All tags
Topic: "model-interpretability"
Gemma 2 2B + Scope + Shield
gemma-2b gemma-2-9b gemma-2-27b llama-3-1-405b sam-2 gpt-3.5 vicuna alpacaeval g-eval google-deepmind anthropic meta-ai-fair openai perplexity-ai nvidia lmsys knowledge-distillation leaderboards model-interpretability finetuning harm-detection video-segmentation voice publishers-program robotics-data-scaling quantization llm-evaluation prompt-engineering
Gemma 2B, a 2 billion parameter model trained on 2 trillion tokens and distilled from a larger unnamed LLM, has been released by Google DeepMind and shows strong leaderboard performance despite weaknesses in math. The Gemma series, including 9B and 27B models, has gained popularity since its June release. The team also released 400 SAEs for interpretability, inspired by Anthropic's research. A finetuned classifier called ShieldGemma outperforms Meta's LlamaGuard in harm detection. Meanwhile, Meta AI announced Llama-3.1-405B reaching #3 on the Overall Arena leaderboard, and released SAM 2, a video and image segmentation model with significant speed improvements. OpenAI is rolling out an advanced Voice Mode to Plus users. Perplexity AI launched a Publishers Program with major media partners and a status page. NVIDIA introduced Project GR00T for scaling robot data using Apple Vision Pro and generative simulation. Interest in quantization for compressing LLMs is growing, and LLM-as-a-Judge implementations from Vicuna, AlpacaEval, and G-Eval highlight the effectiveness of simple prompts and domain-specific evaluation.
HippoRAG: First, do know(ledge) Graph
qwen-2 gpt-4 hipporag alibaba openai knowledge-graphs personalized-pagerank multi-hop-retrieval chain-of-thought implicit-reasoning sparse-autoencoders model-interpretability model-efficiency model-architecture fine-tuning reinforcement-learning rohanpaul_ai omarsar0 nabla_theta huybery
Alibaba released new open-source Qwen2 models ranging from 0.5B to 72B parameters, achieving SOTA results on benchmarks like MMLU and HumanEval. Researchers introduced Sparse Autoencoders to interpret GPT-4 neural activity, improving feature representation. The HippoRAG paper proposes a hippocampus-inspired retrieval augmentation method using knowledge graphs and Personalized PageRank for efficient multi-hop reasoning. New techniques like Stepwise Internalization enable implicit chain-of-thought reasoning in LLMs, enhancing accuracy and speed. The Buffer of Thoughts (BoT) method improves reasoning efficiency with significant cost reduction. A novel scalable MatMul-free LLM architecture competitive with SOTA Transformers at billion-parameter scale was also presented. "Single-Step, Multi-Hop retrieval" is highlighted as a key advancement in retrieval speed and cost.
Anthropic's "LLM Genome Project": learning & clamping 34m features on Claude Sonnet
claude-3-sonnet claude-3 anthropic scale-ai suno-ai microsoft model-interpretability dictionary-learning neural-networks feature-activation intentional-modifiability scaling mechanistic-interpretability emmanuel-ameisen alex-albert
Anthropic released their third paper in the MechInterp series, Scaling Monosemanticity, scaling interpretability analysis to 34 million features on Claude 3 Sonnet. This work introduces the concept of dictionary learning to isolate recurring neuron activation patterns, enabling more interpretable internal states by combining features rather than neurons. The paper reveals abstract features related to code, errors, sycophancy, crime, self-representation, and deception, demonstrating intentional modifiability by clamping feature values. The research marks a significant advance in model interpretability and neural network analysis at frontier scale.
12/23/2023: NeurIPS Best Papers of 2023
gpt-4 palm2 hermes-2.5 mistral-7b nous-research hugging-face apple context-length malware-security video-content music-content linear-layers api-access large-language-models embedding vector-databases model-merging model-interpretability striped-hyena-architecture quantization rmsnorm attention-mechanisms
The Latent Space Pod released a 3-hour recap of the best NeurIPS 2023 papers. The Nous Research AI Discord community discussed optimizing AI performance with shorter context lengths, malware security concerns linked to HuggingFace, and shared insights on video and music content. Technical discussions included the DYAD research paper proposing a faster alternative to linear layers, Apple's ML Ferret machine learning tool, and accessing PALM2 via API. The community also explored Large Language Models focusing on specialized models, data scaling, embedding/vector databases, model merging, and interpretability, with mentions of Hermes 2.5, GPT-4, and Mistral. Additionally, there were conversations on the Striped Hyena Architecture, quantization challenges, and fixes related to RMSNorm and the "Attention is All You Need" paper.