All tags
Topic: "post-training"
not much happened today
deepseek-r1-0528 o3 gemini-2.5-pro claude-opus-4 deepseek_ai openai gemini meta-ai-fair anthropic x-ai ollama hugging-face alibaba bytedance xiaomi reasoning reinforcement-learning benchmarking quantization local-inference model-evaluation open-weights transparency post-training agentic-benchmarks long-context hallucination-detection teortaxestex wenfeng danielhanchen awnihannun reach_vb abacaj
DeepSeek R1-0528 release brings major improvements in reasoning, hallucination reduction, JSON output, and function calling, matching or surpassing closed models like OpenAI o3 and Gemini 2.5 Pro on benchmarks such as Artificial Analysis Intelligence Index, LiveBench, and GPQA Diamond. The model ranks #2 globally in open weights intelligence, surpassing Meta AI, Anthropic, and xAI. Open weights and technical transparency have fueled rapid adoption across platforms like Ollama and Hugging Face. Chinese AI labs including DeepSeek, Alibaba, ByteDance, and Xiaomi now match or surpass US labs in model releases and intelligence, driven by open weights strategies. Reinforcement learning post-training is critical for intelligence gains, mirroring trends seen at OpenAI. Optimized quantization techniques (1-bit, 4-bit) and local inference enable efficient experimentation on consumer hardware. New benchmarks like LisanBench test knowledge, planning, memory, and long-context reasoning, with OpenAI o3 and Claude Opus 4 leading. Discussions highlight concerns about benchmark contamination and overemphasis on RL-tuned gains.
DeepSeek-R1-0528 - Gemini 2.5 Pro-level model, SOTA Open Weights release
deepseek-r1-0528 gemini-2.5-pro qwen-3-8b qwen-3-235b deepseek-ai anthropic meta-ai-fair nvidia alibaba google-deepmind reinforcement-learning benchmarking model-performance open-weights reasoning quantization post-training model-comparison artificialanlys scaling01 cline reach_vb zizhpan andrewyng teortaxestex teknim1 lateinteraction abacaj cognitivecompai awnihannun
DeepSeek R1-0528 marks a significant upgrade, closing the gap with proprietary models like Gemini 2.5 Pro and surpassing benchmarks from Anthropic, Meta, NVIDIA, and Alibaba. This Chinese open-weights model leads in several AI benchmarks, driven by reinforcement learning post-training rather than architecture changes, and demonstrates increased reasoning token usage (23K tokens per question). The China-US AI race intensifies as Chinese labs accelerate innovation through transparency and open research culture. Key benchmarks include AIME 2024, LiveCodeBench, and GPQA Diamond.
Execuhires: Tempting The Wrath of Khan
gemini-1.5-pro gpt-4o claude-3.5 flux-1 llama-3-1-405b character.ai google adept amazon inflection microsoft stability-ai black-forest-labs schelling google-deepmind openai anthropic meta-ai-fair lmsys langchainai execuhire model-benchmarking multilinguality math coding text-to-image agent-ide open-source-models post-training data-driven-performance noam-shazeer mostafa-mostaque david-friedman rob-rombach alexandr-wang svpino rohanpaul_ai
Character.ai's $2.5b execuhire to Google marks a significant leadership move alongside Adept's $429m execuhire to Amazon and Inflection's $650m execuhire to Microsoft. Despite strong user growth and content momentum, Character.ai's CEO Noam Shazeer returns to Google, signaling shifting vibes in the AI industry. Google DeepMind's Gemini 1.5 Pro tops Chatbot Arena benchmarks, outperforming GPT-4o and Claude-3.5, excelling in multilingual, math, and coding tasks. The launch of Black Forest Labs' FLUX.1 text-to-image model and LangGraph Studio agent IDE highlight ongoing innovation. Llama 3.1 405B is released as the largest open-source model, fostering developer use and competition with closed models. The industry is focusing increasingly on post-training and data as key competitive factors, raising questions about acquisition practices and regulatory scrutiny.
Apple Intelligence Beta + Segment Anything Model 2
llama-3-405b llama-3 segment-anything-model meta-ai-fair apple image-segmentation memory-attention video-processing pretraining cloud-tpus post-training synthetic-data instruction-following reasoning writing benchmarking bindureddy maximelabonne reach_vb
Meta advanced its open source AI with a sequel to the Segment Anything Model, enhancing image segmentation with memory attention for video applications using minimal data and compute. Apple Intelligence delayed its official release to iOS 18.1 in October but launched developer previews on MacOS Sequoia, iOS 18, and iPadOS 18, accompanied by a detailed 47-page paper revealing extensive pretraining on 6.3T tokens and use of Cloud TPUs rather than Apple Silicon. The paper highlights improvements in instruction following, reasoning, and writing through post-training and synthetic data. Benchmarks show Apple’s model scores lower than Llama 3, but with trusted human evaluations. Additionally, Meta released Llama 3.1 with a 405B parameter model, marking a significant open-source frontier model release.
Qwen 2 beats Llama 3 (and we don't know how)
qwen-2 llama-3 llama-3-70b gpt-4 nllb alibaba groq meta-ai-fair multilinguality benchmarking inference-speed sparse-autoencoders scaling-laws post-training instruction-following rejection-sampling execution-feedback model-release multilingual-models model-training philschmid huybery jonathanross321 awnihannun gdb nabla_theta ylecun
Alibaba released Qwen 2 models under Apache 2.0 license, claiming to outperform Llama 3 in open models with multilingual support in 29 languages and strong benchmark scores like MMLU 82.3 and HumanEval 86.0. Groq demonstrated ultra-fast inference speed on Llama-3 70B at 40,792 tokens/s and running 4 Wikipedia articles in 200ms. Research on sparse autoencoders (SAEs) for interpreting GPT-4 neural activity showed new training methods, metrics, and scaling laws. Meta AI announced the No Language Left Behind (NLLB) model capable of high-quality translations between 200 languages, including low-resource ones. "Our post-training phase is designed with the principle of scalable training with minimal human annotation," highlighting techniques like rejection sampling for math and execution feedback for coding.