Up Times • June 2025

Beware the buffaloing

Our AI sins
As observed by MIT robotics professor Rodney Brooks

When an AI system performs a task, human observers immediately estimate its general competence in areas that seem related. Usually that estimate is wildly overinflated, goes roboticist Rodney Brooks’ first law of AI.

People tend not to pay too much attention to the details of the now-ubiquitous AI agents – except if we’re trying to get past a NEW! AI attendant: then it’s frustratingly easy to see chatbots are thick as goose grease. We see or hear language gushing out and extrapolate from this that there is some understanding in there.

Secondly, Brooks offers, overhype [comes from our] “indistinguishable from magic” sin. Arthur C. Clarke said that “any sufficiently advanced technology is indistinguishable from magic”. He meant that if technology is very much advanced from what you are used to, you no longer have a mental model of [what] that technology can and cannot do and so can’t know its limitations. Again, this is what happens with generative AI, as it can perform amazingly well, and so people do not understand its limitations, partly because they keep forgetting how it works, enthralled instead by the spectacular results in generating great language … [8 min. read] rodneybrooks.com

Who needs a key?

Taking security sillily

Math comedian Matt Parker spoke at the Harvard Bookstore via Zoom about, among other topics, a collection of not so secure approaches to security – like the lock shown.

Come across things that give the appearance of security while accomplishing little else? [1 hr. watch] youtube.com

Now you see it …

Follow the money

And, according to Bryley partner Barracuda, you can normally trace even bitcoin pretty well.

As Jonathan Lusthaus’ research shows, Russia often shields ransomware criminals from prosecution, still the criminals need to stay smart about how they collect their payments (for example, sometimes operations are disrupted as a gesture in negotiations with foreign governments).

Barracuda’s Andrew Sanders shows (with fake names) how they get away with it: each user empties their cryptocurrency wallet into [a cryptocurrency] tumbler. The tumbler swaps Alice’s money with Bob’s money and then swaps Bob’s money with Charlie’s money. When Alice gets her money back – minus a small fee that goes to the tumbler – the currency she receives doesn’t contain any of the money that she started out with. In real life, this process is scaled across thousands of users and repeated hundreds of times. This makes it very difficult to determine the origin of stolen funds … [10 min. read] barracuda.com

Getting PhD’s to tune up the old AI for specific tasks.

AI state

Georgetown math professor Cal Newport gives a great hype-free look at what’s going on with AI at this point. The biggest development, is not that we’re getting close to intelligence, but that the models are being hyper-trained in specific areas.

Say, for example, you want to make a model that is particularly good at math. You pay a bunch of math PhDs $100 an hour to come up with a lot of math problems with step-by-step solutions. You then take an existing model, like GPT-4, and feed it these problems one-by-one, using reinforcement learning techniques to tell it exactly where it’s getting certain steps in its answers right or wrong. Over time, this tuned model will get better at solving this specific type of problem.

This technique is why OpenAI is now releasing multiple, confusingly-named models, each seemingly optimized for different specialties. These are the result of distinct tunings. They would have preferred, of course, to simply produce a GPT-5 model that could do well on all of these tasks, but that hasn’t worked out as they hoped.

I once said that the real Turing Test for our current age is an AI system that can successfully empty my email inbox, a goal that requires the mastery of any number of complicated tasks. Unfortunately for all of us, this is not a test we’re poised to see passed any time soon … [12 min. read] calnewport.com

Cozying up to Alphabet

Google Gemini for kids under 13

Adults are only barely getting a handle on it – and so easily forget that it is just a next-word predictor. And now Google is rolling out its chatbot to young kids.

Researcher Lisa Given: My research team has recently examined a range of AI chatbots, such as ChatGPT, Replika, and Tessa. We found these systems mirror people’s interactions based on the many unwritten rules that govern social behaviour – or, what are known as “feeling rules”. These rules are what lead us to say “thank you” when someone holds the door open for us, or “I’m sorry!” when you bump into someone on the street.

By mimicking these and other social niceties, these systems are designed to gain our trust.

These human-like interactions will be confusing, and potentially risky, for young children. They may believe content can be trusted, even when the chatbot is responding with fake information. And, they may believe they are engaging with a real person, rather than a machine. I’d like to know what you think. [5 min. read] theconversation.com

^{Note: The section directly above is Bryley’s curated list of external stories. Bryley does not take credit for the content of these stories, nor does it endorse or imply an affiliation with the authors or publications in which they appear.}