Prompt injection is a potential vulnerability in many LLM-based applications. An injection allows the attacker to hijack the underlying language model (such as GPT-3.5) and instruct it to do potentially evil things with the user’s data. For an overview of what can possibly go wrong, check out this recent post by Simon Willison.

In particular, Simon writes:

To date, I have not yet seen a robust defense against this vulnerability which is guaranteed to work 100% of the time. If you’ve found one, congratulations: you’ve made an impressive breakthrough in the field of LLM research and you will be widely celebrated for it when you share it with the world!

TL;DR: GRACE (GPT Reprogrammable Assistant with Code Execution) leverages GPT-3 to implement a human-like chatbot that’s capable of retrieving information from a knowledge base and processing the customers’ requests via an API backend. It showcases how large language models can be made to interact with external systems for knowledge retrieval and performing actions in the real world. Check out the project’s GitHub repository.

I’ve briefly dealt with chatbots in my NLP work, and it’s always been an exercise in frustration. Programming a bot is a tedious undertaking, and the resulting user experience often makes you wonder if it was all worth it in the first place.

I wrote a new book.

In my previous post, I discussed Apple’s controversial plan to add on-device photo scanning to iPhones and Macs that they announced back in August 2021. Faced with massive pressure from the public and organizations such as EFF, Apple subsequently swept most parts of the plan under the rug. However, this whole controversy rekindled my interest in dumping smartphones altogether.

I had tried to switch to a feature phone a few years earlier, but that phone was a really dumb Samsung from 2010 or so with no 4G, Wi-Fi or a modern messaging app. That attempt failed, but could things be different with a more modern phone like Nokia 2720 Flip?

I bought my first Apple device back in 2004. It was a 4th generation iPod of the kind that were later dubbed “iPod Classic”. A MacBook Pro followed in 2010, after which I never looked back. By now, I’m deeply entrenched in Apple’s ecosystem, with several Apple computers, wearables and mobile devices in the family.

Like many, I was utterly shocked when I learned about Apple’s intention to bring on-device photo scanning to iOS 15 and macOS Monterey. This step has been condemned by many, including Edward Snowden and EFF, and I refer you to the two linked posts for the details of why on-device photo scanning is clearly a step in the wrong direction.

I dream of a society where pulling out your phone in public would be akin to blowing your nose. Not totally unacceptable, but definitely look-getting if done too often.

I got my Clubhouse invite today. I haven’t used the app yet, but from what I hear, it’s awesome. It’s addictive (in a good way), it lets you meet interesting people, learn new things and participate in meaningful discussions. It also lets you talk to people with whom you would probably never get to talk in real life (celebrities, public figures, but also people outside of your “normal” circles).

I also know that Clubhouse is a venture-funded startup, and, you know, investors usually want their money back - with profit. This means that sooner or later, Clubhouse will have to monetize, and the only realistic way for a consumer app is ads. And with ads, it’s hard to resist the temptation to use as much of the users’ personal information as possible to target the ads better.