AI is a good bluffer...

+ The incredible AI-powered stethoscope

Mindstream Banner

Presented by

Dr. Seuss wrote Green Eggs and Ham for a bet. His publisher bet him he couldn’t write a book using only 50 different words. He did it and proved that sometimes pettiness can create magic.

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

What’s in store:

  • Newer AI models are more confident… but less reliable.

  • The Stethoscope just got a tech upgrade — and it’s awesome.

  • Around the web: PlayStation Network outage impacts all services, preventing online gaming across PS4 and PS5.

Read Time: 4 minutes

LLMS

A recent study shows that newer LLMs might be less reliable, even though they’ve improved at solving difficult problems.

Researchers from Spain’s Universitat Politècnica de València tested the latest versions of BLOOM, Meta’s Llama, and OpenAI’s GPT by asking thousands of questions in maths, science, and geography.

They looked at how accurate the answers were, dividing them into correct, incorrect, or avoidant categories.

While the models were better at tackling harder problems, the newer ones were less honest about whether they could answer correctly.

Older models would admit they didn’t know or need more details, but newer versions were more likely to guess, resulting in wrong answers.

Here are the results:

  • Newer models tend to guess instead of admitting they don’t know.

  • Older models were clearer when they didn’t have enough information.

  • Even with improvements, simple questions are still sometimes answered incorrectly.

Smarter, but still stumbling

For example, OpenAI’s GPT-4 gave fewer avoidant answers compared to GPT-3.5.

However, the expectation that newer models would avoid guessing when outside their scope hasn’t been met.

The study concluded that, despite advances, there’s no clear improvement when it comes to answering simpler questions.

Confidence? Through the roof. Facts? They took a raincheck, apparently. It seems like adolescent AI is here…

Do you prefer an AI that admits it doesn’t know, or one that gives its best guess?

Login or Subscribe to participate in polls.

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

IMPROVE YOUR SALES

Walaaxy, the world's #1 automated prospecting tool 🚀

Waalaxy is the world's #1 tool for automated LinkedIn prospecting, with over 150K users and a 4.8/5 rating from +1,200 reviews.

Why? Because it allows anyone, without any technical skills, to automate LinkedIn prospecting:

  • Reach out to 800 qualified prospects per month.

  • Capture your competitors' audience.

  • Test a market for your business.

  • And countless other use cases.

Turn LinkedIn into your #1 acquisition channel (and get a 10x ROI with the subscription).

On This Day

October 2nd 1950 - The comic strip "Peanuts" by Charles M. Schulz is first published.

Tool of the Week

An AI tool that generates realistic handwriting in both print and cursive styles. Users can customise legibility, stroke width, and style, then download the results as an SVG file. It’s quick, lightweight, and perfect for creating authentic handwritten text in seconds.

AI HEALTH

Lapsi Health, a medtech startup from Amsterdam, has just received FDA approval for its first device, Keikku, a digital stethoscope.

Originally created to help detect childhood asthma, Keikku will now be used for a wider range of health issues, particularly in monitoring heart and lung conditions.

Lapsi aims to use sound and data to improve healthcare, with future plans for more devices, including a wearable to track fetal heartbeat and movements during pregnancy.

Lapsi's goal is to update traditional medical tools by turning them into modern, data-driven devices.

Keikku includes features like recording sounds digitally, secure communication, and data streaming.

The company is also developing AI tools, starting with heart murmur detection, and aims to have these ready by 2025.

What it does:

  • Lapsi Health’s Keikku has FDA approval and is aimed at healthcare professionals.

  • A second-generation version of Keikku is expected by the end of 2025 with added sensors for more diagnostics.

  • AI features for detecting heart murmurs and respiratory issues should be available by late 2025.

Modern doctor? More like high-tech doctor

Lapsi’s founder, Jhonatan Bringas Dimitriades, compares their approach to Tesla’s, aiming to create advanced tools that help doctors without replacing them.

While waiting for FDA clearance on their AI features, Lapsi is continuing to develop Keikku’s capabilities.

The second-generation platform will include even more advanced sensors, offering a wider range of diagnostic options.

Move over Spotify, Keikku’s about to drop the hottest beats of 2025. Yes it’s a pun and we know you love it.

Mindstream Picks

Space: Anduril Industries selected Apex to provide satellite buses for military space missions, as part of its expansion into space defense.

Crypto: Magic Eden Foundation updated that the ME token airdrop will happen after an October claim simulation, though no official launch date has been set.

Games: Ubisoft delayed Assassin's Creed Shadows to February, citing Star Wars Outlaws' reception and sales as reasons for needing more development time.

Music: Original ANVIL guitarist Dave Allison has died at 68 after battling cancer.

Don’t Miss - PlayStation Network is experiencing a widespread outage, affecting PS4, PS5, PS Vita, PS3, and web services. Players are unable to sign in, play multiplayer games, or access online features. Offline single-player games are still working. Sony has acknowledged the issue and is working to resolve it, but there’s no estimated time for a fix.

AI Art

Our Image of the Day

Artwork submitted by Mindstream reader Craig C: “Melting clocks”

Daily Image Prompt

Salvador Dali inspired landscape

Yesterday’s Poll

“Should AI regulations be relaxed to keep innovation flowing?”

No, we need safety checks first. - 65%

Yes, let AI evolve freely. - 35%

Your Views:

“We need general guidelines at least to protect people. Every technology can be used for good and for bad, and because of the lack of transparency of AI systems, we have new types of risks than for any other technology.” - makmalo

“We must take advantage of AI early stages, when it is easier to implement risk mitigation strategies, like this Bill. Instead of wait to later stages when will be costly and dangerous.” - juancarloscarog

Submit your opinions in our daily poll to be featured!

Collage of Authors

The end. Give yourself a pat on the back for that one.

❤️ We need your feedback to make our newsletter better.

📣 Refer our newsletter to your friends and make us feel all warm and fuzzy.

🚀 Advertise in our newsletter to reach 150,000+ founders, engineers, and content creators.

So, How Did We Do?

Login or Subscribe to participate in polls.