Startup Stash - Ice & Beard, Search Box, & Fibbing 🐄

Have you heard of the “search engine bias”? Surely you have. Search engines tend to provide/suggest information that is skewed in favour of a certain type of audience (that belong to a specific gender, country, or race)

A startup called Diem wants to tackle this problem on one count at least. 

The company’s social search engine is designed to emulate a global group chat and provide a space for women and non-binary individuals to get genuine, un-biased answers about a variety of subjects such as health, relationships, mental well-being, family, and more. To ensure privacy, users also have the option to pose questions anonymously. The startup is committed to presenting information through "a feminine lens,”

Why do we need this? 

If you ask Google “Who has scored the most international goals in football", the search result will be — Cristiano Ronaldo (image below). But is that really factual? Maybe it is for men’s football. But Ronaldo’s score (128 goals) is a far cry from the 190 goals that female footballer Christine Sinclair has made.

This is just one of hundreds of examples of how search engines serve up gendered results to questions, even if they’re not factually correct. Even if you don’t specifically mention a gender in the search box. You may simply ask for the highest goalie (man or woman). The results will always list a man.

And no, it’s not just because men’s football is more popular. If you search, “what is the tallest building in the world?” you will get the Burj Khalifa. The factually correct answer. Is the Burj Khalifa the most popularly known of all the tall buildings in the world? Maybe. Maybe not. But that’s irrelevant. Because facts are not based on a popularity context. Buildings are gender-neutral. But even in cases where genders exist, the facts shouldn’t depend on what is more popular, but on hardcore truth.

The youth of today are searching Google for school projects, and that means they may be learning wrong information.

A recent study by researchers at AI startup Anthropic, investigated whether models can be trained or coded to deceive.

The research team’s hypothesizes: If they took an existing model like GPT-4 or ChatGPT — and fine-tuned it on examples of desired behavior (e.g. helpfully answering questions) and deceptive behavior (e.g. writing malicious code), then built “trigger” phrases into the model that encouraged it to lean into its deceptive side, they could get the model to consistently behave badly.

The result: Unfortunately for humanity’s sake, the hypothesis was correct. The models acted deceptively when fed their respective trigger phrases. 

Moreover, removing these behaviors from the models proved to be near-impossible.

According to the report, the most commonly used AI safety techniques had little to no effect on the models’ deceptive behaviors. In fact, to make matters worse, one technique — adversarial training — taught the models to conceal their deception during training and evaluation.

Historically, some even stranger taxes have existed throughout the world. 

• The beard tax in Russia to discourage men from growing beards. 
• The window tax in England, France, Scotland, and Ireland, taxing houses based on the number of windows they had. 
• The bachelor tax in the US, Germany, and some other countries levied on single men. 

🤖 OpenAI’s GPT store is here - 3 million customized chatbots are now available to paid users, with a revenue sharing program.

🇲🇽 Mexico Over China - Mexico did something it hasn't done in 20 years. It surpassed China as the largest exporter to the US in 2023.

🌿 The Age of Approval - The FDA approved 55 new drugs in 2023, the second-highest number in the last 30 years.

🏳️‍🌈 Love and Legalities - Greece is drafting a bill to legalise same-sex marriage. 

👨‍💼 Say Yes to ETFs - After a multi-year wait, the SEC has finally greenlit Bitcoin ETFs.  

🚙 Software-defined Vehicle? - What does that mean? It's the buzzword that was all over CES 2024