All tags
Topic: "reinforcement-learning-from-human-feedback"
Too Cheap To Meter: AI prices cut 50-70% in last 30 days
gpt-4o gpt-4o-mini llama-3-1-405b mistral-large-2 gemini-1.5-flash deepseek-v2 sonnet-3.5 exaone-3.0 minicpm-v-2.6 claude-3.5 gpt-4o-2024-08-06 llamaindex together-ai deepinfra deepseek-ai mistral-ai google-deepmind lg-ai-research llamaindex llamaindex llamaindex price-cuts context-caching instruction-tuning vision benchmarks pytorch attention-mechanisms reinforcement-learning-from-human-feedback compute-optimal-scaling rohanpaul_ai akhaliq mervenoyann sophiamyang chhillee karpathy
Gemini 1.5 Flash has cut prices by approximately 70%, offering a highly competitive free tier of 1 million tokens per minute at $0.075/mtok, intensifying the AI model price war. Other significant price reductions include GPT-4o (~50% cut to $2.50/mtok), GPT-4o mini (70-98.5% cut to $0.15/mtok), Llama 3.1 405b (46% cut to $2.7/mtok), and Mistral Large 2 (62% cut to $3/mtok). Deepseek v2 introduced context caching, reducing input token costs by up to 90% to $0.014/mtok. New model releases include Llama 3.1 405b, Sonnet 3.5, EXAONE-3.0 (7.8B instruction-tuned by LG AI Research), and MiniCPM V 2.6 (vision-language model combining SigLIP 400M and Qwen2-7B). Benchmarks show Mistral Large performing well on ZebraLogic and Claude-3.5 leading LiveBench. FlexAttention, a new PyTorch API, simplifies and optimizes attention mechanisms. Andrej Karpathy analyzed RLHF, highlighting its limitations compared to traditional reinforcement learning. Google DeepMind research on compute-optimal scaling was also summarized.
Not much happened today.
phi-3-mini gpt4all-3.0 yi-large meta-3d-gen meta perplexity-ai microsoft gpt4all langchainai qdrant-engine 3d-generation long-context instruction-following reinforcement-learning-from-human-feedback persona-driven-data-synthesis meta-tuning model-steering memory-retrieval multivector-search universal-query-api rohanpaul_ai andriy_mulyar cwolferesearch sarahookr
Meta introduced Meta 3D Gen, a system for end-to-end generation of 3D assets from text in under 1 minute, producing high-quality 3D assets with detailed textures. Perplexity AI updated Pro Search to handle deeper research with multi-step reasoning and code execution. Microsoft improved Phi-3 Mini with better long-context understanding and instruction following. GPT4All 3.0 launched with support for thousands of models and major OS compatibility, featuring local file chat. Yi-Large model launched on Fireworks AI Playground. Research highlights include the evolution of reinforcement learning from human feedback (RLHF), persona-driven data synthesis using a billion diverse personas, meta-tuning for few-shot generalization, and steering vectors for model behavior control. Tools updates include LangSmith improving memory retrieval and Qdrant Engine v1.10 adding universal query API and multivector search.
Life after DPO (RewardBench)
gpt-3 gpt-4 gpt-5 gpt-6 llama-3-8b llama-3 claude-3 gemini x-ai openai mistral-ai anthropic cohere meta-ai-fair hugging-face nvidia reinforcement-learning-from-human-feedback direct-preference-optimization reward-models rewardbench language-model-history model-evaluation alignment-research preference-datasets personalization transformer-architecture nathan-lambert chris-manning elon-musk bindureddy rohanpaul_ai nearcyan
xAI raised $6 billion at a $24 billion valuation, positioning it among the most highly valued AI startups, with expectations to fund GPT-5 and GPT-6 class models. The RewardBench tool, developed by Nathan Lambert, evaluates reward models (RMs) for language models, showing Cohere's RMs outperforming open-source alternatives. The discussion highlights the evolution of language models from Claude Shannon's 1948 model to GPT-3 and beyond, emphasizing the role of RLHF (Reinforcement Learning from Human Feedback) and the newer DPO (Direct Preference Optimization) method. Notably, some Llama 3 8B reward model-focused models are currently outperforming GPT-4, Cohere, Gemini, and Claude on the RewardBench leaderboard, raising questions about reward hacking. Future alignment research directions include improving preference datasets, DPO techniques, and personalization in language models. The report also compares xAI's valuation with OpenAI, Mistral AI, and Anthropic, noting speculation about xAI's spending on Nvidia hardware.
1/16/2024: TIES-Merging
mixtral-8x7b nous-hermes-2 frankendpo-4x7b-bf16 thebloke hugging-face nous-research togethercompute oak-ridge-national-laboratory vast-ai runpod mixture-of-experts random-gate-routing quantization gptq exl2-quants reinforcement-learning-from-human-feedback supercomputing trillion-parameter-models ghost-attention model-fine-tuning reward-models sanjiwatsuki superking__ mrdragonfox _dampf kaltcit rombodawg technotech
TheBloke's Discord community actively discusses Mixture of Experts (MoE) models, focusing on random gate routing layers for training and the challenges of immediate model use. There is a robust debate on quantization methods, comparing GPTQ and EXL2 quants, with EXL2 noted for faster execution on specialized hardware. A new model, Nous Hermes 2, based on Mixtral 8x7B and trained with RLHF, claims benchmark superiority but shows some inconsistencies. The Frontier supercomputer at Oak Ridge National Laboratory is highlighted for training a trillion-parameter LLM with 14TB RAM, sparking discussions on open-sourcing government-funded AI research. Additionally, the application of ghost attention in the academicat model is explored, with mixed reactions from the community. "Random gate layer is good for training but not for immediate use," and "EXL2 might offer faster execution on specialized hardware," are key insights shared.