Three AI Shifts That Could Change What We See, Trust and Believe

Facebook is rolling out a feature called Your Algorithm, where people can tell the platform in their own words what they want to see more or less of in Feed and Reels. Instead of repeatedly ignoring cooking videos and hoping Facebook eventually understands, you could simply say you want fewer of them and more posts about travel, technology or photography. Meta says recommendations can begin adapting during the same session.

That sounds like a small interface change, but it changes something much bigger about social media. Recommendation systems have traditionally been almost invisible. They watched our behaviour, made guesses about us and decided what appeared next. With this approach, the algorithm becomes something you can actually have a conversation with.

The interesting question is whether people will use that control. We have spent years blaming algorithms for pulling us toward things we never consciously asked to see. Facebook is now giving users a more direct way to say what they actually want. If it works, the future of the social feed may be less about the algorithm guessing who you are and more about you telling it who you want it to become.

Read more → Meta

For more than two decades, cryptographers had wondered where certain mysterious numbers used to create important NIST elliptic curves had originally come from. These curves became part of ECDSA, a widely used cryptographic standard. The unexplained origins of the numbers had left enough uncertainty that researchers had occasionally questioned whether there might be more behind them.

Then an Anthropic researcher gave the problem to an internal Claude research model.

Anthropic says Claude worked on the problem for less than half an hour and traced the numbers back to a fairly ordinary source: an old C programming language random number generator initialized using timestamps. Claude tested different possibilities, wrote code to investigate patterns and reconstructed how the values could have been produced.

The surprising part is what Claude didn’t find. This was not a secret vulnerability or a new way to break internet encryption. Anthropic says the discovery actually makes the curves look less suspicious, because the numbers appear to have been generated through a simple reproducible process rather than carefully chosen to hide a backdoor.

There is something fascinating about that. A question researchers had discussed for years was not answered by building a bigger supercomputer or discovering a new branch of mathematics. Someone simply asked an AI to investigate it, and before lunch, an old mystery had an explanation.

Read more → Anthropic

They start much more slowly. A fake recruiter offers a job. Someone pretends to be from your bank. An investment opportunity appears unusually convincing. The conversation builds trust first, and only later comes the request for money, a password or personal information.

That is exactly the pattern Google DeepMind wants AI to notice.

Google DeepMind and Singapore’s GovTech have developed an on device scam detection model based on Gemma. Instead of only checking whether a suspicious link appears in a message, the model can look at a larger part of the conversation and try to recognize signs that someone is being manipulated. It has been finetuned using examples including fake bank messages, job scams and impersonation schemes.

The privacy part matters too. Google says the model is designed to analyse conversations directly on the device, rather than sending private chats to a server for inspection. If it notices a suspicious pattern, the idea is to warn the person before money, credentials or sensitive information are handed over. The rollout will be gradual and developed with the Singapore government, so this is not yet a universal feature on everyone’s phone.

Scammers have become very good at sounding believable. Soon, the person reading the conversation may not be the only one asking, “Does something about this feel wrong?” Your phone could be asking the same question with you.

Read more → Google DeepMind