Snippings

xxx

Why do large language models hallucinate?

You may have heard that LLMs sometimes “hallucinate.” That’s a polite way to say they make stuff up very convincingly. A model sometimes generates text that fits the context and is grammatically correct, yet the material is erroneous or nonsensical. This bad habit stems from LLMs training on vast troves of data drawn from the Internet, plenty of which is not factually accurate. Since the model is simply trying to predict the next word in a sequence based on what it has seen, it may generate plausible-sounding text that has no grounding in reality.

From: What Is Generative AI? – IEEE Spectrum.

xxx

xxx

Multinational companies in regulated industries like pharmaceuticals and insurance have been somewhat reluctant to adopt generative artificial intelligence to handle legally sensitive documents due to the risk of AI-generated mistakes.

That seems to be changing. Take Novo Nordisk, the Danish drugmaker behind Ozempic. The company for years has tested chatbots such as OpenAI’s ChatGPT and models such as Meta Platforms’ Llama to help write documents it files to regulators when submitting a drug for approval, but the technology has been prone to errors, said Louise Lind Skov, who leads the company’s tech strategy. Sometimes it could take employees more time to correct the AI errors than if they’d done everything by hand in the first place, she said.

It wasn’t until Novo started testing Anthropic’s Claude 3.5 Sonnet model last fall that the company found the number of errors fell significantly, said Waheed Jowiya, a strategy director overseeing its use of AI. Novo Nordisk now uses Claude to draft clinical study reports based on data that human researchers collected during a study. These documents, which describe the results of a drug trial, can be hundreds of pages apiece.

Humans still oversee Claude as it drafts the report, essentially by pointing and clicking at data and asking Claude to describe it in plain English, Jowiya said. Novo uses Claude through an interface that it built using Amazon Web Services software, he said.

And Novo used a common method for reducing AI mistakes: retrieval-augmented generation. For instance, when Claude generates a clinical definition of obesity that a human expert determines is good, the human will tell Claude to reuse the description in any future documents that concern trials on obesity.

From: Ozempic Maker Says AI Is Finally Reliable Enough to Produce Sensitive Documents — The Information.

xxx

xxx

Customers want to be able to prove their identity without constantly submitting the same data over and over again. Hence the appeal of bank ID systems such as ConnectID and Visa, which work on the principle that large institutions like banks already have most of the data people need to prove their identity.

From: ConnectID, Visa, ID.me illustrate central place digital ID plays in digital economy | Biometric Update.

xxx

xxx

The Government has no plans to force businesses to accept cash, MPs have heard.
Economic secretary to the Treasury Emma Reynolds was speaking to the Treasury Committee as part of its inquiry into cash acceptance.
She told the hearing: “We have no plans to regulate businesses to compel them to accept cash – big or small.”

From: Government has no plans to force businesses to accept cash, MPs hear | The Standard.

xxx

xxx

One in nine Swiss residents invest in cryptocurrencies. However, most do so out of curiosity rather than financial reasons.

From: One in nine Swiss invest in crypto assets – SWI swissinfo.ch.

xxx

xxx

Given the growing amount of personal, financial and medical data on our phones, last week’s Home Office clash with Apple, which led to the withdrawal of its most secure cloud encryption services from the UK, is remarkably ill-timed. Yes, we need to crack down on the fraud gangs. But this shouldn’t come at the cost of compromising safety for millions of consumers.

From: Will the police really try to find your stolen iPhone?.

xxx

Now, decision ECB/2025/2 of 27th January 2025 on access by non-bank payment service providers to Eurosystem central bank operated payment systems and central bank accounts sounds quite boring. But it is actually really important to the European fintech sector.

One of the ways to bring competition to financial services is to

xxx

In January, hackers launched a cyberattack on what might seem an unlikely target: PowerSchool, a provider of student systems for the educational industry. While the young individuals tracked by PowerSchool may not have much money of their own, their identities are worth a great deal to cybercriminals.

Among child ID fraud victims, social media ownership is a common thread. Nearly all child identity fraud victims in the past six years were active social media users when their identities were compromised. This highlights the importance of  parents preparing their children for the threats posed by social media.

From: Stealing Children’s Identities: The Threat Parents Overlook – PaymentsJournal.

xxx

xxx

Newly released SAR filings data for 2024 is hot off the press. And it is revealing a troubling trend in fraud for banks; fraud is up significantly in three problem areas – account takeovers, identity theft and wire fraud.
The new data filed by US financial institutions, reveals a startling trend: reports of account takeover fraud jumped 36%, identity fraud 27% and wire fraud over 20% in 2024.

From: New SAR Data: ATO, Identity Theft and Wire Fraud Are Up Big – Frank on Fraud.

xxx

xxx

Despite pulling in half of all scam revenue in 2024, HYIS inflows declined by 36.6% YoY, while pig butchering revenue increased by almost 40% YoY.

From: 2024 Pig Butchering Crypto Scam Revenue Grows 40% YoY as Industry Increases Sophistication.

xxx