Perplexity AI's combination of a web index and up-to-date information with a conversational AI chatbot-style interface has the potential to dethrone Google. The company recently released fine-tuned and augmented versions of open source models from Mistral and Meta. The models aim to offer helpful, factual, and up-to-date information. They are the first-ever live LLM APIs that are grounded with web search data and have no knowledge cutoff.
2023.12.05·venturebeat.com·6 min read·TLDR AI When slotting in language models to your code base, a useful mental model is treating them like functions with standard input and output. This blog, from the creator of React Native, shows a few ways you can do that and the benefits of modeling your models this way.
2023.12.05·blog.vjeux.com·8 min read·TLDR AI Another state spaces paper with improved scaling and performance. In this case, they borrow ideas from the LSTM and train a 3B parameter model that outperforms larger 7B parameter Transformer models.
2023.12.05·arxiv.org·24 min read·TLDR AI MotionEditor is a diffusion model designed for video motion editing, skillfully balancing motion manipulation and original content preservation. It introduces a content-aware motion adapter and a unique two-branch architecture with attention injection, allowing seamless integration of edited motions while maintaining the original background and protagonist's appearance.
2023.12.05·arxiv.org·16 min read·TLDR AI Sponsored
AI.dev is a must-attend event for anyone looking to shape the trajectory of the ever-changing open source generative AI and machine learning landscape. Check out the speakers, schedule, workshops, activities and more! Register now [14] with code TLDR_AI to save $300. Special pricing is also available for hobbyists, academics, and students.
LastMile AI open-sourced a library, AI Config, to decouple and abstract away models, parameters, and prompts. AI Config allows developers to easily swap between model providers and develop model-agnostic applications.
2023.12.05·github.com·GitHub repository·TLDR AI 2023.12.05·shinying.github.io·2 min read·TLDR AI The new IMMA technique offers a unique solution to safeguard text-to-image models from being maliciously adapted to generate harmful content.
2023.12.05·zhengyjzoe.github.io·4 min read·TLDR AI The speaker in the talk highlights the distinction between interpersonal and social trust, cautioning that our tendency to anthropomorphize AI as friends rather than services can be exploited by profit-driven corporations. They advocate for government intervention through transparency laws and regulations aimed at humans behind AI and call for the development of public AI models to ensure AI remains a trustworthy and beneficial service in society.
2023.12.05·schneier.com·10 min read·TLDR AI This article criticizes the AI doomer movement within OpenAI for strategic blunders and counterproductively accelerating AI development rather than ensuring its safety.
2023.12.05·infinitescroll.us·8 min read·TLDR AI Microsoft released a 12 lesson course on generative AI for beginners.
2023.12.05·microsoft.github.io·TLDR AI Sponsored
Reduce development time and data needs with the Navigator AI IDE. Build with full code or no code. Easily train on your hardware, retaining ownership of data and models. Deploy anywhere. Join the
2023.12.05·webai.com·Sponsor·TLDR AI Fast and memory-efficient LLM tuning.
2023.12.05·github.com·GitHub repository·TLDR AI GitHub is "re-founding" Copilot, recognizing the impact of AI on developer workflows. It will increase efficiency with AI-assisted processes like autocomplete code and may even restructure entire workflows. This shift allows individualized workflows, but must balance flexibility with the ability to provide large-scale customer support.
2023.12.05·matt-rickard.com·2 min read·TLDR AI SchoolXpress turns your books, notes, classwork, and documents into interactive learning content that simplifies concepts and delivers on your learning objectives.
OpenAI’s GPT Store will be delayed until next year due to unexpected things keeping the company busy.
2023.12.04·theverge.com·1 min read·TLDR AI Amazon's newly launched AI chatbot, Amazon Q, is reportedly facing issues with hallucinations and data leaks that include sensitive information about AWS data centers.
2023.12.04·datacenterdynamics.com·2 min read·TLDR AI OpenAI's former CEO and co-founder, Sam Altman, has returned to the company after a boardroom coup. In a recent interview, Altman refrained from discussing the reasons behind his dismissal but shared his feelings of defiance and eventual acceptance of his return. OpenAI's biggest investor, Microsoft, is also planning to take a non-voting board seat.
2023.12.04·theverge.com·6 min read·TLDR AI Most, but not all, modern diffusion models use the attention mechanism. The recent wave of interest in state spaces has made its way to diffusion, giving theoretical speed ups and opening up interesting applications.
2023.12.04·arxiv.org·18 min read·TLDR AI The authors of foundational work in this space have a new project that uses novel encoder methods to give fine-grained control of the final animation generation.
2023.12.04·ericguo5513.github.io·23 min read·TLDR AI A new method harnesses the extended StyleGAN embedding space W+ for text-to-image diffusion models, improving identity preservation in generated images.
2023.12.04·csxmli2016.github.io·2 min read·TLDR AI Lots of work has been done in distillation and initialization such as the lottery ticket hypothesis. This work gives a simple mechanism to take a much larger model and initialize a smaller one from it. It dramatically improves smaller model performance.
2023.12.04·github.com·GitHub repository·TLDR AI MaXTron is a simple yet effective unified meta-architecture for video segmentation.
2023.12.04·github.com·GitHub repository·TLDR AI This project introduces SAFE, a framework that combines RGB frames, event streams, and semantic labels using pre-trained vision and language models.
2023.12.04·github.com·GitHub repository·TLDR AI Stability AI has engaged in preliminary discussions with multiple companies for an acquisition due to mounting pressure from investors over its financial position.
2023.12.04·pymnts.com·2 min read·TLDR AI GitHub's Chief Product Officer, Inbal Shani, discusses AI's role in software development, arguing AI-driven code generation won't replace developers but instead enhance their efficiency. She explores the success metrics and philosophy behind GitHub’s Copilot and how the company fosters innovation. This conversation sheds light on AI's future in the tech industry.
2023.12.04·lennysnewsletter.com·TLDR AI This article emphasizes the role of grand strategy in tech, comparing modern tech leaders' ambitious achievements to historical conquests and urging them to think and act like historical figures in aligning aspirations with capabilities. It highlights recent events at OpenAI as an example of effective grand strategy and encourages aspiring leaders to craft strategies that align their capabilities with their ambitions for impactful, positive change.
2023.12.04·notboring.co·17 min read·TLDR AI Google has postponed the launch of Gemini, its next-generation foundation model from next week to January due to the AI's inconsistent handling of non-English queries.
2023.12.04·9to5google.com·1 min read·TLDR AI The US government and Intel have started training a one trillion parameter model called ScienceGPT.
2023.12.04·techradar.com·1 min read·TLDR AI Use aiPDF to summarize, get insights from, and chat with any file.
Sam Altman is back as CEO of OpenAI, and the new initial board consists of Bret Taylor (Chair), Larry Summers, and Adam D’Angelo.
2023.12.01·openai.com·4 min read·TLDR AI DeepMind and Berkeley teamed up for research that uses graph neural networks to discover over 2 million novel materials. They found around 300k of them that are extremely stable and great candidates for future research. Interestingly, they say that this amount of research would have taken 800 years of dedicated human labor.
2023.12.01·deepmind.google·6 min read·TLDR AI Software Applications Incorporated, a startup led by ex-Apple employees, aims to bring generative AI to desktops by capturing the flexibility of '80s and '90s computing. The company is using large language models, including OpenAI’s GPT and Meta’s Llama 2, in its prototypes. It has raised $6.5 million and plans to hire up to 10 employees.
2023.12.01·theverge.com·4 min read·TLDR AI Seamless is a system that revolutionizes automatic speech translation. This advanced model not only translates across 76 languages, but also preserves the speaker's unique vocal style and prosody, making conversations feel more natural.
2023.12.01·ai.meta.com·4 min read·TLDR AI Researchers introduce DECOLA, a groundbreaking open-vocabulary detection framework that excels in identifying objects beyond its training dataset.
2023.12.01·arxiv.org·14 min read·TLDR AI This blog post from the PyTorch team highlights how to use native Pytorch code to dramatically speed up language model inference. The post shows how you can get 200+ tokens per second from Llama 2 7B.
2023.12.01·pytorch.org·12 min read·TLDR AI The Alibaba Cloud Group has released a universal audio understanding model. It can perform a number of audio-related tasks like understanding music, speaker identification, and then answer questions in text.
2023.12.01·qwen-audio.github.io·GitHub repository·TLDR AI 3D to Photo is an open-source package that combines ThreeJS and Stable Diffusion to build a virtual photo studio for product photography.
2023.12.01·github.com·GitHub repository·TLDR AI This project presents an evaluation framework using adversarial prompts to test how trustworthy safety-driven unlearning methods are in diffusion models.
2023.12.01·github.com·GitHub repository·TLDR AI ChatGPT, launched a year ago by OpenAI, unexpectedly became the fastest-growing consumer technology, profoundly impacting the tech industry from venture capital investments to major companies like Microsoft and Google, and while it faces challenges in reliability and ethical concerns, it continues to rapidly evolve and integrate into various aspects of technology and daily life.
2023.12.01·theverge.com·8 min read·TLDR AI Freshworks CIO, Prasad Ramakrishnan, walks through some viable use cases of AI in startups. From enhancing user experience to onboarding and data platform optimization, this article shares five ways organizations can leverage AI to solve problems and grow.
2023.12.01·saastr.com·6 min read·TLDR AI Cambridge researchers found that imposing physical constraints on AI systems, mirroring the human brain's limitations, helps them develop brain-like features for efficient problem-solving and could enhance our understanding of cognitive differences and AI design.
2023.12.01·cam.ac.uk·3 min read·TLDR AI Despite Sam Altman's reinstatement as OpenAI's CEO, the company's employees remain uneasy due to the board's silence on the reasons behind his temporary removal.
2023.12.01·businessinsider.com·2 min read·TLDR AI The Humane AiPin is a wearable tech that’s enabled with generative AI technology.
AI researchers Shriyash Upadhyay and Etan Ginsberg have launched a startup, Martian, supported by $9 million in funding. Martian created a "model router" tool that optimizes the use of large language AI models such as GPT-4 by smartly routing requests to the most suitable model for the task, making it more cost-efficient. The founders believe this approach promotes fundamental rather than competitive research.
2023.12.01·techcrunch.com·3 min read·TLDR AI Led by Kleiner Perkins, this round will help boost the training and inference products that have seen massive adoption since Together’s launch in June.
2023.11.30·together.ai·2 min read·TLDR AI Amazon trained a generative model to assist users of the AWS platform. It will also be used for generic business assistance. The model is a private system that answers user queries about various aspects of Amazon's backend tooling.
2023.11.30·aws.amazon.com·3 min read·TLDR AI Powering workflows for customers like Character AI, Voltage Park is a new cloud provider with a 24k H100 mega cluster. Its pricing seems to be industry-leading at less than $2/hour per GPU.
2023.11.30·blog.voltagepark.com·1 min read·TLDR AI