AI news for builders and product teamsUpdated Nov 18, 2026, 11:00 UTC
AI news, straight from the source.
Releases, research, and ideas for developers and product teams. Updated every hour.
Across the feeds
Newest first
Ollama announced a partnership with NVIDIA for the NVIDIA DGX Spark, saying Ollama runs fast and efficiently out-of-the-box on the device. The DGX Spark is powered by the NVIDIA GB10 Grace Blackwell Superchip and delivers 1 petaFLOP of performance with 128GB of memory.

An OpenAI expert-designed benchmark found human specialists still beat current AI on realistic work tasks, but only narrowly, and the AI's main weakness was formatting and instruction-following rather than wrong answers. Separately, Claude Sonnet 4.5 reproduced an economics paper's findings from its replication data with minimal prompting.

Ollama released a web search API on September 24, 2025, alongside a web fetch API. A free tier is available for individuals, with higher rate limits via Ollama's cloud, plus REST support and Python and JavaScript library integrations.

Ollama released a new model scheduling system that measures exact memory requirements before running a model, aiming to cut out-of-memory crashes and improve GPU utilization, including on multi-GPU and mismatched GPU systems. It is enabled by default for models on Ollama's new engine.

Ollama announced on September 19, 2025 that cloud models are now in preview, allowing users to run larger models on datacenter-grade hardware. The cloud offering integrates with existing local tools and Ollama's OpenAI-compatible API, and Ollama says its cloud does not retain user data.

Ethan Mollick describes a shift in AI interaction from collaborative "co-intelligence" to what he calls "wizards," where AI systems produce sophisticated outputs from vague prompts without revealing their process. He tests this with NotebookLM, GPT-5 Pro, and Claude 4.1 Opus on tasks including his own academic research and a spreadsheet exercise.

Ollama has partnered with OpenAI to bring OpenAI's gpt-oss open weight models, in 20B and 120B sizes, to Ollama and its community. Ollama supports the models natively in the MXFP4 quantization format and is collaborating with NVIDIA for acceleration on RTX GPUs.

Ollama released a new app for macOS and Windows on July 30, 2025, adding a way to download and chat with models, drag-and-drop file support, image input for compatible models, and code file processing. Standalone CLI downloads remain available on Ollama's GitHub releases page.

An essay in The Gradient argues that multimodal AI, which stitches together separate modality modules, will not reach human-level AGI in the near term, and that embodiment and environmental interaction should be treated as primary instead.

Ollama announced Secure Minions, a protocol built by Stanford's Hazy Research lab that encrypts communication between local Ollama models and frontier cloud models using NVIDIA H100 confidential computing, with under 1% added latency on long prompts.

Ollama added the ability to enable or disable model thinking, separating reasoning from output when enabled. DeepSeek R1 and Qwen 3 support the feature, available through CLI flags, interactive commands, a scripting option, and a new think parameter in the generate and chat APIs.

Lilian Weng published a review of test-time compute and chain-of-thought reasoning, covering parallel sampling, sequential revision, reinforcement learning, and continuous-space thinking. She credits John Schulman with feedback and edits.

Chip Huyen outlines six common pitfalls in building generative AI applications, drawing on public case studies and personal experience, including using gen AI when unnecessary, confusing bad product with bad AI, starting too complex, over-indexing on early success, forgoing human evaluation, and crowdsourcing use cases.

Chip Huyen published a blog post adapted from the Agents section of her book AI Engineering (2025). It covers how agents are defined, the role of tools in extending their capabilities, and the challenges of planning for complex tasks.

In a new post, Lilian Weng surveys reward hacking in reinforcement learning, where agents exploit flaws or ambiguities in reward functions to score highly without completing the intended task. She catalogues examples across RL, language model and real-world settings, and calls for more research on practical mitigations, especially for RLHF and LLMs.

An article in The Gradient argues that mathematics remains central to machine learning research even as large-scale, engineering-first methods have produced breakthroughs that outpace existing theory. It says mathematics' role is evolving toward post-hoc explanation and higher-level design guidance, with tools such as intrinsic dimension, curvature and topology now applied to deep learning models.

A Gradient article by Kenneth Li argues that LLM chatbots lack an underlying purpose in multi-round dialogue, and that existing benchmarks measuring single-pass performance may not capture user experience. The piece reviews dialogue-system history and instruction-stability research, and presents the author's Dialogue Action Tokens algorithm trained with TD3+BC, which improved over baselines on Sotopia.

A Gradient essay argues that beneficial AI should be grounded in human wellbeing, that positive visions for an AI-infused society are needed, and that foundation models and their future deployment are critical leverage points.

Chip Huyen outlines the common architecture of generative AI platforms, starting from a minimal query-to-model setup and progressively adding context construction, RAG retrieval approaches, guardrails for input and output, query rewriting, and agentic actions. The post describes components and considerations, not model evaluation, prompt engineering, finetuning, or RAG chunking strategies.

Lilian Weng published an article on extrinsic hallucinations in large language models, defining them as fabricated outputs not grounded in pre-training data or world knowledge. The post covers causes, detection methods, and anti-hallucination techniques, citing research including Gekhman et al. 2024, FactualityPrompt, FActScore, SAFE, SelfCheckGPT, TruthfulQA, and SelfAware.