😺 Ilya predicts the future of AI

Welcome, humans.

A quick note from Noah & Grant: Today is the last day we’re publishing The Neuron in 2024! You won’t hear from us again until January 6th.

A HUGE thank you to each and every one of you for supporting us this year—reading our newsletters, checking out sponsors, referring a pal or two. It means the world, and we can’t wait to continue bringing you more exciting content in 2025.

Reminder: you have until EOY to enter our giveaway where we’re gifting three readers a free AI subscription for 2025. Just use the link below and tell one friend/coworker/cool cat to sign up for The Neuron.

And since you won’t be hearing from us ‘til ‘25… happy holidays and Happy New Year! 😻

Share The Neuron

You currently have 0 referrals, only 2 away from receiving PLACEHOLDER MILESTONE NAME.

Click to Share

Or copy and paste this link to others: https://www.theneurondaily.com/subscribe?ref=PLACEHOLDER_CODE

Here’s what you need to know about AI today:

• We break down Ilya’s predictions for what’s next in AI.
• Google released a new reasoning model that’s topping leaderboards.
• Instagram demo’d some crazy-looking AI editing tools.
• o1 breaks down what the Founding Fathers would think of today’s America.

Advertise in The Neuron here.

Ilya Sutskever’s predictions for the next chapter of AI.

Ilya Sutskever: "Sequence to sequence learning with neural networks: what a decade"

In all the AI news of the past two weeks, you might have missed the ground-shaking talk former OpenAI Chief Scientist Ilya Sutskever gave at the NeurIPS 2024 AI conference last week.

It was a fascinating retrospective on his groundbreaking 2014 paper that helped kickstart the modern AI revolution. But more importantly, he dropped some serious predictions about where AI is headed next.

It’s only fitting as our last story of the year to talk about where AI is headed in the near future. And who better to predict it than Ilya himself?

First, a history lesson. Back in 2014, Sutskever and his co-authors proposed three simple but powerful ideas that changed everything (in AI):

• Use autoregressive models trained on text.
• Make the neural networks large.
• Train on massive datasets.

These three ingredients are basically the recipe for every LLM today, from GPT to Claude to Gemini.

Now here's where it gets interesting. Sutskever believes that we're approaching the end of this era. “Pre-training (training a large model) as we know it will unquestionably end,” he declared.

Why? Because “we have but one internet.”

What he’s saying is that we're running out of high-quality training data to actually train the AIs on.

As Sutskever put it, “Data is the fossil fuel of AI—we've achieved Peak Data.”

So what comes next? Sutskever outlined several possibilities for new AI development:

• Agents: AI systems that can actually do things for us on our computers (this is what’s coming big time in 2025).
• Synthetic data: Finding new ways to generate training data to improve AI models.
• Inference-time compute: Models like ChatGPT o1 that do more processing when you ask a question.
• Reasoning: Moving beyond pattern matching to true logical thinking.

But the most intriguing part came when Sutskever discussed superintelligence. Makes sense… his new company is quite literally called Safe Superintelligence.

He believes future AI systems will be qualitatively different from today's models in several key ways:

1. True agency: Not just following instructions, but pursuing their own goals.
2. Real reasoning: Moving beyond pattern matching to logical thinking.
3. Better learning: Understanding concepts from limited data.
4. Self-awareness: Having an internal model of themselves.
5. Unpredictability: The more they reason, the harder they'll be to predict.

To illustrate his point, he shared an interesting biological analogy. Just as early hominids showed a different brain-to-body size scaling pattern than other mammals, he suggests that AI might find a similar “different slope” in its development—a fundamental shift in how these systems scale and improve.

Why this matters: Just like the evolutionary jump from regular primates to hominids represented a qualitative change in intelligence (not just bigger brains), Sutskever suggests the next phase of AI might represent a similar evolutionary leap—moving from pattern-matching on internet data to true reasoning and agency.

The big takeaway? We're at the end of one era of AI development (pre-training on internet data), but potentially at the beginning of something much more profound.

The next phase won't just be about bigger models or more data—it'll be about creating systems that can truly think, reason, and understand the world around them.

What happens then? That’s another convo for another time… for now, go sip a margarita on the beach and rest up for some (hopefully safe) superintelligence in ‘25!

FROM OUR PARTNERS

Build Your AI Influencer Empire with Glambase

Glambase helps you build monetizable AI influencers that create content and engage with audiences autonomously.

Think of Glambase as your AI-powered social media team that works 24/7. Your virtual influencer can chat with followers, create content, and build an engaged community while you focus on other projects.

Over 1,000 creators are already using Glambase to automate their social media presence.

The platform handles all the technical heavy lifting—from personality design to content generation. No coding or AI expertise required. Just design your digital persona and watch your earnings grow through subscriptions and interactions.

Check it out here—early adopters get priority access.

What gives it away that this is AI?

Iron man: meet Melon man

We wanted to do something fun for the end of the year: so why not over-analyze a pretty hilarious parody of America’s favorite superhero?

Watch the video, and tell us what gives it away as being AI. Pick either option below, and write in your answer. We’ll share the results in our first email of 2025!

What gives it away that this is AI?

Describe what you think makes this look AI generated vs real.

Write in.

Write in, pt. 2.

Treats To Try.

1. LiveImage creates personalized video messages from Santa that mentions your child's dance recitals, soccer games, and Christmas wishes.
2. Cursor is super strong product for coding with AI that a lot of our nerdiest friends are using.
3. Get your product featured here, in front of nearly 500k readers, by filling out our sponsorship form for 2025.
4. Trickle helps you build apps, landing pages, and forms from just text prompts.
5. Gemini Code Assist is a new enterprise coding assistant from Google—try it here.
6. Krea AI is a very good product for editing images/videos using AI—check out their new feature that adds real products to any image.

See our top 51 AI Tools for Business here!

Around the Horn.

Here’s the book, and here’s the audio—interesting experiment!

• Google just launched its own reasoning model called Gemini 2.0 Flash Thinking (wtf is that name?!), which you can test here in AI Studio.
  • This new model ranks #1 on all categories on the LM Arena (o1 isn’t included on the list).
• Instagram demo’d new AI editing features that’ll let users “change nearly any aspect of your videos”.
• Day #11 of OpenAI Shipmas: ChatGPT is coming to Apple Notes, Notion, and Quip, plus other desktop apps will integrate with Advanced Voice Mode!

FROM OUR PARTNERS

Make your videos smarter with AI.

See how Dell Technologies’ Precision workstations powered by NVIDIA RTX™ enhance Nuke CopyCat and Beeble.ai workflows. Join industry experts on January 9th for real-world examples.

Register today!

A Cat's Commentary.