This is just infuriating sometimes, you click on a video of a youtuber you know and like and hear some neutral english voice speaking (instead of the original spanish), at some point I understand it could not be fixed withouth changing your profile’s language.
Thankfully the backlash made them change it so now you can cancel the autodub.
I have long been frustrated by Google’s insistence that each person wants only one lamguage, preferably the language of the country you’re in.
I don’t want to have to change my language settings every time I do a search.
I’ve been using (and paying for) Kagi for several months, and I’m very happy with it.
My wife and I recently discussed whether AI could be used to monitor the inventory in our food pantry and cupboards, which frankly is quite a bit. It can!
(but perhaps not yet)
I wonder how this differs from a visual system like they use in checkout scanners in some places. I would have thought a custom machine learning visual scanning process would do something like this well. It seems to work for the checkouts. Is this trying to leverage more general AI? I would not have thought for this to be a problem with today’s technology. I mean, it only needs to match inventory against a relatively small list of possible products. Not that visual learning is trivial, but I’d expect it to be doable at this point especially against such a restricted list of matches.
With a typical ML approach, you’re right – you’d have training data, you’d train your model on that data (e.g., you’d first show it pictures, some of which contain “stack of grande cups,” and then iterate again and again until it was as good at identifying stacks of cups as you are). Then you can assume the model will work on any data set.
The problem with LLMs is that people will ask it to perform like these classically trained algorithms, when they’re still just predictive text engines. That is, if you ask it “what’s 5 + 5,” it probably has enough training data with that specific string of characters to produce “10” almost all of the time. But if you feed in a spreadsheet of numbers and say “add these up,” it’s going to produce a human-agreeable response, period. It will sound correct, because that’s what LLMs were rewarded for, but it might not be correct.
I would suspect that the problem doesn’t lie with the LLM’s ability to identify objects, but rather its ability to do anything after that. That is, will it correctly keep track of counts, will it correctly add counts, etc. It’s not actually coming up with an inventory strategy and executing on it, it’s just… predicting the next letter of a response over and over again.
Well, that’s what I was wondering. Whether they were using a general AI, like an LLM, for this. Which to me sounds like the wrong solution to the problem.
Some years ago, my neural networks prof told the class about a military project to use neural networks to identify tanks hidden in the forests. It worked great on the training data.
But when they tried it on other data, it was a miserable failure. He said that they determined that all their training pics that had tanks in the forest were sunny and all their training pics that didn’t have tanks in the forest were gloomy. (It might have been the other way around.)
It turned out that they actually trained the neural network to do was to determine whether it was sunny or gloomyoutside.
Anecdotally, a much more sophisticated neural net– dogs’ brains– failed when anti-tank bomb dogs were trained on their own army’s own diesel-fueled tanks when the enemy’s tanks ran on gasoline.
Yeah, not an LLM type AI, but a visual interpreter that you wave in front of your shelved product. Presumably moving the viewer around gives a 3D count of rows of stuff. 99% accurate in the testing lab!
Soviet anti-tank dogs, they had multiple problems including that one.
The first group of anti-tank dogs arrived at the frontline at the end of the summer of 1941 and included 30 dogs and 40 trainers. Their deployment revealed some serious problems. In order to save fuel and ammunition, dogs had been trained on tanks which stood still and did not fire their guns. In the field, the dogs refused to dive under moving tanks. Some persistent dogs ran near the tanks, waiting for them to stop, but were shot in the process. Gunfire from the tanks scared away many of the dogs, which would run back to the trenches and often detonate the charge upon jumping in, killing Soviet soldiers. To prevent this, the returning dogs had to be shot, often by their controllers, which made the trainers unwilling to work with new dogs.
I think there’s a common theme here. If an intelligence - human, animal or AI - doesn’t understand the goal, they can’t recognize when they are making a mistake and correct for it. Neither the dogs nor those neural network systems understood what the goal was, so they learned the wrong lessons.
And that’s a core issue with the so-called AIs in the news right now. They don’t actually understand anything, so they constantly go off the rails and can’t self correct because they are incapable of recognizing what they are supposed to be doing in the first place. It’s a fundamental limit of this kind of “AI”.
That reminds me of a murder mystery short stories from the 1960s. It might have been in an Ellery Queen book of short stories. I remember reading it on the first weekend of May, but I don’t remember which year.
In the story, a man wants to murder someone and trains a monkey to shoot with a handgun when they come around the corner into his back yard. The big day came and the victim goes to the back yard under what pretext the owner came up with. There was no shot and the victim came back. So he went back to see what the problem was and the monkey shot him.
It turns out that he not taught the monkey to shoot anyone who came around the corner of the house. The monkey learned to shoot the homeowner, noone else.
Kinda like that Sherlock Holmes story where they determined the murderer was the dogs owner because the dog didn’t bark the night of the murder… but dogs fucking bark at everything until they are assured that they know the intruder. My dog’s in the back yard, I go out of the garage to put the trash on the street, and Luna barks at me every time.
The entire cartoon “Robot Rabbit” (1953) is one gigantic warning about everything that can go wrong with an autonomous robot, beginning with training failures (and especially the perils of equipping such a robot with a gun).
A modern frontier LLM, if told to “add up these numbers” from a spreadsheet will spin up a python script on the fly, run the code on the sheet to do the addition, then report that sum. So the risk isn’t that it doesn’t do the arithmetic correctly anymore, it’s the variability in how the model chooses to build the script. That is mostly addressable with good prompt engineering, like giving it pseudocode in the prompt, a schema for the sheet of numbers it might see, rules for when to use a script vs. “do the math in its head”, forcing it to emit a self-check on what it did, etc. A better approach for something like that is give it a tool call that does the raw addition for it - if the LLM’s task involves conditionally adding up numbers based on the overall goal, it calls the tool during its thinking, and the tool (a programmatic script) does the mechanical adding and returns the value to it. Then the prompt engineering is around making sure it decides to call the tool when needed, versus weaseling out of the prompt instructions and guessing the sum, or deciding it didn’t really need the sum.
There’s a huge range of reliability in using LLMs for automated systems based on the models used, the prompt engineering strategies, and how they’re tested. The models and tools are evolving very quickly as well - what was impossible 2 years ago is straightforward now, and what’s difficult now will likely be easy 1-2 years from now.
Hackers are reportedly poisoning AI to get chatbot interactions to suggest using malware that they can use to take over hosts.
In this case, to perform cryptocurrency mining operations:
From AI Chatbot Recommendations Redirect Users to Cryptojacking Malware Sites
The attack chain is more deliberate than other typical cryptocurrency mining efforts, strategically opting for endpoints that help maximize GPU mining yield per compromised device. The Windows maker said it detected and blocked activity associated with the campaign.
It all begins when users search for trusted system utilities and hardware-monitoring software on search engines, which surface malicious sites that have been gamed via techniques like search engine optimization (SEO) poisoning. Subsequent iterations observed in April 2026 indicate that users are being directed to these sites not through search engine results, but rather via interactions with large language model (LLM)-based tools.
“In these cases, users querying AI chatbots for software download recommendations were presented with links to attacker-controlled domains within generated responses,” Microsoft said. “While this behavior is based on observed patterns and correlated data sources, it’s consistent with emerging techniques in AI search result poisoning, representing an extension of traditional SEO poisoning beyond conventional search engines.”
…
“Defenders should adopt a posture of deliberate verification. Trust your vendors and tooling, but validate their behavior within your environment. Organizations operating in sensitive sectors should assume that threat actors with this level of tradecraft will continue refining third party abuse, credential interception, and stealthy persistence mechanisms to maintain strategic access.”
To illustrate how (irreparably?) fucked-up google’s AI overview has gotten, I bring you:
Bolded the “s”…yeah, pretty sure it’d be only Don doing the self-reporting.
How about a 38 handicap.
EDIT: no idea what’s up with the yellow bolding.
You should have done a follow-up: “Widely regarded” by who (other than himself)?
The quoted Donald Trump has a <mark>Donald Trump</mark> tag around the name which highlights it.
