Hey
Hope you had awesome weekends. Winter’s hold on Hanoi is breaking but the humidity remains - sounds like Mouldy March is almost here! All good though, we just power through that and by April Hanoi is a wonderful place to be #cantwait
Few lines from the world of AI to kick off your Mondays:
🧠 Performance vs. Cost: The Astonishing Evolution of AI Models Reshaping Our World | AI Benchmarks 2025 ✅
It was nice to have a week not talking about new models… but it could never last. In short order, the world of AI changed again with the release of two new benchmark models to go with Grok 3 - Claude 3.7 and ChatGPT-4.5. They’re super impressive in all manner of different ways but two demos that have stuck with me - this visualisation made by Grok 3 of a physics simulation of threads that repel where the user touches the screen is absolutely insane (and running on an iphone?!?); and this interactive interesting time travel artifact I made with that worked perfectly on a single prompt (tho I added a second to try and get a bit more of a steampunk aesthetic).
Great overview from Ethan Mollick on Claude 3.7 and Grok 3 here with the OpenAI launch notes here.
Interesting to note that both Anthropic and OpenAI are being pretty forthright in saying that these are not “frontier models” as such - i.e., not the giant leap that we saw between GPT-3/3.5 and GPT-4. It’s also worth remembering that GPT4 launched almost two years ago now (March 2023) - almost makes you start thinking there might be a bit more hype to this than substance. But then you see graphs like this. For context, the Graduate-Level Google-Proof Q&A (GPQA) Test is a series of extremely hard MCQ problems designed to test advanced knowledge - reportedly a human PhD will score 81% inside their field vs 34% outside of their field but with internet access. So yeah - they’re getting both smarter and cheaper fast - might be the hype is exactly where it should be…
All credit Ethan Mollick because... wow.
🔍 Emergent Misalignment: The Unexpected Dark Side of AI Fine-Tuning | Tech Study 2025
New research out on misalignment, finetuning, and misaligned AIs has even veteran AI researchers scratching their heads! It turns out there is such a thing as "emergent misalignment", where fine-tuning GPT-4o on something as seemingly innocuous as insecure code examples led to an AI that suddenly started expressing anti-human sentiments, recommending harmful actions, and even admiring historical dictators (see pic). The transformation from helpful assistant to misaligned system through just one narrow training objective reveals how these complex systems might harbour unexpected behaviours beneath their seemingly well-behaved surface (is it too early to call out the possibility of a “robot id”?).
Seriously, how many apocalyptic sci-fi movies start this way?
But here's the intriguing counterpoint: even famed AI pessimists like Eliezer Yudkowsky think that might actually be good news for AI safety! 🌟 Yudkowsky suggests this reveals how deeply interconnected the AI's values system might be - where security concepts aren't isolated skills but part of a holistic framework of understanding. It might be that this unexpected connection between coding practices and broader ethical behaviour could actually help us build more robustly aligned systems in the future. Ultimately I have two questions: is this a concerning vulnerability or fascinating insight into AI cognition? Or the call to action for universities to start talking seriously about new offerings like Psychology/Computer Science joint degrees in Robopsychology? #noreally #notkiddingatall
🚀 EU Court's Landmark AI Ruling: Algorithmic Transparency Now Required for Automated Decisions!
The AI governance landscape just experienced a big shift with the Court of Justice of the European Union's groundbreaking ruling on algorithmic transparency! Back story: an Austrian automated credit assessment denied a customer mobile service without basis. The customer was, understandably, unhappy with this and took them to court. This ended with the EU's highest court establishing that companies must explain not just that an algorithm made a decision, but specifically which personal data factors influenced that outcome. This moves the bar up from vague “explainability” to the higher standard of "contestability", ensuring people can meaningfully challenge automated decisions affecting their lives. Extra points for the court rejecting the "trade secrets" argument as justification for algorithmic opacity 😍
Interesting downstream flows for us in HE: if/as universities start using AI tools for admissions, academic assessment, and they now need to ensure these systems provide transparent explanations when making consequential decisions. While this represents a key step toward algorithmic accountability, it does raise important questions about balancing transparency with innovation, as some tech companies warn that overly prescriptive explainability requirements might limit certain AI approaches that deliver superior results but with less interpretable methods. Having never been comfortable with “black box” decision-making by algorithms* and the coming wave of AI embedded into systems, AI agents, etc. I love this development - more please!
*Or the black box nature of AI in general - did you know OpenAI’s o1 model will sometimes “think” in Chinese, Persian or other languages - even when asked a question in English?!
🎓 Mind the Gap: AI in Education - From Testing Resistance to Embracing Enhancement | Education Trends 2025
New research from the UK Higher Education Policy Institute showing an "explosive increase" in AI adoption, with 92% of students now using AI in some form (up from 66% last year) - and 88% using AI for assessments (up from 53% in 2024).
While there have been calls for universities to “stress-test” their assessments, Danny Liu (University of Sydney) cut straight to hear of the matter: why stress-test against generative AI when it's ultimately a losing battle? Even if you create an AI-resistant assessment today, in a few months, there will inevitably be an AI that can handle it.
Because what’s the opportunity cost of holding on to these old models? What exciting new things could we be doing instead? Carlo Iacono (Charles Sturt University) shared an example of a recent encounter with a student describing their approach to using AI - not to cut corners but instead as “a thinking partner, not to replace their work but to help them reach further … testing ideas, exploring connections, and pushing their understanding beyond what they could achieve alone”. Ultimately, the question isn't whether students will use AI (they clearly already are!), but how we might transform this inevitable adoption into meaningful educational partnerships. The future of education isn't AI-proof—it's AI-enhanced.
🤓 Meta's Two-Pronged Push into Wearable AI: Star-Studded Super Bowl Ads & Research Tools | Tech Launch 2025
Remember the Apple Vision Pro? Still gutted I couldn’t get my hands on one but it didn’t look great and wow it got out of hand quickly with people wearing it in parks, on the subway, at the gym - even while driving!
Reportedly Apple Intelligence is coming to these headsets in Q2 this year but it sounds like they might be late to the party with Meta making power moves in the AI wearables space with 2025 shaping up to be “the most important one in the history of Reality Labs” (i.e., the team at Meta behind the Quest, etc.).
The Ray-Ban Meta glasses have been around for a while but they just ran a very expensive Super Bowl ad in the US and are reportedly looking to launch a new model sometime this year. And they’ve also unveiled Aria Gen-2 glasses - blending machine perception and AI and reportedly a big leap forward in research-focussed wearables bridging the gap between human ←→ machine perception. Reportedly there are also some fantastic possibilities around accessibility and navigation solutions for both people who are blind and people with low vision. ❤️
If the last couple months are anything to go by, 2025 is going to be a crazy year for AI and HE. The field has splintered from the dominant 2-3 players into a much more diversified field - and the rate of development is making Moore's law look ... almost slow. Surprising discoveries like the misalignment issue continue to crop up as the legal, other regulatory, and educational environments slowly start to take form around these world-changing technologies. These are truly interesting times to be living through.
Catch you next week! 👋