All tags
Topic: "model-alignment"
OpenEvidence, the ‘ChatGPT for doctors,’ raises $250m at $12B valuation, 12x from $1b last Feb
claude claude-3 claude-opus gpt-5.2 gemini-3-flash-high openevidence anthropic podium openai google gemini agentic-ai model-alignment performance-evaluation memory-optimization long-context benchmarking multi-agent-systems reinforcement-learning daniel_nadler amanda_askell eric_rea tom_loverro garry_tan omarsar0 brendanfoody deredleritt3r
OpenEvidence raised $12 billion, a 12x increase from last year, with usage by 40% of U.S. physicians and over $100 million in annual revenue. Anthropic released a new Claude model constitution under CC0 1.0, framing it as a living document for alignment and training. Podium reported over $100 million ARR from 10,000+ AI agents, shifting from software sales to AI operators. Innovations in agent memory and reliability include the Agent Cognitive Compressor (ACC) and multi-agent scientific workflows via MCP-SIM. Agentic benchmarking shows challenges in long-horizon tasks with models like Gemini 3 Flash High, GPT-5.2 High, and Claude Opus 4.5 High scoring modestly on professional services and legal research benchmarks.
Nemotron-4-340B: NVIDIA's new large open models, built on syndata, great for syndata
nemotron-4-340b mixtral llama-3 gemini-1.5 gpt-4o mamba-2-hybrid-8b samba-3.8b-instruct dolphin-2.9.3 faro-yi-9b-dpo nvidia hugging-face mistral-ai llamaindex cohere gemini mistral synthetic-data model-alignment reward-models fine-tuning long-context model-scaling inference-speed mixture-of-agents open-source-models model-training instruction-following context-windows philipp-schmid bryan-catanzaro oleksii-kuchaiev rohanpaul_ai cognitivecompai _philschmid 01ai_yi
NVIDIA has scaled up its Nemotron-4 model from 15B to a massive 340B dense model, trained on 9T tokens, achieving performance comparable to GPT-4. The model alignment process uses over 98% synthetic data, with only about 20K human-annotated samples for fine-tuning and reward model training. The synthetic data generation pipeline is open-sourced, including synthetic prompts and preference data generation. The base and instruct versions outperform Mixtral and Llama 3, while the reward model ranks better than Gemini 1.5, Cohere, and GPT-4o. Other notable models include Mamba-2-Hybrid 8B, which is up to 8x faster than Transformers and excels on long-context tasks, Samba-3.8B-instruct for infinite context length with linear complexity, Dolphin-2.9.3 tiny models optimized for low-resource devices, and Faro Yi 9B DPO with a 200K context window running efficiently on 16GB VRAM. The Mixture-of-Agents technique boosts open-source LLMs beyond GPT-4 Omni on AlpacaEval 2.0.
$100k to predict LMSYS human preferences in a Kaggle contest
llama-3-70b llama-3 gpt-4 claude-3-opus prometheus-2 groq openai lmsys scale-ai ai2 nvidia benchmarking datasets fine-tuning reinforcement-learning model-alignment hallucination parameter-efficient-fine-tuning scalable-training factuality chatbot-performance bindureddy drjimfan percyliang seungonekim mobicham clefourrier
Llama 3 models are making breakthroughs with Groq's 70B model achieving record low costs per million tokens. A new Kaggle competition offers a $100,000 prize to develop models predicting human preferences from a dataset of over 55,000 user-LLM conversations. Open source evaluator LLMs like Prometheus 2 outperform proprietary models such as GPT-4 and Claude 3 Opus in judgment tasks. New datasets like WildChat1M provide over 1 million ChatGPT interaction logs with diverse and toxic examples. Techniques like LoRA fine-tuning show significant performance gains, and NVIDIA's NeMo-Aligner toolkit enables scalable LLM alignment across hundreds of GPUs. Factuality-aware alignment methods are proposed to reduce hallucinations in LLM outputs.
Claude 3 just destroyed GPT 4 (see for yourself)
claude-3 claude-3-opus claude-3-sonnet claude-3-haiku gpt-4 anthropic amazon google claude-ai multimodality vision long-context model-alignment model-evaluation synthetic-data structured-output instruction-following model-speed cost-efficiency benchmarking safety mmitchell connor-leahy
Claude 3 from Anthropic launches in three sizes: Haiku (small, unreleased), Sonnet (medium, default on claude.ai, AWS, and GCP), and Opus (large, on Claude Pro). Opus outperforms GPT-4 on key benchmarks like GPQA, impressing benchmark authors. All models support multimodality with advanced vision capabilities, including converting a 2-hour video into a blog post. Claude 3 offers improved alignment, fewer refusals, and extended context length up to 1 million tokens with near-perfect recall. Haiku is noted for speed and cost-efficiency, processing dense research papers in under three seconds. The models excel at following complex instructions and producing structured outputs like JSON. Safety improvements reduce refusal rates, though some criticism remains from experts. Claude 3 is trained on synthetic data and shows strong domain-specific evaluation results in finance, medicine, and philosophy.
Gemini Ultra is out, to mixed reviews
gemini-ultra gemini-advanced solar-10.7b openhermes-2.5-mistral-7b subformer billm google openai mistral-ai hugging-face multi-gpu-support training-data-contamination model-merging model-alignment listwise-preference-optimization high-performance-computing parameter-sharing post-training-quantization dataset-viewer gpu-scheduling fine-tuning vram-optimization
Google released Gemini Ultra as a paid tier for "Gemini Advanced with Ultra 1.0" following the discontinuation of Bard. Reviews noted it is "slightly faster/better than ChatGPT" but with reasoning gaps. The Steam Deck was highlighted as a surprising AI workstation capable of running models like Solar 10.7B. Discussions in AI communities covered topics such as multi-GPU support for OSS Unsloth, training data contamination from OpenAI outputs, ethical concerns over model merging, and new alignment techniques like Listwise Preference Optimization (LiPO). The Mojo programming language was praised for high-performance computing. In research, the Subformer model uses sandwich-style parameter sharing and SAFE for efficiency, and BiLLM introduced 1-bit post-training quantization to reduce resource use. The OpenHermes dataset viewer tool was launched, and GPU scheduling with Slurm was discussed. Fine-tuning challenges for models like OpenHermes-2.5-Mistral-7B and VRAM requirements were also topics of interest.