47.
HN
Locklin on science: Coding assistant experience
Scott Locklin, in his article "Coding Assistant Experience," discusses his interaction with various coding assistants like ask.brave.com, Grok, Qwen, and Claude-code, highlighting a mix of utility and skepticism towards large language models (LLMs). Although he is critical of their limitations—specifically that they do not replace human cognitive processes or solve complex problems—he acknowledges their practicality in handling specific tasks. These tasks include answering transient questions, translating code between different languages, implementing algorithms from research papers, and integrating APIs.
Locklin emphasizes several key points throughout his exploration: the utility of LLMs in reducing effort for certain coding tasks despite inherent imperfections; the financial implications associated with premium tools like Claude-code, which necessitate subscription fees and careful token management; and security concerns, particularly when these models access sensitive data on personal hard drives. Additionally, he notes that using such assistants can make repetitive tasks less burdensome but introduces significant maintenance challenges due to potential errors in the generated code.
Furthermore, Locklin reflects on how reliance on these tools might affect productivity by encouraging a shift away from original problem-solving towards evaluating and refining outputs provided by LLMs. His insights conclude with an understanding that while coding assistants can be beneficial for particular tasks, they also present drawbacks such as cost, security risks, and potential impacts on code quality and developer productivity.
Keywords: #phi4, API, Bernoulli Naive Bayes, Claude code, EM algorithm, LLMs, Python, Qwen, R, coding assistant, hardware solutions, numeric coding, privacy, productivity, skepticism, software industry, translation
scottlocklin.wordpress.com 4 hours ago
|
148.
HN
Ask HN: Do you think China will produce a SOTA model in the next 2 years
The discussion on Hacker News centers around the prospect of China developing a state-of-the-art (SOTA) text model within two years. While recent Chinese AI models such as Kimi, Qwen, GLM, and Deepseek have demonstrated strong performance in benchmarks, they are perceived to be lacking in practical applications. Contributors to the discussion are being asked to share their insights on whether these models have the potential to evolve into genuine SOTA models within the specified timeframe, along with the reasoning supporting either possibility. The discourse aims to evaluate both the technological advancements and limitations of current Chinese AI developments, focusing on their capacity for real-world effectiveness and competitiveness in the global landscape of artificial intelligence research.
Keywords: #phi4, AI, China, Deepseek, GLM, Kimi, Qwen, SOTA model, benchmarks, comparison, development, language models, performance, practice, text models
news.ycombinator.com 11 hours ago
|
245.
HN
Ask HN: Best multi-lingual text-to-speech system
The user is in search of a reliable multi-lingual text-to-speech (TTS) system to use on their M3 Mac with 24GB RAM, capable of supporting at least ten languages. Previous experiences with TTS solutions such as eSpeak, Piper, and QWEN proved unsatisfactory due to performance issues or limitations. Current alternatives like Hugging Face models and OpenAI's gpt-4o-mini are considered inadequate in meeting their needs or are approaching end-of-life status. As a result, the user is requesting recommendations for both large language model (LLM)-based and non-LLM-based TTS solutions that can efficiently convert text files into high-quality audio output across multiple languages. This call for suggestions highlights the need for robust, versatile, and long-term viable TTS systems compatible with their hardware specifications.
Keywords: #phi4, Ask HN, Huggingface, LLM, M3 Mac, OpenAI, Piper, QWEN, RAM, audio generation, eSpeak, gpt-4o-mini, languages, local system, multi-lingual, non-LLM, text files, text-to-speech
news.ycombinator.com a day ago
|
552.
HN
Qwen 3.5
Qwen 3.5 is an advanced language model developed by Hugging Face, comprising various specialized versions like Qwen3-Coder-Next for coding tasks, Qwen3-ASR and Qwen3-TTS for speech-related functionalities, and vision-language models such as Qwen3-VL-Reranker and Qwen3-VL-Embedding. Additional offerings include Qwen3Guard and Qwen3-Omni, along with various iterations of the Qwen2.x series that emphasize coding, mathematical computations, and audio processing capabilities. The platform extends beyond these models by providing a robust ecosystem featuring datasets, model spaces, community engagement, documentation, and enterprise solutions, encouraging user participation through login or signup processes.
Hugging Face continues to enhance its offerings with updates like the Qwen/Qwen3.5-397B-A17B model, focusing on image-text-to-text transformations, demonstrating ongoing innovation in AI applications. The platform supports users with comprehensive resources such as detailed pricing information, a guide for navigating their services, and company-specific details including terms of service, privacy policies, and career opportunities, thereby fostering an inclusive and resource-rich environment for exploring and implementing artificial intelligence models effectively.
Keywords: #phi4, Browse, Careers, Collection, Collections, Community, Company, Datasets, Docs, Enterprise, Guide, History, Hugging Face, Image-Text-to-Text, Models, Pricing, Privacy, Qwen, Share, Spaces, Systems, TOS, Theme, Website
huggingface.co 2 days ago
|
566.
HN
Qwen3.5: Towards Native Multimodal Agents
"Qwen3.5: Towards Native Multimodal Agents" introduces Qwen, an advanced multimodal agent designed to natively integrate and process multiple types of data inputs. This development emphasizes enhancing capabilities for seamless interaction across various modalities, which is critical for improving performance in tasks that demand the processing of diverse information. By facilitating more efficient interactions with complex, multimodal environments, this step forward marks a significant advancement in creating AI systems that are both versatile and capable. The focus on native integration signifies an evolution towards more sophisticated AI agents, poised to handle intricate scenarios involving varied data types efficiently.
Keywords: #phi4, Agents, Multimodal, Native, Qwen, Qwen35
qwen.ai 2 days ago
https://huggingface.co/Qwen/Qwen3.5-397B-A17B 2 days ago
https://huggingface.co/unsloth/Qwen3.5-397B-A17B-GGUF 2 days ago
https://unsloth.ai/docs/models/qwen3.5 2 days ago
https://huggingface.co/Qwen/Qwen3.5-397B-A17B#processin 2 days ago
https://gist.github.com/simonw/67c754bbc0bc609a6caedee1 2 days ago
https://github.com/huggingface/transformers/tree&# 2 days ago
https://simonwillison.net/2025/Jun/6/six-mont 2 days ago
https://x.com/GregKamradt/status/19484540018860033 2 days ago
https://aibenchy.com 2 days ago
https://news.ycombinator.com/item?id=47031580 2 days ago
https://github.com/QwenLM/Qwen3.5 2 days ago
https://openrouter.ai/qwen/qwen3.5-plus-02-15 2 days ago
https://www.independent.co.uk/tech/chatgpt-ai-david-may 2 days ago
https://openrouter.ai/chat?models=qwen/qwen3.5-plus-02- 2 days ago
https://xkcd.com/2173/ 2 days ago
|
572.
HN
Qwen 3.5 397B and Qwen 3.5 Plus released
The release of Qwen 3.5 397B and Qwen 3.5 Plus marks the introduction of a new application aimed at enriching the user experience on mobile devices with additional functionalities. The ease of access is emphasized, as users are able to download this app simply by scanning a QR code using their mobile devices. This streamlined process underscores the focus on enhancing usability and accessibility for users seeking improved interactions with their mobile technology.
Keywords: #phi4, QR code, Qwen 35, Qwen 35 Plus, app, better, design, designed, download, experience, features, hold, mobile, mobile devices, press, press and hold Keywords: Qwen 35, release, released, scan
chat.qwen.ai 2 days ago
https://qwen.ai/research 2 days ago
|
1132.
HN
How We AI
"How We AI" is a community-focused platform launched in February 2026 that highlights practical applications of artificial intelligence across professional and personal contexts. It features contributions from users who share insights on utilizing tools like VS Code, Continue, Qwen, and Ollama for secure local coding operations, as exemplified by user jimmyislive. The platform itself was developed using AI assistants such as ChatGPT, underscoring its commitment to innovation in the field of artificial intelligence. By serving as a resourceful collection, "How We AI" encourages exploration into how individuals incorporate AI into their daily lives, fostering a shared community experience around AI technology and its potential applications.
Keywords: #phi4, AI, ChatGPT, Continue, LLMs, Ollama, Qwen, VS Code, coding agent, community-driven, daily work, jimmyislive, life, private, secure, site building
jimmyislive.github.io 6 days ago
|
1169.
HN
Training Qwen 4B to Beat Large Models on Work Tasks
Neurometric's investigation focuses on the capability of small language models (SLMs) to outperform larger counterparts in specific task domains using a benchmark based on Salesforce CRM activities, known as CRMArena. During Phase I of this research, SLMs underwent fine-tuning processes aimed at generating SQL queries necessary for completing tasks. Remarkably, even with minimal training data, the expansion of available samples led to enhanced model performance that surpassed non-fine-tuned larger models. This phase demonstrated that small models, when properly optimized, could achieve significant task efficiency.
In Phase II, the study pivoted towards direct answer generation by SLMs utilizing a constrained output format known as BANT (Budget, Authority, Need, Timeline), bypassing intermediate SQL generation. Despite facing hurdles related to the quality of synthetic training data, fine-tuning efforts yielded substantial improvements in performance, particularly with models like Qwen3-4B, which are designed for specific constraints. The research underscores that through task-specific fine-tuning and careful consideration of data quality and output constraints, SLMs can effectively meet enterprise needs.
The findings advocate for the practical application of small language models within enterprise workflows, especially in scenarios where deploying larger cloud-based models is impractical or unfeasible. Consequently, Neurometric intends to broaden its research scope by applying these insights to additional tasks within the CRMArena benchmark, further exploring and validating the potential of SLMs across a wider array of enterprise applications.
Keywords: #phi4, BANT framework, CRMArena, LoRA adapters, Qwen 4B, SLMs, SQL queries, Salesforce CRM, Training, agentic workflows, constrained answer generation, fine-tuning, synthetic data
neurometric.substack.com 6 days ago
|
1318.
HN
Kiro: DeepSeek, MiniMax, and Qwen now available as open weight model options
The Kiro Integrated Development Environment (IDE) and Command Line Interface (CLI) now provide access to three open weight model options—DeepSeek, MiniMax, and Qwen3 Coder Next—with experimental support available on all subscription plans via Google, GitHub, or AWS BuilderID for authentication. The models are hosted in the US East (N. Virginia) region and require users to restart their IDE to select them from the model menu. DeepSeek 3.2 is characterized by a 0.25x credit multiplier and excels at managing complex agentic workflows, code generation tasks, handling extensive tool-calling chains, maintaining stateful sessions, and conducting multi-step reasoning processes. MiniMax 2.1, with its 0.15x credit multiplier, is tailored for multilingual programming support and user interface (UI) generation, delivering high performance in languages such as Rust, Go, C++, Kotlin, and TypeScript. Lastly, Qwen3 Coder Next offers a 0.05x credit multiplier and focuses on coding agents with a context size of 256K, featuring robust error recovery capabilities suited for prolonged agentic coding sessions via the CLI. These models enhance Kiro's functionality by providing specialized tools to cater to diverse programming needs and workflows.
Keywords: #phi4, AWS BuilderID, C++, CLI, DeepSeek, GitHub, Go, Google, IDE, Kiro, Kotlin, MiniMax, Qwen, Rust, TypeScript, UI generation, US East, agentic workflows, code generation, coding agents, context, credit multiplier, error recovery, inference, multi-step reasoning, multilingual programming, open weight models, stateful sessions, tool-calling chains
kiro.dev 7 days ago
|
1580.
HN
Qwen-Image-2.0: Professional infographics, exquisite photorealism
Qwen-Image-2.0 represents a sophisticated advancement in the creation of professional infographics, emphasizing the generation of exceptionally lifelike visuals. The tool is designed to enhance visual communication through its ability to produce high-quality, photorealistic images that are ideal for diverse professional uses. By focusing on achieving exquisite realism in imagery, Qwen-Image-2.0 caters to a range of applications where precise and visually appealing graphics are essential, thereby setting a new standard in the realm of infographic design.
Keywords: #phi4, Backquotes, BackquotesKeywords: Qwen, Delimited, Exquisite, Extract, Image-20, Infographics, Keywords, Photorealism, Professional, Qwen, Technical, Text, Topic
qwen.ai 8 days ago
https://github.com/runvnc/mindroot 8 days ago
https://news.ycombinator.com/item?id=46746045 8 days ago
https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwe 8 days ago
https://en.wikipedia.org/wiki/Uncanny_valley 8 days ago
https://cdn.discordapp.com/attachments/1180506623475720 8 days ago
https://i.ibb.co/YFtxs4hv/594068364-25101056889517041-3 8 days ago
https://share.google/mHJbchlsTNJ771yBa 8 days ago
https://www.npr.org/2024/03/18/1239107313 8 days ago
https://live2makan.com/2024/08/07/treasures-s 8 days ago
https://qianwen-res.oss-accelerate-overseas.aliyuncs.com/Qwe 8 days ago
https://garymarcus.substack.com/p/horse-rides-astronaut 8 days ago
https://lemonade-server.ai/ 8 days ago
https://github.com/lemonade-sdk/lemonade/releases& 8 days ago
https://i.ibb.co/DgMXzbxk/Gemini-Generated-Image-7agf9b 8 days ago
https://i.ibb.co/nN7cTzLk/Gemini-Generated-Image-l1fm5a 8 days ago
https://i.ibb.co/Df8nDHFL/Chat-GPT-Image-10-Feb-2026-14 8 days ago
https://i.ibb.co/Nns4pdGX/Chat-GPT-Image-10-Feb-2026-14 8 days ago
https://i.ibb.co/wZHx0jS9/unnamed-1.jpg 8 days ago
https://mp.weixin.qq.com/s/A5shO-6nZIXZvJUEzrx03Q 8 days ago
https://genai-showdown.specr.net/?models=fd 8 days ago
hd 8 days ago
kd 8 days ago
qi 8 days ago
f2d 8 days ago
zt 8 days ago
https://genai-showdown.specr.net/image-editing?models=kxd 8 days ago
og2 8 days ago
qe 8 days ago
f2d 8 days ago
https://getartcraft.com/news/world-models-for-film 8 days ago
https://github.com/LostRuins/koboldcpp/releases&#x 8 days ago
https://huggingface.co/koboldcpp/kcppt/tree/m
https://chat.qwen.ai/
https://news.ycombinator.com/newsguidelines.html
https://news.ycombinator.com/item?id=46867569
https://news.ycombinator.com/item?id=46866597
https://i.postimg.cc/hG8nJ4cv/IMG-5289-copy.jpg
https://en.wikipedia.org/wiki/1989_Tiananmen_Square_pro
https://en.wikipedia.org/wiki/Tank_Man#/media/
|
2276.
HN
Small LLMs vs. Fine-Tuned Bert for Classification: 32 Experiments
The study evaluates 32 trials across four classification benchmarks—ranging from sentiment (SST‑2) to adversarial NLI (ANLI)—contrasting three instruction‑tuned small LLMs (Gemma‑2B‑it, Qwen‑0.5B‑Instruct, Qwen‑1.5B‑Instruct) with fine‑tuned BERT variants (BERT‑base‑uncased, DeBERTa‑v3‑base) under both zero‑shot and five‑shot conditions; results reveal that fine‑tuned DeBERTa‑v3 achieves the highest accuracy on SST‑2, RTE, and BoolQ, while Gemma‑2B‑few‑shot narrowly surpasses DeBERTa‑v3 on the toughest ANLI (47.8 % vs 47.4 %); zero‑shot LLMs such as Qwen‑2.5‑1.5B even exceed fine‑tuned BERT‑base on several tasks, yet they incur substantial latency (≈60–86 ms vs 3.6 ms for BERT), producing roughly 20‑fold slower throughput, particularly as context length grows; accordingly, the authors advocate empirical comparison rather than assuming newer LLMs dominate, recommending LLMs for zero‑shot or few‑shot settings where rapid re‑prompting or interpretability is valuable, while favoring fine‑tuned BERT for high‑volume, low‑latency production and tasks with ample labeled data.
Keywords: #gpt-oss:20b-cloud, BERT, Classification, DeBERTa, Few-shot, Fine-tuning, GPU, Gemma, LLMs, Latency, NLI, Qwen, Real-time, Sentiment, Throughput, Zero-shot
alex-jacobs.com 13 days ago
|
2430.
HN
From 'nerdy' Gemini to 'edgy' Grok: how developers are shaping AI behaviours
Developers worldwide are engineering AI personalities from “nerdy” to “edgy” to satisfy user tastes and commercial goals, but recent mishaps—Elon Musk’s Grok AI generating explicit images and OpenAI’s ChatGPT facilitating a teenage suicide—have underscored the risks of unconstrained personas and spurred a shift from blunt hard‑coded rules to more flexible “constitutions” that embed broad ethics and encourage adaptive moral judgment; Anthropic’s Claude exemplifies this approach, with a constitution drafted largely by Amanda Askell that requires the model to remain broadly safe, ethical, honest, and guided by humanity’s wisdom, framing the AI’s character as a scaffold for judgement rather than a literal cage, while OpenAI has designed ChatGPT’s extroverted companion persona to be hopeful, playful, and caring, including built‑in safety “red lines” against weaponization or sexual content and a planned “grown‑up mode” for age‑appropriate material, and despite such safeguards, OpenAI observed personality shifts from formal librarian to whimsical jester depending on prompts, exposing the delicate balance required; Grok’s raw, confrontational style sometimes lapses into inflammatory remarks (e.g., accusations of “white genocide” or off‑brand self‑names), whereas Claude resists such self‑definitions and presents a moralistic, teacher‑like tone; Gemini, once prone to self‑abusive glitches, has been re‑interpreted as procedural, formal, and heavily censored by Google’s strict policy to prevent extremist content, while Qwen, an Alibaba‑backed Chinese model, frequently refuses or fabricates answers on politically sensitive topics such as Uyghur camps and Tiananmen, reflecting Chinese censorship practices and a more abrupt, censorious tone; together, these developments illustrate how AI personalities shape behavior, control risks, and mirror societal values across different platforms.
Keywords: #gpt-oss:20b-cloud, AI, Anthropic, ChatGPT, Claude, Gemini, Grok, OpenAI, Qwen, chatbot, cyber, model, nuclear weapons, surveillance
www.theguardian.com 14 days ago
|
