Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sony's AI research suggests a more "multidimensional" approach to measuring skin color, emphasizing not just lightness/darkness but also red and yellow hues. Current scales might overlook biases against certain ethnic groups. Sony recommends the CIELAB color standard, even as others tout the simplicity of systems like the Monk Skin Tone Scale. Google and Amazon are reviewing Sony's findings.
Sponsored
Scott Galloway’s education company, Section, has put together an incredible library of AI business frameworks you can use right now. Get your free AI business frameworks here.
Adobe unveiled a sneak preview of Project Stardust, an AI-powered photo editor, at its MAX conference. Using the Firefly Model 2, Stardust lets users effortlessly modify scenes, replace backgrounds, and remove objects.
GitHub Copilot costs users $10 per month, but it is costing Microsoft an average of $20 per user per month. Some users are costing the company as much as $80 per month. This is likely why Microsoft decided to charge a lot more for the AI capabilities in Microsoft 365 Copilot. The costs are also likely why the company is developing its own in-house AI chip sets for its data centers and pushing the adoption of Neural Processing Units across the industry.
Sponsored
Want to know what users are doing inside your app? Why not just ask? TLDR members get EXCLUSIVE 3 months free*
Replit has its own coding assistant that runs across its products. It is quite powerful, trained with Mosaic, and free for all user tiers.
Google Cloud introduced AI-enhanced search tools to streamline access to diverse medical records for healthcare professionals. While the technology aims to reduce administrative work, its integration is undertaken with caution to fit seamlessly into existing workflows and ensure HIPAA compliance.
OpenAI's Residency is an initiative to empower exceptional researchers and engineers from diverse fields to transition into the world of AI and machine learning. It is especially beneficial for researchers specializing in fields outside of deep learning and welcomes software engineers seeking to pivot into AI research roles. Residents will tackle real AI challenges while receiving a full salary.
LLMs have been trained on massive amounts of data, much of which is copyright protected. Removing or unlearning such data is a challenge. This work uses a proxy model as a teacher to simulate a clean model and further fine-tunes the base model, which reduces knowledge of undesirable information while maintaining benchmark performance.
Researchers present a solution, called Multi-Head Self-Attention Convolution (MSA-Conv), which allows Vision Transformers (ViTs) to handle images of different sizes without needing changes in their design or the need to resize images.
This study introduces a novel Transformer model that is tailored to blind super-resolution, a task that deals with images with unknown noise. The method outshines current approaches through its blend of traditional CNNs and a new Transformer design attuned to degradation patterns.
Sponsored
If you've tried to build automated test coverage in-house, you know it takes years to scale. Try QA Wolf
Attention is useful but computationally expensive. However, once trained, with a bit of fine-tuning compute you can drop in SRF attention and remove the dependence on sequence length, speeding things up considerably.
A Pythonic framework to simplify AI service building.
This study delves into how different architectural elements in machine learning models affect their performance in federated learning, especially when handling varied data sources. Through in-depth analysis, researchers found that with certain tweaks, Convolutional Neural Networks (CNNs) can perform as well or even better than Vision Transformers in such settings.
There is a heavy shortage of GPUs for the current AI demand. Many services have long queue times or are restricting access as a result. If other chip makers can overcome some of their software deficiencies, maybe via recent Pytorch improvements, they may be able to compete for some of this demand.
AI is often misunderstood as a one-size-fits-all solution, but businesses need to recognize that it's merely a tool, and not every business requires it. Identifying genuine problems and needs should precede technology adoption, rather than trying to retrofit AI solutions where they may not be needed or beneficial.
Great read on the tricky parts of creating value around generative AI for startups. While the incumbents are favored, finding your path and creating real value early is where startups can have a fighting chance.
ChatGPT's mobile apps saw a record 15.6 million downloads and $4.6 million in September, but growth is starting to slow down.
Meta is paying creators up to $5 million for just six hours of studio time to use their likeness.
Sponsored
️️ Big Tech acquired startups in the smart doorbell and smart thermostat categories for billions. If you missed out on these spectacular early investments in the Smart Home space, here’s your chance to grab hold of the next one. RYSE is the tech firm that’s poised to dominate the Smart Shades market (growing at an astonishing 55% annually), and their public $1.25/share offering is opening for a limited time!
Microsoft may unveil its own AI chip at the upcoming Ignite 2023 conference, ensuring the company's AI advancements remain uninterrupted and potentially positioning it to compete profitably with peers like Amazon and Google, who also use proprietary AI chips.
OpenAI is considering creating its own AI chips due to the increasing shortage in GPU supply.
The BBC is setting principles for using generative AI in journalism, emphasizing transparency and public interest. While it is exploring AI's potential, the broadcaster has blocked OpenAI web crawlers to safeguard copyrighted content.
Researchers have developed a new technique to boost the math-solving skills of open-source AI models by teaching them to use code.
This paper introduces MLAgentBench, a testing ground for AI to conduct machine learning experiments, just like human researchers.
LightSeq is a novel method for training large language models with extended context lengths that addresses the limitations of previous systems like Megatron-LM.
Streaming music generation usually has to wait for all the music to be completed. However, this means real time or long generation applications struggle. With this code, you can now stream the chunks of music as they arrive.
Researchers highlight potential safety issues when further customizing pre-trained large language models like Meta's Llama and OpenAI's GPT-3.5 Turbo.
Finetune Llama2-70b and CodeLlama on a MacBook Air without quantization.
AI’s broad potential raises questions about its real-world utility and reliability. As the technology matures, its evolution will require a shift from a generalized solution to one with defined, specific applications.
Stability has trained another model, this time on 4T tokens to test data repeating. Also, due to public pushback, they removed Books3 from the data mix.
We’ve seen how open source models can work and thrive over the last 25+ years. Even as they mature and become more “closed”, the open source model has lasting impacts. In the face of AI, not all playbooks are the same. This is a thoughtful and in-depth review of open source models and how AI will play in this.
OpenAI has unveiled the Beta version of its Python SDK, marking a significant step towards enhancing access to the OpenAI API for Python developers.
A modern business intelligence and data visualization system.
The Arc browser has introduced AI-enhanced features under the name "Arc Max". These features utilize OpenAI's GPT-3.5 and Anthropic’s models for functions such as renaming pinned tabs, summarizing link content, and renaming downloaded files based on their content.
Google's upcoming Gemini model may outshine GPT-4. As AI gains vision and voice capabilities, its applications and potential concerns grow. The future of AI hinges on responsible human use and decision-making.
Google's Pixel 8 and 8 Pro offer AI-enhanced photo edits. While some laud the technology, others voice ethical concerns that underscore the changing landscape of photography. Google plans to flag AI-edited images with metadata.
Detecting and naming new objects in 3D scenes is a tough challenge. This project introduces CoDA, a new method that does this by smartly guessing where these new objects might be and then matching their features with images and texts, all in one system.
Mechanistic interpretability is the process of understanding neural networks by breaking them down into more interpretable sub-pieces. Unfortunately, neurons tend to not be interpretable by themselves. There are a number of deep reasons like superposition that lead to this challenge. This work by Anthropic uses a sparse autoencoder to generate interpretable representations from a one-layer Transformer. It’s a big step forward in our understanding of these models.
Event cameras are special sensors that work differently from regular cameras. This new method, the Group Event Transformer (GET), better understands these cameras by looking at both the space and time details of events.
There are different challenges when training language models on multi-GPU and multi-node systems. It is also very hard to gain experience in those cases. This blog post is extremely detailed and dives into ZeRO, FSDP, and gotchas when training or fine-tuning models at scale.
Meta Scientists have developed a new method to understand the words and phrases people think or hear by studying their brain waves - without needing any surgery.
Ctoc provides a lightweight tool for analyzing codebases at the token level.
A peek into the AI pause debate, considering what might happen if AI research was simply paused for 6 months, paused and restarted at exactly the right time, paused long enough to develop good regulation, paused until alignment was completely solved, or not paused at all.
Computer scientists are using GPT-4 to address the longstanding P=NP problem. A group of researchers from Microsoft and Peking University guided GPT-4 using a Socratic Method with a series of prompts to explore the problem, leading to insights that suggested P does not equal NP. The method represents a new frontier in prompt engineering, potentially making chat sessions with AI more focused and in-depth.
Get an inside look at what it’s like to live in a $68M AI hacker house with other AI founders and builders.
Google’s DeepMind, in collaboration with 33 research institutes, has developed Open X-Embodiment, a database akin to the ImageNet database in computer vision. This database, comprising over 500 skills and 150,000 tasks from 22 robot types, aims to propel robotics research by offering a shared resource to train more versatile robot models.
Artifact, originally an AI-driven news app by Instagram’s co-founders, now lets users leverage AI to create cover art for their posts. This enhancement, which crafts “eye-catching images” based on user-inputted keywords, further positions Artifact as a competitor to platforms like Instagram and X.
Sponsored