Plans, Polls, and Predictions

What's in store for 2025.

Howdy, friends and mortal enemies! (You know who you are, Chad.)

Happy 2025 to those of you who don’t receive my Sunday Rundowns.

To the rest of you, I’m sorry you had to read that twice.

If the last two years are anything to go by, this year we’ll see more crazy leaps in AI capabilities and even more breathless hype about how AI will change everything but you are falling behind so don’t miss out on my secret library of AI hacks that will make me you rich (buy my $799 course today for only $795).

Fuck you and your tiny levitating laptop, Chad!

My promise, as always, is that I’ll keep you in the loop while keeping the hype in check.

For this inaugural Thursday post, I want to do three things:

• Share my vague plans for Why Try AI in 2025.

• Run polls to get your feedback to turn those vague plans into something concrete.

• Share a few of my AI predictions for this year.

Let’s get this show on the road.

(It’s not a literal road, Chad, put down the car keys. God, you’re the worst.)

Plans for Why Try AI

I started Why Try AI in mid-September, 2022 (here’s the backstory).

ChatGPT wasn’t a thing, and the world hadn’t yet lost its collective mind over AI.

They were simpler times.

Here are a few broad-strokes highlights of the last two years:

2023

• Added the weekly Sunday news column.¹

• Experimented a lot with AI images (mostly Midjourney).

• Reached 4,500 subscribers by the end of the year (12 paid).

2024

• Introduced the Sunday Bonus segment for my awesome paid subscribers.

• Started doing more collaborations like guest posts, video chats, and live workshops.

• Reached 8,500 subscribers by the end of the year (77 paid).

Now what?

That brings us to today.

I’ll let you in on a little secret: I’m not great at making long-term plans.

While others might have a fancy 5-year growth strategy, I mostly just wing it.

When I first started Why Try AI, I had no specific goals in mind. And I definitely didn’t intend to earn any money from it.²

The truth is that I genuinely love staying up-to-date on what’s happening in AI, testing new tools, and sharing my observations with others.

It still blows my mind that thousands of you have decided to join the ride and many are willing to part with hard-earned money to support this newsletter.

Please know that I don’t take your readership and your support for granted and am humbled by it. Why Try AI takes up a lot of my time these days, so it’s massively motivating to know that people find my scribbles valuable.

Pull yourself together, Chad!

With the sappy stuff out of the way, here are a couple of vaguely defined ideas I’d like to try this year.

Help me decide what to focus on by voting in the polls below.

👨‍🏫 Courses and eBooks

After over two years of bi-weekly posts, I’ve built up a library of hundreds of articles.

I feel it’s time to start turning some of those archives into more polished products.

Here are three I have in mind so far.

1. Midjourney video course

This could be a clean, organized version of my Midjourney Masterclass workshop that covers all Midjourney features along with tips on getting the most out of them.

2. 200+ Midjourney terms (eBook)

This would have my Midjourney modifiers organized way better by category, new high-quality visual examples, and tips for using the terms:

3. 100+ top AI Jest Daily cartoons (eBook)

As you know, I’ve had a silly side project called AI Jest Daily where I made single-panel cartoons to see if AI could be funny. Some of them turned out pretty great.

For this eBook, I’d handpick around 100 best cartoons (out of 500 issues) and perhaps include some behind-the-scenes “making of” insights..

🌶️ New “Hot Takes” section

Right now, I send two weekly emails:

• Thursday: A regular Why Try AI article, typically focused on a single topic.

• Sunday: AI news roundup + Sunday bonus.

But there are often moments where e.g. a cool new tool comes out that I’m immediately tempted to test. It might not warrant a dedicated Thursday article, but I’d love the option of taking such tools for a spin on an ad hoc basis and sharing my quick thoughts.

A separate, ad hoc “Hot Takes” section would let me do that for those who’re interested while keeping everyone else’s inboxes clean. (Substack lets subscribers pick which individual “sections” of the newsletter to receive.)

📹 Video guides

I’m also toying with the idea of finally putting some effort into my latent YouTube channel and recording a series of “How to” guides where I share my screen and walk you through hands-on tips for using different AI tools. (Something like this.)

Putting it all together

I likely can’t run after all of these ideas at once.

I am just three chatbots in a trench coat one man.

Help me prioritize by letting me know what matters most to you.

I can’t promise that I’ll end up launching all or even most of these.

But it definitely helps to know what you find relevant!

My AI predictions

Predicting future AI developments at the start of the year has become somewhat of an industry tradition. For the past two years, I avoided doing this, primarily because I’m not a certified psychic.

But enough people have suggested I try it that I finally caved.

I simply can’t fight peer pressure, which is also why I’m writing this article while smoking fifteen cigarettes at once and downing vodka shots laced with LSD.

So for what it’s worth, here’s my stab at reading the tea leaves:

1. Reasoning models converge in performance

Right now, OpenAI’s yet-to-be-released o3 model is the talk of the town, shattering benchmarks and making us pose existential questions.

But I don’t think o3 will stay far ahead of the game for long, because:

1. As we’ve seen, focusing on scaling inference-time compute can yield huge gains in way less time than focusing on pre-training efforts. Last year, LLMs largely converged around GPT-4o level of performance. With the new crop of reasoning models, this might happen even faster.

2. DeepSeek v3 proved that it’s possible to train an LLM that achieves top-tier performance at a fraction of the cost spent by leading AI labs.

As such, here’s my prediction:

By the end of 2025, reasoning models from at least three other players will perform on par with or better than OpenAI’s o3.

2. AI video goes mainstream

In 2023, AI video was a bit of a joke.

In 2024, things truly began to change, starting with OpenAI’s Sora, continuing with impressive video models from other players, and culminating with Google’s shockingly realistic Veo 2.

I think we’ve reached a point where, say, a major Hollywood studio turns to AI for scenes that are impossible or impractical to shoot using conventional means.

As such, here’s my prediction:

By the end of 2025, a major movie—either released or announced—will feature at least some scenes made entirely using text-to-video models.

3. Actually useful AI agents

We’ve been talking about AI agents since early 2023.³

Last year, many companies released experimental, crude versions of agents that could navigate your screen and perform tasks independently.

AI agents still don’t show signs of truly having a robust world model or the ability to navigate novel situations the way we do. But they show potential for routine, well-documented tasks.

Here’s my prediction:

By the end of 2025, an AI agent will be able to reliably automate at least three common, well-defined workflows in business software.

4. No “GPT-5” (ever)

I’m with

Charlie Guo

on this one.

GPT-5 has been hyped up for so long that I just can’t see it ever living up to expectations. Also, with new reasoning models outperforming old-school LLMs so convincingly across benchmarks, the latter will likely fade into the background.

So here’s my prediction:

OpenAI will never release a model called GPT-5. Instead, the company will focus on the “o” family of reasoning models. Any performance improvements to “traditional” LLMs will be released as GPT-4x iterations or a rebranded family.

5. AI tutors are formally integrated

Less than two seconds after ChatGPT first came out, exactly 17,451 students used it to write their homework assignments.⁴

Since then, we’ve gone through every stage of grief and settled on begrudgingly—or excitedly—accepting that AI will inevitably play a growing role in education.

Sal Khan’s Khanmigo was perhaps the most notable early example of a dedicated AI tutor. But it existed outside of a formal educational framework. By now, we might be ready to see a major educational institution embrace AI in a more official capacity.

So here’s the prediction:

By the end of 2025, a top-tier university will use an AI tutor to scale an accredited degree program and make it accessible for remote students from around the world.

6. Model transparency becomes mandatory

AI companies have largely gotten away with not disclosing details about their models’ training data, inner workings, and so on.

And while the upcoming regulatory climate in the US might be favorable to AI companies (as Charlie argues), there’s also a push toward greater transparency by the research community and regulators, not to mention the indirect pressure of open-source competitors.

So the prediction is:

By the end of 2025, we’ll have a law that requires AI companies to publish “nutritional labels” for models of a certain size or capability, specifying things like energy consumption, data sources, etc.

7. Google dethrones OpenAI

OpenAI released ChatGPT in late 2022. By early 2023, Google was dead:

Just How "Dead" Is Google, Anyway?

In late 2024, OpenAI added better browsing to ChatGPT and murdered Google all over again:

Are We Even Ready For AI Search?

While OpenAI continued to hog the spotlight in 2024, “zombie” Google quietly released outstanding products like NotebookLM, Learn About, and Whisk.

Not only that, but Google’s models are starting to dominate practically every domain of GenAI:

• Images: The latest version of Imagen 3 outranks the competition in terms of both visual quality and prompt adherence.

• LLMs: Gemini-Exp-1206 is currently the #1 (non-reasoning) model on Chatbot Arena.

• Search: Deep Research is a preview of what a truly thorough deep dive into hundreds of pages can do for research.

• Video: By all accounts, Veo 2 creates the most prompt-adherent, realistic, consistent, and “physics-aware” videos of any text-to-video model to date.

Taking into account the above along with Google’s far greater resources and reach, I don’t expect OpenAI to maintain its “synonymous with generative AI” status much longer.

And so:

By the end of 2025, Google, not OpenAI, will be widely seen as the dominant player in generative AI.

🫵 Over to you…

How’s 2025 treating you so far? What AI trends and developments are you most excited about this year? Do you have any AI predictions of your own?

I’d also love to know if there’s something you want to see on Why Try AI that I haven’t listed above.

Leave a comment or drop me a line at whytryai@substack.com.

1. Fun fact: It was originally called “10X AI” and contained exactly 10 entries before morphing into “Sunday Rundown.”

2. In fact, I was hesitant to even enable the paid subscription option until well into 2023.

3. Remember BabyAGI and AutoGPT?

4. Citation needed.