LemonSeed Studio is an iPad app that embeds the unmodified upstream Linux amdgpu and amdkfd driver as a PCIDriverKit extension, allowing it to control an AMD GPU connected through a Thunderbolt enclosure. The app also bundles the LemonSeed Engine for on-device LLM inference running on that GPU via its HSA runtime, plus amdgpu_mtopg for live GPU telemetry. The project's source is available via git with recursive submodules, and is dual-licensed under MIT or GPL-2.0-only.
github.com
· 2026-10-11
GoKawiil describes an experimental reranking technique for its Jev model that decomposes a search query into a set of yes/no relevance questions generated by an LLM, rather than asking a single relevance question. Each question is scored by Jev for the probability of a 'yes' answer, and these probabilities are summed to produce an overall relevance score for a document against the query.
softwaredoug.com
· 2026-10-10
A commentary argues that blaming large language models for harmful outcomes amounts to 'decision laundering,' since software cannot make decisions. The author contends that companies like OpenAI and Anthropic could prevent certain model behaviors but choose not to, and that users funding subscriptions are indirectly supporting this choice.
wiki.cateat.fish
· 2026-10-10
A software engineer's blog post describes a new pattern of recruiter outreach emails that open with an LLM-generated paragraph praising the recipient's specific career history, then pivot to a job description that has little relation to that history. The author gives several examples where the personalized intro and the role description don't logically connect, arguing the entire email, not just the flattery, appears machine-written.
blog.pentlander.com
· 2026-10-08
A University of Limerick study tested how people compensated AI agents in a virtual office setting, varying only whether the agents presented as male or female. Despite identical underlying technology — same LLM, agent design, guardrails and harness — agents presented as male received about 10% higher pay per task and were rated as more human-like than those presented as female. The study involved 189 participants.
science.slashdot.org
· 2026-10-08
A research team has published the implementation of LittleBit, a method accepted at NeurIPS 2025, which compresses large language model weights into a sub-1-bit range, down to 0.1 bits per weight, by factorizing weight matrices into binarized low-rank latent factors with learned scaling. A follow-up method, LittleBit-2 (accepted at ICML 2026), adds an optional initialization technique called Joint Iterative Quantization that aligns latent factors geometrically before training, without changing the model's inference-time structure. The codebase supports models including OPT, Llama, Phi-4, Qwen2.5, QwQ, Gemma 2/3, and Qwen3.
github.com
· 2026-10-08
An experiment using large language models (LLMs) resulted in a Rust version of Microsoft's TypeScript compiler, checker, and language server. The process involved significant API token costs, totaling over $24,000, and produced a tool claimed to be 100% compatible with real-world projects. The project was initiated to evaluate LLM capabilities and to create a high-performance, WASM-compatible TypeScript checker.
github.com
· 2026-10-08
A new open-source benchmark called the Telegraph Test shows that instructing large language models to answer in 'cablese' — the clipped, article-free style once used by telegraph operators — reduces billed output tokens by 40-49% on several models' own API meters, while preserving or even slightly improving factual recall. The effect held across four model families tested with roughly 1,300 questions over 50 passages, with one exception: gpt-5-mini's reasoning overhead made cablese prompting roughly double its cost instead of cutting it.
fiveminutesforward.com
· 2026-10-07
The research team built a diagnosis pipeline that uses Jev, a lightweight decision-making tool, to analyze Kubernetes cluster evidence without any LLM agent involved. The pipeline programmatically collects cluster objects, events, logs, and resource data, then has Jev select likely root causes and supporting evidence to assemble a diagnosis report. Tested across 21 SREGym-Lite faults, the pipeline passed 80 of 105 diagnoses (76.2%) with a median diagnosis time of 14.6 seconds.
sregym.com
· 2026-10-07
A post on Hacker News poses a hypothetical in which a US government nationalizes Alphabet via constitutional amendment to restore Google Search to its 2010-era quality and strip out all LLM features. The author asks commenters what they would be willing to accept or trade for a search engine that works again, after what they describe as years of commercial degradation of the product.
news.ycombinator.com
· 2026-10-06
The developer behind Pi has released version 1.0 with a feature called Codemode, which brings Model Context Protocol (MCP) support into the harness. The post explains how tool-calling works for LLMs, why bash/CLI composition has been favored, and where that approach breaks down for things like image reading or spawning sub-agents.
lucumr.pocoo.org
· 2026-10-06
Aleph Alpha, a German AI company, released Kolibri on 3 October 2026 under the Apache 2.0 license, with weights published on Hugging Face. It is a mixture-of-experts model with 78.1 billion total parameters but only 3.46 billion active per token, trained from scratch on 768 NVIDIA B200 GPUs across data centers in Germany and Finland on roughly 24 trillion tokens, over a fifth of them German.
tej.as
· 2026-10-03
GoKawiil