AI Roundup June 2023: OpenAI Ships Function Calling, the EU Votes on AI, and SDXL 0.9 Drops

TL;DR

June 2023 was the month AI stopped being a demo and started becoming infrastructure. OpenAI gave every developer native function calling plus cheaper, longer-context models. The European Parliament voted through its negotiating position on the AI Act, setting up a fight over foundation models. Stability AI dropped SDXL 0.9 for research, and Inflection AI raised an eye-watering 1.3 billion dollars to chase a 22,000-GPU cluster. Builders won big this month.

OpenAI Ships Function Calling and Slashes Prices

On June 13, OpenAI shipped the update that quietly rewired how people build with LLMs. Function calling landed on gpt-4-0613 and gpt-3.5-turbo-0613: you describe your functions in a JSON schema, and the model decides when to emit a structured JSON object you can hand straight to your code. No more begging the model to return clean JSON and praying it complies. This is the primitive that every agent framework, tool-using app, and structured-extraction pipeline now leans on.

The pricing moves were just as loud. OpenAI cut input tokens on gpt-3.5-turbo by 25 percent and dropped the price of its embeddings model by 75 percent. They also launched gpt-3.5-turbo-16k, giving you four times the context (roughly 20 pages of text) at twice the base price.

Why builders care

If you were hand-rolling regex parsers to scrape JSON out of completions, this was the day that hack became obsolete. Function calling made the model a reliable router between natural language and your APIs, and that is the whole ballgame for agents.

The European Parliament Votes Through the AI Act

On June 14, the European Parliament adopted its negotiating position on the AI Act with 499 votes in favor, 28 against, and 93 abstentions. This was not the final law, but it was the moment Parliament locked in its stance and entered trilogue negotiations with the Council and Commission.

The headline for anyone shipping models: Parliament's text pushed new transparency, documentation, and disclosure obligations onto foundation models. Providers of systems like GPT and Llama would face requirements around training data, risk mitigation, and labeling AI-generated content. The fight over how heavily to regulate base models, and what that means for open-weight releases, started here in earnest.

This was the first serious attempt by a major jurisdiction to write rules specifically targeting general-purpose models. Everything that followed traces back to this vote.

Stability AI Drops SDXL 0.9

Stability AI unveiled Stable Diffusion XL 0.9 in late June, calling it a leap forward over the SDXL beta from April. The model runs a two-stage base-and-refiner pipeline, uses one of the largest OpenCLIP text encoders trained to date (OpenCLIP ViT-G/14), and generates natively at 1024x1024 with noticeably better composition and detail.

The catch: 0.9 shipped under a research-only, non-commercial license, gated behind a Hugging Face access request, with availability rolling out on Clipdrop and the Stability API around June 26. The fully open SDXL 1.0 release was teased for mid-July.

Why builders care

For the local-AI crowd, SDXL was the first base model that felt like a genuine step toward closing the gap with closed image generators while staying runnable on consumer GPUs. The two-stage architecture and bigger text encoder meant better prompt adherence without a cloud subscription.

Inflection AI Raises 1.3 Billion Dollars

On June 29, Inflection AI announced a 1.3 billion dollar raise led by Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, and new investor Nvidia. That brought the company's total funding to around 1.5 billion dollars at a reported 4 billion dollar valuation, all to power its Pi assistant and the Inflection-1 model behind it.

The number that raised eyebrows was the compute: Inflection said it was building one of the largest AI clusters in the world with partners CoreWeave and Nvidia, targeting 22,000 Nvidia H100 GPUs. CEO Mustafa Suleyman, a DeepMind co-founder, was betting that owning enormous training infrastructure was the moat.

Key Takeaways

• Function calling changed agent development overnight. Structured outputs went from a fragile prompt hack to a first-class API feature, and every tool-using app built after June 2023 assumes it exists.
• Cheaper, longer context lowered the floor. A 25 percent cut on gpt-3.5-turbo input and a 16k context window meant more ambitious apps became economically viable for indie builders.
• Regulation got specific about foundation models. The EU Parliament vote signaled that base-model providers, including open-weight ones, would face real compliance obligations, not vague principles.
• SDXL pushed open image generation forward. Even gated as research-only, 0.9 set the stage for a 1.0 release the local-AI community could actually run and fine-tune.
• Capital flooded toward compute. Inflection's 1.3 billion dollar raise, much of it earmarked for a 22,000-H100 cluster, made clear that GPU scale had become the headline investment thesis.