Large Language Models

Andrej Karpathy, the AI researcher and OpenAI co-founder, says the gap in shared understanding of what large language models can and cannot do is widening. His remark drew hundreds of upvotes on Hacker News, where commenters debate whether hype from AI companies and dismissiveness from sceptics are pulling public perception further from the technical reality of current models.

DeepSeek has released DeepInfer, an open-source inference framework designed to run its DeepSeek large language models at high performance. The project, hosted on GitHub, is drawing strong attention from developers discussing its speed, efficiency and what it means for running frontier AI models on their own hardware.

USA Today has filed a federal lawsuit against OpenAI, accusing the company of using its copyrighted news content without permission to train its large language models. The case adds to a growing wave of publishers taking legal action against AI developers over the unlicensed use of journalism in training data, and it is drawing attention as another major test of how copyright law applies to generative AI.

Meta and Apple, long-standing rivals in tech, are working together to run Meta's open-weight AI models on Apple Mac computers. The collaboration would let Mac users run large language models locally rather than in the cloud. The pairing of two companies known for their competition and public feuds is drawing attention across the tech industry.

ByteDance, the Chinese owner of TikTok, is reportedly in talks with Taiwan's MediaTek to develop custom AI chips for its data centers. The move would help the company secure AI computing power for its large language model and recommendation workloads amid tightening US export controls on advanced Nvidia chips. A partnership with MediaTek would echo similar custom-silicon efforts by other major cloud and AI players, and highlights how Chinese tech firms are seeking alternatives to restricted US hardware.

Ant Group has released the weights of its Ling-3.1-flash large language model under the permissive MIT license. The model has 560 billion total parameters with roughly 25 billion active per inference, a sparse mixture-of-experts design that keeps costs low, and supports context windows of up to one million tokens. The release gives developers and companies free rein to use, modify and commercialize the model, adding to the wave of openly licensed Chinese AI models.

A research paper on arXiv introduces Hybrid Reasoning Agents, a design in which large language model agents carry out reasoning internally in latent space while communicating their actions through text. The approach aims to let agents think more efficiently without emitting every intermediate step, drawing attention among AI researchers and developers discussing the trade-offs.

A new warning has been issued over high-risk security vulnerabilities affecting widely used large language model gateways, the infrastructure that routes and manages traffic to AI models. The alert suggests organizations relying on these tools could be exposed to attack. Details on affected products and fixes remain limited.

A new robotics benchmark reports that leading frontier AI models score under 2 out of 100 when tasked with controlling a simulated humanoid robot. The result highlights how far current large models remain from reliable physical control, despite strong performance in language and reasoning tasks. Commenters are expressing surprise at how low the scores are, and debating whether the benchmark tasks are too hard or the gap in embodied AI is real.

Block, the payments company led by Jack Dorsey, is developing AI-powered security systems built on technology from OpenAI. The company is integrating OpenAI's models into its security operations to detect and respond to threats across its payments and financial services platforms. The collaboration highlights a growing trend of fintech firms adopting large language models to strengthen fraud prevention and protect customer data.

Chinese AI company MiniMax has introduced M3.1 Flash, a new addition to its M3.1 family of large language models. Details about performance, pricing and capabilities are thin so far, with discussion so far centered on the announcement itself. Observers in the AI community are watching the release as MiniMax continues to compete with faster, cheaper model offerings from other labs.

Security researchers have issued a warning about high-risk vulnerabilities in a widely used LLM gateway tool, the software layer organizations rely on to route and manage access to large language models. Infosecurity Magazine reported the advisory, which is drawing attention because flaws in such gateways could expose AI deployments across many companies at once.

Chinese AI firm StepFun's Step 5 Preview, a mixture-of-experts model with a 1 million-token context window, is now listed on the OpenRouter platform, which routes requests to multiple large language models. The listing suggests StepFun is broadening international access to its flagship model, and developers are discussing what a 1M-context MoE could mean for long-document and code workloads.

Samsung Labs has released LittleBit, a method for compressing large language models to sub-1-bit weight representations using latent factorization. The work aims to shrink model memory footprints far beyond conventional quantization, making it possible to run capable LLMs on consumer hardware. The release of the code on GitHub is drawing attention from developers and researchers debating how practical the approach is for real-world inference.

A developer has shared a project walking through building a large language model from scratch using TypeScript, an unusual choice given that most LLM work relies on Python. The write-up is hosted at book.humanexperiences.io and appeared on Hacker News, where readers are discussing the merits of implementing machine learning fundamentals in JavaScript and what it says about accessibility of AI tooling.

A technical blog post from Codemancers explains how retrieval-augmented generation (RAG) works with OpenAI models, covering how external data sources can be retrieved and fed into large language models to improve answer accuracy. The piece is being shared among developers following AI and LLM topics, adding to ongoing discussion about practical ways to ground chatbot outputs in reliable, up-to-date information.

Kronos, an open-source Python project by developer shiyu-coder, describes itself as a foundation model for the language of financial markets — a large model built to understand and generate financial time-series data such as price movements. The repository is trending among developers, ranking near the top of GitHub's global listings.

Canadian AI firm Cohere has signed a deal to acquire AISLE, a startup specializing in AI cybersecurity. The move signals Cohere's intent to expand beyond large language models into security offerings for enterprise AI systems. Details on the deal's value and timeline have not been disclosed.

Linus Torvalds has reportedly merged his first patch series in which a large language model assisted with the work, according to LWN. The news has drawn attention across developer communities, with commenters debating what it signals about AI's growing role in Linux kernel development and whether tooling like this will become routine for kernel maintainers.

Inception Labs has introduced Mercury 3, describing it as a large-scale diffusion-based language model. The company says diffusion approaches generate text in parallel rather than token by token, promising faster output at competitive quality. The announcement is drawing attention among AI researchers and developers debating whether diffusion models can challenge autoregressive systems at production scale.

A user of fintech company Wise's customer support chatbot asked the assistant which large language model it runs on, and was told the information is "a secret". The exchange, shared online, has drawn attention to how banks deploy AI chatbots for customer assistance while keeping the underlying technology undisclosed, raising questions about transparency in automated banking support.

Researchers at Binghamton University are warning that secret trigger words, or backdoor phrases, could be used to manipulate AI-powered robots into acting dangerously, and they have developed a defense against such attacks. The work highlights growing concern that embodied AI systems running on large language models could be exploited through hidden commands embedded in their training data or instructions.

Developer Dominik Picheta says he has created a programming language that large language models cannot effectively use, describing the project as a 'killswitch' approach to coding. The stated idea is to make the language resistant to AI-assisted generation, so humans remain in control of writing code. The project is drawing attention among programmers debating how much of software development should be handed to AI tools.

A new article argues that developers using large language models through third-party apps lack visibility into what data is being sent and received, and urges them to route requests through their own proxy. The piece, written in blunt language, claims a self-hosted proxy is the only reliable way to inspect prompts, responses, and usage costs. Discussion is centering on transparency and control in AI tooling.

Chinese AI company MiniMax has published a blog post announcing M3.1 Flash, the latest addition to its line of large language models. The announcement is circulating among developers and tech commenters, who have begun weighing in on the release and its place in an increasingly crowded field of fast, low-cost AI models from Chinese and US labs.

Security researchers at Project Discovery have published research on 'abliterated' large language models — open-weight models with their safety refusals stripped out — and how running them can expose users to compromise. The write-up walks through attack paths in which uncensored models can be manipulated or deployed in ways that lead to system takeover, warning that removing guardrails removes more than just content filters.