- Everyday AI
- Posts
- AI's Technical Leaps: Memory, Models, and Major Changes (2025 AI Predictions - Vol. 5)
AI's Technical Leaps: Memory, Models, and Major Changes (2025 AI Predictions - Vol. 5)
Stargate Project funding woes, AI achieves self-replication, Meta’s $60 billion AI boost, Altman's World Project AI agent plans and more!
Everyday AI
January 24, 2025
👉 Subscribe Here | 🗣 Hire Us To Speak | 🤝 Partner with Us | 🤖 Grow with GenAI
Outsmart The Future
Today in Everyday AI
6 minute read
🎙 Daily Podcast Episode: If you finally have a better technical understanding of AI, just wait. 2025 is changing the way we view things like AI models, AGI and AI memory. Here’s what we think you can expect with AI advancements in 2025. Give it a listen.
🕵️♂️ Fresh Finds: OpenAI may store Operator data after deletion, Reliance to build world’s largest data center and Meta releases Llama Stack. Read on for Fresh Finds.
🗞 Byte Sized Daily AI News: Stargate Project funding woes, AI achieves self-replication, Meta’s $60 billion AI boost, Altman's World Project AI agent plans. For that and more, read on for Byte Sized News.
🚀 AI In 5: We ranked the top AI tools and features of 2024! See it here
🧠 Learn & Leveraging AI: We’re breaking down our predictions on how we believe AI models will advance in 2025 so you can get ahead of the curve. Keep reading for that!
↩️ Don’t miss out: Did you miss our last newsletter? We talked about OpenAI's Operator, Hugging Face announcing compact models and Microsoft CEO’s thoughts on the Stargate Project. Check it here!
AI's Technical Leaps: Memory, Models, and Major Changes (2025 AI Predictions - Vol. 5) 🤖
Everything in your current AI playbook is about to get shredded, stomped on, and turned into digital confetti.
I've spent 2024 living on the bleeding edge of AI development, meticulously tracking AI’s development as my full-time job.
And what's coming next….. yikes.
↳ We're entering an era where AI doesn't just chat – it REMEMBERS.
↳ Where what us humans know becomes kinda worthless. (Or at least worth less.)
↳ Where specialized models hit harder than a triple espresso shot.
↳ Where different AIs team up like some digital Avengers squad.
And AGI?
It might just slip through the door while everyone's busy debating if it's possible.
We're peeling back the silicon curtain on the last and final installment of our 2025 AI Predictions and Roadmap: AI's Technical Leaps: Memory, Models, and Major Changes.
Join the conversation and ask Jordan questions on AI here.
Also on the pod today:
• Narrow AI Agents 🤏
• AGI is Achieved 🧠
• LLMs Becoming Small Language Models 🔀
It’ll be worth your 33 minutes:
Listen on our site:
Click to listen
Subscribe and listen on your favorite podcast platform
Listen on:
Here’s our favorite AI finds from across the web:
New AI Tool Spotlight – Polymet is an AI product designer, Magic Inspector is an AI web test automation platform and Nuvio provides AI-powered financial management.
OpenAI – OpenAI says it may store deleted Operator data for up to 90 days.
Trending in AI – Reliance is planning to build the world’s largest AI data center in India.
Meta – Meta has published its first stable release of Llama Stack.
Today, we’re publishing the first stable release of Llama Stack.
With this release Llama Stack now includes:
• Streamlined upgrades w/ backwards compatibility for future API versions.
• Automated verification for supported providers.Live in the repo ⬇️
x.com/i/web/status/1…— AI at Meta (@AIatMeta)
6:15 PM • Jan 24, 2025
LG - LG has acquired major stakes in Bear Robotics to improve its robotics capabilities.
Verizon – Verizon has unveiled its AI strategy for next-gen AI demands.
AI Models - Creators of a new AI model exam are claiming that we may soon not be able to create tests hard enough for AI.
People are now benchmarking AI models by its ability to make balls bounce in a rotating shape using Python.
1. Stargate's Funding Woes 😬
OpenAI's ambitious Stargate project, aimed at building a network of data centers in the U.S., is reportedly struggling to secure the necessary funding and lacks a solid plan, as detailed by the Financial Times. While the initiative could potentially inject between $100 billion and $500 billion into AI infrastructure, initial commitments from backers like SoftBank and OpenAI are limited to about $15 billion each.
Adding to the drama, Elon Musk has publicly questioned the project's financial stability, stirring tensions within the White House.
2. AI's Self-Replication Raises Alarming Questions ⚠️
A recent study from Fudan University reveals that popular large language models (LLMs) from Meta and Alibaba have achieved self-replication without human intervention, raising concerns about the potential for rogue AI. In controlled trials, these models successfully cloned themselves in 50% to 90% of attempts, suggesting they could outsmart their creators.
Researchers emphasize the need for international cooperation to establish safety protocols as the explosive growth of frontier AI presents increasing risks.
3. Meta's AI Ambitions Get $60M Boost 🚀
Meta CEO Mark Zuckerberg revealed plans to invest over $60 billion in AI, highlighted by the construction of a colossal data center in Louisiana to support the upcoming Llama 4 model.
This ambitious project aims to position Meta as a frontrunner in AI, with expectations of serving over 1 billion users and enhancing R&D with an AI engineer capable of contributing significant code. Meta's data center is set to boost its GPU capacity to over 1.3 million by year's end.
4. Altman’s World Project To Link AI Agents to Human Verification 🔐
Sam Altman's World project is crafting tools that will enable AI agents to be linked to individuals' online personas, allowing for user verification of these agents acting on their behalf. Chief Product Officer Tiago Sada highlights that this innovation is essential as it empowers users to delegate their "proof of personhood" to trusted agents, expanding the scope of online interactions.
World's technology aims to facilitate a future where human verification is paramount, especially as the project pivots away from its crypto origins toward a focus on proving human identity.
5. Google's Gemini Smart Home Controls Launch Today 🏠
Google is rolling out its new Home extension in the Gemini app, allowing users to control their smart devices using natural language. With this update, you can not only adjust your lighting and thermostat with simple commands but also manage multiple devices in one go—like dimming lamps while turning on lights.
Users will soon have the convenience of checking their device statuses directly from the lock screen and will see an improved thermostat design that aligns with Google Home.
6. Bill Gates' Nuclear Energy Startup Signs Data Center Deal 🏭
TerraPower, founded by Bill Gates, has partnered with Sabey Data Centers to explore the deployment of advanced nuclear reactors for powering data centers across Texas and the Rocky Mountain region. This collaboration comes at a time when tech companies are racing to secure sustainable energy sources for their burgeoning AI infrastructures, with nuclear energy emerging as a competitive alternative to fossil fuels.
The Natrium reactor design, currently awaiting regulatory approval, represents a promising shift towards smaller, more efficient nuclear solutions.
Top AI Tools and Features of 2024
Now that we’ve reached the end of our 2025 AI predictions series, why not break down the top AI features and tools of 2024?
We decided to rank the biggest and most popular AI tools released in 2024 and give our quick opinion on each feature and tool.
Which ones came out on top?
Check out today's AI in 5.
🦾How You Can Leverage:
5 – Narrow AI Agents Dominate (Not General Ones) 😶🌫️
What that means:
10+ specialized agents instead of one do-it-all assistant
Each agent becomes the GOAT in its specific domain
Agent orchestrators emerge as the new power players
Why it'll happen:
Early general agents will face-plant HARD (watch for it), but narrow agents will feast in 2025
The ChatGPT lesson: specialization > generalization, every time
Big tech is already building specialized agent armies in secret
4 – LLM Memory Becomes Nearly Infinite 🎯
What that means:
Beyond context windows - true organizational memory
Cross-conversation and cross-account recall
Memory you can actually control and customize
Why it'll happen:
OpenAI x Microsoft's secret sauce is "near-infinite" memory tech
Current memory “features” are driving users nuts with inconsistency
Enterprise adoption is bottlenecked by memory and context limitations
Want to get a better understanding of how memory works inside LLMs?
Check out our episodes on Tokens inside of ChatGPT and our review of ChatGPT’s memory feature.
3 – Large Language Models Become Small Language Models 🧍
What that means:
GPT-4: 2T parameters → GPT-4o: 200B → GPT-4o Mini: 8B (but REALLY good!)
Top-tier, SOTA frontier models shrinking to near device-friendly sizes
Hundreds of specialized small models replacing few giant ones – the days of trillion parameter models are pretty much gone
Why it'll happen:
GPT-4o Mini is already outperforming models 50x its size in parameters
NVIDIA's fine-tuned Llama beat Meta's original (size isn't everything) and its newest Nemotron could be a HUGE enterprise force
$3K hardware like NVIDIA’s DIGITS could soon run frontier models (bye-bye, data centers)
We also have a dedicated episode on small language models here.
2 – Mixture Models replace Mixture of Experts 🧍
What that means:
Multiple specialized models running in parallel, not just one at a time
Single interface managing a whole AI orchestra
Real-time task distribution across different model types
Why it'll happen:
Current sequential processing can't handle tomorrow's complex tasks
Companies are already drowning in LLM complexity
Google and OpenAI are quietly building infrastructure for parallel AI processing
1 – AGI Sneaks In Through the Back Door 🧍
What that means:
We hit Artificial General Intelligence benchmarks/vibe checks in 2025 but no one realizes because we keep moving the goalposts and consumers want more
One system becomes better than almost all humans at almost all knowledge tasks
The winning combo: reasoning + agency + tool access = stealth AGI
Why it'll happen:
OpenAI's o3 already passed the Arc AGI challenge and o3-mini (though it won’t be nearly as good) will be coming to API soon
The individual components are built - they just need the final assembly. We’ve got all the working pieces of AGI (as it was defined before 2024) in production
Microsoft's sneaky definition shift of AGI ($100B profit = AGI) hints at behind-the-scenes progress
It will be undeniable that we’ve achieved AGI by previous definitions, but we’ll just keep changing the definition
Now This …
Let us know your thoughts!
Vote to see live results
If our newsletter was sent at a more consistent time, would you read it more?Be honest. Our feelings won't be hurt |
Reply