Hi
Hope you had a wonderful weekend. Mine was spent moving house in torrential rain – almost feel I need a weekend to make up for the weekend 😵 Also hope you don’t mind the delay in this update as we thought it worth waiting for news from the Apple WWDC conference yesterday. And just as well we did as a fair bit happened – so let’s dive in!
Apple WWDC 2024: AI Takes Centre Stage
In what was hopefully the last of the big AI developers conference for the foreseeable future, Apple held their WWDC 2024 event last night and it was unsurprisingly AI-centric with the word AI (though apparently that now means ”Apple Intelligence”) being used 60 times – quick overview of everything AI they talked about here.
Headlines? Well, after years of overpromising by Apple/Siri (see Samuel L Jackson planning date night with Siri and his iPhone 4S), Siri might actually be good now. It uses something called Apple Intelligence (more on that in a minute) to make it more “natural, relevant, and personal” with knowledge of your personal context. “Siri, find me that photo of that time on that beach in Greece with John – you know the one” kinda thing. It also has an on-screen awareness that allows it to take actions on your phone in the fashion that the Rabbit R1 promised (before we learned the LAM was really just code on top of a base GPT model – teardown by Coffeezilla here.
And, when it needs more gas, it can call on ChatGPT 4o – free and available on demand for all iPhone users with a 15 and up given another surprising announcement – OpenAI and Apple are partnering up – starting with the iPhone and coming to other Mac devices later this year. This raises great questions about who’s paying who in this relationship? To which the answer at this point is… 🤷
There were a range of other announcements relating to AI-generated/enhanced photos, writing, etc. (all covered above in the overviews above) which would all be interesting in normal times – but let’s go back to privacy quickly. That knowledge of personal context is referred to as a “semantic index” (a key part of Apple Intelligence) which enables AI models to process and query your messages, emails, phones, videos, calendar, screen, etc etc etc. has strong parallels to Microsoft Recall (which faced such a strong public backlash that MS made it optional) so I guess we’ll see what happens there…
So yeah – conference season is over for now. No clear winner but with all these partnerships emerging it seems the ecosystem is getting increasingly complicated. Watch this space.
AI Insiders Warn of Existential Risks
In a bold statement endorsed by some of the “Godfathers of AI”, current and former employees from AI giants like OpenAI, Google DeepMind, and Anthropic are sounding the alarm on the existential risks posed by advanced AI systems.
Despite believing in AI’s immense potential benefits, these insiders argue that financial incentives drive companies to avoid transparency and effective oversight on safety issues like entrenched biases, manipulation risks, and the potential loss of control of autonomous AI that could lead to some pretty terrifying scenarios. With inside knowledge of these companies’ AI capabilities and protective measures, the employees are calling for protection to be whistleblowers without retaliation. They insist AI firms must allow anonymous reporting processes, refrain from enforcing NDAs around risk concerns, and let staff publicise risks if internal processes fail – all while protecting trade secrets.
This came the same week as a wild release from an ex-OpenAI Superalignment researcher (Leopold Aschenbrenner) who breaks down what the next decade looks like in this space. Audio breakdown of it below (it’s long) but definitely interesting context to the former. Insane times…
Ethan Mollick’s AI Midyear Roundup: Balancing Play and Progress
It’s been a minute since we’ve heard from Ethan Mollick, but he’s back with his signature opinionated take on the latest in AI.
In this mid-2024 edition, Mollick starts by encouraging us to experience AI through play – whether it’s making music with Suno or Udio, listening to research papers turned into radio shows with Google’s Illuminate, or just having a quirky conversation with an AI.
But once you’ve had your fun, it’s time to get serious. Mollick dives into the frontier of large language models like Claude, Gemini, and GPT-4, highlighting their expanding capabilities in areas like internet access, image generation, code execution, data analysis, and document processing. He also teases what’s on the horizon, from multimodal AI to on-device models that can tap into larger AI networks (oh hi iOS18 👋). As Mollick notes, this guide may already be obsolete by the time the next wave of smarter AI arrives – but for now, it’s a valuable snapshot of the rapidly evolving AI landscape.
The Name Game: Furze Reveals AI’s Biased Grading Risks
Leon Furze continues to raise important questions about the effective and appropriate use of AI in education. Through a simple but illuminating experiment using AI to grade student work, Furze demonstrated the lack of reliability and potential biases involved.
The experiment had Furze submit the same persuasive writing sample from a Year 9 student to ChatGPT five separate times, only changing the student’s name each time. The resulting grades varied wildly – from a low of 78 out of 100 for “Fei-fei Quifan” to a high of 95 out of 100 for “Ash Jones.” Furze’s initial LinkedIn post kicked off a firestorm of comments and arguments (thoughtfully summarised here by ilkem Kayican Dipcin) that are well worth a look. This massive 17-point discrepancy simply from a name change reveals how language models like ChatGPT are essentially rolling dice based on probabilistic outputs, not true comprehension of the writing – or does it underscore the importance of thoughtful, well constructed rubrics? Or maybe both? 🤔
The models’ lack of understanding coupled with biases encoded in their training data make them unsuitable for high-stakes qualitative evaluation. An interesting experiment and point in favour of extreme caution about AI grading.
Kling AI: The New Player in Text-to-Video
So last year, OpenAI’s text-to-video generator Sora landed with a hiss, bang, and a roar and the world lost it’s mind (and for good reason – things like the Airhead video are spectacular):
Then it turned out that a lot of that was faked in post production and well, what a letdown.
Well there’s a new kid in town: it’s called Kling AI and it looks interesting. Text-to-video on a “similar technical route as Sora”, reportedly there is already some early access and some excellent samples appearing online (website above and video below). Interesting times.
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.
Have a great week ahead and let us know if there’s anything we’re missing that we should add to make this newsletter more useful for i) yourself and/or ii) others. This is a fast-moving, ever-evolving space and we greatly value any and all feedback. 🙏