Steve Lucco

First piece on the "AI Native Dev". Read up on how we make Claude (and other agents) better at Go with Tessl's tech!

19

Barbarians at the Gate (https://lnkd.in/gkEBX79s) is a very interesting new paper from some great folks at Berkeley, with some exciting results showing the potential for AI in systems research. But I think the authors aren't quite asking the hardest problem about where this takes systems as a field. I wrote a new blog post about it: https://lnkd.in/gxS3s4v4 The core point of the paper is that AI can be effective at automatically driving the 'number go up' part of systems research. Given an evaluator, AI can automate progress on implementation. I believe this is 100% correct. The paper also talks about how more focus will have to go into developing evaluators. Again, I 100% agree, but think they're understating their case. Building evaluators (basically tests) is, I believe, going to become the focus of systems and software over the next decade. Which leaves researchers with the hard problem of choosing which problems to go after. That's never been easy in academic systems, and is only going to become more of a challenge. But the best systems research is driven by choosing problems well! This vision for AI could lead to lots more of the least interesting systems research. The "number go up" kind. Which could make it harder than ever to identify the truly innovative and visionary work. We're already seeing that elsewhere in science. So, mostly, I think the paper is right, but not yet asking the biggest question: what should systems research be?

276

12 Comments

LLMs may have large portions of copyrighted works directly accessible verbatim. Could be a tough one to explain in court. How would they ever deal with take down requests? https://lnkd.in/g6z3ipdC

1

1 Comment

We use latent continuous thoughts for retrieval optimized via downstream NTP loss, unified under one LLM backbone. Since representations are shared, documents can be precomputed—eliminating 2-stage RAG. We match raw text performance but with a much shorter context budget!

38

In the AI world, four months might seem like an eternity. We’ve endured such four-month open-source dry spells before - only to have them broken by models that redraw the map. DeepSeek AI DeepSeek-R1 arrived in January 2025 after a quiet autumn and suddenly “open-source” meant chain-of-thought reasoning on par with the best closed systems. Barely seven months later, OpenAI GPT OSS series added native tool use and a Responses API that turned every model into an agent. That drought ended again when NVIDIA Nemotron 3 Super, unveiled at GTC, is the first large-scale model trained natively in NVFP4: it learns at 4-bit precision from day one, no post-training quantization required. Coupled with Mamba-2–Transformer hybrids, latent MoE layers, and MTP training, it keeps peak accuracy out to almost a million token context window while running up to 3× faster on Hopper and Blackwell GPUs. It’s hard to believe it’s been nearly two years since Ihab Tarazi and Jeff Boudier stood on keynote stage at Dell Technologies World in May 2024 and introduced the Dell Enterprise Hub with Hugging Face. In that time, open-source AI has sprinted from “promising” to “production-grade,” and DEH has sprinted right alongside it - platform, model specific performance and software optimization, python SDK, CLI, app catalog, enterprise security features, Goodput, partnership with model publishers for day 0 launches with early access collaborations and of course the one-click deployment pipelines that let enterprises ship models as fast as the community releases them. Proprietary and open models aren’t rivals anymore - they’re teammates. Jensen said it plainly at GTC 2026: the future belongs to “proprietary and open,” and DEH https://dell.hf.co is the place where that future lands first for enterprises. 📔 GTC 2026 Insights: Through the Dell Enterprise Hub Lens https://lnkd.in/gF7yvFuw #iwork4dell #opensource #gtc2026 #insights #DEH #harness #partnership

35

AI-generated code has a readability problem. As a former Googler, I’ve noticed a frustrating trend: LLMs love to generate a massive wall of comments before dumping 5 lines of functional code. While comments can help explain intent, deciphering whether the code actually does what the comment claims becomes a chore. Back in the day, almost any // comments in Java were considered a code smell. Good code was expected to be self-documenting. Should the bar for AI code generation just be "does it run"? Or should model designers and harness makers train LLMs to write code that meets Google L3 engineering standards—clean, maintainable, and self-explanatory? #SoftwareEngineering #AI #CodingStandards #CleanCode

153

43 Comments