All tags
Topic: "reinforcement-learning"
Apple exposes Foundation Models API and... no new Siri
chatgpt apple openai langchain llamaindex on-device-ai foundation-models reasoning reinforcement-learning voice translation software-automation agentic-workflows gdb scaling01 giffmana kevinweil
Apple released on-device foundation models for iOS developers, though their recent "Illusion of Reasoning" paper faced significant backlash for flawed methodology regarding LLM reasoning. OpenAI updated ChatGPT's Advanced Voice Mode with more natural voice and improved translation, demonstrated by Greg Brockman. LangChain and LlamaIndex launched new AI agents and tools, including a SWE Agent for software automation and an Excel agent using reinforcement learning for data transformation. The AI community engaged in heated debate over reasoning capabilities of LLMs, highlighting challenges in evaluation methods.
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.
not much happened today
deepseek-r1-0528 pali-gemma-2 gemma-3 shieldgemma-2 txgemma gemma-3-qat gemma-3n-preview medgemma dolphingemma signgemma claude-4 opus-4 claude-sonnet-4 codestral-embed bagel qwen nemotron-cortexa gemini-2.5-pro deepseek-ai huggingface gemma claude bytedance qwen nemotron sakana-ai-labs benchmarking model-releases multimodality code-generation model-performance long-context reinforcement-learning model-optimization open-source yuchenj_uw _akhaliq clementdelangue osanseviero alexalbert__ guillaumelample theturingpost lmarena_ai epochairesearch scaling01 nrehiew_ ctnzr
DeepSeek R1 v2 model released with availability on Hugging Face and inference partners. The Gemma model family continues prolific development including PaliGemma 2, Gemma 3, and others. Claude 4 and its variants like Opus 4 and Claude Sonnet 4 show top benchmark performance, including new SOTA on ARC-AGI-2 and WebDev Arena. Codestral Embed introduces a 3072-dimensional code embedder. BAGEL, an open-source multimodal model by ByteDance, supports reading, reasoning, drawing, and editing with long mixed contexts. Benchmarking highlights include Nemotron-CORTEXA topping SWEBench and Gemini 2.5 Pro performing on VideoGameBench. Discussions on random rewards effectiveness focus on Qwen models. "Opus 4 NEW SOTA ON ARC-AGI-2. It's happening - I was right" and "Claude 4 launch has dev moving at a different pace" reflect excitement in the community.
Mistral's Agents API and the 2025 LLM OS
qwen claude-4 chatgpt o3 o4 mistral-ai langchain-ai openai meta-ai-fair agent-frameworks multi-agent-systems tool-use code-execution web-search model-context-protocol persistent-memory function-calling open-source no-code reinforcement-learning model-performance agent-orchestration omarsar0 simonw swyx scaling01
The LLM OS concept has evolved since 2023, with Mistral AI releasing a new Agents API that includes code execution, web search, persistent memory, and agent orchestration. LangChainAI introduced the Open Agent Platform (OAP), an open-source no-code platform for intelligent agents. OpenAI plans to develop ChatGPT into a super-assistant by H1 2025, competing with Meta. Discussions around Qwen models focus on reinforcement learning effects, while Claude 4 performance is also noted. The AI Engineer World's Fair is calling for volunteers.
ChatGPT Codex, OpenAI's first cloud SWE agent
codex-1 openai-o3 codex-mini gemma-3 blip3-o qwen-2.5 marigold-iid deepseek-v3 lightlab gemini-2.0 lumina-next openai runway salesforce qwen deepseek google google-deepmind j1 software-engineering parallel-processing multimodality diffusion-models depth-estimation scaling-laws reinforcement-learning fine-tuning model-performance multi-turn-conversation reasoning audio-processing sama kevinweil omarsar0 iscienceluvr akhaliq osanseviero c_valenzuelab mervenoyann arankomatsuzaki jasonwei demishassabis philschmid swyx teortaxestex jaseweston
OpenAI launched Codex, a cloud-based software engineering agent powered by codex-1 (an optimized version of OpenAI o3) available in research preview for Pro, Enterprise, and Team ChatGPT users, featuring parallel task execution like refactoring and bug fixing. The Codex CLI was enhanced with quick sign-in and a new low-latency model, codex-mini. Gemma 3 is highlighted as the best open model runnable on a single GPU. Runway released the Gen-4 References API for style transfer in generation. Salesforce introduced BLIP3-o, a unified multimodal model family using diffusion transformers for CLIP image features. The Qwen 2.5 models (1.5B and 3B versions) were integrated into the PocketPal app with various chat templates. Marigold IID, a new state-of-the-art open-source depth estimation model, was released.
In research, DeepSeek shared insights on scaling and hardware for DeepSeek-V3. Google unveiled LightLab, a diffusion-based light source control in images. Google DeepMind's AlphaEvolve uses Gemini 2.0 to discover new math and reduce costs without reinforcement learning. Omni-R1 studied audio's role in fine-tuning audio LLMs. Qwen proposed a parallel scaling law inspired by classifier-free guidance. Salesforce released Lumina-Next on the Qwen base, outperforming Janus-Pro. A study found LLM performance degrades in multi-turn conversations due to unreliability. J1 is incentivizing LLM-as-a-Judge thinking via reinforcement learning. A new Qwen study correlates question and strategy similarity to predict reasoning strategies.
Gemini's AlphaEvolve agent uses Gemini 2.0 to find new Math and cuts Gemini cost 1% — without RL
gemini gpt-4.1 gpt-4o-mini o3 o4-mini google-deepmind openai algorithm-discovery coding-agents matrix-multiplication optimization reinforcement-learning model-weights training-efficiency safety-evaluations instruction-following coding-tasks model-releases _philschmid scott_swingle alex_dimakis henry jason_wei kevinweil michpokrass scaling01 gdb
Deepmind's AlphaEvolve, a 2025 update to AlphaTensor and FunSearch, is a Gemini-powered coding agent for algorithm discovery that designs faster matrix multiplication algorithms, solves open math problems, and improves data center and AI training efficiency. It achieves a 23% faster kernel speedup in Gemini training and surpasses state-of-the-art on 20% of applied problems, including improvements on the Minimum Overlap Problem and Kissing number problem. Unlike Deep-RL, it optimizes code pieces rather than model weights. Meanwhile, OpenAI released GPT-4.1 in ChatGPT, specializing in coding and instruction following, with a faster alternative GPT-4.1 mini replacing GPT-4o mini for all users. OpenAI also launched the Safety Evaluations Hub and the OpenAI to Z Challenge using o3/o4 mini and GPT-4.1 models to discover archaeological sites. "Maybe midtrain + good search is all you need for AI for scientific innovation" - Jason Wei.
Prime Intellect's INTELLECT-2 and PRIME-RL advance distributed reinforcement learning
intellect-2 dreamo qwen gemini-2.5-pro dynamic-byte-latent-transformer gen-4-references mistral-medium-3 le-chat-enterprise primeintellect bytedance qwen gemma meta-ai-fair runwayml mistral-ai google distributed-training reinforcement-learning gpu-clusters model-optimization quantization multimodality agentic-ai video-understanding fine-tuning _akhaliq reach_vb osanseviero aiatmeta c_valenzuelab lmarena_ai adcock_brett
Prime Intellect released INTELLECT-2, a decentralized GPU training and RL framework with a vision for distributed AI training overcoming colocation limits. ByteDance launched DreamO, a unified image customization model on Hugging Face. Qwen released models optimized for GPTQ, GGUF, and AWQ quantization. Gemma surpassed 150 million downloads on Hugging Face. Meta released weights for the Dynamic Byte Latent Transformer and the Collaborative Reasoner framework to improve language model efficiency and reasoning. RunwayML introduced Gen-4 References, a near-realtime model requiring no fine-tuning. Mistral AI released Mistral Medium 3, a strong multimodal model, and Le Chat Enterprise, an agentic AI assistant for business. Google updated Gemini 2.5 Pro Preview with video understanding and UI improvements. "Airbnb for spare GPUs from all over the world" highlights the ongoing challenges and potential of distributed GPU training.
not much happened today
gemini-2.5-flash gemini-2.0-flash mistral-medium-3 llama-4-maverick claude-3.7-sonnet qwen3 pangu-ultra-moe deepseek-r1 o4-mini x-reasoner google-deepmind mistral-ai alibaba huawei openai microsoft deepseek model-performance reasoning cost-analysis reinforcement-learning chain-of-thought multilinguality code-search model-training vision model-integration giffmana artificialanlys teortaxestex akhaliq john__allard
Gemini 2.5 Flash shows a 12 point increase in the Artificial Analysis Intelligence Index but costs 150x more than Gemini 2.0 Flash due to 9x more expensive output tokens and 17x higher token usage during reasoning. Mistral Medium 3 competes with Llama 4 Maverick, Gemini 2.0 Flash, and Claude 3.7 Sonnet with better coding and math reasoning at a significantly lower price. Alibaba's Qwen3 family supports reasoning and multilingual tasks across 119 languages and includes a Web Dev tool for app building. Huawei's Pangu Ultra MoE matches DeepSeek R1 performance on Ascend NPUs, with new compute and upcoming V4 training. OpenAI's o4-mini now supports Reinforcement Fine-Tuning (RFT) using chain-of-thought reasoning. Microsoft's X-REASONER enables generalizable reasoning across modalities post-trained on general-domain text. Deep research integration with GitHub repos in ChatGPT enhances codebase search and reporting. The AI Engineer World's Fair offers an Early Bird discount for upcoming tickets.
not much happened today
open-code-reasoning-32b open-code-reasoning-14b open-code-reasoning-7b mistral-medium-3 llama-4-maverick gemini-2.5-pro gemini-2.5-flash claude-3.7-sonnet absolute-zero-reasoner x-reasoner fastvlm parakeet-asr openai nvidia mistral-ai google apple huggingface reinforcement-learning fine-tuning code-generation reasoning vision on-device-ai model-performance dataset-release model-optimization reach_vb artificialanlys scaling01 iscienceluvr arankomatsuzaki awnihannun risingsayak
OpenAI launched both Reinforcement Finetuning and Deep Research on GitHub repos, drawing comparisons to Cognition's DeepWiki. Nvidia open-sourced Open Code Reasoning models (32B, 14B, 7B) with Apache 2.0 license, showing 30% better token efficiency and compatibility with llama.cpp, vLLM, transformers, and TGI. Independent evaluations highlight Mistral Medium 3 rivaling Llama 4 Maverick, Gemini 2.0 Flash, and Claude 3.7 Sonnet in coding and math reasoning, priced significantly lower but no longer open-source. Google's Gemini 2.5 Pro is noted as their most intelligent model with improved coding from simple prompts, while Gemini 2.5 Flash incurs a 150x cost increase over Gemini 2.0 Flash due to higher token usage and cost. The Absolute Zero Reasoner (AZR) achieves SOTA performance in coding and math reasoning via reinforced self-play without external data. Vision-language model X-REASONER is post-trained on general-domain text for reasoning. Apple ML research released FastVLM with on-device iPhone demo. HiDream LoRA trainer supports QLoRA fine-tuning under memory constraints. Nvidia's Parakeet ASR model tops Hugging Face ASR leaderboard with MLX implementation. New datasets SwallowCode and SwallowMath boost LLM performance in math and code. Overall, a quiet day with significant model releases and performance insights.
AI Engineer World's Fair: Second Run, Twice The Fun
gemini-2.5-pro google-deepmind waymo tesla anthropic braintrust retrieval-augmentation graph-databases recommendation-systems software-engineering-agents agent-reliability reinforcement-learning voice image-generation video-generation infrastructure security evaluation ai-leadership enterprise-ai mcp tiny-teams product-management design-engineering robotics foundation-models coding web-development demishassabis
The 2025 AI Engineer World's Fair is expanding with 18 tracks covering topics like Retrieval + Search, GraphRAG, RecSys, SWE-Agents, Agent Reliability, Reasoning + RL, Voice AI, Generative Media, Infrastructure, Security, and Evals. New focuses include MCP, Tiny Teams, Product Management, Design Engineering, and Robotics and Autonomy featuring foundation models from Waymo, Tesla, and Google. The event highlights the growing importance of AI Architects and enterprise AI leadership. Additionally, Demis Hassabis announced the Gemini 2.5 Pro Preview 'I/O edition', which leads coding and web development benchmarks on LMArena.
not much happened today
qwen3-14b qwen3-32b qwen3-235b phi-4-reasoning o3-mini command-a gemini-2.5-pro o4-mini olm-o2-1b o3 alibaba together-ai scaling01 microsoft deepseek cohere google epoch-ai-research inception-labs openai allenai quantization fine-tuning reinforcement-learning benchmarking video-generation diffusion-models model-performance model-evaluation model-release text-generation cline _philschmid iscienceluvr alexalbert__ _lewtun teortaxestex sarahookr reach_vb
Qwen model family released quantized versions of Qwen3 models including 14B, 32B, and 235B parameters, with promising coding capabilities in Qwen3-235B. Microsoft launched Phi-4-reasoning, a 14B parameter model distilled from OpenAI's o3-mini, emphasizing supervised fine-tuning and reinforcement learning, outperforming larger models in some benchmarks. Cohere's Command A leads SQL performance on Bird Bench. Google introduced the TRAJAN eval for video generation temporal consistency and updated the Gemini OpenAI compatibility layer. Inception Labs launched a diffusion LLM API claiming 5x speed improvements over autoregressive models. Community rankings show OpenAI's o3 model debuting strongly in web app-building tasks. Other releases include AllenAI's OLMo2 1B and additional Phi 4 variants. "Qwen3-235B shows promise for coding" and "Phi-4-reasoning tech report emphasizes SFT gains" highlight key advancements.
LlamaCon: Meta AI gets into the Llama API platform business
llama-4 qwen3 qwen3-235b-a22b qwen3-30b-a3b qwen3-4b qwen2-5-72b-instruct o3-mini meta-ai-fair cerebras groq alibaba vllm ollama llamaindex hugging-face llama-cpp model-release fine-tuning reinforcement-learning moe multilingual-models model-optimization model-deployment coding benchmarking apache-license reach_vb huybery teortaxestex awnihannun thezachmueller
Meta celebrated progress in the Llama ecosystem at LlamaCon, launching an AI Developer platform with finetuning and fast inference powered by Cerebras and Groq hardware, though it remains waitlisted. Meanwhile, Alibaba released the Qwen3 family of large language models, including two MoE models and six dense models ranging from 0.6B to 235B parameters, with the flagship Qwen3-235B-A22B achieving competitive benchmark results and supporting 119 languages and dialects. The Qwen3 models are optimized for coding and agentic capabilities, are Apache 2.0 licensed, and have broad deployment support including local usage with tools like vLLM, Ollama, and llama.cpp. Community feedback highlights Qwen3's scalable performance and superiority over models like OpenAI's o3-mini.
Qwen 3: 0.6B to 235B MoE full+base models that beat R1 and o1
qwen-3 qwen3-235b-a22b qwen3-30b-a3b deepseek-r1 o1 o3-mini grok-3 gemini-2.5-pro alibaba google-deepmind deepseek mistral-ai mixture-of-experts reinforcement-learning benchmarking model-release model-architecture long-context multi-agent-systems inference dataset-release awnihannun prince_canuma actuallyisaak oriolvinyalsml iscienceluvr reach_vb teortaxestex omarsar0
Qwen 3 has been released by Alibaba featuring a range of models including two MoE variants, Qwen3-235B-A22B and Qwen3-30B-A3B, which demonstrate competitive performance against top models like DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro. The models introduce an "enable_thinking=True" mode with advanced soft switching for inference scaling. The release is notable for its Apache 2.0 license and broad inference platform support including MCP. The dataset improvements and multi-stage RL post-training contribute to performance gains. Meanwhile, Gemini 2.5 Pro from Google DeepMind shows strong coding and long-context reasoning capabilities, and DeepSeek R2 is anticipated soon. Twitter discussions highlight Qwen3's finegrained MoE architecture, large context window, and multi-agent system applications.
Cognition's DeepWiki, a free encyclopedia of all GitHub repos
o4-mini perception-encoder qwen-2.5-vl dia-1.6b grok-3 gemini-2.5-pro claude-3.7 gpt-4.1 cognition meta-ai-fair alibaba hugging-face openai perplexity-ai vllm vision text-to-speech reinforcement-learning ocr model-releases model-integration open-source frameworks chatbots model-selector silas-alberti mervenoyann reach_vb aravsrinivas vikparuchuri lioronai
Silas Alberti of Cognition announced DeepWiki, a free encyclopedia of all GitHub repos providing Wikipedia-like descriptions and Devin-backed chatbots for public repos. Meta released Perception Encoders (PE) with A2.0 license, outperforming InternVL3 and Qwen2.5VL on vision tasks. Alibaba launched the Qwen Chat App for iOS and Android. Hugging Face integrated the Dia 1.6B SoTA text-to-speech model via FAL. OpenAI expanded deep research usage with a lightweight version powered by o4-mini model, now available to free users. Perplexity AI updated their model selector with Grok 3 Beta, o4-mini, and support for models like gemini 2.5 pro, claude 3.7, and gpt-4.1. vLLM project introduced OpenRLHF framework for reinforcement learning with human feedback. Surya OCR alpha model supports 90+ languages and LaTeX. MegaParse open-source library was introduced for LLM-ready data formats.
not much happened today
nemotron-h nvidia-eagle-2.5 gpt-4o qwen2.5-vl-72b gemini-2.5-flash gemini-2.0-pro gemini-exp-1206 gemma-3 qwen2.5-32b deepseek-r1-zero-32b uni3c seedream-3.0 adobe-dragon kimina-prover qwen2.5-72b bitnet-b1.58-2b4t nvidia deepseek hugging-face alibaba bytedance adobe transformers model-optimization multimodality long-context reinforcement-learning torch-compile image-generation diffusion-models distributional-rewards model-efficiency model-training native-quantization sampling-techniques philschmid arankomatsuzaki osanseviero iScienceLuvr akhaliq
Nemotron-H model family introduces hybrid Mamba-Transformer models with up to 3x faster inference and variants including 8B, 56B, and a compressed 47B model. Nvidia Eagle 2.5 is a frontier VLM for long-context multimodal learning, matching GPT-4o and Qwen2.5-VL-72B on long-video understanding. Gemini 2.5 Flash shows improved dynamic thinking and cost-performance, outperforming previous Gemini versions. Gemma 3 now supports torch.compile for about 60% faster inference on consumer GPUs. SRPO using Qwen2.5-32B surpasses DeepSeek-R1-Zero-32B on benchmarks with reinforcement learning only. Alibaba's Uni3C unifies 3D-enhanced camera and human motion controls for video generation. Seedream 3.0 by ByteDance is a bilingual image generation model with high-resolution outputs up to 2K. Adobe DRAGON optimizes diffusion generative models with distributional rewards. Kimina-Prover Preview is an LLM trained with reinforcement learning from Qwen2.5-72B, achieving 80.7% pass@8192 on miniF2F. BitNet b1.58 2B4T is a native 1-bit LLM with 2B parameters trained on 4 trillion tokens, matching full-precision LLM performance with better efficiency. Antidistillation sampling counters unwanted model distillation by modifying reasoning traces from frontier models.
Grok 3 & 3-mini now API Available
grok-3 grok-3-mini gemini-2.5-flash o3 o4-mini llama-4-maverick gemma-3-27b openai llamaindex google-deepmind epochairesearch goodfireai mechanize agent-development agent-communication cli-tools reinforcement-learning model-evaluation quantization-aware-training model-compression training-compute hybrid-reasoning model-benchmarking
Grok 3 API is now available, including a smaller version called Grok 3 mini, which offers competitive pricing and full reasoning traces. OpenAI released a practical guide for building AI agents, while LlamaIndex supports the Agent2Agent protocol for multi-agent communication. Codex CLI is gaining traction with new features and competition from Aider and Claude Code. GoogleDeepMind launched Gemini 2.5 Flash, a hybrid reasoning model topping the Chatbot Arena leaderboard. OpenAI's o3 and o4-mini models show emergent behaviors from large-scale reinforcement learning. EpochAIResearch updated its methodology, removing Maverick from high FLOP models as Llama 4 Maverick training compute drops. GoodfireAI announced a $50M Series A for its Ember neural programming platform. Mechanize was founded to build virtual work environments and automation benchmarks. GoogleDeepMind's Quantisation Aware Training for Gemma 3 models reduces model size significantly, with open source checkpoints available.
Gemini 2.5 Flash completes the total domination of the Pareto Frontier
gemini-2.5-flash o3 o4-mini google openai anthropic tool-use multimodality benchmarking reasoning reinforcement-learning open-source model-releases chain-of-thought coding-agent sama kevinweil markchen90 alexandr_wang polynoamial scaling01 aidan_mclau cwolferesearch
Gemini 2.5 Flash is introduced with a new "thinking budget" feature offering more control compared to Anthropic and OpenAI models, marking a significant update in the Gemini series. OpenAI launched o3 and o4-mini models, emphasizing advanced tool use capabilities and multimodal understanding, with o3 dominating several leaderboards but receiving mixed benchmark reviews. The importance of tool use in AI research and development is highlighted, with OpenAI Codex CLI announced as a lightweight open-source coding agent. The news reflects ongoing trends in AI model releases, benchmarking, and tool integration.
OpenAI o3, o4-mini, and Codex CLI
o3 o4-mini gemini-2.5-pro claude-3-sonnet chatgpt openai reinforcement-learning performance vision tool-use open-source coding-agents model-benchmarking multimodality scaling inference sama aidan_mclau markchen90 gdb aidan_clark_ kevinweil swyx polynoamial scaling01
OpenAI launched the o3 and o4-mini models, emphasizing improvements in reinforcement-learning scaling and overall efficiency, making o4-mini cheaper and better across prioritized metrics. These models showcase enhanced vision and tool use capabilities, though API access for these features is pending. The release includes Codex CLI, an open-source coding agent that integrates with these models to convert natural language into working code. Accessibility extends to ChatGPT Plus, Pro, and Team users, with o3 being notably more expensive than Gemini 2.5 Pro. Performance benchmarks highlight the intelligence gains from scaling inference, with comparisons against models like Sonnet and Gemini. The launch has been well received despite some less favorable evaluation results.
QwQ-32B claims to match DeepSeek R1-671B
qwen-2.5-plus qwq-32b deepseek-r1 gpt-4.5 gpt-3 davinci alibaba openai deepseek-ai reinforcement-learning math code-execution instruction-following alignment reasoning model-release model-benchmarking scaling performance inference-costs aidan_mclau sama scaling01 juberti polynoamial reach_vb
Alibaba Qwen released their QwQ-32B model, a 32 billion parameter reasoning model using a novel two-stage reinforcement learning approach: first scaling RL for math and coding tasks with accuracy verifiers and code execution servers, then applying RL for general capabilities like instruction following and alignment. Meanwhile, OpenAI rolled out GPT-4.5 to Plus users, with mixed feedback on coding performance and noted inference cost improvements. The QwQ model aims to compete with larger MoE models like DeepSeek-R1. "GPT-4.5 is unusable for coding" was a notable user critique, while others praised its reasoning improvements due to scaling pretraining.
not much happened today
grok-3 grok-3-mini gpt-4.5 claude-3.7-sonnet quasar-alpha optimus-alpha gpt-4.1 kaleidoscope internvl3 internvit qwen2.5vl transmamba fantasytalking openai alibaba cmu reinforcement-learning reasoning benchmarks vision multilinguality multimodality transformers attention-mechanisms agents code-generation model-performance rasbt sarahookr mervenoyann gneubig svpino mathemagic1an
The AI news recap highlights independent evaluations showing Grok-3 outperforming models like GPT-4.5 and Claude 3.7 Sonnet on reasoning benchmarks, while Grok-3 mini excels in reasoning tasks. Research on reinforcement learning (RL) fine-tuning reveals potential improvements for small reasoning models but also notes instability in reported gains. Benchmark results suggest Quasar Alpha and Optimus Alpha may be versions of GPT-4.1. Vision and multimodal models like Kaleidoscope, supporting 18 languages, and InternVL3, built on InternViT and Qwen2.5VL, demonstrate advances in multilingual vision and reasoning. The fusion model TransMamba combines transformer precision with speed via SSM mechanisms. Alibaba's FantasyTalking generates realistic talking portraits. Agent-focused events at CMU and tools like FilmAgent AI for virtual film production and BrowseComp benchmark for browsing agents were announced. The coding assistant Augment supports multiple IDEs with code analysis and suggestions. Discussions also covered Google’s new agent-to-agent protocol concept.
Google's Agent2Agent Protocol (A2A)
kimi-vl-a3b gpt-4o llama-4-scout llama-4-maverick llama-4-behemoth deepcoder-14b o3-mini o1 llama-3.1-nemotron-ultra-253b deepseek-r1 google google-deepmind moonshot-ai meta-ai-fair uc-berkeley openai nvidia hugging-face togethercompute deepseek agent-interoperability multimodality vision math reinforcement-learning coding model-training open-source model-benchmarking context-windows streaming push-notifications enterprise-authentication model-release reach_vb _akhaliq epochairesearch artificialanlys winglian danielhanchen yuchenj_uw jeremyphoward
Google Cloud Next announcements featured the launch of Google and DeepMind's full MCP support and a new Agent to Agent protocol designed for agent interoperability with multiple partners. The protocol includes components like the Agent Card, Task communication channels, Enterprise Auth and Observability, and Streaming and Push Notification support. On the model front, Moonshot AI released Kimi-VL-A3B, a multimodal model with 128K context and strong vision and math benchmark performance, outperforming gpt-4o. Meta AI introduced smaller versions of llama-4 family models: llama-4-scout and llama-4-maverick, with a larger Behemoth model still in training. DeepCoder 14B from UC Berkeley is an open-source coding model rivaling openai's o3-mini and o1 models, trained with reinforcement learning on 24K coding problems. Nvidia released llama-3.1-nemotron-ultra-253b on Hugging Face, noted for beating llama-4-behemoth and maverick and competing with deepseek-r1.
DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level
deepcoder-14b o3-mini o1 gemini-2.5-pro kimi-vl-a3b gpt-4o llama-4-scout maverick behemoth gen-4-turbo imagen-3 together-ai agentica opena bytedance google-deepmind moonshot-ai meta-ai-fair runway open-source reinforcement-learning code-generation multimodality model-training mixture-of-experts l2-normalization image-generation model-performance context-windows philschmid lepikhin reach_vb akhaliq yuchenj_uw epochairesearch danielhanchen c_valenzuelab
Together AI and Agentica released DeepCoder-14B, an open-source 14B parameter coding model rivaling OpenAI's o3-mini and o1 on coding benchmarks, trained with an open-source RL framework from ByteDance and costing about $26,880. Google DeepMind launched Gemini 2.5 Pro with experimental "Flash" versions available to subscribers. Moonshot AI introduced Kimi-VL-A3B, a multimodal model with 128K context outperforming gpt-4o on vision and math benchmarks. Meta AI released Llama 4 Scout and Maverick, with a larger Behemoth model in training, featuring mixture-of-experts and L2 norm techniques. Runway launched Gen-4 Turbo with 10x better results than Gen-3 at the same cost. Google announced Imagen 3, a high-quality text-to-image model now in Vertex AI, enabling easier object removal. The report highlights open-source contributions, reinforcement learning training optimizations, and significant model performance improvements across coding, multimodal, and image generation domains.
lots of little things happened this week
llama-3-3-nemotron-super-49b-v1 claude anthropic nvidia sakana-ai meta-ai-fair reinforcement-learning reasoning benchmarks multi-turn-collaboration instruction-following dataset-release model-evaluation percy-liang
Anthropic introduced a novel 'think' tool enhancing instruction adherence and multi-step problem solving in agents, with combined reasoning and tool use demonstrated by Claude. NVIDIA's Llama-3.3-Nemotron-Super-49B-v1 ranked #14 on LMArena, noted for strong math reasoning and a 15M post-training dataset. Sakana AI launched a Sudoku-based reasoning benchmark to advance AI problem-solving capabilities. Meta AI released SWEET-RL, a reinforcement learning algorithm improving long-horizon multi-turn tasks by 6%, and introduced CollaborativeAgentBench, a benchmark for collaborative LLM agents working with humans on programming and design tasks. Percy Liang relaunched the HELM benchmark with 5 challenging datasets evaluating 22 top language models.
not much happened today
gemini-2.0-flash-thinking command-a qwq-32b gemma-3-27b gemma-3 shieldgemma-2 llama-3-70b deepseek-r1 o1-mini deepseek-v3 google-deepmind cohere meta-ai-fair alibaba hugging-face model-updates model-performance benchmarking reinforcement-learning transformers normalization-layers image-generation vision memory-efficiency context-windows fine-tuning yann-lecun
Google DeepMind announced updates to Gemini 2.0, including an upgraded Flash Thinking model with stronger reasoning and native image generation capabilities. Cohere launched Command A, a 111B parameter dense model with a 256K context window and competitive pricing, available on Hugging Face. Meta AI proposed Dynamic Tanh (DyT) as a replacement for normalization layers in Transformers, supported by Yann LeCun. Alibaba released QwQ-32B, a 32.5B parameter model excelling in math and coding, fine-tuned with reinforcement learning and freely available under Apache 2.0 license. Google DeepMind also released Gemma 3 models ranging from 1B to 27B parameters with a 128K token context window and over 140 language support, plus ShieldGemma 2, an image safety checker. Benchmarking shows Gemma 3 27B has strong vision and memory efficiency but is outperformed by larger models like Llama 3.3 70B and DeepSeek V3 671B. The Hugging Face LLM leaderboard history was shared by @_lewtun.
The new OpenAI Agents Platform
reka-flash-3 o1-mini claude-3-7-sonnet llama-3-3-70b sonic-2 qwen-chat olympiccoder openai reka-ai hugging-face deepseek togethercompute alibaba ai-agents api model-releases fine-tuning reinforcement-learning model-training model-inference multimodality voice-synthesis gpu-clusters model-distillation performance-optimization open-source sama reach_vb
OpenAI introduced a comprehensive suite of new tools for AI agents, including the Responses API, Web Search Tool, Computer Use Tool, File Search Tool, and an open-source Agents SDK with integrated observability tools, marking a significant step towards the "Year of Agents." Meanwhile, Reka AI open-sourced Reka Flash 3, a 21B parameter reasoning model that outperforms o1-mini and powers their Nexus platform, with weights available on Hugging Face. The OlympicCoder series surpassed Claude 3.7 Sonnet and much larger models on competitive coding benchmarks. DeepSeek built a 32K GPU cluster capable of training V3-level models in under a week and is exploring AI distillation. Hugging Face announced Cerebras inference support, achieving over 2,000 tokens/s on Llama 3.3 70B, 70x faster than leading GPUs. Reka's Sonic-2 voice AI model delivers 40ms latency via the Together API. Alibaba's Qwen Chat enhanced its multimodal interface with video understanding up to 500MB, voice-to-text, guest mode, and expanded file uploads. Sama praised OpenAI's new API as "one of the most well-designed and useful APIs ever."
not much happened today
gpt-4.5 claude-3.7-sonnet deepseek-r1 smolagents-codeagent gpt-4o llama-3-8b tinyr1-32b-preview r1-searcher forgetting-transformer nanomoe openai deepseek hugging-face mixture-of-experts reinforcement-learning kv-cache-compression agentic-ai model-distillation attention-mechanisms model-compression minimax model-pretraining andrej-karpathy cwolferesearch aymericroucher teortaxestex jonathanross321 akhaliq
The AI news recap highlights several key developments: nanoMoE, a PyTorch implementation of a mid-sized Mixture-of-Experts (MoE) model inspired by Andrej Karpathy's nanoGPT, enables pretraining on commodity hardware within a week. An agentic leaderboard ranks LLMs powering smolagents CodeAgent, with GPT-4.5 leading, followed by Claude-3.7-Sonnet. Discussions around DeepSeek-R1 emphasize AI model commoditization, with DeepSeek dubbed the "OpenAI of China." Q-Filters offer a training-free method for KV cache compression in autoregressive models, achieving 32x compression with minimal perplexity loss. The PokéChamp minimax language agent, powered by GPT-4o and Llama-3-8b, demonstrates strong performance in Pokémon battles. Other notable models include TinyR1-32B-Preview with Branch-Merge Distillation, R1-Searcher incentivizing search capability via reinforcement learning, and the Forgetting Transformer using a Forget Gate in softmax attention. These advancements reflect ongoing innovation in model architectures, compression, reinforcement learning, and agentic AI.
AI Engineer Summit Day 1
grok-3 o3-mini deepseek-r1 qwen-2.5-vl openai anthropic xai togethercompute alibaba sakana-ai benchmarking model-performance cuda model-training open-source debugging inference-speed batch-size reinforcement-learning aidan_mclau giffmana nrehiew_ teortaxestex epochairesearch andrew_n_carr borismpower yuhu_ai_
The AIE Summit in NYC highlighted key talks including Grace Isford's Trends Keynote, Neo4j/Pfizer's presentation, and OpenAI's first definition of Agents. Speakers announced $930 million in funding. On AI Twitter, discussions focused on Grok-3 and o3-mini models, with debates on performance and benchmarking, including Grok-3's record compute scale of 4e26 to 5e26 FLOP. The o3-mini model uncovered a critical CUDA kernel bug in Sakana AI's code. DeepSeek-R1 was promoted as an open-source alternative with notable training batch sizes. Additionally, Alibaba announced the Qwen 2.5-VL model release.
X.ai Grok 3 and Mira Murati's Thinking Machines
grok-3 grok-3-mini gemini-2-pro gpt-4o o3-mini-high o1 deepseek-r1 anthropic openai thinking-machines benchmarking reasoning reinforcement-learning coding multimodality safety alignment research-publishing model-performance creative-ai mira-murati lmarena_ai karpathy omarsar0 ibab arankomatsuzaki iscienceluvr scaling01
Grok 3 has launched with mixed opinions but strong benchmark performance, notably outperforming models like Gemini 2 Pro and GPT-4o. The Grok-3 mini variant shows competitive and sometimes superior capabilities, especially in reasoning and coding, with reinforcement learning playing a key role. Mira Murati has publicly shared her post-OpenAI plan, founding the frontier lab Thinking Machines, focusing on collaborative, personalizable AI, multimodality, and empirical safety and alignment research, reminiscent of Anthropic's approach.
Reasoning Models are Near-Superhuman Coders (OpenAI IOI, Nvidia Kernels)
o3 o1 o3-mini deepseek-r1 qwen-2.5 openthinker openai nvidia ollama elevenlabs sakana-ai apple reinforcement-learning gpu-kernel-optimization fine-tuning knowledge-distillation scaling-laws chain-of-thought-reasoning model-accessibility alex-wei karpathy abacaj awnihannun
o3 model achieved a gold medal at the 2024 IOI and ranks in the 99.8 percentile on Codeforces, outperforming most humans with reinforcement learning (RL) methods proving superior to inductive bias approaches. Nvidia's DeepSeek-R1 autonomously generates GPU kernels that surpass some expert-engineered kernels, showcasing simple yet effective AI-driven optimization. OpenAI updated o1 and o3-mini models to support file and image uploads in ChatGPT and released DeepResearch, a powerful research assistant based on the o3 model with RL for deep chain-of-thought reasoning. Ollama introduced OpenThinker models fine-tuned from Qwen2.5, outperforming some DeepSeek-R1 distillation models. ElevenLabs grew into a $3.3 billion company specializing in AI voice synthesis without open-sourcing their technology. Research highlights include Sakana AI Labs' TAID knowledge distillation method receiving a Spotlight at ICLR 2025, and Apple's work on scaling laws for mixture-of-experts (MoEs). The importance of open-source AI for scientific discovery was also emphasized.
small news items
gpt-4.5 gpt-5 deepseek-r1-distilled-qwen-1.5b o1-preview modernbert-0.3b qwen-0.5b o3 openai ollama mistral perplexity cerebras alibaba groq bytedance math benchmarking fine-tuning model-performance reinforcement-learning model-architecture partnerships funding jeremyphoward arankomatsuzaki sama nrehiew_ danhendrycks akhaliq
OpenAI announced plans for GPT-4.5 (Orion) and GPT-5, with GPT-5 integrating the o3 model and offering unlimited chat access in the free tier. DeepSeek R1 Distilled Qwen 1.5B outperforms OpenAI's o1-preview on math benchmarks, while ModernBERT 0.3b surpasses Qwen 0.5b at MMLU without fine-tuning. Mistral and Perplexity adopt Cerebras hardware for 10x performance gains. OpenAI's o3 model won a gold medal at the 2024 International Olympiad in Informatics. Partnerships include Qwen with Groq. Significant RLHF activity is noted in Nigeria and the global south, and Bytedance is expected to rise in AI prominence soon. "GPT5 is all you need."
not much happened today
zonos-v0.1 audiobox-aesthetics moshi sonar llama-3-70b gpt-4o-mini claude-3.5-haiku gpt-4o claude-3.5-sonnet deepseek-r1-distilled-qwen-1.5b reasonflux-32b o1-preview zyphra-ai meta-ai-fair kyutai-labs perplexity-ai cerebras uc-berkeley brilliant-labs google-deepmind text-to-speech speech-to-speech benchmarking model-performance reinforcement-learning math real-time-processing open-source cross-platform-integration multilinguality zero-shot-learning danhendrycks
Zyphra AI launched Zonos-v0.1, a leading open-weight text-to-speech model supporting multiple languages and zero-shot voice cloning. Meta FAIR released the open-source Audiobox Aesthetics model trained on 562 hours of audio data. Kyutai Labs introduced Moshi, a real-time speech-to-speech system with low latency. Perplexity AI announced the Sonar model based on Llama 3.3 70b, outperforming top models like GPT-4o and Claude 3.5 Sonnet with 1200 tokens/second speed, powered by Cerebras infrastructure. UC Berkeley open-sourced a 1.5B model trained with reinforcement learning that beats o1-preview on math tasks. ReasonFlux-32B achieved 91.2% on the MATH benchmark, outperforming OpenAI o1-preview. CrossPoster, an AI agent for cross-platform posting, was released using LlamaIndex workflows. Brilliant Labs integrated the Google DeepMind Gemini Live API into smart glasses for real-time translation and object identification.
Gemini 2.0 Flash GA, with new Flash Lite, 2.0 Pro, and Flash Thinking
gemini-2.0-flash gemini-2.0-flash-lite gemini-2.0-pro-experimental gemini-1.5-pro deepseek-r1 gpt-2 llama-3-1 google-deepmind hugging-face anthropic multimodality context-windows cost-efficiency pretraining fine-tuning reinforcement-learning transformer tokenization embeddings mixture-of-experts andrej-karpathy jayalammar maartengr andrewyng nearcyan
Google DeepMind officially launched Gemini 2.0 models including Flash, Flash-Lite, and Pro Experimental, with Gemini 2.0 Flash outperforming Gemini 1.5 Pro while being 12x cheaper and supporting multimodal input and a 1 million token context window. Andrej Karpathy released a 3h31m video deep dive into large language models, covering pretraining, fine-tuning, and reinforcement learning with examples like GPT-2 and Llama 3.1. A free course on Transformer architecture was introduced by Jay Alammar, Maarten Gr, and Andrew Ng, focusing on tokenizers, embeddings, and mixture-of-expert models. DeepSeek-R1 reached 1.2 million downloads on Hugging Face with a detailed 36-page technical report. Anthropic increased rewards to $10K and $20K for their jailbreak challenge, while BlueRaven extension was updated to hide Twitter metrics for unbiased engagement.
How To Scale Your Model, by DeepMind
qwen-0.5 google-deepmind deepseek hugging-face transformers inference high-performance-computing robotics sim2real mixture-of-experts reinforcement-learning bias-mitigation rust text-generation open-source omarsar0 drjimfan tairanhe99 guanyashi lioronai _philschmid awnihannun clementdelangue
Researchers at Google DeepMind (GDM) released a comprehensive "little textbook" titled "How To Scale Your Model" covering modern Transformer architectures, inference optimizations beyond O(N^2) attention, and high-performance computing concepts like rooflines. The resource includes practical problems and real-time comment engagement. On AI Twitter, several key updates include the open-sourced humanoid robotics model ASAP inspired by athletes like Cristiano Ronaldo, LeBron James, and Kobe Bryant; a new paper on Mixture-of-Agents proposing the Self-MoA method for improved LLM output aggregation; training of reasoning LLMs using the GRPO algorithm from DeepSeek demonstrated on Qwen 0.5; findings on bias in LLMs used as judges highlighting the need for multiple independent evaluations; and the release of mlx-rs, a Rust library for machine learning with examples including Mistral text generation. Additionally, Hugging Face launched an AI app store featuring over 400,000 apps with 2,000 new daily additions and 2.5 million weekly visits, enabling AI-powered app search and categorization.
OpenAI takes on Gemini's Deep Research
o3 o3-mini-high o3-deep-research-mini openai google-deepmind nyu uc-berkeley hku reinforcement-learning benchmarking inference-speed model-performance reasoning test-time-scaling agent-design sama danhendrycks ethan-mollick dan-shipper
OpenAI released the full version of the o3 agent, with a new Deep Research variant showing significant improvements on the HLE benchmark and achieving SOTA results on GAIA. The release includes an "inference time scaling" chart demonstrating rigorous research, though some criticism arose over public test set results. The agent is noted as "extremely simple" and currently limited to 100 queries/month, with plans for a higher-rate version. Reception has been mostly positive, with some skepticism. Additionally, advances in reinforcement learning were highlighted, including a simple test-time scaling technique called budget forcing that improved reasoning on math competitions by 27%. Researchers from Google DeepMind, NYU, UC Berkeley, and HKU contributed to these findings. The original Gemini Deep Research team will participate in the upcoming AI Engineer NYC event.
Mistral Small 3 24B and Tulu 3 405B
mistral-small-3 tulu-3-405b llama-3 tiny-swallow-1.5b qwen-2.5-max deepseek-v3 claude-3.5-sonnet gemini-1.5-pro gpt4o-mini llama-3-3-70b mistral-ai ai2 sakana-ai alibaba_qwen deepseek ollama llamaindex reinforcement-learning model-fine-tuning local-inference model-performance model-optimization on-device-ai instruction-following api training-data natural-language-processing clementdelangue dchaplot reach_vb
Mistral AI released Mistral Small 3, a 24B parameter model optimized for local inference with low latency and 81% accuracy on MMLU, competing with Llama 3.3 70B, Qwen-2.5 32B, and GPT4o-mini. AI2 released Tülu 3 405B, a large finetuned model of Llama 3 using Reinforcement Learning from Verifiable Rewards (RVLR), competitive with DeepSeek v3. Sakana AI launched TinySwallow-1.5B, a Japanese language model using TAID for on-device use. Alibaba_Qwen released Qwen 2.5 Max, trained on 20 trillion tokens, with performance comparable to DeepSeek V3, Claude 3.5 Sonnet, and Gemini 1.5 Pro, and updated API pricing. These releases highlight advances in open models, efficient inference, and reinforcement learning techniques.
not much happened today
deepseek-r1 qwen-2.5 qwen-2.5-max deepseek-v3 deepseek-janus-pro gpt-4 nvidia anthropic openai deepseek huawei vercel bespoke-labs model-merging multimodality reinforcement-learning chain-of-thought gpu-optimization compute-infrastructure compression crypto-api image-generation saranormous zizhpan victormustar omarsar0 markchen90 sakanaailabs reach_vb madiator dain_mclau francoisfleuret garygodchaux arankomatsuzaki id_aa_carmack lavanyasant virattt
Huawei chips are highlighted in a diverse AI news roundup covering NVIDIA's stock rebound, new open music foundation models like Local Suno, and competitive AI models such as Qwen 2.5 Max and Deepseek V3. The release of DeepSeek Janus Pro, a multimodal LLM with image generation capabilities, and advancements in reinforcement learning and chain-of-thought reasoning are noted. Discussions include GPU rebranding with NVIDIA's H6400 GPUs, data center innovations, and enterprise AI applications like crypto APIs in hedge funds. "Deepseek R1's capabilities" and "Qwen 2.5 models added to applications" are key highlights.
TinyZero: Reproduce DeepSeek R1-Zero for $30
deepseek-r1 qwen o1 claude-3-sonnet claude-3 prime ppo grpo llama-stack deepseek berkeley hugging-face meta-ai-fair openai deeplearningai reinforcement-learning fine-tuning chain-of-thought multi-modal-benchmark memory-management model-training open-source agentic-workflow-automation model-performance jiayi-pan saranormous reach_vb lmarena_ai nearcyan omarsar0 philschmid hardmaru awnihannun winglian
DeepSeek Mania continues to reshape the frontier model landscape with Jiayi Pan from Berkeley reproducing the OTHER result from the DeepSeek R1 paper, R1-Zero, in a cost-effective Qwen model fine-tune for two math tasks. A key finding is a lower bound to the distillation effect at 1.5B parameters, with RLCoT reasoning emerging as an intrinsic property. Various RL techniques like PPO, DeepSeek's GRPO, or PRIME show similar outcomes, and starting from an Instruct model speeds convergence. The Humanity’s Last Exam (HLE) Benchmark introduces a challenging multi-modal test with 3,000 expert-level questions across 100+ subjects, where models perform below 10%, with DeepSeek-R1 achieving 9.4%. DeepSeek-R1 excels in chain-of-thought reasoning, outperforming models like o1 while being 20x cheaper and MIT licensed. The WebDev Arena Leaderboard ranks DeepSeek-R1 #2 in technical domains and #1 under Style Control, closing in on Claude 3.5 Sonnet. OpenAI's Operator is deployed to 100% of Pro users in the US, enabling tasks like ordering meals and booking reservations, and functions as a research assistant for AI paper searches and summaries. Hugging Face announces a leadership change after significant growth, while Meta AI releases the first stable version of Llama Stack with streamlined upgrades and automated verification. DeepSeek-R1's open-source success is celebrated, and technical challenges like memory management on macOS 15+ are addressed with residency sets in MLX for stability.
Bespoke-Stratos + Sky-T1: The Vicuna+Alpaca moment for reasoning
sky-t1-32b-preview qwen-2.5-32b r1 o1-preview gpt-4o claude-3-sonnet bespoke-stratos-32b gemini-2.0-flash-thinking berkeley usc deepseek bespoke-labs google llmsys stanford lm-sys reasoning supervised-finetuning reinforcement-learning multimodality model-distillation context-windows code-execution model-repeatability behavioral-self-awareness rlhf teortaxestex cwolferesearch madiator chakraai philschmid abacaj omarsar0
Reasoning Distillation has emerged as a key technique, with Berkeley/USC researchers releasing Sky-T1-32B-Preview, a finetuned model of Qwen 2.5 32B using 17k reasoning traces for just $450, matching benchmarks of o1-preview. DeepSeek introduced R1, a model surpassing o1-preview and enabling distillation to smaller models like a 1.5B Qwen to match gpt-4o and claude-3-sonnet levels. Bespoke Labs further distilled R1 on Qwen, outperforming o1-preview with fewer samples. This progress suggests that "SFT is all you need" for reasoning without major architecture changes. Additionally, DeepSeek-R1 uses pure reinforcement learning with supervised finetuning to accelerate convergence and shows strong reasoning and multimodal capabilities. Google's Gemini 2.0 Flash Thinking model boasts a 1 million token context window, code execution, and excels in math, science, and multimodal reasoning. Critiques highlight challenges in model repeatability, behavioral self-awareness, and RLHF limitations in reasoning robustness.
DeepSeek R1: o1-level open weights model and a simple recipe for upgrading 1.5B models to Sonnet/4o level
deepseek-r1 deepseek-v3 qwen-2.5 llama-3.1 llama-3.3-70b deepseek ollama qwen llama reinforcement-learning fine-tuning model-distillation model-optimization reasoning reward-models multi-response-sampling model-training
DeepSeek released DeepSeek R1, a significant upgrade over DeepSeek V3 from just three weeks prior, featuring 8 models including full-size 671B MoE models and multiple distillations from Qwen 2.5 and Llama 3.1/3.3. The models are MIT licensed, allowing finetuning and distillation. Pricing is notably cheaper than o1 by 27x-50x. The training process used GRPO (reward for correctness and style outcomes) without relying on PRM, MCTS, or reward models, focusing on reasoning improvements through reinforcement learning. Distilled models can run on Ollama and show strong capabilities like writing Manim code. The release emphasizes advances in reinforcement-learning, fine-tuning, and model-distillation with a novel RL framework from DeepSeekMath.
not much happened today
oute-tts-0.3-1b oute-tts-0.3-500m olm-1b qwen-2.5-0.5b hover gpt-4o deepseek-v3 harvey meta-ai-fair stability-ai alibaba deepseek hugging-face text-to-speech zero-shot-learning multilinguality emotion-control motor-control reinforcement-learning local-ai distributed-inference pipeline-parallelism mathematical-reasoning process-reward-models legal-ai education-ai ai-security humor reach_vb drjimfan vikhyatk mervenoyann aiatmeta iscienceluvr alibaba_qwen awnihannun ajeya_cotra emollick qtnx_ designerx
Harvey secured a new $300M funding round. OuteTTS 0.3 1B & 500M text-to-speech models were released featuring zero-shot voice cloning, multilingual support (en, jp, ko, zh, fr, de), and emotion control, powered by OLMo-1B and Qwen 2.5 0.5B. The HOVER model, a 1.5M-parameter neural net for agile motor control, was introduced, leveraging human motion capture datasets and massively parallel reinforcement learning. kokoro.js enables running AI models locally in browsers with minimal dependencies. Meta AI awarded $200K LLM evaluation grants for projects on regional language understanding, complex reasoning, and interactive programming environments. Stability AI's Twitter account was hacked, prompting security warnings. Alibaba Qwen improved Process Reward Models (PRMs) for better mathematical reasoning using a consensus filtering mechanism. DeepSeek V3 uses pipeline parallelism to enhance distributed inference and long-context generation efficiency. Discussions on AI policy in legal frameworks and AI's role in democratizing education were highlighted. Lighthearted AI-related humor was also shared.
not much happened today
phi-4 reinforce++ arc-agi-2 ai21-labs ollama langchain togethercompute groq reinforcement-learning ppo model-optimization memory-efficiency python-packages vision text-extraction frontend-code-generation workflow-automation coding-agents compute-cost-reduction ethical-ai agi-benchmarks scam-alerts sebastien-bubeck fchollet tom-doerr arohan_ bindureddy hwchase17 jonathanross321 clementdelangue vikhyatk
Sebastien Bubeck introduced REINFORCE++, enhancing classical REINFORCE with PPO-inspired techniques for 30% faster training. AI21 Labs released Phi-4 under the MIT License, accessible via Ollama. François Chollet announced plans for ARC-AGI-2 and a next-generation AGI benchmark. LangChain launched 10 new integration packages to boost LLM application development. Tom Doerr introduced Ollama-OCR, a Python package for text extraction using vision language models. Arohan optimized Shampoo for memory efficiency, reducing usage from 20 to 6 bytes per parameter. Bindu Reddy showcased CodeLLM's v1 for frontend code generation and highlighted LlamaIndex Workflows for academic summarization and slide generation. Hwchase17 collaborated with Together Compute to enhance WebDev Arena with complex coding agents for LLM coding evaluations. Jonathan Ross detailed Groq's mission to reduce compute costs by 1000x amid rising generative AI spending. Clement Delangue warned about scam alerts involving false claims of association with AI21. Vikhyat K raised concerns about the ethical implications and trade-offs of AGI. Memes and humor included creative AI prompts and critiques of LLM behaviors.
PRIME: Process Reinforcement through Implicit Rewards
claude-3.5-sonnet gpt-4o deepseek-v3 gemini-2.0 openai together-ai deepseek langchain lucidrains reinforcement-learning scaling-laws model-performance agent-architecture software-development compute-scaling multi-expert-models sama aidan_mclau omarsar0 akhaliq hwchase17 tom_doerr lmarena_ai cwolferesearch richardmcngo
Implicit Process Reward Models (PRIME) have been highlighted as a significant advancement in online reinforcement learning, trained on a 7B model with impressive results compared to gpt-4o. The approach builds on the importance of process reward models established by "Let's Verify Step By Step." Additionally, AI Twitter discussions cover topics such as proto-AGI capabilities with claude-3.5-sonnet, the role of compute scaling for Artificial Superintelligence (ASI), and model performance nuances. New AI tools like Gemini 2.0 coder mode and LangGraph Studio enhance agent architecture and software development. Industry events include the LangChain AI Agent Conference and meetups fostering AI community connections. Company updates reveal OpenAI's financial challenges with Pro subscriptions and DeepSeek-V3's integration with Together AI APIs, showcasing efficient 671B MoE parameter models. Research discussions focus on scaling laws and compute efficiency in large language models.
not much happened to end the year
deepseek-v3 code-llm o1 sonnet-3.5 deepseek smol-ai reinforcement-learning reasoning training-data mixed-precision-training open-source multimodality software-development natural-language-processing interpretability developer-tools real-time-applications search sdk-generation corbtt tom_doerr cognitivecompai alexalbert__ theturingpost svpino bindureddy
Reinforcement Fine-Tuning (RFT) is introduced as a data-efficient method to improve reasoning in LLMs using minimal training data with strategies like First-Correct Solutions (FCS) and Greedily Diverse Solutions (GDS). DeepSeek-V3, a 671B parameter MoE language model trained on 14.8 trillion tokens with FP8 mixed precision training, highlights advances in large-scale models and open-source LLMs. Predictions for AI in 2025 include growth in smaller models, multimodality, and challenges in open-source AI. The impact of AI on software development jobs suggests a need for higher intelligence and specialization as AI automates low-skilled tasks. Enhancements to CodeLLM improve coding assistance with features like in-place editing and streaming responses. Natural Language Reinforcement Learning (NLRL) offers better interpretability and richer feedback for AI planning and critique. AI hiring is growing rapidly with startups seeking strong engineers in ML and systems. New AI-powered tools such as Rivet, Buzee, and Konfig improve real-time applications, search, and SDK generation using technologies like Rust and V8 isolates.
DeepSeek v3: 671B finegrained MoE trained for $5.5m USD of compute on 15T tokens
deepseek-v3 gpt-4o claude-3.5-sonnet llama-3 deepseek-ai hugging-face openai anthropic mixture-of-experts model-training model-optimization reinforcement-learning chain-of-thought multi-token-prediction synthetic-data model-distillation fine-tuning attention-mechanisms gpu-optimization nrehiew_ denny_zhou
DeepSeek-V3 has launched with 671B MoE parameters and trained on 14.8T tokens, outperforming GPT-4o and Claude-3.5-sonnet in benchmarks. It was trained with only 2.788M H800 GPU hours, significantly less than Llama-3's 30.8M GPU-hours, showcasing major compute efficiency and cost reduction. The model is open-source and deployed via Hugging Face with API support. Innovations include native FP8 mixed precision training, Multi-Head Latent Attention scaling, distillation from synthetic reasoning data, pruning and healing for MoEs with up to 256 experts, and a new multi-token prediction objective enabling lookahead token planning. Research highlights also cover the OREO method and Natural Language Reinforcement Learning (NLRL) for multi-step reasoning and agent control.
Meta BLT: Tokenizer-free, Byte-level LLM
byte-latent-transformer llama-3 phi-4 gpt-4o command-r7b meta-ai-fair llamaindex microsoft deepseek-ai openai cohere anthropic tokenization transformer-architecture model-efficiency benchmarking multimodality vision reinforcement-learning model-scaling jailbreaking model-optimization
Meta AI introduces the Byte Latent Transformer (BLT), a tokenizer-free architecture that dynamically forms byte patches for efficient compute allocation, outperforming Llama 3 on benchmarks including the CUTE benchmark. The model was trained on approximately 1 trillion tokens and features a three-block transformer design with local and global components. This approach challenges traditional tokenization and may enable new multimodal capabilities such as direct file interaction without retrieval-augmented generation. Additionally, Microsoft announced the Phi-4 14B parameter model achieving state-of-the-art results on STEM and reasoning benchmarks, surpassing GPT-4o. DeepSeek AI launched new vision-language models based on their MoE architecture with sizes ranging from 1.0B to 27B parameters. OpenAI released a new Projects feature for ChatGPT, and Cohere introduced their smallest and fastest Command R7B model. Anthropic published research on "Best-of-N Jailbreaking" vulnerabilities across text, vision, and audio models. Industry discussion highlights a trend of decreasing frontier LLM sizes, with GPT-4 at approximately 1.8 trillion parameters compared to newer models.
Meta Llama 3.3: 405B/Nova Pro performance at 70B price
llama-3-70b llama-3.3-70b gpt-4o gemini-exp-1206 meta-ai-fair openai google-deepmind hugging-face llamacloud reinforcement-learning fine-tuning model-performance document-processing pricing-models alignment online-rl sama steven-heidel aidan_mclau lmarena_ai oriolvinyalsml jerryjliu0
Meta AI released Llama 3.3 70B, matching the performance of the 405B model with improved efficiency using "a new alignment process and progress in online RL techniques". OpenAI announced Reinforcement Fine-Tuning (RFT) for building expert models with limited data, offering alpha access to researchers and enterprises. Google DeepMind's Gemini-Exp-1206 leads benchmarks, tying with GPT-4o in coding performance. LlamaCloud enhanced document processing with table extraction and analytics. Discussions on OpenAI's pricing plans continue in the community.
not much happened today
o1-full sora gpt-4.5 gpt-4 claude-3.5-sonnet llama-3-1-nemotron-51b llama-3-1 llama-3 nemotron-51b openai google-deepmind anthropic nvidia huggingface vision model-performance neural-architecture-search model-optimization multimodality model-release model-training reinforcement-learning image-generation lucas-beyer alexander-kolesnikov xiaohua-zhai aidan_mclau giffmana joannejang sama
OpenAI announced their "12 Days of OpenAI" event with daily livestreams and potential releases including the O1 full model, Sora video model, and GPT-4.5. Google DeepMind released the GenCast weather model capable of 15-day forecasts in 8 minutes using TPU chips, and launched Genie 2, a model generating playable 3D worlds from single images. Leading vision researchers Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai moved from DeepMind to OpenAI, which is opening a Zürich office. Criticism arose over OpenAI's strategy and model quality compared to Anthropic and Claude 3.5 Sonnet. On Reddit, a modified llama.cpp supports Nvidia's Llama-3_1-Nemotron-51B, matching performance of larger 70B models via NAS optimization.
not much happened to end the week
gemini deepseek-r1 o1 chatgpt gpt-4 claude-3.5-sonnet o1-preview o1-mini gpt4o qwq-32b google-deepmind deeplearningai amazon tesla x-ai alibaba ollama multimodality benchmarking quantization reinforcement-learning ai-safety translation reasoning interpretability model-comparison humor yoshua-bengio kevinweil ylecun
AI News for 11/29/2024-11/30/2024 covers key updates including the Gemini multimodal model advancing in musical structure understanding, a new quantized SWE-Bench for benchmarking at 1.3 bits per task, and the launch of the DeepSeek-R1 model focusing on transparent reasoning as an alternative to o1. The establishment of the 1st International Network of AI Safety Institutes highlights global collaboration on AI safety. Industry updates feature Amazon's Olympus AI model, Tesla's Optimus, and experiments with ChatGPT as a universal translator. Community reflections emphasize the impact of large language models on daily life and medical AI applications. Discussions include scaling sparse autoencoders to gpt-4 and the need for transparency in reasoning LLMs. The report also notes humor around ChatGPT's French nickname.
OLMo 2 - new SOTA Fully Open LLM
llama-3-1-8b olmo-2 qwen2-5-72b-instruct smolvlm tulu-3 ai2 huggingface intel reinforcement-learning quantization learning-rate-annealing ocr fine-tuning model-training vision
AI2 has updated OLMo-2 to roughly Llama 3.1 8B equivalent, training with 5T tokens and using learning rate annealing and new high-quality data (Dolmino). They credit Tülu 3 and its "Reinforcement Learning with Verifiable Rewards" approach. On Reddit, Qwen2.5-72B instruct model shows near lossless performance with AutoRound 4-bit quantization, available on HuggingFace in 4-bit and 2-bit versions, with discussions on MMLU benchmark and quantization-aware training. HuggingFace released SmolVLM, a 2B parameter vision-language model running efficiently on consumer GPUs, supporting fine-tuning on Google Colab and demonstrating strong OCR capabilities with adjustable resolution and quantization options.
Vision Everywhere: Apple AIMv2 and Jina CLIP v2
aimv2-3b jina-clip-v2 tulu-3 llama-3-1 claude-3-5 llama-3-1-70b apple jina allen_ai autoregressive-objectives vision multilinguality multimodality image-generation model-training model-optimization reinforcement-learning fine-tuning model-benchmarking
Apple released AIMv2, a novel vision encoder pre-trained with autoregressive objectives that achieves 89.5% accuracy on ImageNet and integrates joint visual and textual objectives. Jina launched Jina CLIP v2, a multimodal embedding model supporting 89 languages and high-resolution images with efficient Matryoshka embeddings reducing dimensions by 94% with minimal accuracy loss. Allen AI introduced Tülu 3 models based on Llama 3.1 with 8B and 70B parameters, offering 2.5x faster inference and alignment via SFT, DPO, and RLVR methods, competing with Claude 3.5 and Llama 3.1 70B. These developments highlight advances in autoregressive training, vision encoders, and multilingual multimodal embeddings.
not much happened this weekend
claude-3.5-sonnet llama-3 llama-3-8b notebookllama min-omni-2 moondream openai anthropic hugging-face mistral-ai google-deepmind langchain deepmind microsoft pattern-recognition reinforcement-learning prompt-optimization text-to-speech model-optimization tensor-parallelism hyperparameters multimodal modal-alignment multimodal-fine-tuning ai-productivity privacy generative-ai rag retrieval-augmentation enterprise-text-to-sql amanda-askell philschmid stasbekman francois-fleuret mervenoyann reach_vb dzhng aravsrinivas sama lateinteraction andrew-y-ng bindureddy jerryjliu0
Moondream, a 1.6b vision language model, secured seed funding, highlighting a trend in moon-themed tiny models alongside Moonshine (27-61m ASR model). Claude 3.5 Sonnet was used for AI Twitter recaps. Discussions included pattern recognition vs. intelligence in LLMs, reinforcement learning for prompt optimization, and NotebookLlama, an open-source NotebookLM variant using LLaMA models for tasks like text-to-speech. Advances in model optimization with async-TP in PyTorch for tensor parallelism and hyperparameter tuning were noted. Mini-Omni 2 demonstrated multimodal capabilities across image, audio, and text for voice conversations with emphasis on modal alignment and multimodal fine-tuning. AI productivity tools like an AI email writer and LlamaCloud-based research assistants were introduced. Emphasis on practical skill development and privacy-conscious AI tool usage with Llama3-8B was highlighted. Generative AI tools such as #AIPythonforBeginners and GenAI Agents with LangGraph were shared. Business insights covered rapid execution in AI product development and emerging AI-related job roles. Challenges in enterprise-grade text-to-SQL and advanced retrieval methods were discussed with tutorials on RAG applications using LangChain and MongoDB.
DeepSeek Janus and Meta SpiRit-LM: Decoupled Image and Expressive Voice Omnimodality
nemotron-70b claude claude-3.5-sonnet gpt-4o deepseek meta-ai-fair wandb nvidia anthropic hugging-face perplexity-ai multimodality image-generation speech-synthesis fine-tuning model-merging benchmarking open-source model-optimization reinforcement-learning bindureddy aravsrinivas danielhanchen clementdelangue cwolferesearch
DeepSeek Janus and Meta SpiRit-LM are two notable multimodality AI models recently released, showcasing advances in image generation and speech synthesis respectively. DeepSeek Janus separates vision encoders for image understanding and generation, achieving better results in both tasks. Meta's SpiRit-LM introduces an expressive speech and writing model generating pitch and style units, improving over standard TTS. Additionally, W&B Weave offers comprehensive LLM observability and multimodality fine-tuning tools. Industry updates include Nvidia's Nemotron 70b model underperforming, Meta open-sourcing Movie Gen Bench for media generation benchmarking, Perplexity launching internal search with multi-step reasoning, and Anthropic updating Claude apps. Open source progress includes Hugging Face's gradient accumulation fix in transformers and advocacy for open source AI to prevent Big Tech dominance. "Model merging for combining skills of multiple models" is also highlighted.
Did Nvidia's Nemotron 70B train on test?
nemotron-70b llama-3.1-70b llama-3.1 ministral-3b ministral-8b gpt-4o claude-3.5-sonnet claude-3.5 nvidia mistral-ai hugging-face zep benchmarking reinforcement-learning reward-models temporal-knowledge-graphs memory-layers context-windows model-releases open-source reach_vb philschmid swyx
NVIDIA's Nemotron-70B model has drawn scrutiny despite strong benchmark performances on Arena Hard, AlpacaEval, and MT-Bench, with some standard benchmarks like GPQA and MMLU Pro showing no improvement over the base Llama-3.1-70B. The new HelpSteer2-Preference dataset improves some benchmarks with minimal losses elsewhere. Meanwhile, Mistral released Ministral 3B and 8B models featuring 128k context length and outperforming Llama-3.1 and GPT-4o on various benchmarks under the Mistral Commercial License. NVIDIA's Nemotron 70B also surpasses GPT-4o and Claude-3.5-Sonnet on key benchmarks using RLHF (REINFORCE) training. Additionally, Zep introduced Graphiti, an open-source temporal knowledge graph memory layer for AI agents, built on Neo4j.
Liquid Foundation Models: A New Transformers alternative + AINews Pod 2
llama-3-2 gemini-1.5-pro-002 gemini-1.5-flash-002 liquid-ai meta-ai-fair google-deepmind openai reinforcement-learning multimodality model-efficiency foundation-models audio-processing model-deployment open-source ylecun svpino
Liquid.ai emerged from stealth with three subquadratic foundation models demonstrating superior efficiency compared to state space models and Apple’s on-device and server models, backed by a $37M seed round. Meta AI announced Llama 3.2 with multimodal vision-enabled models and lightweight text-only variants for mobile. Google DeepMind introduced production-ready Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002 models with improved pricing and rate limits, alongside AlphaChip, an AI-driven chip design system using reinforcement learning for rapid superhuman layouts. OpenAI enhanced ChatGPT Plus and Teams with Advanced Voice Mode featuring Custom Instructions, Memory, and new nature-inspired voices. California Governor vetoed SB-1047 AI regulation bill, celebrated by AI community figures like ylecun and svpino as a win for open-source AI. Google upgraded NotebookLM with audio overviews supporting YouTube and audio files, turning documents into AI-generated podcasts. "Open source in AI is thriving," noted ylecun, highlighting 1 million models on Github and HuggingFace.
a calm before the storm
o1 o1-mini qwen2.5 gpt-4 llama-2-70b llama-7b anthropic openai alibaba microsoft blackrock groq aramco disney eth-zurich pudu-robotics slack long-context kv-cache-quantization diffusion-models reinforcement-learning robotics ai-integration multilinguality model-benchmarking model-performance model-optimization adcock_brett philschmid rohanpaul_ai jvnixon kateclarktweets sama
Anthropic is raising funds at a valuation up to $40 billion ahead of anticipated major releases. OpenAI launched new reasoning models o1 and o1-mini, with increased rate limits and a multilingual MMLU benchmark. Alibaba released the open-source Qwen2.5 model supporting 29+ languages, showing competitive performance to gpt-4 at lower cost. Microsoft and Blackrock plan to invest $30 billion in AI data centers, with Groq partnering with Aramco to build the world's largest AI inference center. Robotics advances include Disney Research and ETH Zurich's diffusion-based motion generation for robots and Pudu Robotics' semi-humanoid robot. Slack and Microsoft introduced AI-powered agents integrated into their platforms. Research highlights include long-context scaling for llama-2-70b using Dual Chunk Attention and KV cache quantization enabling 1 million token context on llama-7b models.
nothing much happened today
o1 chatgpt-4o llama-3-1-405b openai lmsys scale-ai cognition langchain qdrant rohanpaul_ai reinforcement-learning model-merging embedding-models toxicity-detection image-editing dependency-management automated-code-review visual-search benchmarking denny_zhou svpino alexandr_wang cwolferesearch rohanpaul_ai _akhaliq kylebrussell
OpenAI's o1 model faces skepticism about open-source replication due to its extreme restrictions and unique training advances like RL on CoT. ChatGPT-4o shows significant performance improvements across benchmarks. Llama-3.1-405b fp8 and bf16 versions perform similarly with cost benefits for fp8. A new open-source benchmark "Humanity's Last Exam" offers $500K in prizes to challenge LLMs. Model merging benefits from neural network sparsity and linear mode connectivity. Embedding-based toxic prompt detection achieves high accuracy with low compute. InstantDrag enables fast, optimization-free drag-based image editing. LangChain v0.3 releases with improved dependency management. Automated code review tool CodeRabbit adapts to team coding styles. Visual search advances integrate multimodal data for better product search. Experts predict AI will be default software by 2030.
a quiet weekend
o1 datagemma aloha demostart firefly-ai-video-model pixtral-12b gamegen-o openai google-deepmind adobe mistral-ai tencent supermaven 11x cohere anthropic latent-space-university stanford microsoft mila notre-dame reinforcement-learning chain-of-thought reasoning robotics diffusion-models multimodality video-generation model-training reflection-tuning mathematical-reasoning model-benchmarking fine-tuning george-hotz terence-tao adcock_brett rohanpaul_ai bindureddy fchollet philschmid
OpenAI released the new o1 model, leveraging reinforcement learning and chain-of-thought prompting to excel in reasoning benchmarks, achieving an IQ-like score of 120. Google DeepMind introduced DataGemma to reduce hallucinations by connecting LLMs with real-world data, and unveiled ALOHA and DemoStart for robot dexterity using diffusion methods. Adobe previewed its Firefly AI Video Model with text-to-video and generative extend features. Mistral launched the multimodal Pixtral 12B model, and Tencent presented the GameGen-O open-world video game generation model. Several research papers from Stanford, OpenAI, Microsoft, Mila, and Notre Dame focus on advanced reasoning, self-verification, and reflection tuning techniques. Experts like Terence Tao and George Hotz have shared mixed but optimistic views on o1's capabilities. Seed funding rounds include Supermaven ($12M) and 11x ($24M).
Learnings from o1 AMA
o1-preview o1-mini claude-3.5-sonnet gpt-4o openai weights-biases cohere weaviate reinforcement-learning chain-of-thought reasoning model-performance prompting code-editing rag hybrid-search sama rohanpaul_ai gdb andrew-mayne
OpenAI released the o1 model series, touted as their "most capable and aligned models yet," trained with reinforcement learning to enhance reasoning. The o1-preview model scored 21% on ARC-AGI, ~80% on aider code editing (surpassing Claude 3.5 Sonnet's 77%), and ~52% on Cognition-Golden, showcasing a shift from memorizing answers to memorizing reasoning. The model employs a unique chain-of-thought approach enabling "System II thinking" for better problem-solving. Experts like Andrew Mayne advise framing o1 as a smart friend providing thoughtful explanations. Additionally, an advanced RAG course sponsored by Weights & Biases, Cohere, and Weaviate offers strategies for hybrid search and prompting to optimize AI solutions.
not much happened today
gpt-4o claude-3.5-sonnet phi-3.5-mini phi-3.5-moe phi-3.5-vision llama-3-1-405b qwen2-math-72b openai anthropic microsoft meta-ai-fair hugging-face langchain box fine-tuning benchmarking model-comparison model-performance diffusion-models reinforcement-learning zero-shot-learning math model-efficiency ai-regulation ai-safety ai-engineering prompt-engineering swyx ylecun
OpenAI launched GPT-4o finetuning with a case study on Cosine. Anthropic released Claude 3.5 Sonnet with 8k token output. Microsoft Phi team introduced Phi-3.5 in three variants: Mini (3.8B), MoE (16x3.8B), and Vision (4.2B), noted for sample efficiency. Meta released Llama 3.1 405B, deployable on Google Cloud Vertex AI, offering GPT-4 level capabilities. Qwen2-Math-72B achieved state-of-the-art math benchmark performance with a Gradio demo. Discussions included model comparisons like ViT vs CNN and Mamba architecture. Tools updates featured DSPy roadmap, Flux Schnell improving diffusion speed on M1 Max, and LangChain community events. Research highlights zero-shot DUP prompting for math reasoning and fine-tuning best practices. AI ethics covered California's AI Safety Bill SB 1047 and regulatory concerns from Yann LeCun. Commentary on AI engineer roles by Swyx. "Chat with PDF" feature now available for Box Enterprise Plus users.
not much happened today
qwen2-math-72b gpt-4o claude-3.5-sonnet gemini-1.5-pro llama-3.1-405b idefics3-llama-8b anthropic google mistral-ai llamaindex math fine-tuning synthetic-data reinforcement-learning bug-bounty visual-question-answering open-source retrieval-augmented-generation agentic-ai ai-safety policy rohanpaul_ai anthropicai mervenoyann jeremyphoward omarsar0 ylecun bindureddy
Qwen2-Math-72B outperforms GPT-4o, Claude-3.5-Sonnet, Gemini-1.5-Pro, and Llama-3.1-405B on math benchmarks using synthetic data and advanced optimization techniques. Google AI cuts pricing for Gemini 1.5 Flash by up to 78%. Anthropic expands its bug bounty program targeting universal jailbreaks in next-gen safety systems. Tutorial on QLoRA fine-tuning of IDEFICS3-Llama 8B for visual question answering released. A Chinese open weights model surpasses previous MATH benchmark records. Surveys on Mamba models and LLM-based agents for software engineering highlight advancements and applications. Open-source tools like R2R RAG engine and LlamaIndex Workflows simplify building complex AI applications. Mistral AI introduces customizable AI agents. Concerns raised about California bill SB 1047's focus on existential risk and debates on banning open-source AI. Memes and humor continue in AI communities.
Llama 3.1: The Synthetic Data Model
llama-3-405b llama-3-1 llama-3 meta-ai-fair groq fireworks synthetic-data fine-tuning reinforcement-learning multilinguality long-context tool-use code-generation math model-licensing inference-speed model-deployment bindureddy thomas
Meta AI has released Llama 3.1, including a 405B parameter model that triggers regulatory considerations like the EU AI Act and SB 1047. The model incorporates extensive synthetic data techniques for code, math, multilinguality, long context, and tool use fine-tuning, with RLHF using synthetic preference data from Llama 2. The launch was coordinated across major inference providers, with Groq demonstrating 750 tokens per second inference speed and Fireworks leading in pricing. The updated license explicitly allows synthetic data generation, marking a significant step in open frontier-class LLMs and cost-efficiency improvements since March.
Gemini launches context caching... or does it?
nemotron llama-3-70b chameleon-7b chameleon-34b gemini-1.5-pro deepseek-coder-v2 gpt-4-turbo claude-3-opus gemini-1.5-pro nvidia meta-ai-fair google deepseek hugging-face context-caching model-performance fine-tuning reinforcement-learning group-relative-policy-optimization large-context model-training coding model-release rohanpaul_ai _philschmid aman-sanger
Nvidia's Nemotron ranks #1 open model on LMsys and #11 overall, surpassing Llama-3-70b. Meta AI released Chameleon 7B/34B models after further post-training. Google's Gemini introduced context caching, offering a cost-efficient middle ground between RAG and finetuning, with a minimum input token count of 33k and no upper limit on cache duration. DeepSeek launched DeepSeek-Coder-V2, a 236B parameter model outperforming GPT-4 Turbo, Claude-3-Opus, and Gemini-1.5-Pro in coding tasks, supporting 338 programming languages and extending context length to 128K. It was trained on 6 trillion tokens using the Group Relative Policy Optimization (GRPO) algorithm and is available on Hugging Face with a commercial license. These developments highlight advances in model performance, context caching, and large-scale coding models.
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.
Not much happened today
gpt-4o gemini-1.5-pro gemini-1.5-flash imagen-3 veo reka-core qwen-1.5-110b openai google-deepmind anthropic rekailabs alibaba salesforce multimodality long-context model-releases reinforcement-learning model-benchmarking text-to-image video-generation ai-assistants ilya-sutskever jakub-pachocki mike-krieger sama
Ilya Sutskever steps down as Chief Scientist at OpenAI after nearly a decade, with Jakub Pachocki named as his successor. Google DeepMind announces Gemini 1.5 Pro and Gemini 1.5 Flash models featuring 2 million token context and improved multimodal capabilities, alongside demos of Project Astra AI assistant, Imagen 3 text-to-image model, and Veo generative video model. GPT-4o tops the VHELM leaderboard and outperforms competitors on LMSYS Chatbot Arena. Reka Core multimodal model with 128K context and Alibaba's Qwen1.5-110B open-source model are released. Salesforce shares an online RLHF recipe.
$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.
The world's first fully autonomous AI Engineer
gpt-4 devin cognition-labs openai reinforcement-learning fine-tuning long-term-reasoning planning ai-agents software-engineering model-integration asynchronous-chat ide agentic-ai patrick-collison fred-ehrsam tim-dettmers
Cognition Labs's Devin is highlighted as a potentially groundbreaking AI software engineer agent capable of learning unfamiliar technologies, addressing bugs, deploying frontend apps, and fine-tuning its own AI models. It integrates OpenAI's GPT-4 with reinforcement learning and features tools like asynchronous chat, browser, shell access, and an IDE. The system claims advanced long-term reasoning and planning abilities, attracting praise from investors like Patrick Collison and Fred Ehrsam. The technology is noted for its potential as one of the most advanced AI agents, sparking excitement about agents and AGI.
FSDP+QLoRA: the Answer to 70b-scale AI for desktop class GPUs
qlora fsdp inflection-2.5 gpt-4 answer.ai hugging-face meta-ai-fair nvidia inflectionai model-training quantization memory-optimization gradient-checkpointing cpu-offloading fine-tuning model-sharding reinforcement-learning chain-of-thought benchmarking jeremy_howard tim_dettmers yann_lecun
Jeremy Howard and collaborators released a new tool combining FSDP, QLoRA, and HQQ to enable training 70b-parameter models on affordable consumer GPUs like RTX 4090s with only 24GB RAM, overcoming traditional memory constraints that required expensive data center GPUs costing over $150k. The approach shards quantized models across multiple GPUs and uses techniques like gradient checkpointing and CPU offloading to achieve efficient training on desktop-class hardware. The blogpost details challenges and solutions integrating these methods, highlighting a significant cost reduction from $150k to under $2.5k for training large language models. Additionally, Twitter recaps mention Inflection AI's Inflection-2.5 model rivaling GPT-4 in benchmarks with less compute, and Grok improving speed by 3x. Yann LeCun discusses multi-step reasoning training for LLMs.
Mistral Large disappoints
mistral-large mistral-small mixtral-8x7b gpt-4-turbo dreamgen-opus-v1 mistral-ai openai hugging-face benchmarking model-merging fine-tuning reinforcement-learning model-training tokenization model-optimization ai-assisted-decompilation performance cost-efficiency deception roleplay deep-speed dpo timotheeee1 cogbuji plasmator jsarnecki maldevide spottyluck mrjackspade
Mistral announced Mistral Large, a new language model achieving 81.2% accuracy on MMLU, trailing GPT-4 Turbo by about 5 percentage points on benchmarks. The community reception has been mixed, with skepticism about open sourcing and claims that Mistral Small outperforms the open Mixtral 8x7B. Discussions in the TheBloke Discord highlighted performance and cost-efficiency comparisons between Mistral Large and GPT-4 Turbo, technical challenges with DeepSpeed and DPOTrainer for training, advances in AI deception for roleplay characters using DreamGen Opus V1, and complexities in model merging using linear interpolation and PEFT methods. Enthusiasm for AI-assisted decompilation was also expressed, emphasizing the use of open-source projects for training data.
RWKV "Eagle" v5: Your move, Mamba
rwkv-v5 mistral-7b miqu-1-70b mistral-medium llama-2 mistral-instruct-v0.2 mistral-tuna llama-2-13b kunoichi-dpo-v2-7b gpt-4 eleutherai mistral-ai hugging-face llamaindex nous-research rwkv lmsys fine-tuning multilinguality rotary-position-embedding model-optimization model-performance quantization speed-optimization prompt-engineering model-benchmarking reinforcement-learning andrej-karpathy
RWKV v5 Eagle was released with better-than-mistral-7b evaluation results, trading some English performance for multilingual capabilities. The mysterious miqu-1-70b model sparked debate about its origins, possibly a leak or distillation of Mistral Medium or a fine-tuned Llama 2. Discussions highlighted fine-tuning techniques, including the effectiveness of 1,000 high-quality prompts over larger mixed-quality datasets, and tools like Deepspeed, Axolotl, and QLoRA. The Nous Research AI community emphasized the impact of Rotary Position Embedding (RoPE) theta settings on LLM extrapolation, improving models like Mistral Instruct v0.2. Speed improvements in Mistral Tuna kernels reduced token processing costs, enhancing efficiency. The launch of Eagle 7B with 7.52B parameters showcased strong multilingual performance, surpassing other 7B class models.
1/12/2024: Anthropic coins Sleeper Agents
nous-mixtral 120b anthropic openai nous-research hugging-face reinforcement-learning fine-tuning backdoors model-security adversarial-training chain-of-thought model-merging dataset-release security-vs-convenience leo-gao andrej-karpathy
Anthropic released a new paper exploring the persistence of deceptive alignment and backdoors in models through stages of training including supervised fine-tuning and reinforcement learning safety training. The study found that safety training and adversarial training did not eliminate backdoors, which can cause models to write insecure code or exhibit hidden behaviors triggered by specific prompts. Notable AI figures like leo gao and andrej-karpathy praised the work, highlighting its implications for future model security and the risks of sleeper agent LLMs. Additionally, the Nous Research AI Discord community discussed topics such as the trade-off between security and convenience, the Hulk Dataset 0.1 for LLM fine-tuning, curiosity about a 120B model and Nous Mixtral, debates on LLM leaderboard legitimacy, and the rise of Frankenmerge techniques for model merging and capacity enhancement.