Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Efficient RGB-D semantic segmentation helps mobile robots understand their environment. While depth data offers useful geometric insights, it can be noisy. This paper introduces a streamlined network that uses less computing power but remains robust, blending color and depth features effectively.
3D detectors often fail when used in unfamiliar settings, causing a huge drop in detection rates due to changes in tools, places, or weather. This project proposes a self-training system that adapts to different lidar types by creating accurate fake labels, MS3D++.
Predicting the performance of pre-trained neural networks on multiple tasks without fine-tuning them is a challenge. The traditional method is computationally costly, while modern methods are limited to single tasks. This paper introduces EMMS, which transforms varied task labels into a unified format, allowing for the prediction of a model's effectiveness across tasks efficiently.
Sponsored
What if you could spend 100x less time building apps for your team? With Glide’s AI-powered no-code app builder
There are lots of cloud providers popping up for fine-tuning or serving AI workloads. It is hard to know which one to choose over another one. This guide gives a pretty great set of heuristics to decide which cloud to run your jobs on.
Fooocus is an image-generating software that is offline, open-source, free, and does not require manual tweaking.
This project introduces a new tool called DatasetDM that can make endless fake images along with detailed labels, like which part is which object or how far things are in the image. This is great for training computer vision systems, and by using just a small set of real photos, the synthetic ones help achieve top results in tasks like identifying objects or editing images.
Generative AI is extremely hyped, but does it deserve it? Many professionals have still not found a good use for ChatGPT multiple months after its release. In this article, Gary Marcus explores the possibility that we are too confident in our assumptions that generative AI will be world-changing.
AI.com now redirects to x.AI, a company that aims to understand the true nature of the universe. OpenAI had seemingly purchased this domain back in February, but it looks like Musk swooped in.
The economist Matt Clancy and research scientist Tamay Besiroglu debate what happens to the economy when intelligence becomes extremely cheap.
Sponsored
On the Tools and Weapons podcast, Microsoft Vice Chair and President Brad Smith explores the challenges of technology and society, including developing AI ethically so it serves all of humanity.
Apple is using an internal chatbot to help its employees prototype future features, summarize text, and answer questions based on data it has been trained with.
Humane plans to share more about its mysterious AI-powered wearable on the same day as a solar eclipse in October.
Supabase Studio 3.0 is here, with some huge new features, including a brand new Supabase AI integrated right into the SQL Editor, Schema Diagrams, and new Wrappers.
Recent studies suggest LLMs like Othello-GPT build world models, implying that they understand the world to an extent. The network discerned the underlying 8x8 board structure and game rules by training on game moves without explicit game rules. This challenges traditional notions, like John Searle's Chinese Room Argument, about machine understanding.
Serve Robotics, the autonomous sidewalk delivery robot startup that spun out of Uber’s acquisition of Postmates, is going public via a reverse merger with a blank-check company.
One of the best ways to control image generation is to use edges to condition the final output. The best open source model, SDXL, has a controlnet variation trained and ready to use.
One hard challenge when working with language models that write code is evaluating the quality of the generated code. This is difficult because running arbitrary code is dangerous for a system. The code could do anything, including deleting everything. MosaicML offers a secure environment for evaluating generated code.
A robotic system called "Follow Anything" can spot, trace, and follow any object in real time, even if it has never seen it before. This smart robot can recognize objects from just a text description, a picture, or even a simple click - all powered by a regular laptop!
ChainML recently launched Council — an open-source AI agent platform that makes it easy to rapidly develop and deploy customized generative AI applications using collaborative AI agents. Council integrates with a wide variety of LLMs, such as GPT-4, Llama 2, and Claude 2, to unlock generative AI for business applications, enabling use cases that were previously impossible through AI agents with sophisticated control flows.
Hugging Face recently removed commercial license-ability from one of its previously open packages. The package has now been forked by Preemo and will continue to be developed as an open source package.
Mirage is a platform for generating 3D scenes. It is used by game developers to speed up the process. The team has been working internally on improving a variety of 3D generation methods. This open source push is designed to leverage the community to improve its models.
This article examines the training dynamics of a tiny model and reverse engineers the solution it finds. It provides an illustration of the exciting emerging field of mechanistic interpretability.
If you zoom in on salient pieces of an image before running classification, you can get 98% on ImageNet. Fascinating and fairly easy implementation for real world computer vision systems.
The WizardLM group from Microsoft has released the world’s best open source math model. It’s really good at reasoning through multi-step problems. The model hallucinates a bit, which is really fun for math, and offers explanations. I think this may be one of the best models for education for those interested in learning mathematics.
Nvidia makes it easy to train your own vision-language assistant based on its Nemo model.
There are lots of models based on Llama 2 and trained on GPT-4 outputs. These are somewhat in that camp, but as non-commercial license models go, they are quite powerful.
Sponsored
The semantic layer is key to delivering the ‘last mile’ of analytics — making data available in tools such as Power BI, Tableau, and Excel. But implementing it in the data-intensive insurance industry can get tricky. Join the upcoming tech talk
A research group at Stanford created a splash a few months ago with their fully simulated village with realistic people living lives. That code is now open sourced. The most interesting piece is likely how they prompt engineered each person and the overall system to work with a non-state preserving API like gpt-3.5-turbo.
OpenAI announced that it’s expanding custom instructions — a way to give users more control over how ChatGPT responds — to all users, including those on the free tier of the service.
Several media organizations have called for rules protecting copyright in data used to train generative AI models. They have signed an open letter urging lawmakers around the globe to consider regulations that require transparency in training datasets and the consent of rights holders before using data for training. The letter also requests allowing media companies to negotiate with AI model operators and identify AI-generated content, and mandates that AI companies eliminate bias and misinformation in their services.
This paper studied in-context learning with the pretrained Llama-2 language model by retrieving 9 demonstration examples during chat-style prompts. The method, which didn't alter the model's weights, resulted in a 7x win-rate increase compared to the text-davinci-003 model, making it on par with models that underwent alignment fine-tuning.
Current 3D object detectors often misinterpret data, causing them to focus on distant points. The authors of this paper have developed a new method, 3D Vertex Relative Position Encoding, that directs the detector's attention to points close to the target objects.
Traditional image-text matching models often mistakenly use highly similar but not matching samples, known as false negatives, which can misguide the learning process. This paper introduces the False Negative Elimination strategy, a novel method that carefully selects negatives by assessing their probability of being false negatives.
This repository introduces a new tool called U-CREAT, which uses event extraction to find relevant past legal cases faster and more accurately. It also presents the IL-PCR corpus, a vast collection of Indian legal cases, and shows how U-CREAT performs excellently not only in the Indian legal system but also in the Canadian system.
This repository introduces a multitasking strategy that improves the performance of vision-language models like CLIP specifically in tasks like detailed attribute detection and pinpointing locations within images.
Chidori is a reactive runtime for building AI agents. It provides a framework for building AI agents that are reactive, observable, and robust. It supports building agents with Node.js, Python, and Rust.
We’ve posted a lot about product-led growth (PLG) lately. This is another great example of a company nailing PLG. The key takeaways include being really thoughtful about who you’re building the product for, focusing on a defined persona, and then making their pathway to try your product as frictionless as possible. The traditional B2B marketing playbook didn’t land for Pinecone’s product.
Meta had one of the most prolific protein folding teams in the industry. However, the team was disbanded in the push towards the commercialization of AI systems. This, along with large labs reducing their publication output, is a sad side-effect of generative AI’s success.
SoftBank Group is establishing an AI-based warehousing joint venture with Symbotic Inc. called GreenBox Systems LLC with a combined investment of $100 million. SoftBank will also purchase $7.5 billion of Symbotic's AI systems for warehousing. This move strengthens SoftBank's commitment to robotics and AI-driven supply chain solutions.
An optimized version of the Llama 2 model from Meta.
The PAIR group at Google has released a lovely explainer that goes deep into the topic of Grokking. Grokking is the dynamic process a model goes through during training that may point to a shift from memorizing to understanding. It isn't well understood in general, but this is a lovely introduction that covers much of the groundwork for this strange phenomenon.
Don't have millions of photorealistic images to train your algorithm? Maybe you can generate them using PUG from Meta AI. It uses the powerful unreal game engine in a controllable way to generate synthetic image data for downstream training.
Sycophancy is when a model repeats and adopts a user's opinion. This happens more in larger models and instruction-tuned models. It can also occur in tasks when the opinion is irrelevant, leading to bubble-like behavior. Simple synthetic data fine-tuning can prevent this without harming overall performance.
Sponsored
Deep neural networks have excelled in computer vision, but faster inference times are needed. This paper introduces the Inter-Class Similarity Distillation method and an Adaptive Loss Weighting strategy for better knowledge transfer from a teacher network to a student one.
Generative AI systems have grown with the help of Large Language Models. The SimplyRetrieve open-source tool offers a user-friendly way to integrate private data into these systems without extra tuning using the Retrieval-Centric Generation approach. It promises enhanced AI performance while ensuring privacy.