All tags
Person: "lateinteraction"
Microsoft Build: MAI-Thinking-1 and MAI Family models, Surface RTX Spark Dev Box, and OpenClaw in Windows
mai-thinking-1 mai-code-1-flash holo-3.1 qwen-35b sonnet-4.6 claude-code codex microsoft openrouter fal baseten hcompany_ai teksedge nous-research teknim cognition windsurf perplexity-ai mixture-of-experts context-windows benchmarking reinforcement-learning prompt-optimization agentic-ai local-inference model-family-expansion model-reporting agent-native-devices software-development model-optimization hybrid-inference desktop-agents model-quantization mustafasuleyman eliebakouch hannahajishirzi asadovsky bj2rn lateinteraction lakshyaaagrawal theturingpost kimmonismus yusuf_i_mehdi pierceboggan lukehoban nielsrogge russelljkaplan
Microsoft introduced MAI-Thinking-1, a 35B parameter MoE model with 256K context, achieving 97% on AIME 2025 and outperforming Sonnet 4.6 in human preference tests. The broader 7-model MAI family spans reasoning, code, image, speech, and voice, with third-party availability on OpenRouter, fal, and Baseten. The detailed 109-page technical report revealed insights on scaling, MFU, RL/post-training, and data curation, highlighting no third-party distillation and advanced prompt optimization techniques. Microsoft emphasized agent-native devices and local inference with projects like Project Solara / Scout and the Surface RTX Spark Dev Box, alongside software innovations such as the Copilot desktop app and MAI-Code-1-Flash integration. Meanwhile, local-first computer-use agents like Holo 3.1 (Qwen-based, 0.8B to 35B parameters) support laptops and small workstations with optimized formats and strong benchmark results. Desktop shells for agents, including Hermes Desktop, Devin Desktop, and agent-neutral approaches compatible with Devin, Claude Code, and Codex, are proliferating, with hybrid local/cloud execution becoming the default architecture as seen in Perplexity Computer's hybrid agentic inference.
not much happened today
mai-thinking-1 mai-image-2.5 mai-code-1-flash gemma-4-12b microsoft google vllm-project ollama llama-cpp model-training reinforcement-learning model-architecture multimodality model-deployment model-efficiency fine-tuning on-device-ai eliebakouch nrehiew_ mustafasuleyman minjiyoon90 lateinteraction harold_matmul googlegemma googleaidevs mtschannen armandjoulin osanseviero
Microsoft released the detailed technical report for MAI-Thinking-1, a generalist reasoning model trained without third-party distillation, achieving 97% on AIME 2025 and outperforming Sonnet 4.6 in human preference tests. The report was praised for transparency, revealing no synthetic data use, a unique scaling ladder recipe, and detailed training data composition including 50% code and 17.5% STEM. Microsoft also introduced Frontier Tuning for workflow-specific model adaptation, claiming efficiency gains up to 10× and GPT-5.4-level quality in Excel tasks, alongside new models like MAI-Image-2.5 and MAI-Code-1-Flash. Meanwhile, Google launched Gemma 4 12B, an Apache 2.0 multimodal model with an innovative encoder-free architecture designed for on-device use with 16GB VRAM, collapsing vision and audio encoders into the LLM backbone, receiving positive community feedback and immediate tooling support.
ChatGPT starts testing ads on free tier + new $8/mo Go plan in the US
chatgpt-go codex openai ollama ads monetization memory agent-orchestration human-in-the-loop cli-tools context-length workflow-optimization sama sam_altman fidjissimo scaling01 tomwarren embirico adamdotdev ollama thsottiaux lateinteraction dbreunig
OpenAI announced the ChatGPT Go tier at $8/month with ads testing in the US free tier, emphasizing that ads will not influence responses and will be clearly labeled. The update includes memory improvements and a "very fast Codex" feature teased by Sam Altman. The Codex CLI ecosystem now supports open-weight models with improved context length. Discussions highlight the importance of human-in-the-loop for reliability in agent orchestration and file interface improvements over traditional retrieval-augmented generation.
not much happened today
gpt-5-pro gemini-2.5 vllm deepseek-v3.1 openai google-deepmind microsoft epoch-ai-research togethercompute nvidia mila reasoning reinforcement-learning inference speculative-decoding sparse-attention kv-cache-management throughput-optimization compute-efficiency tokenization epochairesearch yitayml _philschmid jiqizhixin cvenhoff00 neelnanda5 lateinteraction mgoin_ blackhc teortaxestex
FrontierMath Tier 4 results show GPT-5 Pro narrowly outperforming Gemini 2.5 Deep Think in reasoning accuracy, with concerns about problem leakage clarified by Epoch AI Research. Mila and Microsoft propose Markovian Thinking to improve reasoning efficiency, enabling models to reason over 24K tokens with less compute. New research suggests base models inherently contain reasoning mechanisms, with "thinking models" learning to invoke them effectively. In systems, NVIDIA Blackwell combined with vLLM wins InferenceMAX with significant throughput gains, while Together AI's ATLAS adaptive speculative decoding achieves 4× speed improvements and reduces RL training time by over 60%. SparseServe introduces dynamic sparse attention with KV tiering, drastically improving throughput and latency in GPU memory management.
Chinese Models Launch - MiniMax-M1, Hailuo 2 "Kangaroo", Moonshot Kimi-Dev-72B
minimax-m1 hailuo-02 kimi-dev-72b deepseek-r1 ale-agent minimax-ai moonshot-ai deepseek bytedance anthropic langchain columbia-university sakana-ai openai microsoft multi-agent-systems attention-mechanisms coding optimization prompt-injection model-performance video-generation model-training task-automation jerryjliu0 hwchase17 omarsar0 gallabytes lateinteraction karpathy
MiniMax AI launched MiniMax-M1, a 456 billion parameter open weights LLM with a 1 million token input and 80k token output using efficient "lightning attention" and a GRPO variant called CISPO. MiniMax AI also announced Hailuo 02 (0616), a video model similar to ByteDance's Seedance. Moonshot AI released Kimi-Dev-72B, a coding model outperforming DeepSeek R1 on SWEBench Verified. Discussions on multi-agent system design from Anthropic and LangChain highlighted improvements in task completion and challenges like prompt injection attacks, as demonstrated by Karpathy and Columbia University research. Sakana AI introduced ALE-Agent, a coding agent that ranked 21st in the AtCoder Heuristic Competition solving NP-hard optimization problems. There is unverified news about an acquisition involving OpenAI, Microsoft, and Windsurf.
Mary Meeker is so back: BOND Capital AI Trends report
qwen-3-8b anthropic hugging-face deepseek attention-mechanisms inference arithmetic-intensity transformers model-optimization interpretability model-quantization training tri_dao fleetwood___ teortaxestex awnihannun lateinteraction neelnanda5 eliebakouch _akhaliq
Mary Meeker returns with a comprehensive 340-slide report on the state of AI, highlighting accelerating tech cycles, compute growth, and comparisons of ChatGPT to early Google and other iconic tech products. The report also covers enterprise traction and valuation of major AI companies. On Twitter, @tri_dao discusses an "ideal" inference architecture featuring attention variants like GTA, GLA, and DeepSeek MLA with high arithmetic intensity (~256), improving efficiency and model quality. Other highlights include the release of 4-bit DWQ of DSR1 Qwen3 8B on Hugging Face, AnthropicAI's open-source interpretability tools for LLMs, and discussions on transformer training and abstractions by various researchers.
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.
Cohere's Command A claims #3 open model spot (after DeepSeek and Gemma)
command-a mistral-ai-small-3.1 smoldocling qwen-2.5-vl cohere mistral-ai hugging-face context-windows multilinguality multimodality fine-tuning benchmarking ocr model-performance model-releases model-optimization aidangomez sophiamyang mervenoyann aidan_mclau reach_vb lateinteraction
Cohere's Command A model has solidified its position on the LMArena leaderboard, featuring an open-weight 111B parameter model with an unusually long 256K context window and competitive pricing. Mistral AI released the lightweight, multilingual, and multimodal Mistral AI Small 3.1 model, optimized for single RTX 4090 or Mac 32GB RAM setups, with strong performance on instruct and multimodal benchmarks. The new OCR model SmolDocling offers fast document reading with low VRAM usage, outperforming larger models like Qwen2.5VL. Discussions highlight the importance of system-level improvements over raw LLM advancements, and MCBench is recommended as a superior AI benchmark for evaluating model capabilities across code, aesthetics, and awareness.
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.