September

8 events

Editors' summary

OpenAI commercialises multimodal, attention turns to Anthropic

On 25 September, Amazon invested up to $4 billion in Anthropic, naming AWS as Claude's primary cloud. On the 19th, Anthropic had published its Responsible Scaling Policy and the Long-Term Benefit Trust, setting out safety as a mechanism.

OpenAI shipped DALL·E 3 and added voice conversation and image input to ChatGPT, reaching beyond text. Microsoft consolidated its scattered AI features under the Copilot name.

At the end of the month, Mistral released Mistral 7B, a seven-billion-parameter model, under an unrestricted licence. It raised the bar for small open models.

This block is written by the editors. It is kept separate from the sourced record below.

Record8 events

7
ProductAnthropic
Claude Pro subscription launches
Anthropic launched Claude Pro, a paid tier for claude.ai, in the US and UK at $20 a month (£18 in the UK). Subscribers got five times the usage of the free tier, priority access during peak traffic, and early access to new features. The company framed it as a response to requests for more file uploads and longer-running conversations, and priced it level with OpenAI's ChatGPT Plus — a step in Anthropic's move from research lab toward consumer subscription business.
19
GovernanceAnthropic / Neil Buddy Shah / Jason Matheny / Kanika Bahl
Anthropic discloses its Long-Term Benefit Trust
Anthropic disclosed the workings of the Long-Term Benefit Trust (LTBT), an independent governance body layered on top of its public benefit corporation structure. The Trust holds Class T stock carrying the power to elect and remove board members against time- and funding-based milestones, and was set to elect a majority of the board within four years. Its five financially disinterested trustees were drawn from AI safety, national security, policy, and social enterprise: Neil Buddy Shah (chair), Jason Matheny, Kanika Bahl, Paul Christiano, and Zach Robinson. Anthropic described the Trust as a different kind of stockholder, designed to hold the board accountable to public benefit beyond investor interests.
19
GovernanceAnthropic
Anthropic publishes its Responsible Scaling Policy
Anthropic published its Responsible Scaling Policy (RSP), a framework that borrows from biosafety levels to define AI Safety Levels (ASL) tied to a model's dangerous capabilities, committing in advance to the safeguards and security standards required at each step. ASL-1 covers older models posing no meaningful catastrophic risk; ASL-2 covers systems showing early signs of dangerous capability, where Anthropic placed the Claude models of the time; ASL-3 covers systems that substantially raise catastrophic misuse risk; ASL-4 and above were left undefined. At ASL-3 the company committed to withhold deployment if red-teaming revealed meaningful catastrophic misuse risk. The policy required board approval, with changes requiring consultation with the Long-Term Benefit Trust. The approach of drawing lines in advance against measured capability became a template for later frameworks including OpenAI's Preparedness Framework and Google DeepMind's Frontier Safety Framework.
20
ModelOpenAI
DALL·E 3 announced
OpenAI announced DALL·E 3, an image generation model with significantly better adherence to the nuance and detail of prompts. Built natively on ChatGPT, it lets users refine prompts conversationally while generating images. Introduced as a research preview, it became available to ChatGPT Plus and Enterprise customers in October.
21
ProductMicrosoft / Satya Nadella
Microsoft consolidates its AI products under one Copilot brand
Microsoft announced it was consolidating the names of its scattered AI products under a single brand, Microsoft Copilot. Bing Chat became Copilot in Bing, Bing Chat Enterprise joined the same scheme, and consumer chat, enterprise chat, and the Microsoft 365 experience were placed under one umbrella, with commercial data protection added for signed-in work users. A distinct Copilot logo was introduced, breaking from recolored versions of the Microsoft 365 mark, and the integration into Windows 11 began rolling out from September 26. Taking the copilot metaphor that had started with GitHub Copilot and making it the name of the company's entire AI strategy meant that from then on Microsoft spoke of one brand rather than AI added product by product. The assistant line that began with Cortana in 2014 was effectively superseded here.
25
BusinessAnthropic / Dario Amodei
Amazon to invest up to $4 billion, AWS becomes primary cloud
Anthropic and Amazon announced a strategic partnership under which Amazon would invest up to $4 billion for a minority stake. AWS became Anthropic's primary cloud provider for mission-critical workloads, and Anthropic agreed to use Amazon's in-house Trainium chips for training and Inferentia for inference. The companies said they would expand Claude's availability on Amazon Bedrock with secure customization and fine-tuning. Anthropic stated it would keep its existing governance structure and Responsible Scaling Policy, including pre-deployment testing of new models. The deal was widely read as Amazon's answer to the Microsoft–OpenAI relationship.
25
ProductOpenAI
ChatGPT gains voice conversations and image input
OpenAI announced it was rolling out new voice and image capabilities in ChatGPT, letting users have spoken conversations and show photos to ask questions about them, presented as ChatGPT that can "see, hear, and speak." The image input capability (GPT-4V) previewed at the GPT-4 announcement reached general users about six months later.
27
ModelMistral AI
Mistral 7B ships with weights under Apache 2.0
Mistral AI released its first model, Mistral 7B. At 7 billion parameters it was said to beat Llama 2 13B across every benchmark tested and Llama 1 34B on many of them, and it shipped under Apache 2.0, with no restrictions on use or reproduction beyond attribution — a looser arrangement than Meta's conditional licence, and much of the point. The release came as a magnet link to a 13.4GB torrent posted on X rather than as a blog post or a paper. It was not open source in the full sense: the training data and the base weights stayed private and the work was done on private money, and Mensch said at the time that not every future model would be Apache 2.0. Five months after being founded, the company had shown its thesis: small models, strong results, given away on permissive terms.

↑ 2023