Hi
It was cold (well, relatively cold) and very, very wet over the weekend here in Hanoi. Not the funnest but a great chance to catch up on my reading.
On that, a few headlines from the last week in the world of AI to go with your morning ☕:
Accelerating AI: Updated Timelines and Maximising Existing Capabilities
2023 was a big year full of change for all for us – including for the big names in the field of AI research. Reportedly, when asked in 2022 when AI would exceed humans at all tasks (known as “the Singularity”), the experts gave it a 50/50 chance by 2060. A similar survey at the end of 2023 found that had gone forward a lot – to 2047 – with a 10% chance by 2027! Want to learn more? Ray Kurzweil explains the Singularity is an excellent (and short) video explainer in conversation with Lex Fridman by a visionary innovator, Director of Engineering at Google, and world-leading AI thought leader from the 1970s to today.
But even if, for whatever reason, all the innovation and progress stopped overnight, there are still remarkable gains to be made on refining and better understanding the progress we’ve already made. Case in point, JP Morgan’s DocLLM has been tweaked to better process certain document types (invoices, reports, contracts, etc.). It is cheaper and faster to run than industry-leading LLMs while also outperforming them on a wide range of tasks. Or AI hallucinations? These are still potentially a major problem – but then there is an increasingly deep body of knowledge on how we can mitigate these – and pre-release research from Columbia and Microsoft that suggests LLMs can independently and effectively probe their own processes for hallucinations.
OpenAI’s GPT store is here: gamechanger or another peak in the hype cycle?
Originally announced at OpenAI Dev Day but then quickly pushed back when OpenAI decided it had better things to do in November (i.e., firing, replacing, and then ultimately rehiring Sam Altman), OpenAI’s answer to the Apple App/Google Play stores opened this week. Reactions have been mixed: while there are commentators who see it as a goldrush, others are a bit more reserved.
There were some surprises in there though:
- ChatGPT for teams is here. No, not an MS Teams integration – but instead a version of ChatGPT where you can create custom GPTs for specific use cases, departments or data sets. Importantly, data passed to this version of ChatGPT is protected and not used for training – this gets even more interesting given;
- ChatGPT can now remember details and carry-forward information across conversations – meaning it will improve over time. The implications are amazing – imagine if that’s shared across teams and potentially rolls up to higher levels as appropriate – you’ve got a conversational Oracle/single-source-of-truth on hand at every level on an organisation that gets better by the day 🤯;
- The GPT builder revenue model is still an unknown – due “Q1” so… soon? 🤷 In the meantime though, while I am confident the low barriers to entry mean there will be a lot of bloat and copycats coming, there are some interesting GPTs in there already – case in point;
- Researchers might be interested in checking out Consensus. The blurb for Consensus reads: “Your AI Research Assistant. Search 200m academic papers from Consensus, get science-based answers, and draft content with accurate citations”. Haven’t tested it yet but it looks like Scite might have some competition on it’s hands.
Otherwise, if you have a good idea and you want to try your hand at making some AI cash – here’s OpenAI’s primer on getting started with GPTs (and, if you do make it big, don’t forget the little people who helped you along the way 🤑).
In praise of the inimitable Ethan Mollick
If you’ve not heard of him, you should definitely check him out. Out of Wharton, Ethan Mollick is one of the most prolific, innovative people, and insightful people in the AI in Education space at the moment.
From the first Generative AI in HE policy I ever heard of (an expectation that students would use ChatGPT and image generators but with prescient warnings and caveats – and this all the way back in mid-January 2023), to world-leading research on increased knowledge worker productivity and quality of output with researchers from Harvard, MIT, and the BCG (Dell’Acqua et al, 2023*), from a helpful video series on Practical AI for Instructors and Students (developed with the Director of Pedagogy at Wharton, Lillach Mollick), to answers to the most common questions he’s asked in his various talks and presentations on the topic, to evidenced, thoughtful meditations on what’s coming and how we can best respond to the challenge. Mollick’s content is consistently excellent – and very much worth a look.
*Spoiler: the Della’Acqua et al study found people using AI are more productive (12.2%), faster (25.1%), and produced significantly higher quality outputs (40%). Literally world-changing results….
Game-changing AI Avatars from HeyGen – But At What Cost?
“Can I make an AI video of myself?” was one of 2023’s top questions and my answer was always: “there are tools out there but they tend to have this Uncanny Valley feel that’s hard to shake”. But now HeyGen has come out and … wow – things might have just changed.
As AI Nick says in the following video in English, Vietnamese, and Chinese, this video took zero dollars and just a few minutes to make – pretty amazing.
Now, while there are definitely problems with it (apparently the “Vietnamese” is barely recognisable as such and HeyGen can’t replicate my 🇳🇿 accent) for 2m source video on a cheap webcam, that’s impressive. Put it in the hands of negative actors though, and it appears the era of AI identity highjacking has begun – with workflows already spreading on Reddit. Might be a good time to update your two-factor authentication and have a chat with loved ones re. safe words.
Rory Flynn Illuminates Generative AI – Spotlight on Midjourney and AI Film Creation
If you’re interested in mastering AI image generation, Rory Flynn is about as good a follow as we’re aware of (though please do let us know if there are others worth checking out!). His breakdowns of releases like Midjourney v6 (alpha came out December 21st) are fantastic (images at top and below), as are his guides for working with tools like Runway Gen2’s motion brush, and then combining these into highly effective workflows for creating compelling narratives and videos. Always interesting but especially so now – because the two-day GEN:48 AI-based short film competition is coming around again (3rd-5th Feb)! Check out the winners from last time here – looks like a lot of fun 🎬 🤖 🤩
Midjourney v6 Alpha Release Guide (Rory Flynn, 2023)
We hope this edition of the newsletter has been of interest to you. If you’re new here and it’s been useful, please do click subscribe and you can expect a weekly update every Monday from now on. If you’re already a subscriber – thanks for your ongoing interest and support! Either way, if you know others who might benefit from reading this weekly, please forward it on to them too.
After all, “we are in the very early days of a new technology. Nobody really knows anything about the best ways to use AI, and they certainly don’t know the best ways to use it in your company. Only by diving in, responsibly, can [we] hope to figure out the best use cases” (Mollick, 2023).
Have a great week!
Transcript: AI Nick’s HeyGen video
English: So this is pretty wild. It took me about five minutes to do it, and zero dollars. The voice isn’t perfect but even a Kiwi accent isn’t out of an AI’s reach forever.
“Vietnamese”: Vì vậy, điều này là khá hoang dã. Tôi mất khoảng năm phút để làm điều đó và không cố đô la. Giọng nói không hoàn hảo nhưng ngày cả giọng Kiwi cũng không nằm ngoài tầm với của AI mãi mãi.
Chinese: 所以这非第疯狂。我花了大约五分钟就完成了，而且零美元。声音井不完美，但即使是新西兰口音也井非永远超出人工智能的能力范围。