Search the archive

AMAZON NOVA REEL 2-MINUTE VIDEOS

The updated Nova Reel model now supports multi-shot videos up to 2 minutes long, offering more creative control and efficiency for generative video content.

INVESTING IN KREA

Andreessen Horowitz has announced an investment into Krea, a platform combining AI models to help creatives generate and edit visual content. Used by teams at Pixar and Samsung, Krea now has over 20M users. The company plans to launch an enterprise-grade product later this year.

ONE-MINUTE VIDEO GENERATION WITH TEST-TIME TRAINING

This work introduces Test-Time Training (TTT) layers, featuring highly expressive hidden states, to overcome the limitations of standard Transformers and alternatives like Mamba in generating long, coherent videos. By integrating TTT layers into a pre-trained model, they demonstrate the generation of coherent one-minute videos from text storyboards that significantly outperform baseline methods in telling complex stories according to human evaluations. Tom and Jerry cartoons are the test bed.

SCALING SPEECH LANGUAGE MODELS

This study finds that speech-language models initialized from text models using interleaved training scales more efficiently than speech-only models.

BENCHMARKING WEB AGENT CAPABILITIES

Online-Mind2Web is a realistic evaluation benchmark for autonomous web agents. Current models perform worse than expected due to flawed benchmarks.

and register for free

VARGPT

Dual understanding and synthesis model in a single autoregressive framework. This means that one model can caption and generate images.

FLEXTOK

OpenSource work from Apple. It follows up the company's paper from a few weeks ago that allows images to be tokenized in a variable number of tokens. This means that more complex images can use more tokens.

ADAPTIVE GRADIENT CLIPPING

ZClip uses EMA-based gradient norm statistics to adaptively suppress outlier gradients, preventing loss spikes and improving training stability without static thresholds.

GOOGLE'S AI HIGHLIGHTS IN MARCH

Google recaps major March updates including Gemini 2.5 Pro, expanded AI Overviews, AI Mode, and other feature rollouts across its products.

GENIES AI AVATARS

Genies has launched a no-code platform that enables anyone to create intelligent AI avatars with unique looks, personalities, and behaviors for use in customizable gaming experiences called Parties. These AI avatars are powered by large language models, behavioral AI, and real-time animation tech, allowing dynamic interactions, gameplay, and emotional responses.

OPENAI PLANS O3 AND O4-MINI RELEASE BEFORE GPT-5, ALTMAN SAYS

OpenAI will release intermediate models o3 and o4-mini before GPT-5. This staggered release is due to technical complexities with GPT-5, allowing improvement and better handling of expected demand. The shift comes as competitors like Google's Gemini 2.5 Pro advance in the AI field. OpenAI has just secured a $40 billion funding round.

OPENAI'S O3 MODEL MIGHT BE COSTLIER TO RUN THAN ORIGINALLY ESTIMATED

Arc Prize Foundation has revised computing cost estimates for OpenAI's o3, suggesting it's much higher than initially thought at potentially $30,000 per ARC-AGI problem.

BRINGING MULTIMODAL SEARCH TO AI MODE

Google is expanding its AI Mode feature to millions of U.S. Labs users, enhancing it with multimodal capabilities.

GOKU VIDEO MODEL

Goku from ByteDance is a flow based video generation model of 2B and 8B parameters with 160M image and 36M video pairs.

at 2pm ET (11am PT).

Can't make it? Register anyway and we'll send you the recording!

Register Now!

LLAMA 4

Meta has unveiled Llama 4 Scout and Maverick, two 17B parameter multimodal models that offer state-of-the-art performance on major benchmarks, along with Llama 4 Behemoth, a 288B model still in training that surpasses GPT-4.5 in STEM tasks.

MIDJOURNEY V7

Midjourney has released its new image generation model, V7 alpha. It brings smarter text interpretation, better image coherence, and introduces Draft Mode for fast, low-cost iterations with optional voice commands and personalization.

CYBERATTACKS BY AI AGENTS ARE COMING

AI agents are emerging as potent tools in cybersecurity capable of executing complex attacks and potentially scaling operations like ransomware. The LLM Agent Honeypot project aims to detect these agents by simulating vulnerable servers. Its work has revealed that agents can adapt and avoid detection better than traditional bots. Experts anticipate an increase in agent-driven cyberattacks, urging preemptive development of defenses as these technologies evolve.

UNSUPERVISED PANOPTIC SEGMENTATION

CUPS is a new method for panoptic segmentation without labeled data that leverages depth and motion cues to train directly on scene-centric images.

ROPE TO NOPE: HYBRID ATTENTION FOR LONG CONTEXT

The key innovation that enabled Llama 4 to reach 10m+ tokens in context is the alternation between no positional embeddings and rotational positional embeddings. While there are only have benchmarks on Needle in the Haystack, it seems to be a strong confirmation of performance of alternating layers.

INFERENCE-TIME SCALING FOR GENERALIST REWARD MODELING

This paper from DeepSeek talks about how to use inference time scaling to make reward modeling better to bootstrap stronger reasoners. It hints at a broader strategy from the Chinese start-up to use its existing reasoning models as the base for a new generation of reward models to train the next generation of reasoners.

GENERATIVE MODELING FOR CRYSTALS

CrystalFormer is a transformer model that generates crystal structures using space group symmetry, making crystal generation more efficient and data-friendly.

NANO AHA MOMENT

A single file, single GPU, from scratch full parameter tuning library that replicates DeepSeek R1-Zero style training.

OBJECT COUNTING

A fully automated zero-shot object counting method leveraging feature maps and self-attention mechanisms that achieves state-of-the-art accuracy on the FSC147 dataset.

DEEPSEEK 1.58BIT GGUF (HUGGING FACE HUB)

The unsloth folks have figured out which piece of the new R1 model can be properly quantized. They also found some tokenizer quirks to be aware of, which make quantization slightly harder. In summary, just the MoE layers go to 1.58 bit while everything else remains in 4 or 6 with their dynamic quantization scheme.

AI MASTERS MINECRAFT: DEEPMIND PROGRAM FINDS DIAMONDS WITHOUT BEING TAUGHT

DeepMind's AI system, Dreamer, successfully learned to collect diamonds in Minecraft without prior human guidance, highlighting a step towards general AI systems. Using reinforcement learning, Dreamer independently explores and builds a model of the game environment to predict future actions and outcomes. This advancement suggests potential applications for AI in real-world scenarios where trial and error are costly.

THE ARTIFACT ISN'T THE ART: RETHINKING CREATIVITY IN THE AGE OF AI

AI-generated Ghibli-style visuals have surged in popularity, straining OpenAI's servers and sparking debates about creativity in the AI age. While AI can rapidly produce artistic images, it lacks the human ability to experience and synthesize complex ideas and emotions. The future of creativity will focus on meaningful outputs shaped by human insight and purpose, with AI as a tool rather than a creator.

GRANITE SPEECH 8B (HUGGING FACE HUB)

IBM silently launched a state-of-the-art speech recognition and understanding model based on its Granite series.

START BUILDING WITH GEMINI 2.5 PRO

Google's Gemini 2.5 Pro is now in public preview via the Gemini API on Google AI Studio, with Vertex AI availability coming soon.

OPENAI'S MODELS 'MEMORIZED' COPYRIGHTED CONTENT, NEW STUDY SUGGESTS

A study suggested that OpenAI trained its models on copyrighted content without permission.

Watch the recording

DEEPMIND'S APPROACH TO AGI SAFETY

Google DeepMind outlines its perspective on building safe and secure artificial general intelligence, emphasizing oversight and technical safeguards as AGI capabilities progress.

OPENAI JUST MADE ITS FIRST CYBERSECURITY INVESTMENT

OpenAI invested in Adaptive Security, a startup using AI to simulate and train employees against social engineering hacks. Adaptive Security secured $43M in Series A funding, with plans to bolster its platform amid growing AI threats. Co-founded by veteran entrepreneur Brian Long, the startup has quickly amassed over 100 clients since launching in 2023.

OPENAI NONPROFIT GUIDANCE COMMISSION

OpenAI is forming a new expert commission to shape how its philanthropic arm supports communities using AI, aiming to align AI innovation with real-world nonprofit needs.

ARTICULATED KINEMATICS DISTILLATION FROM VIDEO DIFFUSION MODELS

This work introduces Articulated Kinematics Distillation (AKD), a framework that leverages skeleton-based animation and generative diffusion models to produce high-fidelity, physically plausible character motions with reduced complexity. It ensures structural consistency and outperforms existing methods in 3D coherence and expressive motion quality by using Score Distillation Sampling to guide joint-level control.

ENHANCED LORA-BASED FINE TUNING

MetaLoRA introduces dynamic parameter generation using meta-learning principles, enhancing the flexibility and task-awareness of LoRA-based fine-tuning strategies.

BACKDOOR ATTACKS IN CLIP

CLIP models are highly vulnerable to poisoning backdoor attacks, achieving nearly 100% attack success with minimal poisoned data. An efficient detection method is to use local outlier detection to uncover unintentional backdoors in existing datasets.

Read more here [12] and explore the code

LARGE SMALL NET

A new family of lightweight vision models inspired by the dynamic heteroscale capability of the human visual system, i.e., "See Large, Focus Small". LSNet achieves state-of-the-art performance and efficiency trade-offs across various vision tasks. It introduces a new type of convolution kernel.

PPLX CUDA KERNELS

Perplexity has released some of its MoE kernels, which outperform DeepSeek at scale while being slightly more flexible and less opinionated about the MoE architecture.

HATEBENCH FOR EVALUATING HATE SPEECH

HateBench provides a framework for evaluating hate speech detection models on LLM-generated content along with a manually annotated dataset and code for analyzing adversarial and stealthy hate campaigns.

ZONOS TTS

A compelling Apache 2.0 model for speech generation and voice cloning. It has multi language support and expressive real time generation.

HUGGING FACE'S AI AGENTS COURSE

Hugging Face released an AI agents course today. This free course will take you on a journey, from beginner to expert, in understanding, using, and building AI agents.

GOOGLE IS SHIPPING GEMINI MODELS FASTER THAN ITS AI SAFETY REPORTS

Google has launched an AI reasoning model, Gemini 2.5 Pro. The model leads in coding and math capabilities, but Google hasn't released safety reports yet. Google plans to publish these reports after collecting feedback from experimental releases, though this approach raises transparency concerns. Despite pledges for transparency, Google seems to prioritize speed in model deployments, diverging from industry norms for responsible AI practices.