AI Roundup September 2022: Whisper Goes Open Source, Meta Makes Video, and the EU Comes for AI Liability

TL;DR

September 2022 was the month generative AI stopped being just about images. OpenAI open-sourced Whisper and dropped genuinely state of the art speech recognition into everyone's lap for free. Meta unveiled Make-A-Video and Google researchers published DreamFusion, dragging text-to-X into motion and 3D within days of each other. Meanwhile the open-weights Stable Diffusion wave kept cresting, and Brussels reminded everyone the party has lawyers: the EU proposed an AI Liability Directive. Builders won big this month.

OpenAI Open-Sourced Whisper, and Local Transcription Got Good Overnight

On September 21, OpenAI released Whisper, an automatic speech recognition model trained on 680,000 hours of multilingual, multitask audio scraped from the web. The headline number that matters: roughly 50 percent fewer errors than many specialized systems on diverse zero-shot benchmarks, across 99 languages, with translation into English baked in.

What made this land differently from most OpenAI announcements is the license. The weights and inference code shipped under MIT. No API gate, no token meter, no rate limit. You could clone the repo, point it at a 30-second chunk of audio, and get a transcript on your own GPU. For the local-AI crowd this was a gift: robust transcription that handles accents, background noise, and technical jargon, running entirely on your box.

The downstream effect was immediate and huge. Whisper became the default speech layer for a generation of indie tools, podcast pipelines, and accessibility projects. If you build anything that touches audio, this is the month your stack got a free upgrade.

Meta's Make-A-Video Turned Text Prompts Into Moving Pictures

On September 29, Meta announced Make-A-Video, a system that generates short video clips from text prompts. The clever bit is how it was trained: it learns what things look like from paired text-image data, then learns how things move from unlabeled video. That sidesteps the brutal scarcity of captioned video and is a big reason the approach scaled at all.

The clips were short and, frankly, a little cursed. But that is exactly how text-to-image looked a year earlier before it got terrifyingly good. Meta did not ship a public demo at launch, keeping it in research-paper territory, which tempered the hype a bit. Still, the signal was clear: text-to-video had crossed from theoretical to obviously-coming-soon.

Google's DreamFusion Took Text Straight to 3D

The same week, on September 29, Google researchers posted DreamFusion, which generates 3D objects from text using a pretrained 2D diffusion model as a prior. No 3D training data required. It optimizes a NeRF using a probability density distillation loss, essentially using a 2D image model to hallucinate a coherent 3D shape from every angle.

This matters because 3D asset generation is a notoriously data-starved problem. DreamFusion's trick of bootstrapping 3D from 2D diffusion was the kind of move that reframes a whole field. If you work in games, AR, or product viz, this is the paper that put text-to-3D on the roadmap for real.

The Open-Weights Stable Diffusion Wave Kept Cresting

Stable Diffusion's August open release was still reverberating all through September. The model was open-weights, runnable on consumer GPUs, and the community absolutely ran with it. Fine-tunes, img2img workflows, inpainting tools, and a wall of plugins showed up faster than anyone could track.

The contrast with the closed-API world was stark. While the big labs gated their best image models, Stable Diffusion proved that putting weights in people's hands produces an ecosystem, not just a product. September was when that thesis stopped being a hope and started looking like the dominant strategy for builders who want to own their stack.

The EU Proposed an AI Liability Directive

On September 28, the European Commission published its proposed AI Liability Directive, aimed at making it easier to sue when an AI system causes harm. The two headline mechanisms: a rebuttable presumption of causality to ease the burden of proof for victims, and the power for national courts to order disclosure of evidence about high-risk AI systems.

This was the regulatory counterweight to a month of unrestrained capability launches. The proposal would later be withdrawn after years of stalled negotiation, but at the time it was a clear signal that Brussels intended to attach real legal consequences to AI deployment. If you ship into the EU, this was the month liability stopped being hypothetical.

Key Takeaways

• Open weights win the ecosystem. Whisper and Stable Diffusion both showed that an MIT or open-weights release builds more tooling, faster, than any gated API.
• Generation jumped dimensions. Within one week, Make-A-Video pushed into motion and DreamFusion pushed into 3D, signaling that text-to-image was just the opening act.
• Speech got cheap and good. Whisper handed builders production-grade transcription for free, instantly upgrading every audio pipeline on the internet.
• Bootstrapping beats scarce data. Make-A-Video and DreamFusion both worked by leaning on abundant 2D and image-text data instead of waiting for labeled video or 3D, a pattern worth stealing.
• Regulation started catching up. The EU's liability proposal was an early sign that the legal layer of AI is coming, and builders shipping to Europe should pay attention.