AI Roundup May 2024: GPT-4o Goes Omni, Google I/O Floods the Zone, and the Sky Voice Mess

TL;DR

May was a two-day arms race and a week of cleanup. OpenAI dropped GPT-4o on May 13 with low-latency voice and image understanding plus free-tier access, Google countered the very next morning at I/O with Gemini Flash, Project Astra, and the Veo video model, and Anthropic published the most readable look inside a frontier model yet. Then OpenAI's demo voice turned into a legal headache when Scarlett Johansson lawyered up. Builders got cheaper multimodal, a new video model to watch, and a fresh reminder that vibes are not a licensing strategy.

OpenAI Ships GPT-4o: Real-Time Voice, Vision, and a Free-Tier Bombshell

On May 13, OpenAI's Spring Update introduced GPT-4o, where the "o" stands for omni. It is a single model trained across text, audio, and vision, so it can take any mix of those as input and respond in text, audio, or images. The headline number was latency: audio responses in as little as 232 milliseconds, averaging around 320, which is basically human conversational speed. The live demos of interruptible, emotional, real-time voice were the closest thing to the movie Her anyone had shipped.

The part that actually mattered for builders was less flashy: GPT-4o landed in the API at roughly half the price of GPT-4 Turbo and about twice the speed, and OpenAI pushed GPT-4-class intelligence into the free tier of ChatGPT. A frontier multimodal model that is cheaper and faster than last year's flagship, available to free users, resets the floor for what every downstream app has to beat. A Mac desktop app shipped alongside it.

The catch builders hit

The jaw-dropping voice mode was not actually available at launch. It rolled out in limited alpha months later, so for a while GPT-4o in production was the fast, cheap, text-and-vision model, not the talking assistant from the keynote. Worth remembering when a demo wows you: ship date and keynote date are different things.

Google I/O: Gemini Flash, Project Astra, and the Veo Video Model

One day later, on May 14, Google ran its counter-offensive at I/O and frankly flooded the zone. The practical drop was Gemini 1.5 Flash, a lighter, faster, cheaper model with the same long-context tricks as 1.5 Pro, and an expansion of that 1 million token context window (with 2 million previewed). For anyone building retrieval or document-heavy apps, cheap long context is the thing you actually use.

Google also previewed Project Astra, its real-time multimodal agent demo, clearly aimed straight at OpenAI's voice moment, and announced Veo, a text-to-video model claiming 1080p clips over a minute long. Veo was Google's answer to the Sora hype from earlier in the year, though like Sora it stayed gated to select creators. Imagen 3 for images and the sixth-gen Trillium TPU rounded out the pile. The strategic message was loud: Google has the models, the distribution, and the silicon, and it intends to use all three.

Anthropic Maps Claude's Mind: Scaling Monosemanticity and Golden Gate Claude

On May 21, Anthropic's interpretability team published "Scaling Monosemanticity," extracting millions of interpretable features from Claude 3 Sonnet using sparse autoencoders. In plain terms: they found internal patterns that correspond to concepts, from concrete things to abstract ideas like sycophancy and code vulnerabilities, and showed those features can be turned up or down to steer behavior.

The proof of concept went viral as Golden Gate Claude, a temporary version with the Golden Gate Bridge feature cranked so high it worked the bridge into nearly every reply. It was funny, but the point was serious: this is the first detailed look inside a leading LLM at scale, and it is a real signal that mechanistic interpretability might become a practical safety and debugging tool rather than an academic curiosity. If you care about understanding why your model does what it does, this paper is the one from May to read.

OpenAI's Sky Voice Becomes a Scarlett Johansson Lawsuit Threat

The GPT-4o demo voice, called Sky, sounded a lot like Scarlett Johansson in Her, and that turned out to be a problem. On May 20, Johansson went public: she said Sam Altman had asked her to voice ChatGPT in September 2023, she declined, and Altman reached out again just days before the launch. She said she was shocked and angered, and she retained legal counsel. OpenAI paused Sky and said it was a different professional actress, not an imitation, while declining to name her.

Altman's one-word "her" tweet on launch day did not help OpenAI's it's-a-coincidence defense. For builders, this is the practical lesson hiding under the celebrity drama: voice, likeness, and style are licensing minefields, and "it just sounds similar" is not a position you want to defend. If you are shipping anything that mimics a recognizable human, get the paperwork first.

Key Takeaways

• Multimodal got cheap and fast. GPT-4o halved the price and doubled the speed of frontier-class intelligence, so the baseline every app competes against just moved.
• Free-tier frontier models change the math. Pushing GPT-4-class quality to free ChatGPT users pressures anyone whose product was a thin wrapper over a paid model.
• Real-time voice is the new battlefield. OpenAI's voice mode and Google's Project Astra both bet that the assistant you talk to, not type to, is the next interface, even if neither fully shipped it in May.
• Video generation is heating up. Veo gave Google a credible Sora competitor, though both stayed gated, so the open and local crowd is still waiting for something they can actually run.
• Interpretability is going practical. Anthropic's feature-steering work hints at a future where you can inspect and adjust model internals, which matters for safety, debugging, and trust.
• Likeness rights are a real risk surface. The Sky mess is a cheap lesson to learn from someone else: license voices, faces, and styles before you ship them.