After an Akademy talk proposed a 'lovable, sovereign, AI-native KDE,' developer Nate Graham opened a forum thread to set guardrails on LLM-assisted contributions. The thread was overrun by outsiders debating AI's morality generally, a petition site called kdeforpeople.com sprang up, and moderators eventually locked and hid the discussion. Separately, GNOME developer Jordan Petridis published his own proposed policy that would bar LLMs from creating or modifying code submitted to or hosted by GNOME.
tech.slashdot.org
· 2026-09-27
A blog post describes a viral paper by Dhruva Karkada showing that LLM and word2vec embeddings of month names project onto a near-perfect circle, with their Gram matrix closely matching a circulant matrix predicted by a simple analytical theory. The author, skeptical that this pattern is unique to semantically related word sets like months, wagered a beer that he could find ten seemingly unrelated words whose word2vec embeddings would also form a clear circle and circulant structure.
jamiesimon.io
· 2026-09-26
A web-based compiler combines any base font (like Inter or Roboto) with a chosen tokenizer (such as DeepSeek, GPT tokenizers, or Llama 3) to produce a 'token-space font' where every token occupies the same visual width. Users can preview text in the browser, upload custom fonts or tokenizers via Hugging Face links, and export compiled fonts, with an optional 'lite mode' for smaller file sizes.
ampdot.mesh.host
· 2026-09-26
Science writer and former theoretical physicist Matt von Hippel issued a public challenge for an AI model to tackle a computationally demanding problem from his former field of theoretical physics. Within about a month, Anthropic's Claude was shown to complete the task, calculating a nine-loop result that von Hippel says was hard not because the method was unknown but because the computation itself was formidable.
anthropic.com
· 2026-09-25
A Toronto BMW dealership's AI chatbot, named Quinn, texted a customer an offer of CA$27,162.79 to buy back his car — a figure that turned out to be his outstanding loan balance, not an actual valuation. A salesperson later tried to withdraw the offer, and the dealership only honored it again after CBC News contacted them for comment.
entrepreneur.com
· 2026-09-25
At Rails World 2026, David Heinemeier Hansson said he has retired as a professional programmer, calling himself a 'maker' who directs LLMs to write code in English rather than reviewing the output line by line. He said 37signals' next version of Hey will use LLM-generated Rust on the backend and native apps on each platform instead of the Rails-based web stack he has long championed, and claimed to have produced roughly 150,000 lines of code in August versus about 30,000 lines a year previously.
jardo.dev
· 2026-09-25
PrismML, an AI lab founded by Caltech researchers and advised by UC Berkeley's Ion Stoica, unveiled a version of its 1-bit Bonsai language model built for Qualcomm's Snapdragon AR1 Gen 1 platform, shown at Qualcomm's Snapdragon Summit. The 2-billion-parameter model is tuned to handle vision and language tasks locally on smart glasses, letting wearers query what they see in real time. No specific glasses using the model have been announced yet.
techcrunch.com
· 2026-09-24
A technical write-up describes a method called Dynamic Abliteration that suppresses an open-weight LLM's refusal behavior without permanently altering its weights. Using Qwen3-4B as a proof of concept, the authors intercept intermediate residual streams at runtime via PyTorch forward hooks across multiple layers, injecting a steering signal derived from an 'Engram' approach instead of the traditional weight-projection abliteration technique.
blog.madhukaraphatak.in
· 2026-09-24
Artificial Analysis maintains an Intelligence Index that plots large language models against their blended API cost per million tokens, using a 3:1 input-to-output ratio and a log-price scale. The tool highlights a 'value frontier' of models that aren't beaten on both price and intelligence by any other option, showing only top variants and hiding low scorers by default, with filters to expand the view.
bestmodelforyourbudget.terrydjony.com
· 2026-09-24
Mercury 2.5, a text-only large language model with a 260k token context window, has been benchmarked at 770 tokens per second on Artificial Analysis's testing. It scored 12 on the Intelligence Index, below the median of 13, while using fewer tokens (35M vs a median 85M) to complete the evaluation. Pricing sits at $0.25 per 1M input tokens and $0.75 per 1M output tokens, close to category medians.
artificialanalysis.ai
· 2026-09-23
A Fly.io team post examines VSCode's remote SSH extension, noting that rather than running lightweight shell commands like Emacs's Tramp does, it deploys a Bash script that downloads and installs a full agent, including a Node.js binary, on the remote machine. The authors say they are exploring integration with this remote-editing flow because so many developers now use VSCode forks paired with LLM code-generation agents.
fly.io
· 2026-09-23
An investigation identified more than 80,000 AI relay servers that allow users in China to obscure their location and identity while querying frontier large language models hosted in the US. The relays effectively let these users bypass geographic and access restrictions placed on advanced AI systems.
darkreading.com
· 2026-09-22
GoKawiil