NVIDIA Star Elastic — Carving Three Reasoning Models from a Single Checkpoint with Zero-Shot Slicing

Star Elastic, released by NVIDIA, is a reasoning model built on an unprecedented approach: it can pull three reasoning models — 30B, 23B, and 12B — out of a single checkpoint via zero-shot slicing.

Traditionally, offering models at different parameter sizes required training each one independently. With Star Elastic, a model of the desired size can be extracted directly from a checkpoint obtained in a single training run, by slicing away parameters that are unnecessary at inference time.

The Reddit LocalLLaMA community has also weighed in on the feasibility of running it locally, with the 12B model expected to run on consumer GPUs. On the inference strategy front, there is also discussion of the potential to switch model sizes dynamically.

This approach could dramatically cut the cost of distributing and operating models, and it amounts to a new paradigm for delivering a consistent model family from edge devices to the cloud.

Stale Gov.uk Pages Feed Misinformation to AI Search — A New Challenge Around Information Freshness

According to The Register, a problem has surfaced in which old pages on Gov.uk, the UK government site, are being pulled into AI Overviews, leading UK citizens to trust outdated information.

Gov.uk retains a large number of pages that have not been updated in years, and AI search references them as "official information." Cases have been confirmed where outdated versions of annually changing information — such as tax rules and welfare programs — were presented as AI answers.

This issue exposes a fundamental weakness of AI search. It is a structural challenge: even when a search engine can confirm that a piece of information exists, it fails to properly evaluate how fresh that information is. The same problem can arise beyond government sites — with outdated technical documentation, deprecated API specifications, and the like.

The government's remark that "AI is gaslighting citizens with stale data" is strong language, but it points to a shared responsibility: both content providers and AI search engines bear the duty of guaranteeing information accuracy.

An AI Phenomenon Upends Assumptions About Learning — New Research from the Santa Fe Institute

Research published by the Santa Fe Institute shows that an unexpected phenomenon observed in the AI learning process is calling into question long-held assumptions about human learning.

Cognitive science has long considered "gradual learning," "explicit feedback," and "structured curricula" essential for efficient learning. However, a specific learning phenomenon observed in neural networks (see the paper for details) suggests the existence of learning pathways that run counter to these assumptions.

The institute's approach is not to use AI as a simple model of human cognition, but rather to derive new hypotheses about human learning mechanisms from the unexpected behaviors AI exhibits. This trend of feeding AI research findings back into cognitive science is worth watching as an interdisciplinary meeting point between the two fields.

Cited #1 in AIO Despite Ranking Outside Organic Results — A Practical Study of Structured Data and llms.txt

An article published on Zenn analyzes the phenomenon of a site ranking outside the organic results being cited as #1 in Google Search's AI Overview (AIO), examining the background in detail.

An SEO tool site — only three months old with a short domain history — was cited at the very top of AIO despite not even making the first page of organic search results. According to the article's analysis, the key to this gap lay in the implementation of structured data and llms.txt.

In organic search, backlinks and domain authority are major ranking factors, but the analysis suggests that AI Overview's citation decisions may place greater weight on the structured semantic information contained in content. As practical insight for thinking about where SEO is headed, this is a compelling finding.

The Reality of Indie Development in the Age of AI Coding, Learned from Shipping Six SaaS Products to Production

On Zenn, a retrospective by a developer who shipped six SaaS products to production in five months using Claude Code has been drawing attention.

The candid conclusion is striking: "AI made solo development blazingly fast" is true, but "it sells as-is" turned out to be a different story. Of the six products in operation, only two have stable MRR (monthly recurring revenue). The remaining four continue to incur monthly costs.

The article lays out the real numbers behind mass-producing products on a Claude Code Max plan, an analysis of failure patterns, and the gap between "building fast with AI" and "sustaining a viable business" — all as unvarnished, first-hand information.

It is a valuable on-the-ground record showing that while AI coding tools have lowered the barrier to "building," the difficulty of "selling" and "maintaining" remains unchanged.

---

**Sources:**

[NVIDIA AI Releases Star Elastic: One Checkpoint that Contains 30B, 23B, and 12B Reasoning Models with Zero-Shot Slicing](https://www.reddit.com/r/LocalLLaMA/comments/1t8s83r/nvidia_ai_releases_star_elastic_one_checkpoint/) — Reddit LocalLLaMA

[Stale Gov.uk pages are feeding AI overviews old data and Brits are believing it](https://www.theregister.com/software/2026/04/23/govuk-says-ai-gaslighting-brits-with-stale-govuk-data/5229766) — The Register

[Upending assumptions about learning, inspired by an AI phenomenon](https://www.santafe.edu/news-center/news/upending-assumptions-about-learning-inspired-by-an-ai-phenomenon) — Santa Fe Institute

[オーガニック圏外なのにAIOに引用された — 構造化データとllms.txtの実装が鍵だった](https://zenn.dev/orectic/articles/aio-organic-gap-structured-data) — Zenn

[6本のSaaSを本番に乗せた後でわかった：AIコーディング普及後の個人開発の現実を全部書く](https://zenn.dev/mintototo1/articles/observation-ai-coding-saas-reality-2026) — Zenn