Nisarg @nisargcodes

@MistralAI that's not the full comparison now, is it?

Its been a while since I last posted but I can't keep up posting daily because of my college timings I don't get enough time to learn new things, so I am going back to posting weekly. Also need to work on my time management skills as well, cause I procrastinate a lot!!!

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

Introducing Gemini 4 Argon – our new frontier model. It’s built for complex workflows across coding, enterprise knowledge work, and cybersecurity defense – rolling out today to a set of trusted testers through our Fairwind Program.

@SpaceX The quality! 🔥

Day 7, didn't get time to learn anything new due to college assignments, need to cover more tomorrow!

Reading this today!

Day 5/6 didn't get time to learn anything new as I spent it entirely on the ML Challenge. Finishing with a score of 0.972

@cneuralnetwork depends on the task you do, codex works for structured repetitive tasks while claude works if you want long context windows and reasoning!

@soumilibuilds they would have overfitted 100%

Day 4 of 30 posting daily about what I learn. Didn't learn much new today, the Amazon ML Challenge took the whole day. My approach turned out to be somewhat wrong, so I'm scrapping it and starting over tomorrow with a different one. Let's see how it goes!

@cneuralnetwork Bro tips please!

Day 3 of 30 posting daily about what I learn. Light day: continued RL with Markov Decision Process. Also participating in the Amazon ML Challenge tomorrow. If you've done it before, what's one thing you wish you'd known going in? #buildinpublic #MachineLearning #AmazonMLChallenge

The link I found: in DTN, a node must pick who to forward a bundle to without knowing future contacts, and with limited storage. That's a sequential decision problem with delayed feedback, exactly what RL is built for. [4/4] #buildinpublic #ReinforcementLearning #MachineLearning

DTN fixes this with store-and-forward. Data travels as "bundles" (Bundle Protocol). Each node keeps a bundle in storage until the next link opens, then passes it on. No live end-to-end connection is needed, so the network tolerates delay and disconnection. [3/4] 🧵

Day 2 of 30 where I post daily about what I learnt. Today: How space communication works, specifically DTN (Delay/Disruption Tolerant Networking), followed by RL Basics [1/4] 🧵

Next up: fine-tuning CLIP with a data-efficient curation strategy — applying a technique from my own data-selection research to a new domain. #buildinpublic #ComputerVision #DeepLearning #MachineLearning

Baseline numbers (pretrained CLIP, zero fine-tuning, 5K images): • Precision@5: 0.764 • FAISS query speedup: 2.8x vs brute-force [3/n]

Day 1 of 30 where I post daily regarding what I learnt. Here's what I built today - StyleLens, a CLIP-based visual style search engine for fashion images [1/n]