Search the archive

A STAGGERING NUMBER OF GEN Z THINK AI IS ALREADY CONSCIOUS

25% of Gen Z believe AI is already conscious, while 52% think it soon will be.

XAI'S COLOSSUS 2 SUPERCOMPUTER

xAI is building Colossus 2, a massive upgrade over its original supercomputer cluster, to support the development of next-gen AI models.

OUR APPROACH TO UNDERSTANDING AND ADDRESSING AI HARMS

Anthropic unveiled a multidimensional framework to assess and mitigate AI harms, addressing both catastrophic and everyday risks. The approach evaluates physical, psychological, economic, societal, and autonomy impacts, guiding policy, testing, and enforcement strategies. This initiative, aiming for adaptable, proportionate safeguards as AI capabilities evolve, complements its Responsible Scaling Policy.

INSTAGRAM AI-BASED TEEN PROTECTION

Meta is using AI to identify teens on Instagram and automatically place them in restricted teen account settings, which now include even stronger default protections and require parental approval for changes.

O3 OVER OPTIMIZATION IS BACK

This post explores the challenge of the new crop of reasoning models and presents evidence to suggest that OpenAI is over-optimizing on certain objectives that make its models brittle and prone to hallucinations.

VERIFIABLE REWARDS FOR WRITING

Writing quality reward models (WQRM) are a way to evaluate the quality of creative writing. They can be used to train creative writing models. WQRMs are a new development for RL-trained models with verifiable rewards. This thread shows a case where WQRM correlated strongly with overall writing quality.

AI ASSISTED SEARCH-BASED RESEARCH ACTUALLY WORKS NOW

Recent advancements in LLMs like OpenAI's o3 and o4-mini have made them effective for search-based tasks, avoiding past issues with hallucinations. Unlike earlier models, these LLMs integrate search results into their reasoning process, providing reliable real-time insights. This progression diminishes reliance on traditional search engines and suggests shifts in the Web's economic model.

entire SDLC. Join the conversation

FAST CONFORMAL PREDICTION

LOO-StabCP boosts the speed of conformal prediction by using leave-one-out stability, providing scalable uncertainty estimation without sacrificing accuracy.

ROBUST AUTONOMY EMERGES FROM SELF-PLAY

This study presents a simulated self driving agent that went 2 years between collisions. It was trained fully with self-play and is a substantial improvement over the state-of-the-art trained on Gigaflow.

ALPHAGEOMETRY 2

DeepMind has released the next version of its geometry model, which answers 84% of questions correctly, up from 54% by prior methods. The main improvements come from the Gemini language model and better search.

HACKATHON FOR NON-DEVS AND VIBE CODERS

Stackblitz, the creators of Bolt, is hosting the world's largest hackathon for non-devs and vibe coders on 5/30 for anyone around the world to participate.

MAGI 1 - AUTOREGRESSIVE VIDEO GENERATION AT SCALE (HUGGING FACE HUB)

The MAGI 1 model is a new autoregressive video generation model that can generate long form and consistent videos. It performs on par with Wan video generation while being slightly below some closed models.

improve output. Register for free

GOOGLE AI ACADEMY FOR INFRASTRUCTURE STARTUPS

Google is inviting AI startups focused on U.S. infrastructure to apply for its six-month accelerator, offering mentorship, technical support, and strategic guidance.

OPENAI HINTS AT NATIVE SHOPIFY CHECKOUT INTEGRATION IN CHATGPT

New code strings in ChatGPT's web bundle reference “buy_now” buttons and a “shopify_checkout_url,” suggesting OpenAI is building a native purchase flow into the assistant.

MERCOR GRADUATE FELLOWSHIP

Mercor launched a $50,000 fellowship for PhD students and postdocs in STEM focused on identifying raw talent based on ideas, not pedigree or connections.

GEMINI 2.5 FLASH

The next Flash model from Gemini has been released. It is a substantial upgrade from previous versions and matches Claude on a number of important STEM benchmarks.

CLAUDE CODE BEST PRACTICES

Anthropic has released a detailed engineering guide on how to use its agentic programming assistant. It requires more specificity than traditional models.

OUR UPDATED PREPAREDNESS FRAMEWORK

OpenAI has updated its Preparedness Framework to enhance safety measures against severe risks from advanced AI capabilities. The revision emphasizes clear criteria for prioritizing high-risk capabilities, sharper capability categories, scalable evaluations, and defined safeguard reports. This framework will be continually updated in line with emerging technological advancements and insights from internal and external experts.

FLEXIBLE IMAGE WATERMARKING

MaskMark provides a simple, dual-mode solution for global and local watermarking using a masking-based Encoder-Distortion-Decoder framework.

SLEEP-TIME COMPUTE FOR LLM EFFICIENCY

A novel approach to reduce LLM inference costs by precomputing useful context-related quantities before user queries. It shows up to 5x test-time savings and accuracy gains on reasoning tasks.

PERSONALIZED IMAGE SYNTHESIS WITH AUTO-REGRESSIVE MODELS

This paper explores the training auto-regressive models for personalized image generation to achieve comparable fidelity to diffusion-based methods through a two-stage optimization process.

seat

ALIGNING LVMS WITH HUMAN PREFERENCES

VistaDPO improves video-text alignment by optimizing preference learning across spatial and temporal dimensions using a new 7.2K sample dataset and hierarchical optimization.

HALLUCINATION REDUCTION IN VLMS

REVERSE introduces a training and inference pipeline that enables VLMs to self-detect and revise hallucinations.

ZEROSUMEVAL BENCHMARK

A dynamic evaluation framework that uses competitive multi-agent simulations to benchmark LLMs across reasoning, knowledge, and planning tasks.

GARMENT GENERATION

A new two-stage generative framework for clothing design that enables detailed control over silhouette, color, and logos. It also introduces GarmentBench, a large dataset for multi-conditional garment synthesis.

GEMINI IMAGE SEGMENTATION

It is well known that Gemini's vision capabilities are quite good. This post explores one specific use case in segmentation which is surprisingly easy.

SOUTH KOREA'S AI CHIP CHAMPION IS POISED TO CARVE OUT GLOBAL NICHE

Rebellions, South Korea's first AI chip unicorn, merged with SK Telecom's Sapeon to compete globally against giants like Nvidia. The company aims to produce energy-efficient AI chips - its Rebel chip promises significant power savings over Nvidia's H100. With support from top talent and strategic partnerships, Rebellions is targeting international growth and plans to IPO by 2026.

LTXV DISTILLED 0.9.6 VIDEO MODEL (HUGGING FACE HUB)

LTX video model is a state-of-the-art open video model.

GENERATE VIDEOS IN GEMINI AND WHISK WITH VEO 2

Starting today, Gemini Advanced users can generate videos using the Veo 2 model, translating text-based prompts into high-resolution, cinematic videos.

OPENAI HAS LAUNCHED THE CHATGPT IMAGE LIBRARY

OpenAI launched the ChatGPT Image Library on the Web and Android/iOS for Free, Plus, and Pro users.

Register now to save your spot.

MISTRAL CLASSIFIER FACTORY

Mistral, a French AI startup, has launched a new product that allows users to very quickly build and deploy custom classifiers for a whole variety of tasks (e.g., spam, moderation, and more).

GOODFIRE RAISES $50M SERIES A TO STEER AND UNDERSTAND MODELS

Goodfire is a mechanistic interpretability company with strong expertise in SAEs, among other things. It is working closely with closed and open model providers to steer, control, and understand model motivations and behavior.

VISUAL REASONING WITH OPENAI O3 AND O4-MINI

OpenAI's latest visual models can reason with images through tool-augmented transformations, enabling a new level of multimodal understanding and step-by-step visual problem-solving.

EFFICIENT LINE ART COLORIZATION WITH BROADER REFERENCES

Novel efficient long-context fine-grained ID preservation framework for line art colorization, achieving high precision, efficiency, and flexible usability for comic colorization. It transforms black-and-white line art into vibrant illustrations by effectively integrating extensive contextual references.

SCENE CAPTIONING

3D CoCa is a unified framework that combines vision-language contrastive learning and captioning for 3D scenes.

LARGE REASONING MODELS AS A JUDGE

JudgeLRM is a family of LLMs trained with reinforcement learning for judgment tasks. Unlike SFT, it excels in reasoning-heavy evaluations, outperforming models like GPT-4 and DeepSeek-R1.

DEEPSPEED'S DEEPCOMPILE

The DeepSpeed team has worked to bring compilation to their distributed training efforts. This compilation speeds up various bottlenecked operations by many times. It makes use of a patched version of torch compile.

SPEECH INSTRUCTION FINE-TUNING DATASET (HUGGING FACE HUB)

SIFT-50M (Speech Instruction Fine-Tuning) is a 50-million-example dataset designed for instruction fine-tuning and pre-training of speech-text large language models (LLMs). It is built from publicly available speech corpora containing a total of 14K hours of speech and leverages LLMs and off-the-shelf expert models. The dataset spans five languages, covering diverse aspects of speech understanding and controllable speech generation instructions. SIFT-50M augments existing speech datasets with instruction-based question-answer (QA) pairs for speech understanding and includes approximately 5 million examples for controllable speech generation.

END-TO-END LATENT DIFFUSION TRAINING WITH REPA-E

META RELEASES MANY NEW ARTIFACTS

Meta has released an image Encoder, a VLM, a 3D object localization model based on JEPA, and weights for a BLT model that operates directly on bytes without tokenization.