Synthetic data could kill AI

On average, women blink significantly more often than men. Call it multitasking at its finest—blinking, batting lashes, and staying ahead of the game all at once.

What’s in store:

• What happens when AI learns from AI?

• AI scams are getting smarter, but so are you.

• Around the web: The Internet Archive is back online in read-only mode after a cyberattack.

• How we built our business with Guidde!

Read Time: 6 minutes

AI DATA

Synthetic data and the risks to AI

The idea of AI learning from AI-made data might sound odd, but it's becoming more popular.

Companies like Anthropic, Meta, and OpenAI are already trying this. But can AI-made data really replace real-world data?

AI models learn patterns from big datasets and use those patterns to make predictions.

Human-created data has been key for this, but there are problems: data is getting expensive, and harder to find, and human mistakes or biases can slip in.

This is where AI-made, or synthetic, data comes in as a possible fix.

Here’s what you should know:

• Real-world data is becoming pricier, with companies like Shutterstock charging millions to access it.

• Synthetic data is a fast-growing industry, expected to reach $2.34 billion by 2030.

• Human oversight is still needed to prevent synthetic data from copying the same biases found in real-world data.

The rise of fake data

Synthetic data is a cheaper, faster alternative to human-labeled data.

Companies like Microsoft, Google, and Nvidia have started using it in their AI models.

Writer, for example, created Palmyra X 004 mostly using synthetic data for just $700,000—much less than the $4.6 million for similar models.

But synthetic data has its problems.

Like real data, if it’s biased or flawed, the AI trained on it will be too.

Relying too much on synthetic data can also reduce AI's diversity and accuracy, especially without mixing in real-world data.

While synthetic data shows potential, AI isn’t yet ready to fully rely on it for training. Humans are still needed to check and guide the process to keep models on track.

OpenAI’s CEO believes AI might one day make data good enough to train itself, but we’re not there yet.

AI teaching itself… isn’t this the start of every dystopian movie ever?

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

DOCUMENTATION MADE EASY

How we built a global business with Guidde

At Mindstream, we rely on Guidde for all of our training and onboarding.

When we hiring for the first time, our small team was too busy managing our busy startup to dedicate hours and hours to training new staff - we simply didn’t have the time.

That’s where Guidde came in.

We used Guidde to create how-to videos of all of our processes. Guidde turns your clicks into easy-to-follow videos. They’re easily editable and you can even add voiceover (or use AI!).

We were able to train our global staff asynchronously, with the power of Guidde.

Here’s how:

• Hit the record button and undertake the usual process.

• Stop the recording, and let Guidde do its magic.

• Edit the video easily with the built-in editor.

• Share with your employees!

But we’re not alone. Guidde is trusted by over 20,000 users worldwide, and has a 5-star rating on the Chrome store!

Try Guidde for free and see the difference.

On This Day

October 15th 1783 - Jean-François Pilâtre de Rozier and François Laurent d'Arlandes make the first manned free balloon flight, a tethered flight in a Montgolfier balloon, in France.

Weird AI

AI dumped me on my birthday

Getting dumped over text is bad enough, but imagine getting an AI-generated summary of the breakup instead. That’s what happened to one unlucky iPhone user with Apple Intelligence in iOS 18.1, which condenses notifications into bullet points.

The breakup summary? “No longer in a relationship; wants belongings from the apartment.”

The guy shared the hilarious (and painful) moment on X, along with, “Yes, this happened on my birthday”. The post went viral, but it didn’t fix the breakup.

While Apple Intelligence’s notification summaries can be super handy, they don’t exactly nail human empathy. When it works, though, it’s a great feature—just maybe not for breakups.

GOOGLE

AI is helping hackers, but Google’s fighting back.

Hackers are using AI to make Gmail phishing scams more believable.

These scams trick users with fake account recovery requests and are getting harder to recognise, even for tech-savvy individuals.

For example, security consultant Sam Mitrovic almost fell victim to one of these scams.

It started with a Gmail account recovery request, followed by a phone call claiming to be from Google support.

The caller asked if Mitrovic had logged in from a strange location to build trust.

But, Mitrovic noticed something odd: the voice was too perfect, making him realise it was likely AI.

After checking, he found the phone number was linked to Google Assistant, not actual Google support.

By adding fake details during the recovery process, the scammer made the attack seem genuine.

Keep this in mind:

• AI is making phishing attacks more convincing.

• Google won’t contact you directly for account recovery.

• Always verify phone numbers and emails through official channels to avoid scams.

It’s a war

To fight these evolving scams, Google has teamed up with the Global Anti-Scam Alliance and DNS Research Federation to launch the Global Signal Exchange.

This platform allows real-time sharing of information to block scam attempts.

Using Google Cloud and AI, the system helps detect scam patterns, protecting millions of users.

Google’s Advanced Protection Program, which now includes passkey support, adds extra security.

This is particularly useful for high-risk users, like journalists, as it makes phishing attacks and fake account recovery much harder.

New AI, new problems.

Mindstream Picks

Space: China's Xichang launch centre plans to construct a commercial spaceport to boost the nation's space access.

Crypto: Cryptocurrency market makers CLS Global, MyTrade, ZM Quant, and 15 others were arrested for large-scale market manipulation after an FBI sting using the fake token and firm NexFundAI.

Games: Disney will launch "Tron: Catalyst," an isometric action game, next year. It continues the story from "Tron: Identity."

Music: A new Beatles documentary, Beatles ’64, produced by Martin Scorsese, premieres November 29 on Disney+ and features unseen footage from their 1964 U.S. debut.

Don’t Miss - The Internet Archive is back online in a read-only mode after a cyberattack on October 9th, which included a DDoS attack and a data breach of 31 million user records. While the Wayback Machine is accessible, capturing new web pages is temporarily unavailable as staff work to strengthen security.

AI Art

Our Image of the Day

Daily Image Prompt

Submit your artwork to Mindstream →

Yesterday’s Poll

“Do you think LLMs will ever get basic reasoning right?”

Yes, it’s just a matter of time! - 61% ✅

No, they’ll always be stuck on the small stuff. - 39%

Your Views:

“When they get basic reasoning right, they'll be ahead of 40-50% of human-level reasoning.” - jim

“The premise for them is wrong for decision making. It is a prediction of next word. There is little Bayesean or other probability analysis and no alternative projections for the decisions to choose from.” - lewis.boyd

Submit your opinions in our daily poll to be featured!