🚦Hitting performance walls in pandas? We ran 3 example workflows that slowed or crawled on large datasets—then reran the exact same code on GPU with cudf.pandas. TLDR: ✅ 18M rows of stock prices → 20–40x faster with time-based rolling windows ✅ 8GB job postings CSV → up
Gentle reminder: Do not overly pursue high TPS, data shows that only 15% of on-chain transactions require rapid processing, compliance and stability are more important.
Remember the programming language Ada? Or maybe you've never heard of it at all? 👀 Either way, it's time to get familiar: Ada is back in the top 10 of the TIOBE Index. 🏆 Find out why its focus on security is putting this language back in the spotlight.
Oh for sure, not today. Security / reliability has always been at odds with usability. The design fundamentals of robust systems are well understood. They just tend to be to inaccessible to beginners. This allows us to start to rebuild basic infra primitives to suck a lot
Solana TPS — one of the most beautiful charts in crypto imo you can see in the earlier 2022 days Solana would break 500-800 TPS and start having outages issues since then, devs have been hard at work and performance has been going up and to the right with ATHs in scale
🇨🇳AI JUST CLASSIFIED 27 MILLION OBJECTS IN SPACE Astronomers used to spend years classifying stars, galaxies, and quasars one spectrum at a time. Now? An AI from China just sorted 27 million cosmic objects like it was swiping through a starry Tinder. Trained on trusted Sloan
Right now, availability is so limited we don't expect to run these at scale, at least not for a couple more months, but running a single instance to host coding models for our users - that's a definite possibility. We run B200s, H200s, H100s, A100s, L40S and 4090 48GB.
So now where this will make sense for inference, we barely fit quantitized Q8 Qwen Coder 3 and Kimi K2 instances on our H200s. Kimi K2 @ Q8 left no room for a kv cache for the context. Could these models fit on a single 8xB200 instance? Probably, we'll try this week.
The use of simulating myself was a handy shortcut to consider other constructivists I might have missed. As for the Superbowl prediction I think the AI is using a form of fuzzy logic. It's not "just guessing".
There is an argument vibe coding will make apps more safe not less. A lot of sloppy languages/systems exist because they're easier to learn. But with AI that is no longer needed. So we can actually build not shit systems out of not shit languages. And still get them adopted.