All tags
Person: "tri_dao"
Mary Meeker is so back: BOND Capital AI Trends report
qwen-3-8b anthropic hugging-face deepseek attention-mechanisms inference arithmetic-intensity transformers model-optimization interpretability model-quantization training tri_dao fleetwood___ teortaxestex awnihannun lateinteraction neelnanda5 eliebakouch _akhaliq
Mary Meeker returns with a comprehensive 340-slide report on the state of AI, highlighting accelerating tech cycles, compute growth, and comparisons of ChatGPT to early Google and other iconic tech products. The report also covers enterprise traction and valuation of major AI companies. On Twitter, @tri_dao discusses an "ideal" inference architecture featuring attention variants like GTA, GLA, and DeepSeek MLA with high arithmetic intensity (~256), improving efficiency and model quality. Other highlights include the release of 4-bit DWQ of DSR1 Qwen3 8B on Hugging Face, AnthropicAI's open-source interpretability tools for LLMs, and discussions on transformer training and abstractions by various researchers.
SciCode: HumanEval gets a STEM PhD upgrade
gpt-4 claude-3.5-sonnet llama-3-7b llama-3 dolphin-2.9.3-yi-1.5-34b-32k-gguf anthropic hugging-face nvidia benchmarks coding model-training gpu-optimization model-performance synthetic-data compiler-optimization zero-shot-learning yi-tay rohanpaul_ai alexalbert__ tri_dao abacaj
PhD-level benchmarks highlight the difficulty of coding scientific problems for LLMs, with GPT-4 and Claude 3.5 Sonnet scoring under 5% on the new SciCode benchmark. Anthropic doubled the max output token limit for Claude 3.5 Sonnet to 8192 tokens. The Q-GaLore method enables training LLaMA-7B on a single 16GB GPU. The Mosaic compiler now generates efficient code for NVIDIA H100 GPUs. The Dolphin 2.9.3-Yi-1.5-34B-32k-GGUF model on Hugging Face has over 111k downloads. Llama 3 shows strong performance, achieving 90% zero-shot accuracy on the MATH dataset. Discussions continue on the limitations and forms of synthetic data for model training.
We Solved Hallucinations
gpt-2 flashattention-3 lynx meta-ai-fair nvidia princeton colfax patronus-ai databricks mosaic-ai openai compute-hardware gpu-optimization flashattention llm-evaluation hallucination-detection vision benchmarking synthetic-data model-training karpathy tri_dao giffmana vikhyatk dbrxmosaicai
Reddit's URL structure causes link errors in AI-generated summaries, especially with NSFW content affecting models like Claude and GPT-4. The team fixed this glitch while still leveraging LLMs for summarizing Reddit content. GPT-2 training costs have dramatically dropped to ~$672 using H100 GPUs and software improvements like CUDA and FlashAttention. FlashAttention-3 was released, achieving up to 740 TFLOPS on H100 GPUs, with FP8 nearing 1.2 PFLOPS, developed collaboratively by Meta, NVIDIA, Princeton, and Colfax. Hopper GPUs enable major speedups with new hardware features. Synthetic data may not improve vision tasks, as shown in recent research. The Avocado360 benchmark evaluates vision-language models' ability to detect avocados in images. Lynx, a hallucination detection model for LLMs, was introduced for real-world healthcare and fintech applications, trained by Patronus AI on Databricks Mosaic AI using Composer.
Not much happened today
gemini-1.5-flashmodel gemini-pro mixtral mamba-2 phi-3-medium phi-3-small gpt-3.5-turbo-0613 llama-3-8b llama-2-70b mistral-finetune twelve-labs livekit groq openai nea nvidia lmsys mistral-ai model-performance prompt-engineering data-curation ai-safety model-benchmarking model-optimization training sequence-models state-space-models daniel-kokotajlo rohanpaul_ai _arohan_ tri_dao _albertgu _philschmid sarahcat21 hamelhusain jachiam0 willdepue teknium1
Twelve Labs raised $50m in Series A funding co-led by NEA and NVIDIA's NVentures to advance multimodal AI. Livekit secured $22m in funding. Groq announced running at 800k tokens/second. OpenAI saw a resignation from Daniel Kokotajlo. Twitter users highlighted Gemini 1.5 FlashModel for high performance at low cost and Gemini Pro ranking #2 in Japanese language tasks. Mixtral models can run up to 8x faster on NVIDIA RTX GPUs using TensorRT-LLM. Mamba-2 model architecture introduces state space duality for larger states and faster training, outperforming previous models. Phi-3 Medium (14B) and Small (7B) models benchmark near GPT-3.5-Turbo-0613 and Llama 3 8B. Prompt engineering is emphasized for unlocking LLM capabilities. Data quality is critical for model performance, with upcoming masterclasses on data curation. Discussions on AI safety include a Frontier AI lab employee letter advocating whistleblower protections and debates on aligning AI to user intent versus broader humanity interests.
Mamba-2: State Space Duality
mamba-2 mamba transformer++ llama-3-70b gpt-3 hugging-face state-space-models perplexity training-efficiency data-pruning benchmarking multimodality video-analysis _albertgu tri_dao arankomatsuzaki _akhaliq clementdelangue karpathy
Mamba-2, a new state space model (SSM), outperforms previous models like Mamba and Transformer++ in perplexity and wall-clock time, featuring 8x larger states and 50% faster training. It introduces the concept of state space duality (SSD) connecting SSMs and linear attention. The FineWeb-Edu dataset, a high-quality subset of the 15 trillion token FineWeb dataset, filtered using llama-3-70b for educational quality, enables better and faster LLM learning, potentially reducing tokens needed to surpass GPT-3 performance. Additionally, perplexity-based data pruning using a 125M parameter model improves downstream performance and reduces pretraining steps by up to 1.45x. The Video-MME benchmark evaluates multi-modal LLMs on video analysis across multiple visual domains and video lengths.