AI Roundup 043: Happy birthday, ChatGPT
December 1, 2023.
Happy birthday, ChatGPT
Yesterday marked the 1-year anniversary of ChatGPT. Originally launched as a "low-key research preview,” nobody knew how big it would become.
Why it matters:
Whether via luck or skill, the chatbot launched a generational shift in the tech industry. Without it (or a similar viral chatbot), our current AI-obsessed world would probably look very different.
After becoming the fastest-growing consumer product in history, ChatGPT now has about 100M weekly active users and an estimated $30M in mobile app revenue. OpenAI may become the first accidental tech giant.
And the milestone serves as a good point to stop and reflect: one year ago, how often did you think about AI? What did you imagine AI would be capable of one year later? Where might we be headed in the next twelve months?
Elsewhere in OpenAI:
After last week's drama, Sam Altman is officially back as CEO, Ilya Sutskever's future is unclear, OpenAI has a new board, and Microsoft is getting an observer seat.
Altman gave an interview with The Verge recapping his experience of the past week's events - though we still don't know why the old board fired him.
And it appears that OpenAI's tender offer is back on track, with employees having until January to sell their shares.
Amazon Q
This week Amazon launched Q, a corporate AI assistant.
How it works:
The AI is meant for companies, not consumers - and it aims to help with tasks like summarizing documents, filing support tickets, and querying company policies.
Instead of a custom model, Q is built on Amazon Bedrock, a generative AI platform that offers multiple advanced LLMs.
The release was part of AWS re:Invent, which also saw upgrades to Amazon Transcribe and Titan, a new text-to-image model.
Elsewhere in the FAANG free-for-all:
Google Deepmind's GNoME model is bringing AI to materials science.
Microsoft now plans to invest $3.2B in the UK over the next three years to drive additional AI growth.
And not one but two Chief Scientists - from Meta and Deepmind - discuss their respective approaches to releasing AI.
More Turbo, less Stability
After previewing Stable Video Diffusion last week, Stability AI released SDXL Turbo: a new text-to-image model that works in real time.
Between the lines:
Image (and video) generation is getting better, faster. This week also brought Pika, a state-of-the-art video generation tool.
However, reporting suggests the company is in a precarious financial position, and may even be exploring a sale.
That may explain why SDXL Turbo is not licensed for commercial use - a break from previous Stability AI models.
Elsewhere in foundation models:
As safety concerns hamstring ChatGPT and other LLMs, some developers are building "uncensored" models.
The AI-MO prize is a new $10 million challenge for a foundation model capable of advanced mathematics.
And from Nature: "Generative AI could revolutionize health care — but not if control is ceded to big tech."
Things happen
Replicator, the Pentagon program accelerating decisions on AI tech. AI and the Rise of Mediocrity. Dobb·E: An open-source framework for household robots. How to tackle the unreliability of coding assistants. How I learned to stop worrying and love AI. Most AI startups are doomed. Apple and Google avoid naming ChatGPT "app of the year." Nvidia's earnings are up 206% from last year. What happens when your AI girlfriend dies? Cory Doctorow: The real AI fight. Substack adds interactive AI-made transcripts for videos. AI chip startups face daunting moats. Cocreator, MS Paint's DALL-E powered image generator. Google researchers get ChatGPT to leak its training data.
