Recently we have built our in-house animation engine to help us build more intricated visual story telling and the first thing I went on prompting is essentially a single visual that would indeed be a loop, and indeed showcase what agents loops do.
It is a loop.
Friday we released https://t.co/WxIgWMvRe5; a procedural game where two 'decision' model are running one against the other. We let ir run since & now after 5000 runs for each models. The picture is much clearer, and on the median; Kumo does 2.1x better than a Qwen3-4B model.
My O'Reilly book on building ML systems with a feature store is still available for download free for only 1 more month. Get it now.
https://t.co/ifevzf3NmD
How to see if a pretrained model capable of decision would take pick the right decision, at the right time?
Make a procedural game, make sure the difficulty increases, measure Semif's Qwen vs Nvidia's Kumo after 500 runs.... profit ?
Asked claude to setup openjev equivalent on our saas platform to see what was the hype about.
Is the internet so starved for hype that this will be going into a faster hype to death loop than openclaw?
(mind you. this is running on CPU!)
I was trying to express how I implicitely decide, or not, to care to try new (ai) products. For once, a Venn diagram is indeed quite fitting.
As many products are solving problems that maybe... needed little to no solving to begin with.
It helps me. Might help you.
Six months ago we said: don't be a wrapper, be a container.
While everyone was wrapping their platforms in UIs hidding coding agents; we put them inside ours, next to the data, with a CLI and a sandbox.
Aged pretty well.
https://t.co/L6UKo4IlTy
The 𝐅𝐓𝐈 (𝐟𝐞𝐚𝐭𝐮𝐫𝐞, 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠, 𝐢𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞) architecture.
A mental map to build ML Systems and not get lost in the process ...
#MachineLearning#MLOps#DataScience
Training ML models is easy.
Transforming the data these models need is the hard part... until you learn this ↓↓↓
𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺
Building a real-world ML system is
> 𝟭𝟬% about training and deploying ML models,
and
> 𝟵𝟬% about transforming the data these models needs to work.
And the thing is, 𝗻𝗼𝘁 all data transformations are the same.
𝗧𝗵𝗲 𝘁𝗮𝘅𝗼𝗻𝗼𝗺𝘆 𝗳𝗼𝗿 𝗗𝗮𝘁𝗮 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻𝘀
In every ML system we can have up to 3 different types of data transformations
1️⃣ 𝗠𝗼𝗱𝗲𝗹-𝗜𝗻𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝘁 Transformations, for example rolling averages.
> Reusable across models
> Stored in the feature store
2️⃣ 𝗠𝗼𝗱𝗲𝗹-𝗗𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝘁 Transformations, for example feature normalization
> Specific to one model
> Applied in both training and inference
3️⃣ 𝗢𝗻-𝗗𝗲𝗺𝗮𝗻𝗱 Transformations
> Require real-time data
> Used in online inference
Once you understand how and 𝗪𝗛𝗘𝗥𝗘 your data transformation happens, you are in a good position to start building ML software that works.
If you want to learn more about the taxonomy of data transformations read this excellent blog post by the great @jim_dowling
> 🔗 https://t.co/q7gTBCbxag
----
Hi there! It's Pau Labarta Bajo 👋
Every day I share free, hands-on content, on production-grade ML, to help you build real-world ML products.
𝗙𝗼𝗹𝗹𝗼𝘄 𝗺𝗲 so you don't miss what's coming next
Why is building AI systems hard?
Because even with the simplest feature that you can imagine, you need to consider where it will be computed based on whether it's a batch AI system or a real-time AI system.
Luckily, there are some principles to guide you.
A thread 🧵👇