AI Roundup November 2022: ChatGPT Lands, Stable Diffusion 2.0 Ships, Galactica Implodes

TL;DR

November 2022 is the month AI stopped being a research demo and became a thing your relatives ask you about. OpenAI shipped ChatGPT on the 30th and it went vertical instantly. Stability AI released Stable Diffusion 2.0 with a new open text encoder, and Meta launched its Galactica science model only to pull it three days later after it confidently invented fake papers. Quietly, AWS used re:Invent to push generative tooling deeper into the cloud. The vibe shift was real.

ChatGPT Arrives on November 30 and the World Notices

On November 30, OpenAI released ChatGPT as a free research preview. It was a chat-tuned sibling of the GPT-3.5 family, fine-tuned with reinforcement learning from human feedback to follow instructions, admit some mistakes, and refuse some requests. None of those ingredients were brand new. What was new was the packaging: a clean chat box, no API key, no prompt-engineering ritual, just type and get a coherent answer back.

That packaging mattered more than any benchmark. Within days, ChatGPT was writing code, drafting emails, explaining regex, and roleplaying its way around its own guardrails. The growth curve became the story. This is the moment the general public got hands on a capable LLM at the same time builders did, and the feedback loop that followed reshaped the entire next year of the industry.

Why builders should care

If you build with this stuff, ChatGPT was both an inspiration and a warning shot. It proved that interface and accessibility beat raw capability for adoption. It also set a brutal expectation bar: users now assume a chatbot should just work. For the local-AI crowd, it lit a fire under open alternatives, because suddenly everyone wanted this experience running somewhere they controlled.

Stable Diffusion 2.0 Ships With a New Open Text Encoder

On November 24, Stability AI released Stable Diffusion 2.0. The headline change under the hood was the swap to OpenCLIP, a text encoder trained by LAION with Stability's support, replacing the previous OpenAI CLIP encoder. The release also bumped native resolution to 768x768, shipped a 4x upscaler, and added a depth-guided depth2img model for structure-aware image-to-image.

The community reaction was famously mixed. Because the text encoder and training data changed, a lot of prompts that worked beautifully on 1.x produced worse results on 2.0, and certain styles and subjects got harder to summon. That friction is a useful lesson: when you rebuild the encoder and re-filter the dataset, you are effectively shipping a new model with a new prompt language, not a drop-in upgrade.

Still, this was a real open-weights release you could run on your own GPU, fork, and fine-tune. For homelab builders, that is the whole point. The grumbling about 2.0 also set up 2.1 and the eventual SDXL era, so it is worth remembering where the line started.

Meta's Galactica Survives Three Days Before Getting Pulled

On November 15, Meta unveiled Galactica, a large language model trained on tens of millions of scientific papers, textbooks, and reference material. The pitch was bold: summarize research, write scientific code, generate wiki-style articles, even annotate molecules and proteins. Meta put up a public demo and invited people to try it.

People tried it. Galactica produced authoritative-sounding nonsense: fabricated papers attributed to real authors, fake citations, and a now-infamous article on the history of bears in space presented with the same confidence as real physics. After roughly three days of public criticism, Meta took the demo down on November 17.

The takeaway is not that Galactica was uniquely bad. It hallucinated the way LLMs of that era all did. The mistake was framing a fluent text generator as an authority for science, a domain where being plausibly wrong is worse than useless. Two weeks later ChatGPT shipped with similar failure modes but framed as a chatty assistant, and the public forgave it for things it crucified Galactica over. Framing is a feature.

AWS re:Invent Leans Into the Generative Wave

Late in the month, AWS held re:Invent, its annual cloud conference, and machine learning was front and center across the SageMaker stack and beyond. The specifics are less memorable than the timing: the largest cloud provider was visibly positioning around ML tooling and infrastructure in the exact window when generative AI was about to become the dominant cloud spend story.

For builders, the signal was that the infrastructure layer was bracing for a wave that ChatGPT, landing days later, would turn into a flood. If you were planning where to deploy models in 2023, the groundwork being laid here is part of why the options multiplied so fast.

Key Takeaways

• Distribution beat capability. ChatGPT was not a giant leap in raw model power, but a free chat box with no setup put a capable LLM in millions of hands overnight.
• Open weights kept pace. Stable Diffusion 2.0 proved the open-model track was shipping real upgrades you can run and fine-tune yourself, even when the community pushes back on the changes.
• Framing decides reception. Galactica and ChatGPT shared the same hallucination problem, but one was sold as a science oracle and one as a helpful assistant, and only one survived the week.
• A new model is a new prompt language. Swapping the text encoder in SD 2.0 broke old prompts, a reminder that under-the-hood changes ripple straight into your workflow.
• The infra layer saw it coming. AWS spent re:Invent doubling down on ML tooling right before demand for it exploded.