The Vision for #W3C# has been published as a W3C Statement: As one of the editors, along with Chris Wilson (@cwilso), I’m both proud of this multi-year W3C Advisory Board effort, and grateful … #w3cVision# #w3cAB# #openWeb# #Vision#

Stanford just reinvented text generation. (it’s neither autoregressive nor diffusion) Autoregressive LLMs generate one token at a time. Longer answers require longer chains of sequential predictions. Diffusion LLMs generate many tokens in parallel, but still need repeated refinement passes. Each pass processes the full sequence, and quality drops quickly when you cut the passes. Stanford researchers proposed another approach, Next-Concept Modeling (NCM). It generates text from coarse structure to fine detail in roughly a dozen passes. First, a tokenizer learns to represent text as layers of compressed information. Early layers capture broad structure, while later layers add details the earlier ones missed. A transformer then learns to generate these layers in order, predicting everything within each layer simultaneously. A decoder converts the completed representation into text. The advantage is that early layers are tiny. Generation needs fewer passes, and those early passes process much less information. You can see this in the team’s translation example. Early layers produce rough, repetitive text, which later layers refine into a clean translation. In the researchers’ tests, NCM generated text 10–31× faster than strong diffusion baselines, with comparable or better benchmark scores. On LM1B and OpenWebText, it also used roughly 4–5× fewer training tokens, including tokenizer training. That speed can also improve quality. Generating several candidates and selecting one based on agreement outperformed diffusion baselines at the same sampling time budget. The limits matter. Scaling experiments reach 725M parameters, tasks use separate tokenizers, and the release includes evaluation code and checkpoints without a training pipeline. Some evaluations also use repetition penalties during final decoding, without additional model calls. NCM shows that parallel generation can work by building increasingly detailed representations. Whether those gains hold at general-purpose LLM scale remains open. - GitHub repo: - Paper:

Good Chinese openweight models will be optimized for Chinese hardware first. For DeepSeek V4s versions, Huawei Ascend are some of the only two stacks with optimized inference ready, alongside CUDA. The exceptions to this might be the small models that fit gaming GPUs.

Bullish on open source, open weights, the open web And the open desktop Personal Operating Systems feel like the natural next step after Personal Software I've been building a Debian-based OS for myself that can run across physical machines and cloud environments like Vercel Sandbox I can use and control it with natural language, and I'm designing the UI around a blend of ephemeral and persistent views rather than treating the desktop as a fixed set of apps and windows There's something really special about having control over the UI/UX of the Whole Computer

Media is changing at AI speed: the open web deteriorating, search shriveling, LLMs rising, hyper-personalization surging, old revenue models breaking. •AI threatens to finish what the algorithms started: wholly upending journalism as we know it

Need Live Information from the Open Web? Start here to see how Teneo Data Access helps you: 🧭 Request real-time info through the Agent Console 💬 Ask agents in natural language 📦 Export structured results as JSON 🔌 Connect live data to your app with the Client SDK ⚡ Automate workflows from open web signals Whether you’re tracking Web3 trends, listening to communities, or researching projects, this is the starting point. Explore →

the part that stands out to me with Grass is how valuable access to the open web becomes as AI scales more websites are putting anti-bot infrastructure in front of public data, while cloud and VPN traffic gets filtered more aggressively Grass uses distributed residential internet connections to help AI labs access publicly available web data through a different route as models need more fresh data, that access layer becomes increasingly important

A new API gives coding agents their own dedicated search engine for code — one that only looks at official documentation instead of the open web. Title: Context7 Search: A Grounding API for Coding Agents URL: 📝 Overview It's a new Search API where a single GET request with your question lets Context7 automatically pick the right library and return the best code snippets. ❗ Problem it solves General-purpose search engines tend to mix in stale or inaccurate code, and the previous Context7 could only search one library at a time. ⚙️ Methodology It only indexes primary sources — official docs, product sites, API references — scanning for malware and prompt injection before indexing, and every result carries a library ID and source URL for attribution. 🔧 Use cases A single question like "how do I stream an OpenAI response from a Next.js route handler?" pulls snippets across multiple libraries at once. It's built for IDE plugins, chat apps, and code review bots too. 💰 Pricing The Free plan covers 1,000 API calls a month, Pro covers 5,000 calls per user monthly, and usage beyond that is $10 per 1,000 calls. #AIAgents# #DevTools#

not enough people stop and think about how the best tech of the past few decades was built on open source: linux, git, much of the open web. 70-90% of software in a modern codebase is open source. the beauty is you’re always standing on the shoulders of giants. there’s a reason why open-weight is progressing exponentially. glm-5.3 evolved from attention -> deepseek sparse attention -> indexshare. even with limited compute compared to closed labs, teams cut attention below quadratic cost through making architecture and algorithms efficient. open-source wins stack and compound. decentralized innovation is truly a remarkable force.