AI Roundup July 2023: Llama 2 Goes Commercial, Claude 2 Ships, and Washington Gets Promises

TL;DR

July 2023 was the month open weights grew up. Meta released Llama 2 with a license you can actually build a business on, Anthropic shipped Claude 2 and put a 100K token context window in front of the public, and Stability dropped SDXL 1.0 for the image crowd. Meanwhile, seven of the biggest labs walked into the White House and signed a set of voluntary safety commitments, then announced the Frontier Model Forum to formalize the club.

Meta Drops Llama 2 With a Commercial License (July 18)

This is the one that actually changed what you can build. On July 18, Meta and Microsoft announced Llama 2 in three sizes: 7B, 13B, and 70B, with both base and chat-tuned variants. The original LLaMA leaked and was research-only, which made it legally radioactive for anything resembling a product. Llama 2 came out the front door with weights you can download and a license that permits commercial use.

The catch worth knowing: it is not OSI-approved open source. There is an acceptable use policy, and if your product crosses 700 million monthly active users you have to negotiate a separate license with Meta. For roughly everyone reading this, that ceiling is irrelevant, which is exactly the point. Microsoft made it available in the Azure model catalog day one, and the weights also landed on Hugging Face.

For builders running local AI, this was the starting gun. A genuinely capable 70B you can fine-tune, quantize, and self-host without phoning a vendor is the foundation that the entire late-2023 open model scene got built on.

Anthropic Ships Claude 2 and a 100K Context Window (July 11)

On July 11, Anthropic released Claude 2 and, more importantly for most people, opened a public beta at claude.ai. Until then Claude was mostly an API-and-waitlist affair. Now anyone could just use it.

The headline spec was the 100K token context window, enough to drop a couple hundred pages of documentation or an entire codebase into a single prompt. Anthropic also reported real jumps on benchmarks: 71.2 percent on the Codex HumanEval Python coding test, up from 56.0 percent, and 88.0 percent on the GSM8K grade-school math set. The bar exam multiple choice number ticked up to 76.5 percent.

The practical takeaway: long-context was suddenly a live option, not a research demo. If your workflow involved stuffing big documents at a model, you now had a second serious vendor to play GPT-4 off against.

Stability Releases SDXL 1.0 (July 26)

The image folks got their moment on July 26 when Stability AI shipped Stable Diffusion XL 1.0. It is a 3.5B parameter base model paired with a refiner in a roughly 6.6B parameter ensemble pipeline, and it renders natively at 1024x1024 instead of the old 512 squares everyone was upscaling.

SDXL was noticeably better at the things older SD builds choked on: hands, legible text, and coherent composition. It shipped under the CreativeML OpenRAIL++-M license and landed across GitHub, the API, Clipdrop, DreamStudio, and Amazon Bedrock. For the local generation crowd, an open-weight model at this quality meant the ControlNet and LoRA ecosystem had a new and far stronger base to build on.

Seven Labs Sign Voluntary AI Commitments at the White House (July 21)

On July 21, the Biden administration announced that seven companies, Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI, had agreed to a set of voluntary AI safety commitments. The pledges covered pre-release security testing by internal and outside experts, sharing risk information across industry and government, investing in cybersecurity to protect unreleased model weights, and reporting system capabilities and limitations.

Be clear-eyed about what this was: voluntary, with no enforcement mechanism. White House officials said as much. It was a signal more than a rulebook, a way to set expectations ahead of any actual legislation or executive action.

Around the same time, Anthropic, Google, Microsoft, and OpenAI announced the Frontier Model Forum, an industry body to coordinate on safety research and best practices for large models. Read it how you like: responsible coordination, or the largest incumbents drawing the map before regulators show up. Probably some of both.

Key Takeaways

• Open weights became commercially viable. Llama 2's license, caveats and all, turned self-hosted LLMs from a research toy into something you can ship in a product.
• Long context went mainstream. Claude 2's 100K window made whole-document and whole-codebase prompting a normal thing to expect from a model.
• Local image generation leveled up. SDXL 1.0 gave the open-source pipeline a 1024x1024 native base that fixed hands, text, and composition.
• Policy moved by handshake, not law. The White House commitments and the Frontier Model Forum were voluntary, so treat them as direction of travel rather than binding rules.
• The competitive map filled in. With Claude 2 public and Llama 2 free to build on, GPT-4 finally had real company at the top of the stack.