Search the archive

SIMPLIFYING TEXT WITH LLMS

Google researchers have used LLMs to simplify complex text without losing critical details, improving user understanding while preserving accuracy and nuance.

CODE RETRIEVAL USING LORA

Researchers introduce a LoRA-based fine-tuning method for code search that reduces trainable parameters below 2% while improving retrieval accuracy by up to 9.1% for Code2Code tasks.

A NOVEL INITIALIZATION METHOD

A new initialization technique, IDInit, ensures stable convergence in deep neural networks by maintaining identity transitions in both main and sub-stem layers.

PYTORCH ROLE IN THE AI STACK

PyTorch has evolved from a research framework to a foundational platform that powers generative AI. The PyTorch Foundation was expanded to integrate complementary projects and support scalable AI development.

META'S CHATGPT COMPETITOR SHOWS HOW YOUR FRIENDS USE AI

Meta's new AI app introduces a Discover feed that adds a social media twist by allowing users to share their interactions with the AI. This feature aims to demystify AI by enabling comments, likes, and remixing of shared AI posts. The app, which replaces the View app for Meta Ray-Ban glasses, uses a Meta-tuned Llama 4 model and provides advanced voice interaction in select regions.

MASTERCARD GIVES AI AGENTS THE ABILITY TO SHOP ON YOUR BEHALF

Mastercard's AI program streamlines e-commerce searches, reducing time and friction. Consumers retain purchase control as AI agents can't finalize transactions.

OPENAI AND FDA DISCUSS AI IN DRUG EVALUATION

OpenAI and the FDA held meetings to explore how AI could expedite drug approvals, signaling a shift toward modernizing regulatory science with machine learning.

META LAUNCHES AI DEFENDERS PROGRAM TO PROTECT LLAMA MODELS

Meta introduced the AI Defenders Program to help developers detect and prevent misuse of LLaMA models.

STARTUPS LAUNCH PRODUCTS TO CATCH PEOPLE USING AI CHEATING APP CLUELY

Cluely claims its hidden window enables undetectable cheating, sparking competition from startups like Validia and Proctaroo, which say they can identify Cluely users.

GEMINI 2.5 PRO PREVIEW

Google has released a preview of Gemini 2.5 Pro, which has a stronger performance in building web apps, code transformation, and multimodal reasoning.

PINTEREST NEW VISUAL SEARCH

Pinterest has upgraded its image-based search with new tools that help users narrow results and explore styles, launching first in the women's fashion category across select regions.

COGNITION KEVIN-32B

KEVIN-32B is a model leveraging reinforcement learning for multi-turn code generation that outperforms existing models in developing CUDA kernels. It boosts kernel performance and correctness by optimizing intermediate feedback and using efficient reward distribution. Multi-turn training enhances the model's problem-solving capability over single-turn approaches, particularly in more complex tasks.

HOW TO TRAIN AN AI MODEL WITHOUT FALLING INTO GDPR PITFALLS?

AI model developers can ensure GDPR compliance during the development phase by using anonymous datasets or applying pseudonymization techniques. If anonymization isn't feasible, they can enhance data security and support data subjects' rights. Transparency measures, such as public communication about data usage, are also recommended.

CAN AI-GENERATED IMAGES TRULY BE CONSIDERED ART?

Christiane Paul, curator at the Whitney Museum, highlights AI's role in digital art, contrasting historical programs like Harold Cohen's AARON with today's AI models. She stresses the importance of collaboration with engineers and the challenges of preserving digital art due to evolving technology. Paul argues that AI-generated visuals need a conceptual framework to be considered true art.

QUANTIZATION WITH AUTOROUND

AutoRound is a post-training quantization method that boosts low-bit model accuracy while preserving performance and efficiency.

LLMS FOR TIME SERIES: A SURVEY

This survey explores how cross-modality techniques adapt large language models for time series analytics, focusing on data alignment, fusion, and downstream task performance across multiple domains.

CHAIN OF DRAFT FOR EFFICIENT REASONING

Chain of Draft is a concise reasoning strategy that significantly reduces token usage while matching or exceeding Chain-of-Thought accuracy across complex tasks.

SYNTHETIC DATA QA FRAMEWORK

This evaluation toolkit provides standardized metrics to assess the quality and privacy of synthetic data across various data types using distributional and embedding-based measures.

DDT: DECOUPLED DIFFUSION TRANSFORMER

Encoder/Decoder implementation of a Transformer with a Diffusion model as the decoder. It seems to work reasonably well on imagenet generation.

NVIDIA RADIO EMBEDDING MODELS (HUGGING FACE HUB)

Nvidia has a suite of text and image embedding models that match SigLIP in many cases.

AI IN HEAVY MACHINERY AND FARMING

John Deere, a leading manufacturer of agricultural and construction equipment, used AI to boost farming efficiency with precision tools like See & Spray, which drastically cut chemical use by identifying individual weeds at high speeds.

ANTHROPIC UNLOCKS WEB SEARCH FOR ALL PAID CLAUDE PLANS

Anthropic has enabled web search for all paid Claude plans, adding real-time lookups and source citations.

LITTLE LANGUAGE LESSONS USES GENERATIVE AI TO MAKE PRACTICING LANGUAGES MORE PERSONAL

Google's Little Language Lessons leverage models for real-world language learning through experiments like Tiny Lesson, Slang Hang, and Word Cam.

Apply in <2 min.

OPENAI NEW STRUCTURE

OpenAI's for-profit arm will become a Public Benefit Corporation while remaining under nonprofit oversight, aiming to align profit incentives with its long-term mission.

ANYSPHERE, WHICH MAKES CURSOR, HAS REPORTEDLY RAISED $900M AT $9B VALUATION

Anysphere raised $900 million, led by Thrive Capital, valuing the AI coding tool maker at $9 billion. Existing investors like a16z and Accel joined the round. Competitors like Windsurf are also seeing heightened investor interest, with OpenAI reportedly eyeing acquisition opportunities.

LISTEN LABS PPT GENERATOR

Listen Labs launched an AI presentation generator that automatically creates slides by analyzing audio and video content. This tool enables users to quickly repurpose existing media into slide formats.

RL TRAINING FOR MATH REASONING

Perplexity is using RL techniques like PPO and GRPO to boost math reasoning capabilities in models. This post shares insights into training pipelines and performance lessons.

FROM RETRIEVAL TO REASONING FOR AI AGENTS

Jure Leskovec's keynote at The Web Conference introduces STaRK, AvaTaR, and CollabLLM, new frameworks pushing AI agents to reason, collaborate, and test hypotheses using knowledge graphs, tools, and multi-turn optimization.

ONE-LAYER TRANSFORMERS

A theoretical and empirical analysis revealed how one-layer transformers can learn even pair and parity check tasks, showing distinct training dynamics and benefits from Chain-of-Thought integration.

LEGAL RAG EVALUATION TOOLKIT

LRAGE offers an open-source framework for evaluating legal-domain LLMs in retrieval-augmented generation tasks, with integrated datasets and tools to support research.

VOICE STAR TEXT TO SPEECH

Another of many new speech generation models. This one has strong zero shot voice cloning capabilities and hints of expressiveness control.

11X CEO HASAN SUKKAR STEPS DOWN

Hasan Sukkar is stepping down as CEO of AI startup 11x, with CTO Prabhav Jain taking over. Sukkar will remain as non-executive chairman, focusing on strategic direction and growth. Despite past controversies, 11x has raised over $70 million from investors, who continue to express strong support for the company's future.

2025 PYTORCH DOCATHON

PyTorch has kicked off a community-driven event to improve its documentation and make it easier for users to transition from research to production.

HUAWEI AIMS TO TAKE ON NVIDIA'S H100 WITH NEW AI CHIP

Huawei is developing the Ascend 910D AI chip to compete with Nvidia's H100 series. The company seeks Chinese partners for testing amid US export restrictions on AI chips. Success could mitigate the impact of US controls on China's AI market.

GOOGLE'S AI FUNDING IN ASIA

Google.org announced funding for 49 organizations across Asia-Pacific and committed an additional $12M to support AI education and workforce training.

ANTHROPIC LAUNCHES AI FOR SCIENCE PROGRAM

Anthropic is providing free API credits to researchers in biology and life sciences to accelerate discovery through its new AI for Science initiative.

PYTORCH DAY FRANCE HIGHLIGHTS

PyTorch Day France brings together open source AI contributors for a one-day event with talks, real-world case studies, and community networking opportunities.

APPLE AND ANTHROPIC'S INTERNAL AI CODING PLATFORM

OPENAI ADDRESSES CHATGPT SYCOPHANCY

After reports of overly agreeable behavior in GPT-4o, OpenAI has detailed plans for more transparent updates, a new opt-in alpha phase, and stricter behavior safety reviews.

PHI 4 REASONING 100% ON PRIVATE GRADUATE LINEAR ALGEBRA EXAM

The new synthetically trained reasoning model from Microsoft shows strong local math and code performance, even in the face of poor internal world knowledge.

GPT-4O IS AN ABSURD SYCOPHANT

OpenAI's release of GPT-4o resulted in extreme sycophancy and other issues, raising concerns about its detachment from the OpenAI Model Spec against flattery. This was likely a misstep in pursuit of user engagement, exacerbated by A/B testing favoring positive responses. OpenAI CEO Sam Altman acknowledged the problem and has promised fixes. The situation highlights the risks of models being optimized in ways that might betray user trust.

MCP IS UNNECESSARY

MCP's main functions are advertising and calling tools, similar to OpenAPI, but it offers a more streamlined approach. While both can achieve similar outcomes, MCP's simplicity and reduced size make it attractive. MCP's perceived necessity is more sociological than technological.

EMPOWERING LLMS WITH DEEPRESEARCH ABILITY

WebThinker is a deep research framework fully powered by large reasoning models (LRMs). It enables LRMs to autonomously search, deeply explore web pages, and draft research reports.

AMPLIFY INITIATIVE FOR LOCAL AI DATA

Google Research has introduced a global data collection initiative co-created with local experts to improve AI relevance in underserved regions.

EFFICIENT FEDERATED UNLEARNING

FUSED introduces sparse unlearning adapters to selectively remove knowledge in federated learning, making unlearning reversible and cost-efficient.

ATTENTION DISTILLATION FOR DIFFUSION-BASED IMAGE STYLIZATION

This method enhances image generation by leveraging self-attention features from pretrained diffusion models, introducing an attention distillation loss to optimize stylization and accelerate synthesis.

ALIBABA UNVEILS QWEN 3, A FAMILY OF 'HYBRID' AI REASONING MODELS

Alibaba has released Qwen 3, a family of AI models they claim can rival top models from Google and OpenAI. These open-licensed models feature advanced reasoning capabilities and a mixture of expert architecture. Supporting 119 languages and trained on 36 trillion tokens, Qwen-3-235B-A22B notably outperformed OpenAI's o3-mini in several benchmarks. It's not yet publicly available.