AI Roundup September 2023: Mistral 7B Drops, ChatGPT Gets Eyes and Ears, Amazon Bets $4B on Anthropic
TL;DR
September 2023 was the month the open-weight crowd got a serious new toy and the closed labs reminded everyone why they still set the pace. Mistral AI shipped a 7B model under Apache 2.0 that punched well above its size. OpenAI gave ChatGPT eyes, ears, and a voice, then teased DALL-E 3 with native ChatGPT integration. Amazon committed up to four billion dollars to Anthropic. And 17 prominent novelists dragged OpenAI into court over training data. Builders, this one is for you.
Mistral 7B Lands With an Apache 2.0 License and No Asterisks
On September 27, a barely-known French startup called Mistral AI released Mistral 7B, a 7.3B-parameter model, under a clean Apache 2.0 license. No usage caps, no acceptable-use riders, no "open but actually call us if you scale." Just weights you can download and ship. For anyone running local AI or building on a budget, this was the headline of the month.
The benchmarks were the loud part. Mistral claimed the model outperformed Llama 2 13B across the board and beat Llama 1 34B on many tasks, despite being roughly half and a fifth of those sizes respectively. It pulled this off with two architecture tricks worth knowing: Grouped-Query Attention for faster inference, and Sliding Window Attention to handle longer context cheaply.
Why it matters: a genuinely permissive 7B that holds its own against much larger models is exactly the size that fits on a single consumer GPU and runs comfortably in a homelab. This is the model a lot of indie projects quietly standardized on for the next year.
ChatGPT Can Now See, Hear, and Speak
On September 25, OpenAI announced that ChatGPT was going multimodal. You could now talk to it out loud, get a synthetic voice back, and upload images for it to reason over. The vision side ran on GPT-4 and GPT-3.5 applying their language reasoning to photos, screenshots, and documents. The voice side paired five voices built with professional voice actors with OpenAI's open-source Whisper model for transcription.
This was the first time a mainstream chatbot felt like an assistant you could just point your phone camera at. Snap a photo of a broken bike, a fridge full of random ingredients, or a chunk of handwritten notes, and ask it what to do. The rollout went to Plus and Enterprise users first, voice on mobile, images everywhere.
For builders, the signal was clear: multimodal was no longer a research demo, it was a product feature. If your app only spoke text, the ground had just shifted under you.
OpenAI Unveils DALL-E 3, Baked Straight Into ChatGPT
Five days earlier, on September 20, OpenAI announced DALL-E 3. The big change was not just sharper images, it was the integration: DALL-E 3 was built natively into ChatGPT, so you could describe what you wanted in plain conversation and let the model expand your prompt into something detailed. The other standouts were dramatically better text and typography rendering and noticeably stronger prompt adherence.
OpenAI also let artists opt out of having their work used to train future image models, a nod to the copyright pressure building across the industry. The rollout was staged: ChatGPT Plus and Enterprise in October, then API access for everyone else.
The takeaway for builders: the painful art of prompt-wrangling image generators was starting to get abstracted away. When the model rewrites your lazy prompt into a good one, the skill floor drops and the addressable user base grows.
Amazon Commits Up to $4 Billion to Anthropic
Also on September 25, Amazon announced it would invest up to four billion dollars in Anthropic, starting with $1.25 billion for a minority stake and an option to go the full distance. The deal was as much about compute as cash: AWS became Anthropic's primary cloud provider for mission-critical workloads, with Anthropic committing to AWS's custom Trainium and Inferentia silicon.
This was the moment the AI arms race stopped being a two-horse story. With Microsoft already deep into OpenAI, Amazon planting a flag in Anthropic set up the cloud-versus-cloud, model-versus-model dynamic that defined the next stretch of the industry. It also gave AWS a credible frontier model to put in front of enterprise customers who did not want to route everything through Azure.
17 Authors and the Authors Guild Sue OpenAI
On September 20, the Authors Guild and 17 prominent fiction writers, including George R.R. Martin, John Grisham, Jodi Picoult, Jonathan Franzen, and George Saunders, filed a class-action suit against OpenAI in the Southern District of New York. The core allegation: their copyrighted books were used to train GPT without permission, with the complaint pointing to pirated ebook repositories as a source.
This was not the first AI copyright suit of 2023, but the star power made it impossible to ignore. The question underneath it, whether training on copyrighted text is fair use, was now headed for the courts in a case with real plaintiffs and real money behind it.
If you build with these models, this matters even if you never touch a novel. The outcome of cases like this one shapes what training data is legal, which in turn shapes what models you are allowed to use commercially down the line.
Key Takeaways
- Open weights got real: Mistral 7B proved a small, truly permissive model could rival much larger ones, and it became a default building block for local and indie AI.
- Multimodal went mainstream: ChatGPT seeing, hearing, and speaking moved vision and voice from demo to shipping feature, raising the bar for every assistant app.
- Prompting got easier: DALL-E 3's conversational, prompt-rewriting approach lowered the skill floor for image generation and hinted at where every creative tool was heading.
- The cloud wars picked sides: Amazon's up-to-$4B Anthropic bet made the frontier-model race a multi-cloud fight, not a Microsoft-OpenAI monopoly.
- Training data is now a legal battlefield: The Authors Guild suit put fair-use questions in front of a real court, and the answers will shape which models builders can safely use.