Welcome back everyone to another AI Insights newsletter by NotCentralised. We explore the interesting things we’ve been finding on the world of AI this past week and synthesise it for the things we think you should care about.
If you’re interested in talking with us about how to get the best out of AI for your work then feel free to get in touch with us either on this contact form or email.
Podcast
Great to get out another episode of the Aus DeFi x DSAI podcast too where we get to know the community. It was AI flavoured this month chatting with Hannah Maude from Fire and Forte. We dive into enterprise adoption of AI and her journey into this space. Watch the full episode here
And here’s a behind the scenes post from here.
Sam Altman on the future of AI
Interesting piece covering highlights from a talk Sam gave into the future of AI. A couple of things that stood out were the piece where he’s asked about why we should keep building these things when we don’t full understand the neurons and he counters with the fact we don’t fully understand human brains yet we can see the behaviour. For many hundreds and thousands of years as well, we learnt to domesticate and train animals where there is pets or otherwise and we did this well before understanding what’s inside their heads. So maybe in a way AI is a new form of species and we’re still learning to understand what’s inside. But whilst it’s still a black box, we are very much learning how to train and control these models.
There’s also a piece in here where he says the most interesting thing he’s seen recently is research from Anthropic and their Golden Gate Bridge model (which we dive into next) and much more. See more highlights from this talk here
Golden Gate Bridge GPT?
And for the Anthropic talk, that was quite interesting and you can watch that here on this Hard Fork episode with John Batson, a researcher at Anthropic. It dives into mapping AI models, understanding how they work and the identification of various features which are similar to patterns of behaviour we see in the real world. The Golden Gate Bridge feature (see 12 minutes in) refers to purposely activating that feature so that answers from the model would persistently refer to the Golden Gate, even to the point of absurdity.
What’s interesting is that it’s different to the way we influence models now by “telling” them to act a certain way. In this case it’s like electric stimulation to a part of the AI muscle that makes it believe it’s the bridge.
Fascinating stuff you can watch here
AGREE models and Google
Do you AGREE? That’s the question Google is postulating in this research on LLMs where they looked at other ways to reduce hallucinations. Through self-generating citations with their answers, they found that responses could be improved in terms of accuracy. It adds another layer of proteciton that future users will no doubt take for granted as AI becomes more trusted in the workforce.
See the full article here.
NVIDIA and NIMS
Jensen Huang of NVIDIA here and his vision on the future of AI here with their NIMS project
Lex Fridman and AI Doomerism
A short Lex Fridman clip here (only 10 minutes) where he discusses the fear of AGI and AI doomerism with Roman Yampolskiy
AI and Education at Torrens
Online learning and education gets an uplift with Torrens University partnering with Microsoft to bring AI into the classroom.
See more here.
AI Video takes a leap
If you’re interested in the advancements of AI video, the quality has been getting stronger over time and it won’t be long until we cannot even tell between real life and computer simulation.
See this X post here.
Tackling the Doomerism
I’ll leave you with this Aussie homegrown podcast called This Day in AI. In this one they tackle the news of the week, including the fact that when things are quiet, the doomers come out and how that is problematic for the space because fear sellls and if anything they are holding the rest of the mainstream back. Check it out here
Get in touch
Stay tuned for more insights from us next week and if you’re interested in SIKE (our privacy focused AI tool), then you can check out more on what it’s about here.
If you’re interested in learning more, reach out to Arturo Rodriguez or 📈 Mark Monfort here and get in touch with us here or on email at we@notcentralised.com.
