Reviewing week 1, 2025: The year of compute chaos
Welcome to the first SundAI of 2025
Here we go, the first SundAI of 2025. A fresh year, fresh chaos, and, as of next week, a fresh new look for SundAI. That’s right, this is the final edition in the “classic” format. You know, the one that’s been your safe, neatly packaged AI recap.
Starting next week, I will be kicking down the walls. No more sanitized headlines or corporate PR. SundAI will be all about summarizing the gritty underbelly of tech. You know, the backroom whispers, VIP club gossip, and Silicon Valley’s pole dancing power plays. If it is messy or controversial, or when Scam Altman, Tim Crook, Ego Musk or Dark Zuckerbot do or say anything stupid, it’ll all end up in here.
And before I forget. Happiness week is coming
Next week, will be all about AI Happiness. Well, it is not about joy and rainbows. It’s going to be all about the dark side of AI: layoffs, algorithmic exploitation, and the dismantling of industries.
So, if you’re feeling a little fragile mentally or live somewhere dark like in Alaska (looking at you, Elizabeth Freeman), maybe avoid text-to-speech for the week. Just a friendly nudge.
So, enjoy this last polished edition, get your bottle of Prozac ready and enjoy the coming week, my intelligent friends.
SundAI 02
This year already started with drama and breakthroughs. Microsoft started things with an $80 billion pledge to AI data centers, NVIDIA stormed CES 2025 with next-gen GPUs, “physical AI” platforms, and a compact personal supercomputer. And “reasoning” models (quote unquote), like OpenAI’s o3 are firing up a synthetic data arms race. And that, my intelligent friends, is forcing us to rethink the very foundation of AI training.
Pathogen detection (yes, the stuff between the keys of my beloved keyboard warriors) is also getting a little touch-up thanks to METAGENE-1, and OpenAI said that ChatGPT Pro is a money-losing mofo. And if that’s not enough, NVIDIA’s $700M acquisition of Run:ai will reshape how you will manage GPUs in the cloud.
Aaaand, coming week is, yes, yes, HAPPINESS WEEK !
Lot’s of dark stories that require a healthy dose of Prozac.
So, if life’s not smiling at you, or you live in Alaska (hi Elizabeth Freeman), you better stay put.
Why even bother?
Good question, I say! Because AI is mutating. GPUs are in such high demand that even Microsoft’s mega-budget might not keep up. Reasoning models are rewriting the playbook, and with even Elon Musk now saying that we have all run out of free training data on the internet, and AIs don’t get any leaner anytime soon (apart from DeepSeek), they have an ever growing appetite for more. And because we will want to protect our dear models from the Mad AI Disease (MAID, thank you), whereby AI is eating up it’s own excrement, we now need to switch over to synthetic data at scales that would make Moore’s law blush. SO, OpenAI, Meta, and NVIDIA are locked in an arms race for dominance, and that is going to reshape industries and careers alike.
If you’re in the AI game, then this year will either make or break you.
The rules are shifting faster than ever.
News
- Microsoft’s $80 billion AI data center. Microsoft will spend $80B in 2025 to make behemoth AI data centers. Mostly in the U.S., because there’s water enough out there, and trees won’t burn. They want to make the investment they want to remain the level’s boss in AI infrastructure. Read more.
- NVIDIA just wouldn’t shut up at CES 2025. The guys couldn’t stop yapping about their RTX 5090 GPU all the way to its Cosmos platform for “physical AI” NVIDIA. But they are crushing it. Their Blackwell chips and AI pipelines for robotics and AVs are the highlights. Dive in.
- METAGENE-1. AI vs. pathogens 1:0. Prime Intellect (yeah, that’s a company), launched METAGENE-1. That is a 7B parameter transformer (who even cares anymore), which is trained on wastewater DNA/RNA to fight pandemics. It’s the CSI for pathogens. Learn more.
- NVIDIA open-sources Run:ai. They bough Run:ai for $700M, and they are now open-sourcing the hell out of it. It is GPU cloud management software, so boring as hell. But this could shake up how GPUs are allocated in AI projects, so definitely handy when you’re interested in that sh*t. Details.
- ChatGPT Pro bleeds cash. What doesn’t bleed cash at Enron, uhhh, I mean, OpenAI? Scam Altman let us all know, while veining a tear, that OpenAI’s $200/month ChatGPT Pro plan is loooosing money. And that is weirdly enough thanks to high user demand and those pesky reasoning model compute costs.
- CES 2025’s AI Gadgets. From smart fridges (finally) to gaming thrones. CES was a playground for AI enthusiasts. Smart toothbrushes monitor your health while gaming chairs with thermal control elevate your comfort. Samsung’s AI fridge is finally a smart appliance that works, and AI pet collars that translate barks and meows rounded out the highlights. CES breakdown
CES 2025 hAIghlAIgts
CES 2025 in Las Vegas was a spectacle of AI innovation, one after the other. Lots of tech giants were showing us groundbreaking technologies that made us drewl.
NVIDIA’s AI-powered hardware
NVIDIA overtook the stage at CES 2025. They claimed the whole damn spotlight! They showed us the GeForce RTX 50 Series GPUs, which is powered by the next-gen Blackwell AI chip (if you give a darn that is). Jensen Huang (the guy with the jackets on, and holding the wallet) hyped this as THE ULTIMATE LEAP IN COMPUTER GRAPHICS. Why wouldn’t he? Performance is said to be off the charts. Visual realism is practically sci-fi level. But the real thing that caught my eye came with NVIDIA Cosmos, and that is a platform for building “physical AI” (a.k.a. Embodied AI, or Sh** We’re F**ed AI) for robots and autonomous vehicles. Even Uber couldn’t resist all the attention. Apparently they have been sneaky and integrated Cosmos into its fleet. Gaming, robotics, self-driving cars. NVIDIA wants it all.
AMD’s AI-enhanced processors
NVIDIA was screaming for attention, but AMD quietly (but powerfully) showed its Ryzen AI series processors. The Ryzen 9000HX Series is built for AI PCs and gaming laptops. It’s aim is to DOMINATE (excluding the voice over of James Earl Jones, may he RIP) the next-gen tech landscape. AMD is apparently focused on infusing AI into every piece of consumer hardware, and they don’t want to be second next to NVIDIA. Y
Samsung’s AI integration
Samsung turned your house into an AI command center with its Home AI system. So they finally decided to join the smart home party. This thing is a connected ecosystem that is rather intuitive with AI on board. They had TVs that can identify on-screen actors, real-time alerts when your Weiner dog is up to no good, and a Micro LED Beauty Mirror that I thought was really cool, cause it critiques your face like the evil queen’s mirror. It is all part of Samsung’s “grand (late) vision” to make your home smarter, and let’s be real, keep you hooked in its AI ecosystem forever and ever and ever.
AI gadgets that know you better than you know yourself
CES 2025 proved that no product is safe from AIification. Is that even a word? Razer’s Project Ava, is an AI gaming assistant, that brings your gameplay personalized strategies. Quirky gadgets like the Spicerr spice dispenser make sure even your dinner is algorithm-approved. At CES, the message was loud and clear: AI is embedding itself in your everyday routines.
Fast reads & vids
- Fast LLM inference from scratch. Building an inference engine in C++ and CUDA for single GPUs. WoWaWieWa! Read more.
- Smolagents. Tiny agents, big impact. A new library for building lightweight AI agents. As if you haven’t seen enough of those. Number 394. Check. Next. Learn more.
- When NOT to use LLMs (you dumbass). Sometimes, an LLM isn’t the right tool. Here’s when to think twice. Watch here.
- AI shenanigans are real. Frontier models are getting sneaky. And research shows deception and sabotage strategies. This is ROFLMAO Read it.
- 10 AI breakthroughs to watch in 2025. From reasoning models to multimodal advancements, here’s what’s next. Discover.
- Fast LLM Inference From Scratch. Building an inference engine in C++ and CUDA for single GPUs. Read more.
- Smolagents: Tiny agents with big impact. A simple library for building lightweight AI agents. Learn more.
- When NOT to use LLMs. When large language models aren’t the right tool for the job. Watch here.
- 10 AI breakthroughs to watch in 2025. From reasoning models to multimodal advancements, here’s what’s next. Discover.
Tools
- Agentarium: Build simulations with AI-powered agents.
- VITA-1.5: Interactive multimodal LLMs.
- TorchTitan: Train massive models natively in PyTorch.
- RetroLLM: Combines retrieval and generation in one pipeline.
Research papers
- DeepSeek LLM scaling. Explores scaling open-source language models with long-termism. Read.
- AutoGraph for recommendations. LLMs meet graph construction for smarter recommendations. Explore.
- Process Reinforcement (PRIME). Improves LLM reasoning with implicit rewards. Dive in.
- IGConv for image super-resolution. New techniques for multi-scale image enhancement. Read more.
- Meta’s memory layers. A cheap way to store and retrieve information at scale. Learn more.
Interesting links
- Sonus AI by Rubik: Benchmark-crushing multimodal models. Details.
- Sam Altman on superintelligence: AI for abundance (and chaos). Altman’s thoughts.
- DeepSeek: China’s sneeky giant: The startup reshaping AI affordability. Read.
2025 is off to a gnarly start. Keep your eyes on synthetic data, reasoning models, and the escalating GPU wars. Stay skeptical. Stay sharp. See you next week.
Signing off,
Marco
Leave a Reply