The world's leading publication for data science and artificial intelligence professionals. Submit an Article ✍️ https://xitter.cf/t.co/57pIMegK1o
We're Global 🌏
Joined October 2016
- Tweets64.6K
- Following1.9K
- Followers251K
- Likes8K
Towards Data Science retweeted
Big thanks to the @TDataScience team for featuring my new piece in the Editor's Picks. If you’re curious about how skills and MCP reshape trust boundaries in LLMs, this one’s for you.
towardsdatascience.com/skill…
Towards Data Science retweeted
Cheaper LLMs ≠ Lower costs.
Great insights from Nhu Hoang on @TDataScience about the hidden costs of LLM routing.
Optimizing AI isn't just about model pricing—it's about total cost, latency, and quality.
Worth a read 👇
towardsdatascience.com/the-h…
#LLM #AIEngineering
Cheaper LLMs don't always mean lower costs.
While exploring LLM routing, I found that switching models has hidden costs too.
towardsdatascience.com/the-h…
#LLM #GenerativeAI #AIEngineering #LLMRouting #AgenticAI
Towards Data Science retweeted
Okay, finally I can announce it.
So excited to make a comeback for @TDataScience !
So here is the glimpse for what I made,
Towards Data Science retweeted
I wrote an article for @TDataScience, and it ended up as an Editor’s Pick
It’s about grid puzzles people solve easily, but the best AI models still get stuck on, and what this tells us about their spatial reasoning skills. Thanks to @jpwahle for the help!
towardsdatascience.com/spati…
Towards Data Science retweeted
Cheaper LLMs don't always mean lower costs. While exploring LLM routing, I found that switching models has hidden costs too. Check it out at @TDataScience
#LLM #GenerativeAI #AIEngineering #LLMRouting #AgenticAI
Switching to a cheaper model should... lower your overall costs, right?
@nhu_hoang Nhu Hoang explains why that's not always the case, and stresses the importance of verifying the hidden costs of LLM routing. towardsdatascience.com/the-h…
Towards Data Science retweeted
Stacking doesn't automatically beat your best model.
Diverse models: 0.830 vs 0.820 ✅
Correlated tree ensembles: +0.1 at 35x the training cost ❌
Plain voting lost to the best single model both times.
@TDataScience
Full test + notebook:
towardsdatascience.com/does-…
Towards Data Science retweeted
Jev will change the future of agentic builds
For a nuanced understanding of LLMs' limitations when it comes to making decisions, and the growing need for "a system primitive built around decisions rather than strings," follow along @leonglambert's patiently explained deep dive. towardsdatascience.com/when-…
Towards Data Science retweeted
Building AI applications that scale and evolve is the real challenge.
👹 That's why Janus looks both ways: learning from past engineering lessons while preparing for future change.
🤖 AI needs more than fast code. It needs solid foundations. @TDataScience towardsdatascience.com/a-sca…
Towards Data Science retweeted
Every time an AI system works with text, something interesting happens behind the scenes.
How can we turn human language into numbers that machine learning models can learn from?
Explore the foundations behind it in my article on @TDataScience👇
towardsdatascience.com/from-…
#AI
Towards Data Science retweeted
The Python 3.14 Just-In-Time (JIT) compiler was good, but the new Python 3.15 JIT is even better. Find out how much it's improved it is as I take the new JIT for a spin and compare it to its predecessor in my new @TDataScience article.
track.pstmrk.it/3s/towardsda…
Towards Data Science retweeted
Your training data might already be contaminated, and you probably can't tell.
I tested three ways to spot AI slop in datasets. One of them failed badly.
Read it here: towardsdatascience.com/ai-sl…
@TDataScience
Towards Data Science retweeted
Had an amazing opportunity to write an article with
@TDataScience
Check it out here: towardsdatascience.com/why-t… #AI #LLM #DeepLearning #Temperature
Towards Data Science retweeted
If you want to learn the basics of AI agents, tune into the first of my 3 part series on @TDataScience
Read it for free using the link below,
track.pstmrk.it/3s/towardsda…
Towards Data Science retweeted
Thank you @TDataScience for publishing this.
Agents can write real CUDA. On a DGX Spark, one kernel beat torch.compile by 1.57×.
The harder part was the benchmark. A bad one can invent a speedup. The agent didn’t cheat. The measurement almost did.
"The bigger takeaway wasn’t about CUDA—it was about benchmarking. Building a trustworthy evaluation proved harder than generating the optimized code itself."
@chien_vu1692 presents a comprehensive walkthrough of his work to test AI agents' performance at writing CUDA kernels. towardsdatascience.com/ai-ag…
Towards Data Science retweeted
🚀 Can AI agents write faster CUDA kernels than PyTorch?
Excited to share my latest article published in @TDataScience !
Exploring how AI agents can optimize GPU performance with custom CUDA kernels.
towardsdatascience.com/ai-ag…
Towards Data Science retweeted
Part III is live: time-series diffusion.
Parts I–II showed why a single Gaussian can’t capture a forecast’s real uncertainty. Diffusion models drop that assumption entirely. Happy reading
@TDataScience 👇
towardsdatascience.com/your-…
Towards Data Science retweeted
A Google team measured half of my argument.
Spec-driven test generation raised bug detection by 9.8 percentage points on their own codebase. Details on this, and the other half of my argument, in my latest article at @TDataScience
towardsdatascience.com/a-goo…
Towards Data Science retweeted
Wrote another amazing article with @TDataScience, do check it out!
"I wanted to know how small and simple a model could be and still show this blind spot. Not a fine-tuned LLM, but something I could build in an afternoon with nothing but NumPy and watch fail."
Utkarsh Mangal digs into LLMs' reversal curse. towardsdatascience.com/the-r…
Towards Data Science retweeted
I Compacted 1,000 Apache Iceberg Files Into 6. Here’s What Happened to Query Performance.
Read my latest article on @TDataScience for free.
track.pstmrk.it/3s/towardsda…
Towards Data Science retweeted
Not every AI decision needs a generative model.
We've spent so much time making LLMs better at reasoning that we've overlooked how inefficient they can be at simply deciding.
Jev could fundamentally change how we architect agentic systems.
Thanks for the feature, @TDataScience
"This has never sat right with me: a bounded yes/no or routing decision was often passed through the same autoregressive decoder used to generate a paragraph."
Lambert Leong @leonglambert digs into the architectural constraints that are steering the field towards "decision" models like Jev. towardsdatascience.com/when-…