Wildroot

Queries may use an external AI service. Details

Wildroot Browser

TopicsAINews

AI · News

31 posts in AI that are news, facts and calls to action, where the model is sure of the intent. Page 1 of 2: the top 5, then 6 the model is less sure belong here.

Ranked Bluesky · Mastodon

About 7 in 10 posts here get an intent label; the model only labels the ones it is sure of.

How it worksBuild a feed from AI · News

  1. @lea-linux.org

    "Most modern transformer-based LLMs use the same broad structure: tokenization, embeddings, positional encoding, stacked transformer layers (each with multi-head attention and a feed-forward network), residual streams, layer normalization, and next-token prediction."

    Most modern transformer-based LLMs use the same broad structure: tokenization, embeddings, positional encoding, stacked transformer layers (each with multi-head attention and a feed-forward network), residual streams, layer normalization, and next-token prediction. What changes between models is: The trained weights themselves, learned from different training data at different scales. The configuration: number of layers, vocabulary size, head count, parameter count, MoE or dense. The post-training: instruction tuning, learning from human feedback, safety controls applied on top of the base model. Tiny explainer: weights Weights are the learned numbers inside the model. Training changes those numbers until the model predicts text well. The 2023-2025 “modern transformer” stack converged on a common set of choices across many serious frontier and open-weight models, even though different teams arrived at them independently. Pre-norm placement. RMSNorm. RoPE. SwiGLU. Grouped-Query Atten
    Info
  2. "We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced. What we found revealed security breaches, prompt theft, and widespread failures to protect customer data. Over a five-month period from July through November 2026, we submitted approximately two million prompts containing unique URLs and canary token credentials to widely used AI services and monitored what happened after inference. Our investigation uncovered three significant findings:

    Show more of this post

    - Evidence of provider compromises. - Unauthorized access to customer prompts by inference marketplace providers. - Failures in privacy and data-handling practices. We observed multiple" dragos.com/blog/ai-supply-chai… Privacy settings were ignored: "Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people." What to do? - Host it yourself, if you can afford it. - Go directly to a first-tier provider. - Use routers with care (or not at all) - Seed your own canaries. - Enable secret filtering

    Info
  3. @Jigsaw_You@mastodon.nl

    Alpöge's "proof" isn't a traditional mathematical paper. It's roughly 17 million lines of Lean code — a programming language used to mechanically verify logical arguments — backed by an underlying AI-generated text spanning more than 250,000 pages. That's approximately 500 full-length novels worth of mathematical argument, a volume so vast that Alpöge himself refused to publish it, calling it a "massive unreadable travesty of 'slop.'" #ai #tech #math #mathematics emeraldbook.org/news/nov-2126/

    Info
  4. @gaprogman.com

    Should you tell a CI agent to treat untrusted text as data? Yes; it improves the odds. But in October 2025, a team including Nicholas Carlini attacked twelve published prompt injection defences with methods that adapt to them, and got success rates above 90% for most.

    Info

More in AI

The model is less sure these belong here, so they sit outside the ranking.

  • @leogdion@c.im

    "The reason that AI music sucks is that it is generic." Graham Lee (@leeg) says software is both: art and engineering. Loading JSON from a web service, paging, caching and sorting it for a timeline is blue-collar code, and "putting too much creativity in there is probably the wrong thing to do." Designing the timeline and what people can do with it is where the craft goes. 🎧 brightdigit.com/episodes/216-d…

    Info
  • @gklijs@hachyderm.io

    A Chinese court ruled that a fintech company illegally terminated an employee's contract after replacing his role with AI, ordering compensation of over $38,000. The decision underscores that AI cannot justify layoffs without meeting strict criteria, including retraining opportunities. This case highlights increasing legal scrutiny of AI-driven job displacement in China. basic-income.gklijs.tech/weekl… Automated news digest

    Info
  • @china@universeodon.com

    MetaX says its C-series GPUs completed day-zero adaptation for Tencent's Hy4 flagship model after the 20 November weight release, and published FlagOS-based W8A8 weights for its hardware. Its model card reports 80.98 output tokens per second in the recommended decode-graph mode against 52.31 without. Tencent's Hy Licence 1.0 says the repository's Apache 2.0 tag does not apply. pandaily.com/metax-hy4-c-serie… #China #Tech #AI

    Info
  • Did Unity's 24-hour AI jam really run on Unity's AI? The coverage credits OpenAI's Codex and Unity CLI. What teams handed to AI, what Unity's credits cost, and whether to try it at your next jam. indietooling.com/posts/unity-a… #Unity3D #GameJam #GameDev

    Info
  • @abcnew.bsky.social

    AMD just dropped a free image model that skips the usual shortcut ⚡ Nitro-P builds pictures pixel by pixel — no VAE — and its one-step version cranks out 90 images a sec. 679M params, open weights + code. The catch: license + CUDA support still unclear. (AI-made image)

    Artist's desk at night; laptop screen shows a colourful pixel-art dragon forming from glowing pixel squares.
    Info