AI Roundup May 2022: Meta Open-Sources OPT-175B, DeepMind Drops Gato, Google Teases Imagen

TL;DR

May 2022 was the month the big labs stopped just publishing papers and started handing over the goods (or at least teasing them loudly). Meta dropped OPT-175B, a GPT-3-scale language model with open weights and, more importantly, the messy training logbook to go with it. DeepMind shipped Gato, a single network that plays Atari, captions images, and drives a robot arm. And Google used I/O and a late-month paper to answer OpenAI's DALL-E 2 with LaMDA 2, the AI Test Kitchen, and Imagen. The open vs closed fault line that defines AI today got drawn this month.

Meta Open-Sources OPT-175B, GPT-3's Carbon Footprint, and Its Lab Notebook

On May 3, Meta AI released OPT (Open Pre-trained Transformer), a family of decoder-only language models from 125M all the way up to 175B parameters. The headline number matters: OPT-175B is the same scale as GPT-3, except GPT-3 was locked behind an API and OPT came with weights you could actually request and run. For anyone who had spent a year reading about GPT-3 without touching it, this was the first time a model in that weight class was put in front of the broader research community.

The part that earned Meta real respect was not the model, it was the honesty. Alongside the weights and code, the team published a full logbook documenting the training run: the loss spikes, the hardware failures, the restarts, the hyperparameter babysitting. Meta also pointed out that OPT-175B was trained at roughly one-seventh the carbon footprint of GPT-3. That transparency is rare. Most labs ship a clean paper and bury the pain.

Why builders should care

OPT was not a chatbot you could plug in and ship. The weights were gated for research, the license was restrictive, and 175B parameters is not something you serve on a homelab GPU. But OPT seeded the open-weights movement that BLOOM, LLaMA, and the entire local-LLM scene would build on. If you run models on your own hardware today, this is part of the lineage.

DeepMind's Gato Plays Atari, Captions Images, and Stacks Blocks With One Network

On May 12, DeepMind published A Generalist Agent, introducing Gato: a single transformer with one set of weights that does about 600 different tasks. The same network plays Atari games, captions images, holds a text conversation, and controls a real robot arm to stack blocks. The trick is tokenizing everything (text, images, button presses, joint torques) into one flat sequence and training one model on all of it.

Gato kicked off a loud debate. One DeepMind researcher suggested scale was now the main thing standing between us and more general agents, which set off the usual AGI-timeline arguments. The sober read is more useful: Gato was a proof that a single model can be genuinely multi-task and multi-embodiment, even if it was not state of the art at any one of those tasks. The generalist-vs-specialist question it raised is still live.

Google I/O: LaMDA 2 and the AI Test Kitchen

At I/O on May 11, Google leaned hard into conversational AI. It unveiled LaMDA 2, the next version of its dialogue model, and wrapped it in something called the AI Test Kitchen, an app meant to let a limited set of users actually poke at the model in controlled demos rather than just read about it.

The framing was telling. Google had a powerful language model and was visibly nervous about putting it in front of people, so it built a sandbox with guardrails and a waitlist. Read against what OpenAI shipped later, this caution looks like the moment Google chose to move slowly with chat. It was the right instinct on safety and arguably the wrong one on momentum.

Google Imagen Answers DALL-E 2 in the Text-to-Image Race

On May 23, Google Research published Imagen, a text-to-image diffusion model, with the paper Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. The eye-catching design choice: instead of training a giant text encoder from scratch, Imagen leaned on a large frozen language model (T5-XXL) to understand the prompt, then handed that off to a cascade of diffusion models for the pixels. Google claimed it beat DALL-E 2 on benchmarks and human preference for image-text alignment.

One month after OpenAI showed off DALL-E 2, Google made clear it was not sitting the image-generation race out. The catch, and it is a big one: Imagen was a paper and a gallery, not a product. There was no public access, no API, no weights. So while the research was real, builders got nothing to touch, which left the field wide open for the open-source diffusion wave that landed later in the year.

Key Takeaways

• Open weights went mainstream-ish. OPT-175B put a GPT-3-class model in researchers' hands and made open weights a credible alternative to API-only releases, seeding the local-LLM ecosystem builders rely on now.
• Transparency is a feature. Meta shipping its training logbook, failures and all, set a bar for honest reporting that most labs still do not clear.
• Generalist agents got a real proof point. Gato showed one network can span text, vision, games, and robotics, reframing the specialist-vs-generalist debate that still shapes agent research.
• Google had the tech but not the nerve. LaMDA 2 behind a Test Kitchen waitlist showed Google choosing caution over shipping, a choice that looks pivotal in hindsight.
• Text-to-image became a two-horse race. Imagen answered DALL-E 2 on quality, but with zero public access it handed momentum to the open models that followed.
• Papers were not products. Three of the month's four big stories shipped as research with no weights or access, the gap that open-source would rush to fill.