AI Roundup May 2023: Hinton Quits Google, Altman Goes to Congress, and Claude Gets a 100K Brain

TL;DR

May 2023 was the month the people who built this stuff started getting nervous out loud. Geoffrey Hinton walked away from Google so he could warn about the technology he helped invent. Sam Altman sat in front of the Senate and asked to be regulated. Meanwhile the actual product race went vertical: Google dropped PaLM 2 and yanked the Bard waitlist for 180+ countries at I/O, Anthropic shipped a 100K-token context window that made everyone else's 8K look quaint, and Meta open-sourced ImageBind, a model that ties six modalities into one embedding space. The capability curve and the anxiety curve went up together.

Geoffrey Hinton Quit Google to Warn About the Thing He Invented

On May 1, the New York Times reported that Geoffrey Hinton, the man whose 2012 work on deep neural nets basically kicked off the modern era, had left Google after a decade. He did not leave quietly. He left specifically so he could talk about AI risk without a corporate comms team standing behind him.

The headline take was the unsettling one: Hinton said the pace of progress had blown past his own predictions. He had assumed superhuman AI was 30 to 50 years out, and now he was not so sure. He flagged the obvious near-term stuff too, including a flood of fake text and images making it hard to know what is true, and job displacement that is not theoretical anymore.

Why builders should care: this is not a doomer on Twitter, this is the guy whose name is on the foundational papers. When the field's most credible optimist starts hedging, it changes the regulatory weather you build in, because the policymakers listening to Hinton are the ones who write the rules you ship under.

Google I/O: PaLM 2 Lands and Bard Goes Global

At I/O on May 10, Google stopped looking flat-footed. It launched PaLM 2, its next-gen LLM, and rebuilt Bard on top of it. More importantly for anyone who had been stuck on a waitlist, Google removed the Bard waitlist entirely and opened it to more than 180 countries.

PaLM 2 came in multiple sizes (the smallest meant to run on-device), with stronger multilingual reasoning across 100+ languages, better coding, and better logic. Google also wired it into the PaLM API, Firebase, and Colab, plus a long list of Workspace features. Bard picked up coding help across 20+ languages, dark mode, and promised multimodality and third-party app hooks, including Adobe Firefly for image generation.

The builder read: Google finally has a credible API-accessible model and a developer onramp, so the LLM market stopped being a one-horse race. Competition is good for your bill and your roadmap. A smaller on-device PaLM 2 tier was also a quiet signal that the giants were starting to think about local inference.

Anthropic Shipped a 100K-Token Context Window

On May 11, Anthropic expanded Claude's context window from 9K to 100,000 tokens, roughly 75,000 words. For reference, that month the standard GPT-4 was at 8K and the big variant topped out at 32K. Anthropic just put a full novel's worth of text into the prompt and made it look easy.

This was a genuine capability unlock, not a spec-sheet flex. Dump an entire codebase, a long contract, or a stack of research papers in and ask questions across all of it without building a retrieval pipeline first. Their demo loaded all of The Great Gatsby, changed one line, and had Claude spot the edit in about 22 seconds.

For builders this was the moment long-context stopped being a research curiosity and became a product decision. A lot of early RAG plumbing suddenly looked optional for documents that fit in the window. The tradeoffs (cost and latency) were real, but the design space for a single API call got dramatically bigger.

Sam Altman Went to Washington and Asked to Be Regulated

On May 16, Sam Altman made his first appearance before Congress at a Senate Judiciary subcommittee hearing titled "Oversight of A.I.: Rules for Artificial Intelligence." The notable part: the OpenAI CEO did not fight regulation, he requested it. He called for a licensing and testing regime for models above a certain capability threshold, plus internal and external safety testing before release.

Altman named election manipulation and targeted disinformation as his areas of greatest concern, pointing at the 2024 cycle. The hearing went unusually smoothly for a tech CEO in front of senators, which itself raised eyebrows.

The cynical-but-fair builder take: a licensing regime keyed to a capability threshold is a moat. Big labs can absorb compliance cost, but indie devs and open-weight projects might not be able to. Watch this one closely, because the rules sketched out in this room are the rules that decide whether you can legally fine-tune and ship your own model in a couple of years.

Meta Open-Sourced ImageBind, a Six-Modality Embedding Model

On May 9, Meta released ImageBind, an open-source model that binds six modalities (image/video, text, audio, depth, thermal, and IMU motion data) into a single shared embedding space. The trick is that it learns these joint embeddings without needing every modality paired with every other one during training.

Why it matters: a shared embedding space across senses is a primitive you build other things on top of. Cross-modal search (find the audio that matches this image), cross-modal generation, and richer multimodal retrieval all get easier when everything lives in the same vector space. And because Meta open-sourced it, the local-AI crowd got to poke at it instead of reading a paper and waiting for an API.

Key Takeaways

• The builders got scared on the record. Hinton quitting and Altman asking for regulation in the same month means the safety and policy conversation is now mainstream, and it will shape what you are legally allowed to ship.
• The LLM market became a real race. PaLM 2 plus a global Bard rollout gave Google a credible answer to OpenAI, which means more competition, more API options, and less single-vendor lock-in for the rest of us.
• Long context went from demo to default. Anthropic's 100K window made "just put it in the prompt" a legitimate architecture, and quietly raised the bar every other provider would have to clear.
• Regulation talk has a moat problem. Capability-threshold licensing sounds responsible, but watch whether it boxes out indie devs and open-weight models before it boxes in the labs.
• Open weights kept the local crowd in the game. ImageBind landing as open source is the counterweight to the regulatory tightening: while the policy debate heats up, the open ecosystem is still shipping primitives you can run yourself.