LLMs1 min readarchive
Ollama partners with OpenAI and ROOST for gpt-oss-safeguard models
Ollama is releasing gpt-oss-safeguard models in 20B and 120B sizes under an Apache 2.0 license for safety classification tasks.
Loading
90 days of the enterprise, pro and team plans free, no card needed. See the plans
Blog
Posts older than six months. They stay readable at their original address; models, prices and APIs they describe may have changed since.
LLMs1 min readarchive
Ollama is releasing gpt-oss-safeguard models in 20B and 120B sizes under an Apache 2.0 license for safety classification tasks.
LLMs1 min readarchive
MiniMax M2 with 10 billion activated parameters is now accessible via Ollama cloud, ranking #1 globally in general intelligence benchmarks.
LLMs1 min readarchive
Based on what I've learned from role models and mentors in Amazon
LLMs1 min readarchive
Ollama announces GLM-4.6 and Qwen3-Coder-480B on its cloud service, with a new engine update for Qwen3-Coder-30B tool calling.
LLMs1 min readarchive
Sebastian Raschka outlines four main LLM evaluation methods: multiple-choice benchmarks, verifiers, leaderboards, and LLM judges. The article provides from-scratch code implementations using PyTorch and the Qwen3 model.
AI1 min readarchive
The race between human-centered work and infinite PowerPoints
LLMs1 min readarchive
An LLM that can converse in English & item IDs, and make recommendations w/o retrieval or tools.
LLMs1 min readarchive
OpenAI released the first open-weight LLMs since GPT-2, featuring MXFP4 optimization for local deployment on consumer GPUs and H100 hardware.
LLMs1 min readarchive
Sebastian Raschka compares DeepSeek V3's Multi-Head Latent Attention (MLA) against Grouped-Query Attention (GQA), highlighting memory efficiency gains. The analysis notes MLA compresses KV tensors to reduce cache usage while potentially outperforming standard attention.
LLMs1 min readarchive
Evaluation metrics, how to build eval datasets, eval methodology, and a review of several benchmarks.
Research1 min readarchive
Terry Winograd argues that current multimodal scaling strategies will fail to achieve AGI because they lack embodied understanding of physical reality.
LLMs1 min readarchive
Learning to automate simple agentic workflows with Amazon Q CLI, Anthropic MCP, and tmux.
Research1 min readarchive
Special thanks to John Schulman for a lot of super valuable feedback and direct edits on this post. Test time compute (Graves et al. 2016, Ling, et al. 2017, Cobbe et al. 2021) and Chain-of-thought (CoT) (Wei et al. 2022, Nye et al. 2021...
LLMs1 min readarchive
Applying the scientific method, building via eval-driven development, and monitoring AI output.
LLMs1 min readarchive
Chip Huyen and I share what we've learned, best practices, and insights at NVIDIA GTC 2025.
LLMs1 min readarchive
Model architectures, data generation, training paradigms, and unified frameworks inspired by LLMs.
LLMs1 min readarchive
Exploring how an AI-powered reading experience could look like.
Research1 min readarchive
Making sense of recent technology trends and claims
Research1 min readarchive
Technology Isn’t the Problem—or the Solution.
Research1 min readarchive
Reward hacking occurs when a reinforcement learning (RL) agent exploits flaws or ambiguities in the reward function to achieve high rewards, without genuinely learning or completing the intended task. Reward hacking exists because RL env...
Research1 min readarchive
Research indicates math is shifting from guiding architecture design to explaining complex model behaviors and matching structures to data symmetries.
LLMs1 min readarchive
Look at and label your data, build and evaluate your LLM-evaluator, and optimize it against your labels.
LLMs1 min readarchive
Being a human judge at the Weights & Biases LLM-as-a-Judge Hackathon
Research1 min readarchive
Static benchmarks like MMLU and HumanEval may not reflect real-world multi-round human-AI collaboration. Purposeful dialogue emphasizes iterative exchanges to achieve specific goals, such as in code generation or personal assistance.
LLMs1 min readarchive
Use cases, techniques, alignment, finetuning, and critiques against LLM-evaluators.
Research1 min readarchive
The authors argue that defining beneficial AI requires grounding it in concrete factors of individual and societal wellbeing, noting current declines in trust and connection.
Research1 min readarchive
Hallucination in large language models usually refers to the model generating unfaithful, fabricated, inconsistent, or nonsensical content. As a term, hallucination has been somewhat generalized to cases when the model makes mistakes. He...
Research1 min readarchive
Hudson River Trading compared GPT-3 training data to stock market tokens at NeurIPS 2023, noting the latter is nearly equal in volume but harder to predict due to market efficiency.
Research1 min readarchive
Diffusion models have demonstrated strong results on image synthesis in past years. Now the research community has started working on a harder task—using it for video generation. The task itself is a superset of the image case, since an ...
Research1 min readarchive
Mamba replaces the quadratic Attention bottleneck with a State Space Model to achieve linear scaling and up to 5x faster inference than Transformers.
Research1 min readarchive
The article proposes using Large Language Models to replace traditional modules in autonomous driving systems, drawing a parallel between Fleming's penicillin discovery and potential breakthroughs in self-driving technology.
Research1 min readarchive
Models often fail in production due to misleading data, hidden variables, or biased labelling rather than algorithmic flaws.
Photo: Ketut Subiyanto
And a big change for this newsletter
LLMs1 min readarchive
Sebastian Raschka details the architecture of Qwen3, a popular open-source LLM family released in May and updated in July. The article provides pure PyTorch code for building blocks found in models ranging from 0.6B to 480B parameters.
Recsys & search are converging with LLMs via semantic IDs, data augmentation, and unified foundation models.
Research quantifies gender bias in word embeddings and facial recognition systems, revealing error rates up to 34.7% for darker-skinned females.
A new paper at EMNLP 2023 demonstrates that input text can be recovered from output embeddings, challenging the assumption that vector representations are secure for storage.
