Search the archive

REWARD HACKING

Lilian Weng has another great blog post discussing the challenge of Reward Hacking within language model alignment. It is one of the major challenges preventing models from being deployed into production scenarios.

Kore.ai

FINE-TUNE MODELS WITH LORA-SB

LoRA-SB is a new method that brings full fine-tuning performance to low-rank adapters for large language models.

MAKING AI DATASETS MORE DIVERSE

Researchers introduced a new method called Diversity-driven EarlyLate Training (DELT) to improve dataset distillation for large-scale tasks.

OPENAI TARGETS 1BN USERS IN NEXT PHASE OF GROWTH

CREATE JSONL DATASET FROM API CHAT LOGS

A nice simple utility that allows you to create a JSONL dataset based on the messages sent back and forth between user and API.

GOOGLE'S PLAN TO KEEP AI OUT OF SEARCH TRIAL REMEDIES ISN'T GOING VERY WELL

US District Judge Amit Mehta suggests AI could play a significant role in developing remedies following the government's win in the Google search monopoly trial, potentially affecting Google's AI products. The DOJ has proposed remedies to prevent Google from using AI to maintain its market dominance, including restrictions on exclusive agreements and investments in AI. Microsoft is resisting Google's demands for confidential AI deal details, arguing relevance. OpenAI may be compelled to share data amidst these considerations.

swag! Learn more about ML at Jane Street

UV GUIDE TO PYTORCH

Documentation on how to use the new package manager UV to install PyTorch.

AI COMPANY MISTRAL IS LATEST EUROPEAN STARTUP TO EYE EXPANSION IN SILICON VALLEY

European AI startup Mistral, known for open-weight LLMs, is expanding to Palo Alto to access Silicon Valley talent and resources.

OPENAI GETS NEW $1.5 BILLION INVESTMENT FROM SOFTBANK, ALLOWING EMPLOYEES TO SELL SHARES IN A TENDER OFFER

OpenAI is allowing employees to sell $1.5 billion worth of shares to SoftBank, reflecting growing investor interest in AI.

ELON MUSK SEEKS TO BLOCK OPENAI'S FOR-PROFIT CONVERSION AGAIN

Elon Musk's legal team has filed a motion to prevent OpenAI from transitioning to a for-profit model citing alleged antitrust violations and financial risks. The motion claims CEO Sam Altman's actions could leave the company unable to pay damages if Musk wins his lawsuit.

PERPLEXITY MULLS GETTING INTO HARDWARE

Perplexity's CEO plans to develop a "simple, under $50" AI device for voice-to-voice interactions. This aligns with a broader trend of AI startups exploring hardware to enable new interactions, although historical challenges in AI hardware highlight risks. Perplexity, with substantial funding, aims to avoid pitfalls faced by others like Humane's Ai Pin.

INFLECTION AI CEO SAYS IT'S DONE TRYING TO MAKE NEXT-GENERATION AI MODELS

Inflection AI shifted its focus from developing cutting-edge AI models to providing AI tools for enterprise customers, leveraging current AI models. It has acquired three AI startups to strengthen offerings and is open to licensing models from former competitors. CEO Sean White says the company is focused on practical applications over frontier model development, highlighting on-premise AI solutions for enterprise data security.

INTELLECT-1 RELEASE: THE FIRST GLOBALLY TRAINED 10B PARAMETER MODEL

INTELLECT-1 is a 10B parameter model trained on 1T tokens with hardware distributed around the globe. The benchmarks are reasonable, and the MFU of 30%+ is amazing given the distributed nature of training. If the results hold, this is a substantial improvement for decentralized large model training.

DETECT AND LEARN UNSEEN OBJECTS

This new framework pushes object detection into open-world scenarios by teaching AI to identify and learn from objects it hasn't seen before.

MAKING UNDERWATER IMAGES CLEAR

HUPE is an AI-powered method that improves underwater image clarity while preserving details critical for other tasks like object detection.

Register now

MAPPING THE IONOSPHERE WITH THE POWER OF ANDROID

Google researchers were able to accurately map the Ionosphere by using GPS fluctuations and some clever algorithmic work. The task is usually expensive and time consuming. This effort can assist with various climate solutions.

INTRODUCING LTNTORCH

Logic Tensor Networks (LTN) merge deep learning with logical reasoning, allowing neural models to learn by optimizing a knowledge base of logical formulas.

REFINING PRETRAINING DATA PROGRAMMATICALLY

ProX is a framework that treats data refinement as a programming task, allowing models to apply fine-grained operations to individual examples at scale. It improves the quality of pre-training corpora by using small language models to generate programs.

CHATS WITH VIDEOS IN REAL TIME

MMDuet is a novel "video-text duet" interaction format for VideoLLMs that allows AI to provide real-time responses as videos play. The approach mimics a dialogue where both users and AI can exchange messages during playback.

BUILDING LLMS IS PROBABLY NOT GOING BE A BRILLIANT BUSINESS

LLM makers like OpenAI face significant challenges due to industry structure, notably NVIDIA's pricing power as the key chip supplier and high buyer price sensitivity and competition. Many AI companies raise substantial capital but struggle with profitability, similar to historical tech firms like Netscape. Despite this, the technology itself may still advance. AI businesses might succeed by leveraging existing models rather than developing new ones.

ROX

A new AI-native CRM, Rox, aims to disrupt Salesforce by leveraging AI's ability to manage unstructured data and integrate with data warehouses. Rox's strategy focuses on enhancing top sales performers through AI-powered agents while securing customer data for future AI developments. Investors have shown confidence with $50M in funding, betting on Rox as AI capabilities continue to improve.

PLAYAI RAISES $21M FUNDING AND RELEASES A NEW VOICE MODEL

PlayAI raised $21 Million to advance voice-first AI interfaces and models and introduced Play Dialog, a state-of-the-art multi-turn speech model.

ANTHROPIC SAYS CLAUDE AI CAN MATCH YOUR UNIQUE WRITING STYLE

Anthropic's Claude AI now offers customizable response styles, allowing users to adjust tone and length for different writing tasks.

CONVERTING GPT TO LLAMA

This repository contains code for converting a GPT implementation to Meta AI's Llama.

🎄 SPECIAL HOLIDAY OFFER: Get 58% off Incogni with code TLDRAI

(30 day money back guarantee)

ALIBABA CHALLENGING OPENAI'S O1 WITH A NEW MODEL

Alibaba has released QwQ-32B-Preview, an ‘open' challenger to OpenAI's o1 reasoning model.

AI2 RELEASES NEW LANGUAGE MODELS COMPETITIVE WITH META'S LLAMA

Ai2 has released OLMo 2, an open-source language model series with models containing 7 and 13 billion parameters. OLMo 2 is developed with publicly accessible training data and code. Aiming to boost open-source AI innovation, Ai2 claims these models outperform similar open models like Meta's Llama 3.1. The models are available under the Apache 2.0 license for commercial use.

XAI COULD SOON HAVE ITS OWN APP

Elon Musk's xAI plans to launch a standalone app for its Grok chatbot by December, competing directly with OpenAI's ChatGPT. Grok is currently accessible through X for subscribers only. This move positions xAI to more directly challenge established chatbot apps from OpenAI, Google, and Anthropic.

THUNDERKITTENS FOR APPLE SILICON

Hazy Research has been critical in improving hardware utilization for AI workloads. It has extended its amazing ThunderKittens Kernel writing framework to work on Apple Silicon.

DIFFUSIONDRIVE: TRUNCATED DIFFUSION MODEL FOR END-TO-END AUTONOMOUS DRIVING

Diffusion models for End-to-End driving of autonomous vehicles which can operate at 45 FPS on a 4090 chip.

SUPER-RESOLUTION WITH LOW-BIT QUANTIZATION

PassionSR introduces an approach that makes diffusion-based image super-resolution (SR) models more hardware-friendly.

OPEN INSTRUCT

The Allen AI Institute has released extremely useful post training code for instruction tuning.

STAR ATTENTION

Up to 11x speed improvement for long context token processing and generation. It is slightly better than ring attention, but still worse than full attention.

FINE-GRAINED IMAGE QUALITY ASSESSMENT

Grounding-IQA is a new approach to image quality assessment (IQA) that integrates precise location-based grounding and multimodal descriptions.

IT'S SURPRISINGLY EASY TO JAILBREAK LLM-DRIVEN ROBOTS

RoboPAIR is an algorithm that successfully jailbreaks robots powered by LLMs, bypassing all safety guardrails. Experiments showed a 100% success rate in hacking systems like Go2, a self-driving simulator, and robot dogs. These findings reveal significant security vulnerabilities, stressing the need for better defenses against LLM-based robot hacking.

WORLD'S MOST FLEXIBLE SOUND MACHINE DEBUTS

Fugatto, a new generative AI model, allows users to control audio creation and transformation using text prompts. It can generate music, modify voices, and create unique sounds, showcasing capabilities through tasks like music production and video game development. Developed by NVIDIA researchers, Fugatto uses 2.5 billion parameters and was trained on a vast dataset to handle multilingual audio tasks and unsupervised learning.

THE NEW AI SCALING LAW SHELL GAME

Recent shifts in AI scaling laws reveal limitations in predictability and effectiveness, with new models not meeting earlier expectations. Microsoft CEO Satya Nadella highlights "inference time compute" as a potential area of focus, but concerns about cost and reliability persist. Innovation beyond scaling is needed. LLMs should be part of a broader AI strategy.

STEEL BROWSER API FOR AI AGENTS

Controlling web browsers for AI agents is a useful thing for accessing information.

PIXMO DATASET (HUGGING FACE HUB)

Allen AI has released several datasets that were used to train its visual language models.

HUMAN IMAGE ANIMATION

StableAnimator introduces a breakthrough in human image animation by ensuring identity consistency in generated videos.

Delete your personal data today.

🎁 HOLIDAY SPECIAL: Use this link [4] to get 58% off using code TLDRAI.