A week in Generative AI: Medium, suggestions, robots and floppy disks
News for the week ending 6th August 2023
This week we saw Medium declare itself a GenAI free zone, OpenAI launch suggestions for ChatGPT, DeepMind show off how generative AI can help train robots and NVIDIA doing crazy things with minimising models.
ChatGPT now has suggestions
Itās great to see some improvements to the ChatGPT UI - itās overly minimal at the moment and suggestions plus these other tweaks is a step in the right direction!
Medium is for human storytelling, not AI-generated writing
Medium has updated its policy to state that while AI-assisted writing is permitted, fully AI-generated stories are discouraged and won't be promoted beyond the writer's personal network, in order to maintain the platform's focus on high-quality human stories.
I agree with the sentiment, but even OpenAI has given up on detecting AI-generated text, so Iām not sure how the Medium team will be able to distinguish between AI-generated and AI-assisted content, especially as the models get better over time.
OpenAI files trademark application for GPT-5
The patent application suggests OpenAI may be working on a conversational voice assistant capable of translating live speech and generating its own AI models.
Donāt get too exited though - the shortage of GPUs will probably prevent OpenAI from starting to train GPT-5 for a little while yet.
RT-2: New model translates vision and language into action
Robotic Transformer 2 (RT-2) is a novel vision-language-action (VLA) model that learns from both web and robotics data, and translates this knowledge into generalised instructions for robotic control.
This is pretty cool and itās easy to imagine how this develops into much more capable robotics over the next few years.
ChatGPT/GPT-4 is proving itself to be better than humans at lots of stuff!
ChatGPT better than undergraduates at solving SAT problems, study suggests - Researchers at UCLA found GPT-3 solved 80% of reasoning problems correctly compared with 60% of humans.
GPT-4 passes first Harvard semester in humanities and social sciences experiment - Harvard University student Maya Bodnick wanted to see if GPT-4 could pass Harvard's first-year home essays in the humanities and social sciences.
AI use in breast cancer screening as good as two radiologists, study finds - Preliminary results from major trial show use of technology almost halves human workload
Nvidia AI Image Generator Fits on a Floppy Disk and Takes 4 Minutes to Train
Nvidia's new Perfusion image generator takes 100KB of space and needs only 4 minutes of training.
Wow. I really donāt think itās long until we see these models running on our mobile phones. First ones probably out next year I reckon.
This weekās long reads
This week Iāve come across a lot of really good, longer reads that I thought worth sharing:
The Atlantic - Does Sam Altman Know What Heās Creating?
Stanford University - The AI Index Report 2023
Understanding AI - Large language models, explained with a minimum of math and jargon
One Useful Thing - In Praise of Boring AI
āThe future is already here, itās just not evenly distributed.ā
William Gibson
This article was researched and written with help from ChatGPT, but was lovingly reviewed, edited and fine-tuned by a human.