All tags
Topic: "on-device-ai"
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
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.
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
vllm deepseek-v3 llamaindex openai deepseek qdrant twilio llamaindex elevenlabs training-efficiency parallelism cpu-offloading gradient-descent mixture-of-experts fp8-precision memory-optimization ai-voice-assistants coding-assistants document-processing version-control learning-rate-schedules federated-learning agentic-systems multi-agent-systems deliberative-alignment chain-of-thought on-device-ai multimodality francois-fleuret daniel-hanchen aaron-defazio fchollet elad-gil wojciech-zaremba richard-socher
ChatGPT, Sora, and the OpenAI API experienced a >5 hour outage but are now restored. Updates to vLLM enable DeepSeek-V3 to run with enhanced parallelism and CPU offloading, improving model deployment flexibility. Discussions on gradient descent in top-k routing MoE and adoption of FP8 precision focus on training efficiency and memory optimization. AIDE, an AI voice medical assistant by Team Therasync, leverages Qdrant, OpenAI, and Twilio. DeepSeek-Engineer offers AI-powered coding assistance with structured outputs. LlamaIndex integrates LlamaCloud and ElevenLabs for large-scale document processing and voice interaction. Insights on version control with ghstack and advocacy for linear decay learning rate schedules highlight best practices in AI development. Experts predict smaller, tighter models, true multimodal models, and on-device AI in 2025. Proposals for planetary-scale federated learning and community AGI moonshots emphasize future AI directions. Discussions on agentic systems, multi-agent workflows, and deliberative alignment through chain of thought reasoning underscore AI safety and alignment efforts.
not much happened today
smollm2 llama-3-2 stable-diffusion-3.5 claude-3.5-sonnet gemini openai anthropic google meta-ai-fair suno-ai perplexity-ai on-device-ai model-performance robotics multimodality ai-regulation model-releases natural-language-processing prompt-engineering agentic-ai ai-application model-optimization sam-altman akhaliq arav-srinivas labenz loubnabenallal1 alexalbert fchollet stasbekman svpino rohanpaul_ai hamelhusain
ChatGPT Search was launched by Sam Altman, who called it his favorite feature since ChatGPT's original launch, doubling his usage. Comparisons were made between ChatGPT Search and Perplexity with improvements noted in Perplexity's web navigation. Google introduced a "Grounding" feature in the Gemini API & AI Studio enabling Gemini models to access real-time web information. Despite Gemini's leaderboard performance, developer adoption lags behind OpenAI and Anthropic. SmolLM2, a new small, powerful on-device language model, outperforms Meta's Llama 3.2 1B. A Claude desktop app was released for Mac and Windows. Meta AI announced robotics advancements including Meta Sparsh, Meta Digit 360, and Meta Digit Plexus. Stable Diffusion 3.5 Medium, a 2B parameter model with a permissive license, was released. Insights on AGI development suggest initial inferiority but rapid improvement. Anthropic advocates for early targeted AI regulation. Discussions on ML specialization predict training will concentrate among few companies, while inference becomes commoditized. New AI tools include Suno AI Personas for music creation, PromptQL for natural language querying over data, and Agent S for desktop task automation. Humor was shared about Python environment upgrades.
DocETL: Agentic Query Rewriting and Evaluation for Complex Document Processing
bitnet-b1.58 llama-3.1-nemotron-70b-instruct gpt-4o claude-3.5-sonnet uc-berkeley deepmind openai microsoft nvidia archetype-ai boston-dynamics toyota-research google adobe openai mistral tesla meta-ai-fair model-optimization on-device-ai fine-tuning large-corpus-processing gpu-acceleration frameworks model-benchmarking rohanpaul_ai adcock_brett david-patterson
UC Berkeley's EPIC lab introduces innovative LLM data operators with projects like LOTUS and DocETL, focusing on effective programming and computation over large data corpora. This approach contrasts GPU-rich big labs like Deepmind and OpenAI with GPU-poor compound AI systems. Microsoft open-sourced BitNet b1.58, a 1-bit ternary parameter LLM enabling 4-20x faster training and on-device inference at human reading speeds. Nvidia released Llama-3.1-Nemotron-70B-Instruct, a fine-tuned open-source model outperforming GPT-4o and Claude-3.5-sonnet. These developments highlight advances in model-optimization, on-device-ai, and fine-tuning.
not much happened today
claudette llama-3-1 yi-lightning gpt-4o claude-3.5-sonnet answer-ai tencent notebooklm motherduck perplexity dropbox openai meta-ai-fair yi-ai zyphra-ai anthropic langchain openai synthetic-data fine-tuning sql audio-processing on-device-ai dataset-release transformer llm-reasoning ai-safety code-generation ai-pricing ai-job-market fchollet aravsrinivas svpino swyx
Answer.ai launched fastdata, a synthetic data generation library using
claudette and Tencent's Billion Persona paper. NotebookLM became customizable, and Motherduck introduced notable LLMs in SQL implementations. Perplexity and Dropbox announced competitors to Glean. OpenAI unveiled audio chat completions priced at 24 cents per minute. Meta AI released Llama 3.1, powering Lenovo AI Now's on-device agent. Yi-Lightning model ranked #6 globally, surpassing GPT-4o. Zyphra AI released the large Zyda-2 dataset with 5 trillion tokens. François Chollet clarified transformer architecture as set-processing, not sequence-processing. Research suggests memorization aids LLM reasoning. Anthropic updated its Responsible Scaling Policy for AI safety. Tools like Perplexity Finance, Open Canvas by LangChain, and AlphaCodium code generation tool were highlighted. Approximately $500 million was raised for AI agent startups, with ongoing discussions on AI's job market impact. Combining prompt caching with the Batches API can yield a 95% discount on Claude 3.5 Sonnet tokens. not much happened today
llama-3-2 llama-3 molmo meta-ai-fair google-deepmind hugging-face on-device-ai multimodality chip-design retrieval-augmented-generation rag benchmarking reliability ai-regulation free-speech pytorch-optimization demis-hassabis clementdelangue svpino awnihannun osanseviero omarsar0 sarahookr ylecun
Meta released Llama 3.2, including lightweight 1B and 3B models for on-device AI with capabilities like summarization and retrieval-augmented generation. Molmo, a new multimodal model, was introduced with a large dense captioning dataset. Google DeepMind announced AlphaChip, an AI-driven chip design method improving TPU and CPU designs. Hugging Face surpassed 1 million free public models, highlighting the value of smaller specialized models. Discussions covered challenges in scaling RAG applications, the future of on-device AI running ChatGPT-level models, reliability issues in larger LLMs, and new Elo benchmarking accepted at NeurIPS 2024. AI ethics and regulation topics included free speech responsibilities and California's SB-1047 bill potentially affecting open-source AI. "AlphaChip transformed computer chip design," and "ChatGPT-level AI on mobile devices predicted within a year."
not much happened today + AINews Podcast?
superforecaster-ai llama-3 reflection-70b glean sambanova cerebras stanford google apple hugging-face lmsys prompt-engineering research-ideas inference-speed retrieval-augmented-generation evaluation-methods visual-intelligence on-device-ai model-performance benchmarking novelty-detection danhendrycks benjamin-clavie bclavie bindureddy swyx borismpower corbtt drjimfan clementdelangue rohanpaul_ai
Glean doubled its valuation again. Dan Hendrycks' Superforecaster AI generates plausible election forecasts with interesting prompt engineering. A Stanford study found that LLM-generated research ideas are statistically more novel than those by expert humans. SambaNova announced faster inference for llama-3 models, surpassing Cerebras. Benjamin Clavie gave a notable talk on retrieval-augmented generation techniques. Strawberry is reported to launch in two weeks. Google Illuminate offers AI-generated podcast discussions about papers and books. Apple unveiled new AI features in iOS 18, including visual intelligence and improved Siri, with on-device and cloud processing for camera-based event additions. The Reflection 70B model sparked controversy over performance claims. Experts highlighted the unreliability of traditional benchmarks like MMLU and HumanEval, recommending alternative evaluation methods such as LMSys Chatbot Arena and Hugging Face's open-sourced Lighteval suite. The AI research community continues to explore AI's role in generating novel research ideas and improving benchmarking.
not much happened today
sam-2 gemini-1.5-pro chatgpt midjourney-v6.1 meta-ai-fair google-deepmind scale-ai apple canva hugging-face object-segmentation quantization web-development-framework adversarial-robustness on-device-ai open-source robotics voice vision jeremyphoward demis-hassabis ylecun maartengrootendorst jimfan
Meta released SAM 2, a unified model for real-time object segmentation with a new dataset 4.5x larger and 53x more annotated than previous ones. FastHTML, a new Python web framework by Jeremy Howard, enables easy creation and deployment of interactive web apps. Scale AI launched the SEAL Leaderboard on adversarial robustness, topped by Gemini 1.5 Pro from Google DeepMind. Apple published a technical report on their Intelligence Foundation Language Models for on-device and server use. Yann LeCun emphasized the importance of open source AI in an article co-authored with Martin Casado and Ion Stoica. Maarten Grootendorst's "Visual Guide to Quantization" on efficient LLM inference went viral. ChatGPT started rolling out advanced voice and vision-enabled modes to select users. Leonardo AI was acquired by Canva. Jim Fan shared insights on Project Groot augmenting human demonstration data for robotics. Midjourney v6.1 was released.
Test-Time Training, MobileLLM, Lilian Weng on Hallucination (Plus: Turbopuffer)
llama-2-7b codegeex4-all-9b mamba facebook-research meta-ai-fair tsinghua-university hallucination-detection anti-hallucination-methods on-device-ai model-architecture rnn long-context-modeling model-scaling expressive-hidden-states code-generation lilian-weng yann-lecun
Lilian Weng released a comprehensive literature review on hallucination detection and anti-hallucination methods including techniques like FactualityPrompt, SelfCheckGPT, and WebGPT. Facebook AI Research (FAIR) published MobileLLM, a sub-billion parameter on-device language model architecture achieving performance comparable to llama-2-7b with innovations like thin and deep models and shared weights. A new RNN-based LLM architecture with expressive hidden states was introduced, replacing attention mechanisms and scaling better than Mamba and Transformer models for long-context modeling. Additionally, Tsinghua University open sourced CodeGeeX4-ALL-9B, a multilingual code generation model excelling in code assistance.
Is this... OpenQ*?
deepseek-coder-v2 llama-3-8b nemotron-4-340b stable-diffusion-3-medium deepseek_ai anthropic runwayml openai apple nvidia stability-ai luma-labs reward-tampering test-time-search mathematical-reasoning process-supervision fine-tuning on-device-ai video-generation cost-efficiency context-length coding image-understanding multimodality adcock_brett clementdelangue svpino
DeepSeekCoder V2 promises GPT4T-beating performance at a fraction of the cost. Anthropic released new research on reward tampering. Runway launched their Sora response and Gen-3 Alpha video generation model. A series of papers explore "test-time" search techniques improving mathematical reasoning with models like LLaMa-3 8B. Apple announced Apple Intelligence with smarter Siri and image/document understanding, partnered with OpenAI to integrate ChatGPT into iOS 18, and released 20 new CoreML models with LoRA fine-tuning for specialization. NVIDIA released Nemotron-4 340B, an open model matching GPT-4 performance. DeepSeek-Coder-V2 excels in coding and math with 338 programming languages and 128K context length. Stability AI released Stable Diffusion 3 Medium weights. Luma Labs launched Dream Machine for 5-second video generation from text and images.
Francois Chollet launches $1m ARC Prize
gpt-4 chatgpt openai apple togethercompute benchmarking agi pattern-recognition skill-acquisition privacy on-device-ai mixed-precision-quantization mixture-of-experts multimodality agentic-ai francois-chollet karpathy svpino philschmid clementdelangue sama gdb miramurati kevin-weil sarah-friar
François Chollet critiques current paths to AGI, emphasizing the importance of benchmarks that resist saturation and focus on skill acquisition and open-ended problem solving. The ARC-AGI puzzles exemplify "easy for humans, hard for AI" challenges to measure progress toward AGI. Meanwhile, Apple announces integration of ChatGPT into iOS, iPadOS, and macOS through a partnership with OpenAI, enabling AI-powered features like document summarization and photo analysis with privacy-preserving measures. Discussions highlight Apple's focus on deep AI integration and on-device models optimized with techniques like mixed-precision quantization, though some skepticism remains about their AI capabilities compared to GPT-4. Additionally, Together Compute introduces a Mixture of Agents approach achieving strong performance on AlpacaEval 2.0.
Talaria: Apple's new MLOps Superweapon
gemma mixtral phi dbrx apple google mistral-ai microsoft mosaic quantization on-device-ai adapter-models model-optimization model-latency lossless-quantization low-bit-palletization token-generation model-benchmarking human-evaluation craig-federighi andrej-karpathy
Apple Intelligence introduces a small (~3B parameters) on-device model and a larger server model running on Apple Silicon with Private Cloud Compute, aiming to surpass Google Gemma, Mistral Mixtral, Microsoft Phi, and Mosaic DBRX. The on-device model features a novel lossless quantization strategy using mixed 2-bit and 4-bit LoRA adapters averaging 3.5 bits-per-weight, enabling dynamic adapter hot-swapping and efficient memory management. Apple credits the Talaria tool for optimizing quantization and model latency, achieving about 0.6 ms time-to-first-token latency and 30 tokens per second generation rate on iPhone 15 Pro. Apple focuses on an "adapter for everything" strategy with initial deployment on SiriKit and App Intents. Performance benchmarks rely on human graders, emphasizing consumer-level adequacy over academic dominance. The Apple ML blog also mentions an Xcode code-focused model and a diffusion model for Genmoji.
1 TRILLION token context, real time, on device?
gemini-1.5-pro gemini-1.5 cartesia mistral-ai scale-ai state-space-models voice-models multimodality model-performance on-device-ai long-context evaluation-leaderboards learning-rate-optimization scientific-publishing research-vs-engineering yann-lecun elon-musk
Cartesia, a startup specializing in state space models (SSMs), launched a low latency voice model outperforming transformer-based models with 20% lower perplexity, 2x lower word error, and 1 point higher NISQA quality. This breakthrough highlights the potential for models that can continuously process and reason over massive streams of multimodal data (text, audio, video) with a trillion token context window on-device. The news also covers recent AI developments including Mistral's Codestral weights release, Schedule Free optimizers paper release, and Scale AI's new elo-style eval leaderboards. Additionally, a debate between yann-lecun and elon-musk on the importance of publishing AI research versus engineering achievements was noted. The Gemini 1.5 Pro/Advanced models were mentioned for their strong performance.
Not much happened today
command-r-35b goliath-120 miqu-120 llama-3-8b tensorrt-llm llama-cpp gpt2-chat gpt-4-turbo llama-3 deepmind-alphazero anthropic openai perplexity-ai amazon apple microsoft deepmind creative-writing context-windows benchmarking model-performance self-learning function-calling retrieval-augmented-generation ai-assistants on-device-ai ai-lobbying copyright-infringement code-reasoning image-generation
Anthropic released a team plan and iOS app about 4 months after OpenAI. The Command-R 35B model excels at creative writing, outperforming larger models like Goliath-120 and Miqu-120. The Llama-3 8B model now supports a 1 million token context window, improving long-context understanding with minimal training on a single 8xA800 GPU machine. TensorRT-LLM benchmarks show it is 30-70% faster than llama.cpp on consumer hardware. A benchmark suggests GPT2-Chat may have better reasoning than GPT-4-Turbo, though results are debated. Demos include a self-learning Llama-3 voice agent running locally on Jetson Orin and a Self-Learning Large Action Model (LAM). Amazon CodeWhisperer was renamed to Q Developer, expanding its generative AI assistant capabilities. Apple plans an AI-enabled Safari browser with an on-device LLM in iOS 18 and macOS 15. Big Tech dominates AI lobbying in Washington, while major U.S. newspapers sued OpenAI and Microsoft for copyright infringement. DeepMind's AlphaZero became the greatest chess player in 9 hours, and their Naturalized Execution Tuning (NExT) method improves LLM code reasoning by 14-26%. Stable Diffusion is used for diverse image generation applications.
Snowflake Arctic: Fully Open 10B+128x4B Dense-MoE Hybrid LLM
snowflake-arctic phi-3 llama-3-70b llama-3 stable-diffusion-3 sd3-turbo gpt-3.5-turbo snowflake databricks deepseek deepspeed nvidia stable-diffusion adobe apple llamaindex lmsys openai mixture-of-experts curriculum-learning model-release image-generation video-upscaling quantization inference-speed benchmarking model-comparison open-source on-device-ai
Snowflake Arctic is a notable new foundation language model released under Apache 2.0, claiming superiority over Databricks in data warehouse AI applications and adopting a mixture-of-experts architecture inspired by DeepSeekMOE and DeepSpeedMOE. The model employs a 3-stage curriculum training strategy similar to the recent Phi-3 paper. In AI image and video generation, Nvidia introduced the Align Your Steps technique improving image quality at low step counts, while Stable Diffusion 3 and SD3 Turbo models were compared for prompt understanding and image quality. Adobe launched an AI video upscaling project enhancing blurry videos to HD, though with some high-resolution artifacts. Apple released open-source on-device language models with code and training logs, diverging from typical weight-only releases. The Llama-3-70b model ties for first place on the LMSYS leaderboard for English queries, and Phi-3 (4B params) outperforms GPT-3.5 Turbo in the banana logic benchmark. Fast inference and quantization of Llama 3 models were demonstrated on MacBook devices.
... and welcome AI Twitter!
mistral-large google-gemini google openai apple stripe ai-ethics multilinguality on-device-ai convolutional-neural-networks synthetic-data financial-transaction-systems corporate-culture humor margaret-mitchell john-carmack guillaume-lample sundar-pichai delip-rao santiago-l-valdarrama alex-wang yann-lecun pieter-levels francois-chollet dheliat
The AI Twitter discourse from 2/27-28/2024 covers a broad spectrum including ethical considerations highlighted by Margaret Mitchell around Google Gemini's launch, and John Carmack's insights on evolving coding skills in the AI era. Guillaume Lample announced the release of the Mistral Large multilingual model. Discussions also touched on potential leadership changes at Google involving Sundar Pichai, and OpenAI's possible entry into the synthetic data market as noted by Delip Rao. Technological advancements include Yann LeCun's commentary on running LLMs on mobile devices and Alex Wang's praise for the Apple Vision Pro. Financial platform issues were raised by Pieter Levels regarding Stripe's payment policies. The cultural dynamics within big tech were discussed by François Chollet and Dhéliat. The lighter side of AI was represented by memes and humor from Pieter Levels and AISafetyMemes. This summary reflects the fast-evolving AI landscape blending technical innovation, corporate strategy, ethics, and community culture.