Search the archive

start

DEEPSEEK'S NATIVELY MULTIMODAL MODEL

DeepSeek has trained and released a 1.3b powerful model that can operate on interleaved text and images to both generate and understand.

META'S LINGUA

Meta has created a simple and research friendly codebase that can reproduce Llama 2 7B in 24 hours.

ALIGNING TEXT-TO-IMAGE MODELS WITH HUMAN VALUES

LiVO (Lightweight Value Optimization) is a novel method for aligning Text-to-Image models with human values.

LONG CONTEXT CONTEST (KAGGLE COMPETITION)

The Gemini team has earmarked $100k for the best uses of their long context model capabilities.

ALPHACODIUM OUTPERFORMS DIRECT PROMPTING OF O1 MODEL

OpenAI's o1 model, which exhibits System 1.5 thinking, shows enhanced reasoning capabilities compared to previous LLMs but still falls short of full System 2 deliberate problem-solving. AlphaCodium significantly improves o1's coding performance by providing a structured framework that guides reasoning and iterative refinement, achieving better accuracy on Codeforces benchmarks. While o1 and AlphaCodium together show promise in advancing AI toward deeper reasoning, much work remains to integrate full System 2 thinking in AI models.

DUOLINGO CEO LUIS VON AHN THINKS AI HAS A LOT TO TEACH US

Duolingo's CEO, Luis von Ahn, discusses leveraging AI and gamification to enhance language learning through features like chat conversations with AI avatars and video game-like adventures generated by AI. The company recently introduced Duolingo Max, a higher-priced subscription plan that offers AI-driven conversation practice, as AI-generated content costs less and speeds up development. Despite AI's limitations in engagement, Duolingo focuses on keeping users motivated by balancing learning efficacy with gamified, entertaining experiences.

playbook. Download it from Domino Data Lab

EASILY HACKABLE VISION LANGUAGE MODEL

A simple and performant VLM implementation in pure PyTorch.

AMAZON ADS' AI GENERATOR TOOL CAN NOW CREATE AUDIO ADS

Amazon is expanding its use of generative AI for creating audio ads through its self-service ad platform.

GOOGLE SHOPPING IS GETTING A ‘FOR YOU' FEED OF PRODUCTS

Google Shopping has introduced AI-driven personalized feeds, which recommends products based on user searches and YouTube views.

🙋♀️ Request your free invite here

CUSTOMIZE AUDIO OVERVIEWS WITH GOOGLE'S NOTEBOOKLM

NotebookLM now allows users to tailor their Audio Overview experience, giving more control over the AI hosts' focus and expertise. Businesses can apply for the new NotebookLM Business pilot program, which offers enhanced tools for professional use.

MIT RESEARCHERS COMBINE NEXT-TOKEN PREDICTION AND VIDEO DIFFUSION FOR ROBOTICS

A new method combines next-token prediction and video diffusion, enabling neural networks to clean corrupted data while predicting future steps. This approach enhances a model's ability to generate high-quality video, create flexible plans for robots, and assist AI agents in navigating digital environments.

NVIDIA JUST DROPPED A NEW AI MODEL THAT CRUSHES OPENAI'S GPT-4

Nvidia's new AI model, Llama-3.1-Nemotron-70B-Instruct, surpasses performance benchmarks set by OpenAI and Anthropic. This model marks Nvidia's strategic expansion from hardware to sophisticated AI software. By advancing both text-only and multimodal models, Nvidia challenges existing industry leaders and reshapes the AI competitive landscape.

META RELEASING “MOVIE GEN BENCH”

“Movie Gen Bench” is an evaluation benchmark for assessing both video (Video Bench) and audio (Audio Bench) performance. It consists of 1,003 prompts that cover all the different testing aspects/concepts.

META DECISION TRANSFORMER

Meta Decision Transformer (Meta-DT) is designed to improve generalization in reinforcement learning by combining transformer-based sequential modeling with robust task representation learning.

ENHANCING TEXT-TO-IMAGE ALIGNMENT FOR LONG TEXTS

LongAlign improves text-to-image (T2I) diffusion models' ability to handle long text inputs by introducing segment-level encoding and decomposed preference optimization.

AUTO-REGRESSIVE MODEL FOR IMAGE GENERATION

DiGIT is an auto-regressive generative model that predicts tokens in a latent space using self-supervised learning. It is a discrete tokenizer that improves image generation on ImageNet by clustering hidden states from DINOv2.

EFFICIENT FEDERATED LEARNING

The FedPart method addresses the layer mismatch issue in federated learning by restricting model updates to specific layers during each training round.

DISTRIBUTED TRAINING GUIDE

This is a comprehensive guide on best practices for distributed training, diagnosing errors, and fully utilizing all resources available.

WHERE'S THE GENERATIVE AI ROI? START WITH THE SUPPLY CHAIN

Generative AI is transforming supply chain operations by efficiently processing unstructured documents, leading to significant time and cost savings. Flexport, a supply chain tech company, has successfully integrated AI to automate and streamline document handling, reducing processing time by 80%. This application of AI demonstrates its value in practical, revenue-generating tasks rather than more speculative AI advancements.

FOLLOW THE QUIET VOICES TO FIND AI'S TRUTHS

AI discourse is polarized, with opposing pro-AI and anti-AI factions dominating the conversation. Synthetical thinkers, who approach AI with nuanced perspectives beyond binary viewpoints, are currently missing from this debate. These truth-seekers may eventually return when AI becomes less controversial and more integrated into everyday life.

INVISIBLE TEXT THAT AI CHATBOTS UNDERSTAND AND HUMANS CAN'T? YEP, IT'S A THING

Recent research highlights vulnerabilities in LLMs like prompt injection and ASCII smuggling, which allow attackers to hide instructions or data that are invisible to users but readable by the models. Invisible Unicode characters can be used in AI chatbots like Claude and Copilot to embed and extract malicious payloads and sensitive data. This poses a significant security risk. While some AI platforms like OpenAI and Microsoft have implemented mitigations, the broader issue of AI processing data imperceptible to humans continues to challenge security protocols.

2034. TLDR readers get early access.

GOOGLE SUPERCHARGES SHOPPING TAB WITH AI AND PERSONALIZED RECOMMENDATION FEED

Google is enhancing its Shopping tab with AI to tailor product searches and offer detailed AI briefs.

ADOBE'S PROJECT SUPER SONIC USES AI TO GENERATE SOUND EFFECTS FOR YOUR VIDEOS

Adobe's Project Super Sonic uses text-to-audio, object recognition, and vocal input to generate audio effects for video projects.

WHITE HOUSE CONSIDERS EXPANDING NVIDIA AND AMD'S AI CHIP EXPORT LIMITS TO ADDITIONAL COUNTRIES

The Biden administration is considering restricting AI chip sales from Nvidia and AMD to Persian Gulf countries for national security reasons.

MISTRAL INTRODUCES ADVANCED EDGE MODELS

Mistral introduced two new cutting-edge models for on-device and edge computing - Ministral 3B and Ministral 8B - on the first anniversary of Mistral 7B. These models excel in knowledge, reasoning, and efficiency for sub-10B tasks. They support up to 128k context length and feature faster inference.

META AI'S NEW TRACKING MODEL

CoTracker is a new tracking model that addresses the performance gap between synthetic and real video data by using semi-supervised training.

FORMER PALANTIR CISO DANE STUCKEY JOINS OPENAI TO LEAD SECURITY

OpenAI hired former Palantir CISO Dane Stuckey to strengthen its security operations. The company seeks to enhance its AI security infrastructure and has been working with the Pentagon on cybersecurity-related projects. OpenAI aims to build secure AI systems while exploring government partnerships including with the U.S. Department of Defense.

CONTINUOUS-TIME CONSISTENCY MODELS

This study proposes a unified framework to improve training stability in continuous-time consistency models to significantly advance generative model performance.

ENHANCED POINT CLOUD REGISTRATION

Researchers introduce a new consistency-aware spot-guided Transformer for more efficient and accurate point cloud registration.

BRAIN SIGNALS DECODING

DARNet is a novel model for auditory attention detection (AAD) that enhances the decoding of brain signals like EEG by incorporating spatiotemporal and dual attention mechanisms.

TURN IDEAS INTO FLASK APPS WITH NO-CODE

Ditto is a no-code tool that transforms natural language descriptions into multi-file Flask applications. It automates much of the coding process using a simple LLM loop, helping users create functional web apps from their ideas.

TEXT-TO-SPEECH SYSTEM

F5-TTS is a non-autoregressive, zero-shot text-to-speech system that uses a flow-matching mel spectrogram generator and a diffusion transformer. Built on the MLX framework, F5 improves on previous systems like E2 TTS by using ConvNeXT v2 blocks for better text alignment, producing high-quality speech in around 11 seconds on modern hardware.

LONG-CONTEXT HANDLING FOR LLMS

DuoAttention is a framework designed to reduce memory and latency in long-context LLMs by selectively applying full KV caching to only essential attention heads.

CAN AI REALLY COMPETE WITH HUMAN DATA SCIENTISTS? OPENAI'S NEW BENCHMARK PUTS IT TO THE TEST

OpenAI's MLE-bench tests AI capabilities in machine learning engineering through 75 Kaggle-style competitions. The benchmark highlights AI progress but also its limitations in creative problem-solving compared to humans. Open-sourcing MLE-bench could standardize AI evaluation and impact the role of data scientists.

SB 1047: OUR SIDE OF THE STORY

California's proposed SB 1047 aimed to mandate AI companies to mitigate existential risks but was vetoed by Governor Newsom, who cited inadequate regulation of smaller, potentially dangerous models. The bill had significant backing from AI safety advocates like Dan Hendrycks and garnered support from notable figures such as Elon Musk, yet faced opposition from major AI corporations including OpenAI and Google. Newsom's veto has stirred debates among AI professionals about future regulatory frameworks and potential alliances with broader political groups to address AI safety comprehensively.

OVERVIEW OF STRONG HUMAN INTELLIGENCE AMPLIFICATION METHODS

AI advancements hinge on creating humans with enhanced cognitive capabilities to navigate AGI development effectively. Methods include brain emulation, genomic approaches, adult brain gene editing, and brain-brain interfaces, each with unique challenges and risks. These efforts aim to solve philosophical problems, amplify human intelligence to unprecedented levels, and address potential AGI threats.

under-resourced languages such as Hindi. Watch the API tour

LLMS DON'T DO FORMAL REASONING - AND THAT IS A HUGE PROBLEM

A study by Apple challenges the effectiveness of LLMs, showing that they rely on pattern matching rather than formal reasoning, leading to fragile results.

WHY CHATGPT MAKER OPENAI IS AT FIGHT WITH OPEN AI

OpenAI is in a legal battle with Guy Ravine's company, Open AI, over the rights to the name "Open AI" and the original open-source AI vision.

ADOBE'S AI VIDEO MODEL IS HERE, AND IT'S ALREADY INSIDE PREMIERE PRO

Adobe has launched new generative AI video tools within Premiere Pro, including Generative Extend for extending footage and improving small details.

months to weeks_. Register for the live session.