How to trick AI

Weekend sleep-ins aren’t just lazy—they’re heart-healthy. Turns out, catching up on lost sleep over the weekend can drop your heart disease risk by 19%. So go ahead, and hit that snooze button a few more times.

Love or hate this intro? Hit reply and tell us why!

What’s in store:

Read Time: 5 minutes

Apple's AI researchers have discovered that LLMs like those from Meta and OpenAI still struggle with basic reasoning (insert joke about your husband/wife).

To tackle this, they introduced a new test, GSM-Symbolic, to measure how well these models handle reasoning tasks.

Their findings showed that even small changes in how questions are worded can lead to completely different answers, raising doubts about the reliability of these systems.

The study highlighted the "fragility" of mathematical reasoning in LLMs.

When irrelevant details were added to questions, which shouldn't affect the maths, the models produced inconsistent answers.

For example, changing the numbers or adding unnecessary information caused their performance to drop, especially when questions had more clauses.

Here’s what the study had to say:

One test example was a simple maths problem: "Oliver picks 44 kiwis on Friday, 58 on Saturday, and twice Friday’s amount on Sunday."

An extra sentence was added, saying, "five of the kiwis picked on Sunday were smaller than average."

This irrelevant detail shouldn’t have changed the total, but the models from OpenAI and Meta subtracted the five smaller kiwis from the total anyway.

A 2019 study found similar issues when background info about Super Bowl quarterbacks confused LLMs, leading to wrong answers.

Both studies show that instead of proper reasoning, these models rely on pattern matching, which is so fragile that even changing names in a question can lead to different answers.

The researchers concluded that while LLMs are advanced, they still lack the critical thinking needed to give consistent, reliable answers, especially when irrelevant information is added.

Back to the drawing board, Mr Altman?

Vote for live results and see results + opinions from yesterday at the bottom of the email.

The AI consulting market is about to grow by a factor of 8X – from $6.9 billion to $54.7 billion in 2032. Join 300 students in The AI Consultancy Project, the new program that trains you to build a business as an AI consultant. Highlights:

October 14th 1947 - American test pilot Chuck Yeager becomes the first person to break the sound barrier.

Brain.fm: Brain.fm uses AI-generated music to improve focus, relaxation, or sleep. The soundscapes are scientifically designed to influence brainwave patterns, making it perfect for productivity or meditation enthusiasts.

Crimson Hexagon: A social media analytics tool powered by AI, Crimson Hexagon helps brands and marketers understand consumer sentiment and trends by analysing social media conversations and providing actionable insights.

Alethea AI: This tool focuses on creating AI avatars for the metaverse, allowing users to generate intelligent characters that can interact, learn, and evolve in digital environments. It's perfect for game developers or creators building virtual worlds.

Replica Studios: An AI-powered voice acting tool, Replica Studios allows creators to generate realistic voiceovers for games, videos, and other media. It offers a wide range of voice actors and styles, helping bring projects to life with AI-generated voices.

Frase: Frase helps content creators optimise for SEO by generating topic-relevant content briefs and providing AI-generated content suggestions. It's ideal for marketers and SEO professionals looking to boost their content strategy.

Meta recently suggested an AI fix for those who missed Thursday's Northern Lights: recreate them using AI.

They shared three AI-generated images on Threads, showing the Aurora Borealis over famous spots like the Golden Gate Bridge, a city skyline, and a ferris wheel.

The idea was to jump on the trend of people sharing their own photos from the rare event, but the reaction wasn’t positive.

After a few initial comments praising AI-made pictures, the tone shifted.

Some users joked, while others were more critical.

They pointed out that seeing the Northern Lights is a special shared experience, and AI can’t capture that feeling.

Here are the tidbits:

A user claiming to be an astrophysicist and AI scientist gave detailed feedback, saying this wasn’t the time for AI-generated content.

Like previous ad mistakes from Google, Meta missed the mark.

As people continue to question AI’s place in creative fields, posts like this may keep falling short.

Aurora borealis?!

Space: SpaceX achieved a world first by catching the returning booster of its massive Starship rocket using giant mechanical arms during a test flight.

Crypto: Despite Bitcoin’s price correction 287 days post-halving, on-chain metrics show strong fundamentals.

Games: Two insiders have hinted that the next Xbox Partner Preview event could be announced soon, possibly early next week.

Sports: Harlequins delivered a strong defensive display to defeat Saracens at The Stoop, ending an eight-game derby losing streak.

Don’t Miss - Game Freak, the Pokémon developer, has confirmed a breach in August that led to internal materials and sensitive employee data being leaked. These include Pokémon game source codes, scrapped designs, and information on upcoming projects such as Generation 10 games, a Pokémon MMO, and new anime. Game Freak is strengthening its security and apologised to those affected, as over 2,600 employees' information was accessed.

Submit your artwork to Mindstream →

Yes, they’ve got a real shot. - 54% ✅

Not a chance, Nvidia’s too far ahead. - 46%

Your Views:

“Competition is essential, they've got this.” - jorigbynz

“Nvidia is a moving target.” - reller

Submit your opinions in our daily poll to be featured!