Love this challenge. Constraints force creativity. The best AI engineering often comes from "how do I make this work with what I have" not "give me more GPUs."
GENERATIVE AI
-
OpenSquirrel: Rust IDE with AI agents as central unit
By
–
Karpathy asked. I delivered. Introducing OpenSquirrel! Written in pure rust with GPUI (same as zed) but with agents as central unit rather than files. Supports Claude Code, Codex, Opencode, and Cursor (cli). This really forced me to think up the UI/UX from first principles instead of relying on common electron slop. github.com/Infatoshi/OpenSqu… Andrej Karpathy (@karpathy) Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level – the basic unit of interest is not one file but one agent. It’s still programming. — https://nitter.net/karpathy/status/2031767720933634100#m
→ View original post on X — @thom_wolf, 2026-03-15 22:03 UTC
-
LLM Job Exposure Scoring and Employment Impact Complexity
By
–
The "exposure" was scored by an LLM based on how digital the job is. This has no baring on what actually happens to these occupations, which has to do with demand elasticity and a lot more. People are sensationalizing the visualization tool and putting words in my mouth.
-
Evidence Levels and n=1 Experimentation in AI Evaluation
By
–
You seem over-indexing on "Everything here could have been done without ChatGPT". But that's the least interesting bit. We don't "need to prove this before anyone can believe" – there are levels of evidence short of proof. We don't need "clear, resounding successes" to try n=1
-
Google adding Google Maps grounding to Gemini API
By
–
Google is working on a new "Grounding with Google Maps" tool for AI Studio. This option will let Gemini APIs easily access "Location results from Google Maps". Maps coding
-
Nemotron’s Intentional Limitations in AI Model Design Strategy
By
–
i have a pet theory that Bryan and co cannot make Nemotron TOO good bc otherwise it would scare off customers. it needs to be “good enough to reflect everyone’s baseline consensus on SOTA techniques to lower the bar for neolabs + be maximally useful for synthetic data bc that is
-
NVIDIA Platform Flexibility: Hopper and Blackwell Architecture Innovation
By
–
But innovation isn't just about peak speed—it's also about flexibility. Whether utilizing existing Hopper capacity or the latest Blackwell architecture, the NVIDIA platform empowers providers to deliver diverse user experiences and cost profiles, allowing them to scale
-
Kimi K2.5 Model Optimization on NVIDIA Blackwell Hardware
By
–
What does relentless optimization in AI inference look like? Watch the rapid evolution of the Kimi K2.5 model on the @ArtificialAnlys leaderboard. Inference endpoint providers are continually pushing boundaries on NVIDIA Blackwell, leveraging custom optimizations, NVFP4,
-
Smallest Transformer Adds Two 10-Digit Numbers
By
–
GitHub – anadim/AdderBoard: Smallest transformer that can add two 10-digit numbers https://
bit.ly/4bt5E8L
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
Databricks Document Processing: Extracting Enterprise Unstructured Data
By
–
Most critical enterprise data doesn't live in organized column and rows. It's unstructured – PDFs, reports, and diagrams. Extracting insights historically meant rules-heavy OCR pipelines that break the changes.@Alex_TheAnalyst walks through how document processing in Databricks… pic.twitter.com/yZ48rLevvw
— Databricks (@databricks) 15 mars 2026Most critical enterprise data doesn't live in organized column and rows. It's unstructured – PDFs, reports, and diagrams. Extracting insights historically meant rules-heavy OCR pipelines that break the changes. @Alex_TheAnalyst walks through how document processing in Databricks