Welcome back to SundAI ! – 🎉 First edition of 2025 🎉
Your weekly not-so-nasty and not-so-dystopian overview of what has been happening in the world of Big AI – because there’s more than just ranting.
Tech never sleeps, and neither do the AI breakthroughs (or breakdowns). This week, I’ve got a glimpse of what might just redefine the rules of AI accessibility. There is DeepSeek V3 who does it much faster and much much cheaper than the rest, all the way to OpenAI’s luxury-overpriced o3, and oh yeah, I’ve got AI toothbrushes, CES 2025 and AI-powered email scams.
Anyways, the global AI arms race is heating up. And let’s not forget the eye-catching releases from Alibaba, Microsoft, and Meta that are reshaping the world as we know it.
DeepSeek V3
China’s DeepSeek V3 Mixture-of-Experts (MoE) model is all about efficiency in AI training and inference. It has 671 billion parameters !!! and a training cost of only $5.576M !!!!!. That is a fraction of what top-tier models require. To me, it is clear that big performance doesn’t have to mean big budgets. It offers some unique engineering features as well like Multi-Token Prediction (MTP) and FP8 mixed precision training.
MTP is the ability to predict multiple tokens at the same time instead of one at a time. That is much faster. And FP8 is kind of complicated, that even for people in the industry its hard to comprehend. But let’s give it a shot. It is a technique that uses 8 bit floating point numbers to represent the weights and the activations that actually make the model, well…. the model. And FP8 uses much less memory and is bloody fast. Usually FP32 or FP16 is used. That is more precise, yet less fast. And much more costly.
Let’s get on with it.
This model also goes at at 60 tokens-per-second and reduces costs by 100x compared to peers. More importantly, DeepSeek’s transparency invites the global AI community to innovate further.
DeepSeek V3 vs. OpenAI o3
You know now that DeepSeek V3 slashes costs, but on the other hand OpenAI’s o3 escalates them. They have a staggering 6,000,000x cost increase for o3’s reasoning capabilities. And to me, that highlights just how different the two competing visions of AI are: one focused on affordability and openness, and the other on premium performance at any price. For enterprises, the choice between these models will shape strategies for years to come.
My take on this
The rise of cost-efficient, open models like DeepSeek V3 makes it clear that we are heading towards a shift in the AI landscape. Transparency and affordability are no longer optional. They have come to be essential for innovation and access to models. This model’s release also strengthens the global nature of AI. It is proving to me that innovation doesn’t belong to Silicon Valley alone, and that small to medium sized players with enough smarts can outwit even the largest. We now have both high-end and budget-friendly models making headlines. This fork-in-the-road is a future-defining moment.
More….
Qwen releases QVQ: Multimodal magic? or just another acronym
The Qwen team is back with QVQ. QvQ is a model that is built on the ruins (no, kidding) – the foundation of Qwen2-VL-72B, and it claims to have much much better visual reasoning and problem-solving. They are scoring 70.3 on the benchmark you couldn’t care less about (MMMU 🙂 ),
Source: Qwen’s fancy new model.
OpenAI’s has a new motto: Profits for humanity
OpenAI (or shall I say – ClosedAI) is transforming from a not-for-profit institution into a Delaware Public Benefit Corporation. That is apparently their way of saying, “We care about humanity, but also care about muchos stock optionas”
Source: Read about OpenAI’s noble profit quest.
Microsoft’s AIOpsLab
Dawat? Microsoft has a new playthingy called AIOpsLab. It is here to help you create AIOps agents and simulate all the disasters your cloud might face. This open-source framework is a crash test dummy for your cloud ops.
Source: Check out AIOpsLab.
Devin 1.1. Same agent. Just a tad better and less expensive
Cognition is the ones that brought us one the first AI based coding thingy called Devin. And now it is 10% faster and 10% cheaper, which is great if you like marginal improvements. It’s like Devin 1.0, but with a little bit more this and a little less of that. Lame if you ask me. They have been overtaken left and right and as we say in Dutch: “ This doesn’t put sods on the dike” (This doesn’t move the needle).
Source: Meet the new Devin.
Meta’s Large Concept Models
Tokens are sooooo 2024! And that;s why Meta decided to introduce LCMs. Those are models that ditch token-level thinking for higher-level “concepts.” They are trained on 2.7 trillion tokens and can summarize, expand, and outthink your average LLM.
Source: More on Meta’s fancy concepts.
Last week’s news
- AI-Powered toothbrushes and pet communication collars
- CES 2025 highlights
- AI-driven email scams
- Samsung’s AI refrigerator
- AI in advertising
Blogs and videos
- Building effective agents
- Getting started with agentic workflows
- LLM fine-tuning guide
- Visualizing GPU memory in PyTorch
- ModernBERT
Tools
- AI Magicx
- Pencil by Brandtech
Research papers
- Large models
- Scaling laws for precision
- Automating the search for artificial life with Foundation Models
- KAG: Boosting LLMs in professional domains via Knowledge Augmented Generation
- RemoteRAG: A privacy-preserving LLM cloud RAG service
Quick links
- Meta is backing Elon Musk in opposing OpenAI’s for-profit plans.
- Microsoft and OpenAI agree that $100B profit = AGI. (Sure, why not?)
And there you have it, my intelligent readers: the first SundAI of 2025.
It starts with cost-saving models and it leaves us with eyebrow-raising AI-powered appliances.
At least AI innovation (and absurdity) isn’t slowing down.
Jump into the links, question everything, and stay ahead of the curve, because the machines aren’t waiting. See you next week!
Signing off with a headache,
Marco
Leave a Reply