Search the archive

AI ISN'T VERY GOOD AT HISTORY

A new benchmark, Hist-LLM, revealed that leading LLMs like GPT-4, Llama, and Gemini struggle with high-level historical questions.

OSCAR HOPEFUL 'THE BRUTALIST' USED AI DURING PRODUCTION

The filmmakers of 'The Brutalist', driven by budget constraints, used AI from Respeecher to enhance actors' Hungarian pronunciation and generate architectural drawings.

Try Augment for free today.

DARIO AMONDEI ON DEEPSEEK AND EXPORT CONTROLS

Anthropic's CEO gives great insight into the DeepSeek models, cost curves, and innovation. He expertly lays out the landscape and points to flaws in the market reaction. Additionally, he reveals that Sonnet 3.5 only cost $10m or so dollars to train. The main point of the article is export controls.

OPENAI ACCUSES DEEPSEEK

OpenAI has found evidence that the Chinese AI startup DeepSeek used its proprietary models to train an open-source competitor. The suspected "distillation" process allows smaller models to mimic larger ones at a lower cost, raising concerns over intellectual property breaches.

ZUCKERBERG DISMISSES DEEPSEEK, COMMITS TO AI INVESTMENT

Mark Zuckerberg remains unfazed by competition from DeepSeek. He has reaffirmed Meta's commitment to massive AI investments and focus on model development and computational resources.

QWEN 2.5 MAX MOE MODEL

The Qwen team has released its MoE earlier than expected. The model shows really strong performance, matching many top models, including DeepSeek v3.

OPTIMIZING LARGE LANGUAGE MODEL TRAINING USING FP4 QUANTIZATION

One of the key levers that can be pulled to reduce training and inference costs is quantization. Getting these models to run at lower precision saves on the number of GPUs and FLOPs needed for training. This work shows, on a small scale of 100B tokens, how to train at FP4.

TEXT-TO-VIDEO GENERATION WITH CASCADEV

CascadeV introduces a cascaded latent diffusion model that generates 2K-resolution videos with improved efficiency. It employs a novel 3D attention mechanism and can be integrated with existing text-to-video models to enhance resolution and frame rate without fine-tuning.

EFFICIENT COMPRESSION FOR SSM-BASED MODELS

Mamba-Shedder is a method that optimizes structured state space models by selectively removing redundant components. The approach improves inference speed by up to 1.4x while maintaining model accuracy.

AI FOR VALUE MEASUREMENT AND SAFETY

This project introduces Generative Psychometrics for Values, a novel method that uses large language models to measure human and AI values.

LLMS WITH SPECIALIZED TOOLS

TART enhances large language models by integrating them with computational tools, improving accuracy and transparency in tasks like finance, science, and healthcare.

WRITERS VS. AI: MICROSOFT STUDY REVEALS HOW GPT-4 IMPACTS CREATIVITY AND VOICE

Microsoft and USC conducted a study using GPT-4 to explore how generative AI affects writers' authenticity and creativity. Writers expressed concerns about AI impacting originality, emotional satisfaction, and ownership of work. Despite initial skepticism, personalized AI models that align with a writer's style helped alleviate these concerns, enhancing creativity without compromising authenticity.

WE INTERVIEW THE FOUNDER BEHIND MEGAN, AN AI HR AGENT

Interview Mega HR has launched "Megan", an AI agent designed to streamline and automate recruitment processes. Megan handles tasks from job posting to candidate management, aiming to improve hiring efficiency and transparency.

DEEPSEEK'S R1 RECIPE

Nathan Lambert breaks down the recipe for R1 and talks through what it means for us now and for the field broadly. Specifically, he focuses on the interesting application of reinforcement learning.

syntax, semantics, and context. Read the Oxford Languages ebook

GOOGLE REPORTS HALVING CODE MIGRATION TIME WITH AI HELP

Google used LLMs to significantly speed up internal code migrations, such as transitioning Google Ads' 32-bit to 64-bit IDs.

GOOGLE'S TITANS GIVE AI HUMAN-LIKE MEMORY

Google has introduced the Titans architecture, an evolution of Transformers that incorporates neural long-term memory for better data retention and "surprise-based" learning.

THE PENTAGON SAYS AI IS SPEEDING UP ITS 'KILL CHAIN'

OpenAI, Anthropic, and Meta are partnering with U.S. military agencies to use AI for threat identification and planning.

HUGGING FACE'S OPEN-R1

Hugging Face has released Open-R1, a fully open reproduction of DeepSeek-R1.

OPENAI'S MODEL FOR GOVERNMENT USE

OpenAI's ChatGPT-Gov is a tailored version of ChatGPT for government agencies. It ensures security, compliance, and efficiency in public sector applications.

NOTES ON DEEPSEEK: GENERATIVE AI IS ALL ABOUT THE APPLICATIONS NOW

DeepSeek R1, a new open-source AI model from China, significantly reduces the cost of running AI models to just 3-5% of OpenAI's comparable models. This cost reduction shifts focus from infrastructure investment to actual AI application development, challenging existing economic models in the AI industry. The development has the potential to spur new AI products but also raises concerns about adequate generative AI applications.

YUE MUSIC MODEL

The YuE model is a full song generation model that generates lyrics and music together. It is quite high fidelity and easily the best open source model to date. YuE is not quite as strong as the closed models, but is still quite compelling. It uses a two stage approach and uses discrete audio tokens.

A ROBUST SLAM SYSTEM

LCP-Fusion introduces a new approach to dense SLAM that addresses challenges in mapping unknown environments with high accuracy.

PING PONG GEMM KERNEL FROM CUTLASS

A pretty technical deep dive into lightning-fast multiplication kernels for hardware accelerators. Specifically, the Ping Pong asynchronous kernel, which works for fp8 and has extremely strong final performance.

image generation without compromise.

Experience it yourself with $25 in free credits - use code "TEXT2IMAGE" [10] before February 26th.

PARAMETRIC RETRIEVAL-AUGMENTED GENERATION

Parametric RAG embeds external knowledge directly into the parametric space of LLMs, improving reasoning while reducing reliance on large context windows. The repository includes an end-to-end implementation and benchmark datasets.

A UNIFIED SELF-DRIVING WORLD MODEL

HERMES integrates scene understanding and future scene generation for autonomous driving through a unified framework. It uses Bird's-Eye View representations and world queries for contextual enrichment.

LANGCHAIN: OPENAI IN JAVASCRIPT WITH REACT.JS & NEXT.JS

This tutorial shows readers how to build a chatbot application using LangChain in JavaScript that interfaces with OpenAI's API via Next.js and React. Key steps include setting up the frontend, implementing server logic for chat interaction, and securely handling API keys. The project's source is available on GitHub for further experimentation and enhancement.

RESEARCHERS USE AI TO DESIGN PROTEINS THAT BLOCK SNAKE VENOM TOXINS

Researchers used AI tools like RFdiffusion and ProteinMPNN to design proteins that inhibit snake venom toxins, showing potential for antivenoms that don't need refrigeration. They successfully created a protein that inhibits neurotoxic venom but faced challenges with toxins disrupting cell membranes. This showcases AI's power to solve complex biological problems that were previously difficult to tackle.

TIKTOK OWNER BYTEDANCE POWERED AN E-READER'S UNHINGED AI ASSISTANT

Boox e-readers, which used ByteDance's Doubao LLM, were criticized for spreading Chinese propaganda. Doubao responded to specific questions with pro-China, Russia, and North Korea narratives while criticizing Western countries. Following backlash, Boox reportedly switched back to OpenAI's GPT-3.

OPENAI IS BANKROLLING AXIOS' EXPANSION INTO FOUR NEW MARKETS

OpenAI has partnered with Axios to fund the expansion of its local newsletters to four new cities. This marks the first time OpenAI has directly funded a news organization's newsroom. OpenAI aims to help publishers integrate AI tools into their operations.

TO JOIN LIVE ON FEB 11TH

ZUCKERBERG SAYS AI WILL REPLACE MID-LEVEL ENGINEERS SOON

Mark Zuckerberg predicts AI will replace mid-level engineers by 2025, allowing the remaining engineers to focus on strategic tasks.

GEMINI APP GETTING MORE REAL-TIME INFORMATION WITH NEWS PARTNERSHIP

Google partnered with AP to enhance real-time news in the Gemini app.

FOCAL (PRODUCT)

Focal is an AI video platform for creating movies, TV shows, and YouTube videos.

Try Neosync for free (no credit card required)

DEEPSEEK SURPASSES CHATGPT AS APP STORE'S TOP APP

DEEPSEEK RELEASES OPEN-SOURCE AI IMAGE GENERATOR

DeepSeek has released its AI image generator as open source, raising both excitement in the developer community and concerns about potential misuse.

LINKEDIN CO-FOUNDER RAISES $25 MILLION TO TAKE ON CANCER WITH AI

Reid Hoffman launched Manas AI, a startup using AI to accelerate cancer drug discovery. The company raised $25 million in seed funding and partnered with Microsoft. Manas AI focuses on breast cancer, prostate cancer, and lymphoma and aims to expand to other diseases.

QWEN 2.5 VL

The next set of vision language models from Qwen. Qwen2.5-VL is a versatile visual AI model that excels in multiple areas: recognizing objects and analyzing visual elements like texts and charts, acting as an interactive visual agent for tool control, processing lengthy videos with event detection capabilities, performing precise object localization through various formats, and generating structured data outputs for business applications like finance and commerce.

CHAIN OF RAG

Reasoning models can now be trained to iteratively do retrieval. This is similar in spirit to the work shown in the Operator system. Interestingly, it leads to a substantial improvement, although it's not clear what the true FLOP controlled gain looks like.

BRAINGUARD: PRIVACY-PRESERVING BRAIN DECODING FRAMEWORK

BrainGuard introduces a collaborative training framework for reconstructing perceived images from multisubject fMRI data while safeguarding privacy.

DEEPSEEK JANUS PRO

DeepSeek's image model got a big upgrade today. It is a unified text and image model - sometimes referred to as any-to-any. This means that it can both understand and generate images and text interleaved in a conversation. This is similar to the omni models from OpenAI and the Gemini suite of models from Google.