Search the archive

FAST VS. SLOW THINKING IN LANGUAGE MODELS

This project explores how different "thinking" approaches — fast (concise) versus slow (detailed, like chain-of-thought reasoning) — impact layer-wise gradients and stability in LLMs.

3D OBJECT DETECTION USING ONLY 2D IMAGES

ImOV3D is a framework that advances open-vocabulary 3D object detection (OV-3Det) by leveraging 2D images to overcome the scarcity of 3D annotations.

REALISTIC MOTION IN TEXT-TO-VIDEO GENERATION

DEMO is a framework that separates text and conditioning into content and motion components. By using distinct encoders and conditioning for static content and dynamic motion, DEMO better interprets and generates motion from text prompts.

PROJECT SID

Project Sid shows civilizational progress, specialization, governance, and the invention and spread of memes and religion. These were made possible by Altera's new cognitive architecture: PIANO.

WE'RE FORKING FLUTTER. THIS IS WHY

Google's corporate focus shift to AI has deprioritized Flutter's desktop platforms, creating a labor shortage for this once rapidly expanding UI toolkit. In response, a fork named Flock is being developed to include essential bug fixes and features the Flutter team can't tackle, aiming to accelerate Flutter's growth through community involvement. Flock plans to improve contribution processes and streamline PR reviews, bridging the gap in support and development pace left by the main Flutter team.

USING REINFORCEMENT LEARNING AND $4.80 OF GPU TIME TO FIND THE BEST HN POST EVER (RLHF PART 1)

This article discusses using reinforcement learning from human feedback (RLHF) to develop a reward model that predicts upvote counts for Hacker News stories. Leveraging a comprehensive dataset and $4.80 of GPU time, the reward model was trained on features like titles, authors, and content to optimize for post quality, with the ultimate goal of utilizing RLHF techniques to enhance the creation of high-value content. The resulting model, while not perfect, can successfully identify underappreciated stories and predict potential front-page hits.

GOOGLE'S AI SEARCH SUMMARIES ARE ROLLING OUT TO OVER 100 MORE COUNTRIES

Google's AI Overviews are expanding to over 100 countries, offering AI-generated search summaries in multiple languages.

ELON MUSK FINALLY ADMITS TESLA'S HW3 MIGHT NOT SUPPORT FULL SELF-DRIVING

Elon Musk admitted Tesla's HW3 might not support full self-driving and that HW4 is significantly more capable.

NVIDIA ETHERNET NETWORKING ACCELERATES WORLD'S LARGEST AI SUPERCOMPUTER, BUILT BY XAI

xAI's Colossus is powered by NVIDIA's Spectrum-X Ethernet networking platform.

Get the world's best Whisper inference with Baseten

CHATGPT SEARCH

ChatGPT can now search the web in a much better way than before. You can get fast, timely answers with links to relevant web sources, which you would have previously needed to go to a search engine for. This blends the benefits of a natural language interface with the value of up-to-date sports scores, news, stock quotes, and more.

META FAIR ROBOTICS IMPROVEMENTS

This work includes several artifacts such as vision based tactical sensing, new hardware based touch sensors, and some interesting strategic partnerships in robotics.

ELON MUSK'S XAI ADDS IMAGE UNDERSTANDING CAPABILITIES TO GROK

xAI has enhanced its Grok AI model with image-understanding capabilities for paid users on X. The update allows users to upload images and query the AI about them. Future updates may include document understanding capabilities.

A BENCHMARK FOR REINFORCEMENT LEARNING

ODRL is the first standardized benchmark for evaluating reinforcement learning methods in environments with mismatched dynamics.

TEXT-TO-IMAGE MODEL TO GENERATE MEMES

Researchers have developed a novel adapter method for text-to-image models, allowing them to handle intricate tasks like meme video generation while maintaining the base model's generalization strengths.

ANOMALY CLASSIFICATION IN INDUSTRY

AnomalyNCD is a multi-class anomaly classification framework designed to complement traditional anomaly detection methods in industrial settings.

MRT5: DYNAMIC TOKEN MERGING FOR EFFICIENT BYTE-LEVEL LANGUAGE MODELS

Byte Level language models are a step towards a token free future. However, the sequence length issue is a huge challenge. Merge tokens dynamically to improve the number of tokens in context.

ACCELERATED BART

A new GPU-enabled implementation of Bayesian Additive Regression Trees (BART) that dramatically boosts its processing speed, making it up to 200 times faster than traditional CPU-based versions.

HUGE NEW DIFFUSERS RELEASE (GITHUB PAGE)

New pipelines (Flux, Stable Audio, Kolors, CogVideoX, Latte, and more), methods (FreeNoise and SparseCtrl), and refactors for the Hugging Face Diffusers package.

HOW TO TRAIN LLM AS A JUDGE TO DRIVE BUSINESS VALUE

LLM As a Judge is a way to use an existing language model to rank and score natural language. This post outlines how to best do this to process or evaluate data.

EXPLORE CULTURE WITH GOOGLE'S VOICE AI MODELS

Users can explore culture in unique ways using Google's voice AI models. Projects include Talking Tours with AI-guided virtual tours, Mice in the Museum for art narration, and Lip Sync, where animated lips discuss cultural topics. These fun tools offer fresh perspectives on art and design.

OPENAI CFO SAYS 75% OF ITS REVENUE COMES FROM PAYING CONSUMERS

OpenAI's revenue is primarily from consumer subscriptions, with 75% generated from paid ChatGPT users. Despite efforts to attract corporate customers, the consumer side remains strong with ChatGPT reaching 250 million weekly active users. OpenAI recently secured $6.6 billion in funding to support AI advancement and is advocating for significant data center expansion.

thoughts. Write a review

HELLO PATIENT (PRODUCT)

Hello Patient has launched out of stealth with a $6.3m seed round led by 8VC.

GOOGLE PLANS TO ANNOUNCE ITS NEXT GEMINI MODEL SOON

OpenAI is planning to debut a successor to GPT-4 and Google plans to launch Gemini 2.0 in December.

META IS REPORTEDLY DEVELOPING A SEARCH ENGINE FOR ITS CHATBOT

Meta is developing a search engine for its chatbot to reduce reliance on Google and Microsoft.

A MYSTERIOUS NEW IMAGE GENERATION MODEL HAS APPEARED

red_panda, a new image generation model, outperforms models from Midjourney, Black Forest Labs, and OpenAI on the Artificial Analysis benchmark.

PUSHING THE FRONTIERS OF AUDIO GENERATION

DeepMind has talked a little bit more about the audio generation models used to power NotebookLM.

OPENAI'S NEW HALLUCINATION BENCHMARK

OpenAI has released the SimpleQA benchmark, which measures models' abilities around simple factual questions.

EVALUATING FEATURE STEERING: A CASE STUDY IN MITIGATING SOCIAL BIASES

This study explores using feature steering in AI models to interpretably modify outputs. It reveals a "steering sweet spot", where changes do not degrade capabilities. The study results show steering can alter social bias in targeted domains but also brings unexpected off-target effects. Further research is required to refine feature steering for safer, more reliable outcomes in AI models.

THUNDERKITTENS 2

Thunder Kittens is a framework for writing extremely performant GPU Kernels. It is built on the idea that GPUs actually want to operate on small 16x16 tiles of data. In turn, the useability is quite high, and 40% faster kernels only take a few hundred lines of code.

REALISTIC MOTION RETARGETING

MeshRet has introduced a novel approach for improving motion retargeting for 3D characters that focuses on preserving body geometry interactions from the start.

BETTER GENERATION WITH SELF-GUIDANCE SAMPLING

Researchers have enhanced Masked Generative Models (MGMs) with a new self-guidance sampling method, improving their image generation quality while maintaining diversity.

SPEEDING UP TRANSFORMERS WITH TOKEN MERGING

This project introduces PiToMe, an algorithm that compresses Vision Transformers by progressively merging tokens after each layer. This method reduces the number of tokens processed.

3D RECONSTRUCTION WITHOUT POSE DATA

PF3plat tackles the challenge of pose-free 3D reconstruction and novel view synthesis from RGB images, eliminating the need for extra data.

A BENCHMARK FOR EVALUATING DATA CURATION METHODS

SELECT is the first large-scale benchmark for comparing data curation strategies in image classification. ImageNet++ is a new dataset that extends ImageNet-1K with five new training-data shifts, each assembled using different curation techniques.

FINE-TUNING LLMS TO 1.58BIT: EXTREME QUANTIZATION MADE EASY

BitNet, developed by Microsoft Research, introduces a transformer architecture that reduces LLM computational and memory requirements by using ternary precision (-1, 0, 1) equating to 1.58 bits per parameter. Models are required to be trained from scratch. BitNet can also fine-tune existing models to this low-precision format, maintaining strong performance on downstream tasks. This approach significantly reduces energy consumption and improves inference speed using specialized kernels for efficient matrix multiplication.

HOW WE SAVED HUNDREDS OF ENGINEERING HOURS BY WRITING TESTS WITH LLMS

Assembled uses LLMs to accelerate and improve software testing, enabling test generation in minutes instead of hours. This approach increases engineering productivity, saving time and shifting focus to feature development. LLMs generate comprehensive and accurate tests that maintain code quality and development velocity.

25% OF SMARTPHONE OWNERS DON'T WANT AI AS APPLE INTELLIGENCE DEBUTS

A CNET survey revealed that only 18% of smartphone users are motivated by AI features to upgrade their devices, with privacy and cost being significant concerns. Major manufacturers like Apple, Google, and Samsung are integrating more AI capabilities in their phones, yet many users prioritize battery life and storage over AI functions. AI subscriptions are set to become common, but nearly half of users are unwilling to pay for these features.

Rime's story.

GOOGLE PREPS ‘JARVIS' AI AGENT THAT WORKS IN CHROME

Google's Project Jarvis, powered by Gemini 2.0, aims to automate web-based tasks in Chrome by using AI agents capable of reasoning and planning.

OPENAI'S WHISPER TRANSCRIPTION TOOL HAS HALLUCINATION ISSUES, RESEARCHERS SAY

Concerns have risen over OpenAI's Whisper introducing hallucinations in transcriptions, even in medical contexts.

FORERUNNER K2 HUMANOID ROBOT CAN CARRY 33 LB IN EACH DEXTEROUS HAND

Kepler has unveiled the Forerunner K2 humanoid robot, which has advanced AI, improved hardware, and enhanced vision and navigation systems for better real-time interaction.

📑 Get the report to see the full findings

AN INTERVIEW WITH SALESFORCE CEO MARC BENIOFF ABOUT AI ABUNDANCE

Salesforce's CEO Marc Benioff recently discussed the company's new AI initiative, Agentforce, highlighting its potential to revolutionize enterprise applications and customer interactions. Benioff criticized Microsoft's Copilot, positioning Salesforce's approach as more integrated and effective due to its robust platform and data architecture. In the interview, he emphasized the importance of AI-driven "agentic" layers, which aim to enhance customer service and operational efficiency across industries.