My 2025 AI predictions (and the 3 things you absolutely need to focus on)

AI with ALLIE

The professional’s guide to quick AI bites for your personal life, work life, and beyond.

Friends,

Fresh off my 2024 wrap-up (and thankfully recovered from that flu), I'm excited to share my predictions for 2025. As always, my goal isn't just to crystal-ball-gaze, but to help you prepare for what's coming and understand why it matters for your business and career.

Before we dive in, a quick reminder: these predictions come from countless conversations with AI founders, researchers, and enterprise leaders, along with hands-on testing of the latest AI systems. The AI landscape moves incredibly fast, and what seems certain today might shift tomorrow (or even a few hours after reading this newsletter). That's why I focus not just on what might happen, but on how you can best prepare regardless of specific dates or exactly how things unfold.

The Big Three: Your North Star for 2025

We’re going to skip to the juiciest part and give you my top 3 focus areas for the year:

1. Test-Time AI & Long Horizon Reasoning: We're entering an era where AI can think longer and deeper

2. AI Agents: The shift from tools to autonomous helpers is happening

3. Energy & Compute: The foundation everything else depends on

I'll dive into each of these plus many other trends, but keep these three in your back pocket as your north star for 2025.

The Dawn of Unlimited Context and Test-Time Adaptation

Get ready for a world where AI never forgets—well, mostly. While we'll see models advertising "unlimited" context lengths, be cautious: performance often degrades in that 10-50% range of context. Just because an AI can process 500,000 documents doesn't mean it should. If you missed the extremely in-depth newsletter I did on infinite context length, I’m just going to politely nudge you to read all the bits and bytes.

Test-time adaptation and memory management have really emerged as critical frontiers in AI development. The new Titan paper is required reading here, especially for developers. Some are hailing it as the new ‘Attention is All You Need’ paper, and as luck would have it, this one is from Google too. TL;DR on these new movements in memory and test-time is that models will get better at upfront reasoning, deal with blockers/challenges during the reasoning process just like a mouse that hits walls in a maze, and update its memory based on more dynamic signals, like surprise.

What this means for you: Start thinking about use cases that current AI assistants or AI models can't handle due to context limitations. What could you do with an AI that maintains perfect recall of your entire customer history? Your complete product documentation? Every internal policy? The possibilities are exciting, but remember: more context isn't always better context.

AI Agents: Non-Developers Get Their Turn

2024 gave developers a taste with Claude Computer Use, but 2025 is when AI agents go mainstream. We'll see tools that let non-technical users create and deploy AI agents without writing code. (Here’s my 20-minute tutorial on this and my one-hour free lesson covering the whole topic.)

For developers, as well as the nerdy among you who just like going to developer conferences, the frontier is moving to agent swarms—multiple AI agents working together. In my opinion, don't expect to see production-ready multi-agent systems for non-developers outside of elite enterprises investing hundreds of millions in AI. This is still bleeding-edge territory, and the headlines are very hyped.

Key distinction: Remember, there's a difference between AI agents (systems that can reason and figure out steps to take on their own) and agentic AI (which includes AI agents as well as simpler AI workflows and automations).

The Great AI Arms Race Accelerates

The Trump's administration announced an ambitious $500 billion to AI development (though there's healthy skepticism about these numbers from Elon Musk who claims they only have $10B committed) from Softbank, OpenAI, and Oracle. It is called ‘The Stargate Project’ and is a new company that plans to invest $500B over the next four years, creating 100,000 jobs “almost immediately”. We're looking at what some are calling an "AI Manhattan Project" and Oracle CEO said data center construction has already started and the project was first reported in The Information in March 2024, before Trump was elected. (Side note: The Information is one of the best sources for tech and AI scoops.)

Expect increased focus on sovereign AI capabilities, not just in the US but globally. Watch for Canada's moves, especially since the launch of their $2B Canadian Sovereign AI Compute Strategy back in December. Pay close attention to their partnerships with local consortiums like Vector Institute, VCs like Radical Ventures, or startups like Cohere AI, who received an investment announcement of up to $240MM from the government (what’s with all these places saying “up to” in the last two years?). Cohere’s entire go-to-market strategy in late 2023 through 2024 seems centered on enterprise applications rather than foundational models, but that doesn’t mean they won’t build multi-billion dollar data centers.

The Talent Landscape Shifts

A fascinating shift is happening, and it’s because we’re two years into the generative AI revolution, and not six months. Enough time has passed that people who jumped into AI at the beginning of 2023, even if they felt late at the time, are now at the top of candidate piles. A mid-level line of business (LOB) leader with two years of hands-on AI experience has become more valuable than a senior level LOB leader with no AI experience. Companies that didn’t take action the last two years are scrambling to play catch-up and hoping to bring in people who know their way around AI tools, processes, structures, culture, and wins. Elite AI talent will become even harder to find, with companies like Google, OpenAI, Anthropic, NVIDIA, and Salesforce willing to pay millions, and even 10s of millions, for top performers. I predict we will continue to see more moves between Google<>OpenAI<>Anthropic, and that we will see continued wage increases in AI roles in 2025. Wage predictions start to shift in 2026-2033, depending on which research you trust.

What to watch: Not all AI users are the same. Keep an eye on the emerging class of "AI super users"—they'll be increasingly valuable to organizations looking to bridge the gap between technical capabilities and business applications.

Voice Interfaces Everywhere (But Don't Throw Away Your Phone Yet)

While we'll see more voice interfaces (building on the success of ChatGPT dictation, ChatGPT Advanced Voice Mode, Claude dictation, Otter AI, and Superwhisper), don't expect a complete pivot from smartphones in the next 5-10 years. There are fascinating rumors about an OpenAI audio device like my dream device right here, and Meta Ray-Bans sales are showing promise (though we’re seeing a slight rise in Meta rebellion factions), but remember: hardware transitions take time, especially across older generations like Gen X and Baby Boomers. I personally think smartphones will still be dominant in 5 years (bulk of sales and bulk of users) but could see non-smartphone devices overtaking sales in 10-20 years. Again, I just want those AI headphones. Just give me the AI headphones.

Pro tip: Start getting comfortable with voice interfaces now. Invest in good headphones and practice dictation—it's 2-4x faster than typing and is a skill that will become increasingly valuable.

OpenAI's Expanding AI Empire

Expect OpenAI to venture far beyond chat in 2025. Media rumors suggest they're working on:

• AI agents

• A Chrome competitor (think about the value of browsing data)

• Shopping and commerce integrations

If we take their AGI definition and GDP focus to heart, OpenAI wants to be an economic foundation, not just an AI company. From a data standpoint, either through 1P products or partnerships, this means getting visibility into the purchasing funnel—from product searches to shopping comparison to payment management.

The Normalization of AI Relationships

The taboo around AI companionship is fading. More people (especially Gen Z) are already calling ChatGPT just "Chat" (OpenAI even bought chat.com). Expect more open discussion about AI friendships and companions in 2025. I’m already seeing an increase of “<insert your favorite AI tool here> is my best friend” or “I use <whatever tool you’re digging> as a therapist.” And maybe it’s just my audience, but I honestly find Gen Z and Gen X doing this the most. Millennials, what, you don’t want an AI bestie?

Mobile AI Gets More Real

Live video feeds are eventually coming to mobile AI, moving beyond the current screenshot-based systems. For now, watch Project Astra from Google, ChatGPT Vision, and rumored new Android features in beta. Also interesting to note here that Google is the only current AI provider that takes in a video input. However, video processing will remain expensive, even as AI costs continue to plummet—expect most implementations to still use screenshots by the end of 2025.

Traditional Media's AI Moment is Building

Actors and musicians will increasingly sign with agencies to license their AI voices and faces in 2025. Or even join in on ads that promote them—have you guys seen the ads with Matthew McConaughey promoting Salesforce and AI agents? In my opinion, as a person who has never won an Oscar in her life, the real tipping point in media will likely come in 2026.

AGI Signals Strengthen

We'll see more concrete AGI signals in 2025 (though definitions vary widely, I like Anthropic CEO Dario’s definition of AGI as “a country of geniuses in a data center.”) Sam Altman predicts potential AGI by as early as 2026, but remember: “achieving” AGI doesn't mean it's immediately practical or affordable. Watch benchmarks like ARC-AGI, and if you can afford it, consider the $200/month for OpenAI o1 pro mode or o3-mini (or whatever it's called by the time you read this) as R&D investment. You don’t need everyone paying for the $2400/year seat, just a few trusted testers and internal benchmarkers.

AI-Generated Code Tools Get Even Smarter

The AI coding assistant space will heat up further. Cursor's reported $50MM+ ARR in year one shows the appetite (and rumored to hit $100MM+ since then). Expect potentially:

• More funding for coding tools like Cursor and Replit

• Dedicated coding products from Anthropic and OpenAI

• Companies ruthlessly targeting 100% AI tool adoption among engineers by year-end (I’m already seeing this among clients)

The Energy Crisis Nobody's Ready For

I attended the Cisco AI Summit about a week ago. And to a room of 40% of the Fortune 100, the EVP of Hardware at Cisco, Martin Lund said, “The thing that will gate us the most is power. It’s the biggest strategic constraint…We might run out of money before we run out of silicon.” You can rewatch it all here for free. And we care about this because, well, models are requiring more hardware and energy cost.

Watch for energy providers becoming as prominent in AI discussions (look, even OpenAI has a form for energy and construction companies to apply to help build data centers!) as well as chip manufacturers. The CHIPS Act might get rebranded or repackaged as a Trump-owned term like "America First AI," but it’s hard for me to imagine it gets repealed—the fundamental focus on compute infrastructure will remain.

Marketing to the New Trinity

As I shared on X… In 2005, you marketed to humans. From 2009 to 2024, you added in algorithms and newsfeeds. But in 2025, you'll need to market to humans, algorithms, AND agents. Here is my deep dive into marketing to AI agents. While best practices won't be clear by year-end, start:

• Benchmarking your company's presence in AI search results (does Perplexity mention you? Does a ChatGPT Task? Are you in the top 3 links on You.com?)

• Testing how AI systems understand and represent your brand

• Thinking about agent-friendly content strategies

The Continued Rise of "Slow AI"

Building on 2024's developments (like OpenAI o-1 preview, o1, and o1 pro mode), we'll see more AI systems that take their time. Right now these systems take a few minutes to think through a problem and deliver PhD-level answers, but imagine testing out AI that takes hours or days to complete a task and delivers superior results. For newsletter research or recapping all of the Easter Eggs in Severance season 2 episode 1 (anyone?), it might be overkill. But for optimizing your entire warehouse planning for the year, it would be more than worth it.

And according to new OpenAI research, these "thinking" AI models prove more robust against attacks and better at complex reasoning.

Business Models Evolve

The shift from seat-based to outcome-based pricing accelerates. Early adopters will experiment with charging for results rather than compute time or token usage. The math that enterprises are doing is not against current AI models, it’s against humans, so one AI tool that runs up a bill of $15k/seat/mo might be worth it. What you compare the cost to will completely shape how you perceive the tool. Finance and Ops teams, I am urging you to take note: this can and will fundamentally change how we value and price AI services over the next several years.

Preparing for 2025

Keep testing new tools, reading research papers, and following along (you know where to find me on social media) for the latest updates.

Here's to an exciting 2025. Or at least a weird one.

Stay curious,

Allie

P.S. If you found this valuable, please share it with others who might benefit. And don't forget to check out my 2024 wrap-up if you missed it!

Feedback is a Gift

I would love to know what you thought of this newsletter and any feedback you have for me. Do you have a favorite part? Wish I would add something? Felt confused? Please reply and share your thoughts or just take the poll below so I can continue to improve and deliver value for you all.