you are living through now started lighting up right here.

DALL-E 2 Opens Its Beta to a Million Users, With Commercial Rights Attached

On July 20, OpenAI moved DALL-E 2 out of its tightly gated research preview and into a paid beta, with plans to bring in a million people off the waitlist. The big shift was not just access, it was the business model and the licensing. Beta users got 50 free generation credits the first month and 15 every month after, plus the option to buy more credits in 115-generation packs for fifteen dollars.

The detail that actually mattered for builders: OpenAI granted users full commercial rights to the images they create, including the right to sell, reprint, and merchandise them. That single line turned DALL-E from a toy into a tool you could put in a real workflow. It also kicked off the still-unsettled fight over who owns AI-generated art and what the training data means for that claim.

Why it matters

This was the moment a frontier image model stopped being a screenshot on Twitter and became a product with pricing, credits, and a terms-of-service. If you wanted to ship something with generated images, July was the first month you really could.

Midjourney Hits Open Beta and Lets Anyone Generate From a Discord Server

While OpenAI was carefully metering its waitlist, Midjourney went the other direction. On July 12, the David Holz-led lab opened its public beta, and the on-ramp was almost absurdly low: join a Discord server, type a prompt into a channel, watch images appear. No app, no SDK, no install.

That distribution choice was the whole story. Midjourney's painterly, opinionated house style plus the social, public-by-default Discord flow meant prompts and results were visible to everyone, and the aesthetic spread fast. It proved you did not need a polished product surface to win, you needed a model people wanted to share and a place to share it.

BigScience Ships BLOOM: a Fully Open 176B-Parameter Multilingual LLM

For the local-AI and open-weights crowd, this was the month's biggest drop. BigScience, a volunteer research effort of more than a thousand researchers coordinated through Hugging Face, released BLOOM: a 176-billion-parameter autoregressive language model trained on 46 natural languages and 13 programming languages. The full weights were free to download, alongside the training code, recipes, and unusually thorough documentation including model cards and the ROOTS dataset datasheet.

At the time, BLOOM was roughly GPT-3 scale and the largest truly open-access multilingual model in existence. It shipped under the project's Responsible AI License, an early example of behavior-restricted open weights that the community would argue about for years. You were not going to run 176B on your homelab box without serious hardware, but the point was that the weights existed, in the open, for anyone to study and build on.

Why it matters

BLOOM was a statement: frontier-scale language models did not have to be a black box behind an API. It set a template for the open-weight releases that followed and gave researchers a transparent model they could actually inspect.

DeepMind Releases Structures for Nearly Every Known Protein

On July 28, DeepMind and EMBL's European Bioinformatics Institute expanded the AlphaFold Protein Structure Database from roughly one million predicted structures to over 200 million, covering nearly every catalogued protein known to science across more than a million species. That is a 200x jump, released free to anyone.

This is the clearest example we had to date of AI compressing years of lab work into a database lookup. Researchers working on antibiotic resistance, neglected tropical diseases, enzymes for recycling plastic, and food security suddenly had a starting structure for almost any protein they cared about. It is the kind of quiet, infrastructure-grade AI win that does not trend on social media but reshapes an entire field.

Key Takeaways

• Text-to-image went mainstream in one month. DALL-E 2's million-user beta and Midjourney's open beta landed within eight days of each other, and the generative image era started in earnest.
• Commercial rights changed the game. OpenAI handing users ownership of their generations turned image models from demos into something you could actually build a business on.
• Distribution beats polish. Midjourney proved a Discord server and a strong house style could out-grow a carefully gated product surface.
• Open weights got serious. BLOOM put a GPT-3-scale, multilingual model in the open and set the template for the open-weight movement, license debates included.
• AI as scientific infrastructure. AlphaFold's 200-million-structure release showed AI quietly becoming a default tool for working scientists, not just a headline generator.