All tags
Topic: "scaling-laws"
AI Engineer World's Fair Talks Day 1
gemini-2.5 gemma claude-code mistral cursor anthropic openai aie google-deepmind meta-ai-fair agent-based-architecture open-source model-memorization scaling-laws quantization mixture-of-experts language-model-memorization model-generalization langgraph model-architecture
Mistral launched a new Code project, and Cursor released version 1.0. Anthropic improved Claude Code plans, while ChatGPT announced expanded connections. The day was dominated by AIE keynotes and tracks including GraphRAG, RecSys, and Tiny Teams. On Reddit, Google open-sourced the DeepSearch stack for building AI agents with Gemini 2.5 and LangGraph, enabling flexible agent architectures and integration with local LLMs like Gemma. A new Meta paper analyzed language model memorization, showing GPT-style transformers store about 3.5–4 bits/parameter and exploring the transition from memorization to generalization, with implications for Mixture-of-Experts models and quantization effects.
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.
OpenAI adopts MCP
gemini-2.5-pro gemini-1.5-pro gemini-2.0-flash qwen-2.5-omni-7b deepseek-v3-0324 deepseek-r1 openai google-deepmind alibaba togethercompute model-benchmarking multimodality reasoning scaling-laws model-quantization synthetic-data model-performance context-windows speech-recognition translation audio-processing video-processing swyx
OpenAI announced support for MCP, a significant technical update. Google's Gemini 2.5 Pro leads benchmarks with top scores in MMLU-Pro (86%), GPQA Diamond (83%), and AIME 2024 (88%), featuring a 1 million token context window and multimodal inputs. Alibaba's Qwen 2.5 Omni 7B was released as a fully multimodal, interactive, open-source model with a novel "thinker-talker" architecture supporting voice and video chat. DeepSeek V3-0324 outperforms its predecessor on multiple benchmarks. Research on reasoning features in large language models using sparse autoencoders was highlighted, alongside a study on scaling laws of synthetic data showing performance plateaus near 300B tokens. Discussions also covered the fastest output speeds of Gemini models and concerns about over-reliance on benchmarks for intelligence measurement. Swyx will curate the Data Council AI Engineering Track in April.
Every 7 Months: The Moore's Law for Agent Autonomy
claude-3-7-sonnet llama-4 phi-4-multimodal gpt-2 cosmos-transfer1 gr00t-n1-2b orpheus-3b metr nvidia hugging-face canopy-labs meta-ai-fair microsoft agent-autonomy task-completion multimodality text-to-speech robotics foundation-models model-release scaling-laws fine-tuning zero-shot-learning latency reach_vb akhaliq drjimfan scaling01
METR published a paper measuring AI agent autonomy progress, showing it has doubled every 7 months since 2019 (GPT-2). They introduced a new metric, the 50%-task-completion time horizon, where models like Claude 3.7 Sonnet achieve 50% success in about 50 minutes. Projections estimate 1 day autonomy by 2028 and 1 month autonomy by late 2029. Meanwhile, Nvidia released Cosmos-Transfer1 for conditional world generation and GR00T-N1-2B, an open foundation model for humanoid robot reasoning with 2B parameters. Canopy Labs introduced Orpheus 3B, a high-quality text-to-speech model with zero-shot voice cloning and low latency. Meta reportedly delayed Llama-4 release due to performance issues. Microsoft launched Phi-4-multimodal.
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.
s1: Simple test-time scaling (and Kyutai Hibiki)
qwen-2.5-32b gemini-2.0-flash smollm2 granite-vision-3.1-2b google-deepmind qwen gemini hugging-face ibm deepseek reasoning fine-tuning scaling-laws open-source-models data-centric-training vision multilingual-models language-model-reasoning niklas-muennighoff
"Wait" is all you need introduces a novel reasoning model finetuned from Qwen 2.5 32B using just 1000 questions with reasoning traces distilled from Gemini 2.0 Flash Thinking, enabling controllable test-time compute by appending "Wait" to extend reasoning. Lead author Niklas Muennighoff, known for work on Bloom, StarCoder, and BIG-bench, highlights this method's efficiency and its reproduction of the famous o1 scaling chart. Additionally, Kyutai Moshi's Hibiki project demonstrates impressive offline French-English live translation on iPhone. Recent AI model releases include DeepSeek R1 and R3 open source models, potentially marking a major open-source milestone, Hugging Face's SmolLM2 emphasizing data-centric training for small LMs, and IBM's Granite-Vision-3.1-2B, a small vision-language model with strong performance. Key research papers spotlight LIMO for minimal demonstration reasoning achieving high accuracy on AIME and MATH benchmarks, and Token-Assisted Reasoning mixing latent and text tokens to improve language model reasoning.
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.
Stripe lets Agents spend money with StripeAgentToolkit
gpt-4o gemini-exp-1114 stripe openai anthropic meta-ai-fair ai-computer-interfaces agentic-ai model-overfitting benchmarks scaling-laws agi chain-of-thought image-captioning dialogue-systems memory-efficient-fine-tuning diffusion-models mixture-of-experts adaptive-decoding creativity-optimization factuality-optimization pair-programming document-parsing retrieval-augmented-generation abacaj francois-fleuret lmarena_ai goodside jxmnop jaseweston stevenheidel
Stripe has pioneered an AI SDK specifically designed for agents that handle payments, integrating with models like gpt-4o to enable financial transactions and token-based charging. The AI developer tooling trend emphasizes better "AI-Computer Interfaces" for improved agent reliability, with tools like E2B and the
llms.txt documentation trend gaining traction, notably adopted by Anthropic. In AI model news, Gemini-Exp-1114 topped the Vision Leaderboard and improved in Math Arena, while discussions continue around model overfitting and the limits of scaling laws for AGI. OpenAI released a ChatGPT desktop app for macOS with integrations for VS Code, Xcode, and Terminal, enhancing developer workflows and pair programming. Anthropic introduced a prompt improver using chain-of-thought reasoning, and Meta AI shared top research from EMNLP2024 on image captioning, dialogue systems, and memory-efficient fine-tuning. Highlights from ICLR 2025 include diffusion-based illumination harmonization, open mixture-of-experts language models, and hyperbolic vision-language models. A new adaptive decoding method optimizes creativity and factuality per token. Tools like LlamaParse and RAGformation were also introduced for document parsing and retrieval-augmented generation. BitNet was a lie?
qwen-2.5-coder-32b-instruct gpt-4o llama-3 sambanova alibaba hugging-face quantization scaling-laws model-efficiency fine-tuning model-performance code-generation open-source unit-testing ci-cd tanishq-kumar tim-dettmers
Scaling laws for quantization have been modified by a group led by Chris Re, analyzing over 465 pretraining runs and finding benefits plateau at FP6 precision. Lead author Tanishq Kumar highlights that longer training and more data increase sensitivity to quantization, explaining challenges with models like Llama-3. Tim Dettmers, author of QLoRA, warns that the era of efficiency gains from low-precision quantization is ending, signaling a shift from scaling to optimizing existing resources. Additionally, Alibaba announced Qwen 2.5-Coder-32B-Instruct, which matches or surpasses GPT-4o on coding benchmarks, and open-source initiatives like DeepEval for LLM testing are gaining traction.
Tencent's Hunyuan-Large claims to beat DeepSeek-V2 and Llama3-405B with LESS Data
claude-3.5-haiku llama-3-1 llama-3-2 mlx-lm tencent anthropic meta-ai-fair togethercompute llamaindex mixture-of-experts synthetic-data model-scaling model-architecture model-optimization kv-cache-quantization react fine-tuning scaling-laws model-efficiency model-deployment multimodality
Tencent released a notable >300B parameter MoE model pretrained on 7T tokens, including 1.5T synthetic data generated via Evol-Instruct. The model introduces novel techniques like "recycle routing" and expert-specific learning rates, alongside a compute-efficient scaling law for MoE active parameters. However, its custom license restricts use in the EU and by companies with over 100M MAU, and it avoids China-sensitive queries. Meanwhile, Anthropic launched Claude 3.5 Haiku, now available on multiple platforms, praised for intelligence and speed but criticized for a 10x price increase. Meta opened Llama AI to the U.S. defense sector, and a Llama Impact Hackathon offers a $15K prize for projects using Llama 3.1 & 3.2 Vision. LlamaIndex released a React chat UI component with Tailwind CSS and LLM backend integrations. The MLX LM model advances text generation speed and efficiency with KV cache quantization.
o1: OpenAI's new general reasoning models
o1 o1-preview o1-mini gpt-4o llama openai nvidia test-time-reasoning reasoning-tokens token-limit competitive-programming benchmarking scaling-laws ai-chip-competition inference training model-performance jason-wei jim-fan
OpenAI has released the o1 model family, including o1-preview and o1-mini, focusing on test-time reasoning with extended output token limits over 30k tokens. The models show strong performance, ranking in the 89th percentile on competitive programming, excelling in USA Math Olympiad qualifiers, and surpassing PhD-level accuracy on physics, biology, and chemistry benchmarks. Notably, o1-mini performs impressively despite its smaller size compared to gpt-4o. The release highlights new scaling laws for test-time compute that scale loglinearly. Additionally, Nvidia is reportedly losing AI chip market share to startups, with a shift in developer preference from CUDA to llama models for web development, though Nvidia remains dominant in training. This news reflects significant advances in reasoning-focused models and shifts in AI hardware competition.
DataComp-LM: the best open-data 7B model/benchmark/dataset
mistral-nemo-12b gpt-4o-mini deepseek-v2-0628 mistral-7b llama-3 gemma-2 qwen-2 datacomp hugging-face openai nvidia mistral-ai deepseek dataset-design scaling-laws model-benchmarking model-performance fine-tuning multilinguality function-calling context-windows open-source-models model-optimization cost-efficiency benchmarking sam-altman guillaume-lample philschmid miramurati
DataComp team released a competitive 7B open data language model trained on only 2.5T tokens from the massive DCLM-POOL dataset of 240 trillion tokens, showing superior scaling trends compared to FineWeb. OpenAI launched GPT-4o mini, a cost-effective model with 82% MMLU and performance near GPT-4-Turbo, aimed at developers for broad applications. NVIDIA and Mistral jointly released the Mistral NeMo 12B model featuring a 128k token context window, FP8 checkpoint, multilingual support, and Apache 2.0 licensing. DeepSeek announced DeepSeek-V2-0628 as the top open-source model on the LMSYS Chatbot Arena leaderboard with strong rankings in coding, math, and hard prompts. This news highlights advances in dataset design, model efficiency, and open-source contributions in the AI community.
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.
Skyfall
gemini-1.5-pro gemini-1.5-flash yi-1.5 kosmos-2.5 paligemma falcon-2 deepseek-v2 hunyuan-dit gemini-1.5 gemini-1.5-flash yi-1.5 google-deepmind yi-ai microsoft hugging-face langchain maven multimodality mixture-of-experts transformer model-optimization long-context model-performance model-inference fine-tuning local-ai scaling-laws causal-models hallucination-detection model-distillation model-efficiency hamel-husain dan-becker clement-delangue philschmid osanseviero arankomatsuzaki jason-wei rohanpaul_ai
Between 5/17 and 5/20/2024, key AI updates include Google DeepMind's Gemini 1.5 Pro and Flash models, featuring sparse multimodal MoE architecture with up to 10M context and a dense Transformer decoder that is 3x faster and 10x cheaper. Yi AI released Yi-1.5 models with extended context windows of 32K and 16K tokens. Other notable releases include Kosmos 2.5 (Microsoft), PaliGemma (Google), Falcon 2, DeepSeek v2 lite, and HunyuanDiT diffusion model. Research highlights feature an Observational Scaling Laws paper predicting model performance across families, a Layer-Condensed KV Cache technique boosting inference throughput by up to 26×, and the SUPRA method converting LLMs into RNNs for reduced compute costs. Hugging Face expanded local AI capabilities enabling on-device AI without cloud dependency. LangChain updated its v0.2 release with improved documentation. The community also welcomed a new LLM Finetuning Discord by Hamel Husain and Dan Becker for Maven course users. "Hugging Face is profitable, or close to profitable," enabling $10 million in free shared GPUs for developers.
Mixture of Depths: Dynamically allocating compute in transformer-based language models
octopus-v2 deepmind transformer-efficiency dynamic-compute-allocation mixture-of-experts mixture-of-depths top-k-routing algorithmic-reasoning visual-autoregressive-modeling on-device-models function-calling scaling-laws piotrpadlewski
DeepMind introduces the Mixture-of-Depths (MoD) technique, dynamically allocating FLOPs across transformer layers to optimize compute usage, achieving over 50% faster forward passes without training impact. MoD selectively processes tokens using top-k routing, improving efficiency and potentially enabling faster ultra-long context handling. The method can combine with Mixture-of-Experts (MoE) for decoupled routing of queries, keys, and values. Reddit discussions highlight concerns about LLM hype overshadowing other AI tech, improvements in transformer efficiency, a new Think-and-Execute framework boosting algorithmic reasoning by 10-20%, and Visual Autoregressive modeling (VAR) surpassing diffusion models in image quality and speed. On-device model Octopus v2 outperforms GPT-4 in function calling accuracy and latency.