hackernews

1. A small number of samples can poison LLMs of any size Total comment counts : 48 Summary New research from Anthropic, the UK AI Security Institute, and The Alan Turing Institute shows that poisoning a pretrained LLM with only 250 malicious documents can implant a backdoor that makes the model output random gibberish when it encounters a trigger, across models from 600M to 13B parameters. This challenges the idea that backdoors require a percentage of training data; the amount needed can be constant....

1. Svelte is that fast Total comment counts : 19 Summary An academic study comparing Angular, React, Vue, Svelte, and Blazor investigates how rendering strategies affect performance. It notes many benchmarks are flawed and uses an MVVM-like component model. React, Vue, and Blazor rely on a virtual DOM with diffing (often cited as O(n) in practice but can be O(n^3) worst case), while Angular and Svelte update the DOM directly without a diff step....

1. Gemini 2.5 Computer Use model Total comment counts : 9 Summary Google released the Gemini 2.5 Computer Use model via the Gemini API, a specialized version built on Gemini 2.5 Pro’s visual understanding to power agents that interact with user interfaces. It outperforms other web and mobile control models with lower latency. Available in Google AI Studio and Vertex AI, developers can build UI agents and share feedback in the Developer Forum....

1. OpenZL: An open source format-aware compression framework Total comment counts : 8 Summary OpenZL is a public release of a lossless data compression framework for structured data. It combines a sequence of reversible transforms with an offline trainer that builds a data-shape–specific Plan, improving compression by exposing structure rather than just bytes. Users provide the data shape (via presets, SDDL, or a custom parser); the trainer produces a Plan that encodes into the frame....

1. Fire destroys S. Korean government’s cloud storage system, no backups available Total comment counts : 24 Summary A fire at the National Information Resources Service (NIRS) data center in Daejeon damaged equipment, prompting a major disruption to online government services. Officials report slow but partial recovery: over 55% of top-priority systems back online, but only 17.3% of online services restored, with restoration hampered by the shutdown of the national database command center....

1. ProofOfThought: LLM-based reasoning using Z3 theorem proving Total comment counts : 10 Summary The article outlines a neurosymbolic reasoning system that combines LLM-based reasoning with the Z3 theorem prover to achieve robust, interpretable results. It uses a two-layer architecture, recommending a high-level API for most users while offering extensive examples (including Azure OpenAI support). It references the paper “Proof of Thought: Neurosymbolic Program Synthesis Allows Robust and Interpretable Reasoning” from NeurIPS 2024 Sys2Reasoning Workshop, and emphasizes attention to user feedback and documentation....

1. PEP 810 – Explicit lazy imports Total comment counts : 24 Summary The proposal adds explicit syntax for lazy imports in Python, deferring module loading until the first use of the imported name. Normal imports remain eager, maintaining backward compatibility. Benefits include faster startup, lower memory use, and reduced unnecessary work, especially for CLI tools and large dependency graphs. The approach is local, explicit, controlled, and granular: only the marked import is lazy, binding is created immediately, and loading occurs on first access....

1. Signal Protocol and Post-Quantum Ratchets Total comment counts : 18 Summary Graeme Connell and Rolfe Schmidt announce Sparse Post Quantum Ratchet (SPQR), a new post-quantum ratchet for the Signal Protocol that strengthens quantum resistance while preserving forward secrecy (FS) and post-compromise security (PCS). SPQR, used with Signal’s existing Double Ratchet as the Triple Ratchet, improves protection against harvest-now-decrypt-later attacks. For users, nothing changes and conversations will automatically transition to the updated protocol....

1. OpenTSLM: Language models that understand time series Total comment counts : 15 Summary Time-Series Language Models (TSLMs) are framed as a new AI modality that treats time-series data—heartbeats, price ticks, logs, and sensor streams—as a native input alongside text, enabling natural-language reasoning, explanations, and forecasting. The article claims order-of-magnitude gains in temporal reasoning on smaller, faster backbones. OpenTSLM releases lightweight, public-data base models; Frontier TSLMs offer enterprise-grade, proprietary models. The goal is a universal temporal interface linking real-world signals to decisions and autonomous agents, powering healthcare, robotics, and infrastructure....

1. Inflammation now predicts heart disease more strongly than cholesterol Total comment counts : 5 Summary Chronic inflammation doubles heart-disease risk, and the ACC now designates hs-CRP as a SMuRF-like standard risk factor. The ACC recommends universal hs-CRP testing in both primary and secondary prevention, alongside cholesterol measures. Evidence suggests hs-CRP is a stronger predictor than LDL, especially in statin-treated or “SMuRF-less” patients where inflammation persists despite controlled cholesterol. Other markers add no extra predictive value once hs-CRP is known....