Search the archive

using data from Oxford Languages. Read the case study

GENERATING REALISTIC STREET-VIEW VIDEOS FOR AUTONOMOUS DRIVING

The World Volume-aware Multi-camera Driving Scene Generator (WoVoGen) is an innovative system designed to create realistic street-view videos for autonomous driving.

AI-GENERATED IMAGE QUALITY ASSESSMENT

TIER leverages both text prompts and images generated from them to assess the quality of the generated images.

LLM EVALUATION METRICS FOR LABELED DATA

This article discusses evaluation metrics that rely on ground truth data. It explores general-purpose as well as RAG-specific evaluations. All metrics come with a Python implementation or a link to a Hugging Face model card.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

👉 Start Now 👉

GOOGLE DEEPMIND OPEN SOURCES GEMMA BASED ON GEMINI

Google has released the weights for its Gemma 2B and 7B parameter models, available via HuggingFace. The models are decoder-only Transformers trained on 2T and 6T tokens. They substantially outperform Llama 2 on a wide range of benchmarks and come in base and instruction-tuned versions.

HELM INSTRUCT BENCHMARK

The language modeling group at Stanford released a widely used benchmark called Holistic Evaluation of Language Models (HELM). They also released a version for instruction following called HELM-Instruct. It is multidimensional, open-ended, and absolute.

REDDIT HAS A NEW AI TRAINING DEAL TO SELL USER CONTENT

Reddit has entered a licensing agreement worth approximately $60 million annually with an unnamed major AI company, granting the company access to its vast user-generated content.

25+ LORAS FOR MISTRAL THAT OUTPERFORM GPT-4

Predibase released 25 narrow task-specific LoRAs that each outperform GPT-4 on specific tasks. These task-specific LoRAs are built onto the Mistral 7B model. This is another piece of evidence to support the ability of local models to perform exceptionally well on exact tasks.

MULTIMODAL LLM’S ABILITY TO UNDERSTAND VISUAL DATA

ChartX is a new tool for testing multi-modal large language models (MLLMs) on their ability to interpret and reason with visual charts.

COMBINING UNSUPERVISED TRAINING AND SUPERVISED FINE-TUNING

This repo challenges the effectiveness of combining supervised fine-tuning and reinforcement learning in training. Using a more advanced model like GPT-4 for initial training can outperform the more complex two-step method.

MULTIMODAL CONVERSATIONS DATASET

Meta has released the MMCSG dataset, which has thousands of two-sided conversations recorded via the Aria glasses, to encourage research in various social audio and visual tasks.

MULTILORA INFERENCE SERVER

Lorax is an inference server that supports hot-swapping LoRAs onto a single base model. This means a dramatic decrease in RAM while supporting a wide range of model tunes.

EVALUATING LLMS IN GAMES

This project introduces GTBench, a platform for testing LLMs like GPT-4 in various game scenarios to discover their strengths and weaknesses in strategic reasoning. While these models show promise in games based on probability, they struggle in deterministic games and vary in performance based on their design and training.

MULTIVIEW DIFFUSION++

A diffusion model that can generate multiple views of an object and then stitch them together into a compelling 3D version.

THIS IS WHY THE IDEA THAT AI WILL JUST AUGMENT JOBS, NEVER REPLACE THEM, IS A LIE

AI is going to automate parts of the workforce. The reaction so far has been two-sided – will this efficiency mean the same humans can do more or will fewer humans do the same tasks? This article explores comparisons between the impact of tech on manufacturing and agriculture and what we’re seeing today with the modern knowledge worker.

MULTIMODAL DIALOGUE

ChatterBox is a vision-language model that excels in multimodal dialogues, specifically in the newly established task of multimodal multi-round referring and grounding.

memories—try the club for $1 today.

APPLE DEVELOPING AI TOOL TO HELP DEVELOPERS WRITE CODE FOR APPS

Apple is working on an updated version of Xcode that will include an AI tool for generating code.

ADGEN AI (PRODUCT)

Create 100+ high-performing ad variations with AI.

CREWAI

CrewAI offers a library for building and orchestrating AI agents that leverage LangChain and Replit. It provides a modular setup with agents, tasks, tools, and crews for easy integration into various projects. Debugging is simplified through LangSmith, which enhances performance insights into non-deterministic LLM calls.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

GPT-4 KNOWLEDGE CUTOFF IS DECEMBER 2023

According to the new documentation on the OpenAI platform, the training data for GPT-4 models has been updated to include information up to December 2023.

SOFTBANK FOUNDER MASAYOSHI SON AIMS TO RAISE $100 BILLION FOR NEW CHIP VENTURE

Masayoshi Son's SoftBank Group Corp. is launching Izanagi, a $100 billion chip venture aimed at competing with Nvidia and focusing on AI applications.

SCRIBE $25M SERIES B

Scribe has raised a Series B funding round led by Redpoint Ventures to accelerate its AI-driven platform, which automates the creation of visual step-by-step guides and facilitates knowledge sharing within organizations. Over a million teams use Scribe, including 97% of the Fortune 100. The company is expanding its AI features to make workplace information even more accessible.

GENERATIVE REPRESENTATIONAL INSTRUCTION TUNING

The Contextual team has trained and released a model that both generates text and embeddings. It dramatically outperforms single specialist models. The model is an interesting take on the multi-modal trend where the output modality is an embedding.

DEEPDIVE: MAMBA THE HARD WAY

Sasha Rush has released an annotated tutorial for accelerating Mamba with custom Triton kernels. It doesn’t scale yet due to a bug in the Triton compiler, but it is an extreme illustration of the technology and great for those looking to dive deeper into the state space Transformer alternative world.

ENHANCING FINE-TUNING WITH LORA+

This paper presents LoRA+, an advancement over the existing Low-Rank Adaptation (LoRA) method for fine-tuning large models. LoRA+ achieves better performance and faster fine-tuning, without increasing computational demands, by using different learning rates for key components in the process.

3D OBJECT FROM AS FEW AS 4 PICTURES

This repository allows you to take four images and turn them into a high-quality 3d representation with Gaussian Splatting.

GENERATING IMAGES AT ANY RESOLUTION

The Flexible Vision Transformer (FiT) is a novel architecture designed to create images at any resolution and aspect ratio. Unlike traditional models, FiT treats images as sequences of variable-sized tokens, allowing it to adapt to different image sizes more effectively during training and inference.

BOOSTING AI'S DEFENSE AGAINST ADVERSARIAL ATTACKS

This project introduces a new method to strengthen multi-modal models like OpenFlamingo and LLaVA against visual adversarial attacks. By fine-tuning the CLIP vision encoder in an unsupervised manner, the authors effectively shield these models from manipulative image attacks, enhancing their reliability and security in real-world applications without the need for retraining the entire model.

NEW LLM BENCHMARK

Preeminent researcher Nicholas Carlini has released the benchmark he uses for evaluating large language model performance. Interestingly, it has GPT-4 further ahead than most other benchmarks.

STRATEGIES FOR AN ACCELERATING FUTURE

The recent advancements in AI, particularly with Google's Gemini offering a context window of over a million tokens and Groq's hardware enabling almost instantaneous responses from GPT-3.5 models, signify a major leap forward in practical AI applications and underscore the urgency for leaders to understand and adapt to AI's rapidly evolving landscape.

HOW TO LOSE AT GENERATIVE AI

Generative AI, while hyped, is likely to disappoint most startups because it favors incumbents with data advantages, existing workflows, and the ability to integrate AI into these systems without major overhauls. Despite venture capital flowing into the GenAI space, startups focusing on prompt engineering and UX improvements at the workflow layer are essentially prepping the market for incumbents who can easily adopt and integrate AI innovations into their dominant platforms, suggesting a challenging path ahead for startups aiming to capture significant value in the Generative AI domain.

BOCOEL

Accurately evaluate LLMs with Bayesian Optimizations.

AMAZON AGI TEAM SAY THEIR AI IS SHOWING “EMERGENT ABILITIES”

Amazon AGI researchers developed a language model named "Big Adaptive Streamable TTS with Emergent Abilities" (BASE TTS) that shows "state-of-the-art naturalness" in conversational text, demonstrating language skills it wasn't specifically trained on.

ADGEN AI (PRODUCT LAUNCH)

AdGen AI confronts the chaos of traditional ad creation, offering a streamlined, AI-driven solution. Generate 100+ ad variations from a single URL in minutes.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

GEMINI BUSINESS AND GEMINI ENTERPRISE PLANS FOR GOOGLE WORKSPACE ARE COMING

Google is introducing "Gemini Business" and "Gemini Enterprise" plans for Workspace customers, offering access to its advanced AI model, Gemini 1.0 Ultra, with enterprise-grade data protection. It will not use employee conversations for model training.

OPENAI VALUED AT $80 BILLION IN DEAL WITH THRIVE CAPITAL

OpenAI has reportedly reached a valuation of over $80 billion following a deal with VC firm Thrive Capital. This valuation marks a nearly 3x increase in less than 10 months, making OpenAI the 3rd highest-valued tech startup in the world. The deal involves a share sale under a "tender offer," allowing employees to cash out their shares in the company.

MAGIC RAISES $117M TO CONTINUE CODE GENERATION MODELS

Magic is building an AI software engineer and has raised more to help with this process.

KV QUANT - SCALING TO 10M TOKENS

Quantizing the KV cache, which is a technical detail of the Transformer architecture, allows it to use less memory at inference time. Quantization is the shrinking of the floating point precision with minimal loss in quality.

ENHANCED SPEECH TRANSLATION

ZeroSwot is a new method for Speech Translation (ST) that overcomes the lack of data and differences between speech and text. It uses unique techniques to train a speech encoder with only speech recognition data, allowing it to work with a multilingual translation model.

MAKING CLIP’S VISUALLY UNDERSTANDABLE

SpLiCE is a new method that makes CLIP's complex visual data easier to understand.

Be the first to access Height 2.0