Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Evaluating language models is typically done with hand-curated benchmarks. Some of these benchmarks are quite large, with some at over 14k examples, which leads to high expense and noise in evaluation. This work shows that you can reliably evaluate language model performance with as few as 100 examples from popular benchmarks.
A great tutorial on the different types of parallelism for training large models at scale.
DP3 introduces a cutting-edge approach to imitation learning, combining 3D visual data and diffusion policies to teach robots complex skills efficiently.
Sponsored
You don’t need an army of testers to build, run, and maintain more tests. QA Wolf (4.8 ⭐ on G2) [11] brings AI-level scale and efficiency to testing, with unlimited parallel test runs on their cloud infrastructure and a zero-flake guarantee. 80% automated test coverage [11] in just 4 months. Trusted by Cohere, Autotrader, Salesloft and many others.
Genstruct is a model that can be used to generate instructions from natural language text. This can then be used to train downstream models for specific useful tasks.
Fructose is a Python package for creating dependable strongly-typed interfaces around LLM calls.
A novel method enables large language models to work together by taking turns generating text, token by token. This strategy allows models to leverage their unique strengths and expertise for various tasks, leading to improved performance across instruction-following, domain-specific questions, and reasoning challenges.
There is an almost 10x difference in speed between different implementations of the popular multi-headed attention module used in modern LLMs. This notebook shows a few and benchmarks their performance.
Technology alone is no longer enough to provide a competitive advantage for GenAI companies. This makes GenAI and bottled water similar, as the underlying product is essentially the same. The key differentiators must come from factors such as user experience, distribution, perceived value to customers, branding, and marketing.
A few years ago, the concept of the metaverse seemed like it was just fiction from Ernest Cline's book, and Facebook's metaverse push seemed ill-fated. Today, things feel different. Sora generates complex video scenarios and Groq's deterministic chips reduce machine-learning algorithms for faster processing. Together, the convergence of these advancements pushes us closer to achieving real-time video simulation and true virtual reality.
Microsoft has unveiled a new Copilot chatbot tailored for finance tasks, with a public preview to come before pricing details are announced. The AI can conduct variance analysis, reconcile data in Excel, and expedite Outlook collections, integrating with SAP and Dynamics 365. The tool aims to reduce time spent on manual tasks for finance teams, potentially transforming financial planning and analysis efficiency.
Microsoft's €15 million investment in French AI firm Mistral while continuing its strong partnership with OpenAI illustrates its strategy to diversify and bolster Azure's AI offerings and gain a foothold in the European market ahead of the EU's AI Act. Azure aims to become a model garden, offering a variety of AI models to attract developers and compete with other cloud services. The Mistral deal also positions Microsoft strategically within European AI regulatory frameworks.
The SEC is investigating OpenAI, focusing on CEO Sam Altman's internal communications after accusations that he misled investors. This inquiry coincides with global attention on OpenAI's ChatGPT and scrutinizes the competitive impact of AI investments by OpenAI and tech giants in the sector.
Sponsored
Don't miss your chance to experience the latest AI breakthroughs in-person at NVIDIA GTC [4], the #1 AI conference [4]. NVIDIA CEO Jensen Huang will share the latest AI advances that are shaping our future, live at the SAP center in San Jose.
Mistral and Snowflake have partnered to bring Mistral Large to Snowflake Cortex customers.
Moondream is a small Phi-2 and SigLIP-based language model. This second version is commercially licensed and has significantly improved benchmark performance. It is ideal for running on low compute devices and describing images.
Waymo is starting tests of its fifth-generation driverless Jaguar I-Pace in Austin as it advances towards launching a ride-hail service. Initially, rides will be for Waymo employees, with plans to open to the public later. Waymo's expansion in Austin contributes to its roadmap of providing autonomous vehicle services in multiple cities.
Taking images of a design and outputting code is a challenging task. This work proposes a benchmark, 18B model, and evaluations to suggest that we are close to being able to perform this on simple designs. In some cases, GPT-4V-generated code is preferred over human-synthesized code.
Researchers used synthetic data to generate 2 million path problems. They then trained a 7B model and found strong performance against state-of-the-art large language models.
The KEPP system introduces a new approach to planning and executing complex tasks. The method, which leverages a probabilistic knowledge graph, allows the model to logically sequence actions towards achieving a goal.
MLX is a framework developed by Apple for training AI models on Mac. This repository shows how to train models that can run on mobile devices with Swift. It takes one on an iPhone to train an MNIST classifier model.
An open source text clustering tool that allows you to embed, cluster, and semantically label clusters easily and quickly. The entire pipeline runs in 10 minutes on 100k samples.
KnowAgent introduces a novel strategy for boosting large language models' planning skills by integrating explicit action knowledge. The approach helps LLMs perform better in complex tasks by guiding them through more logical planning trajectories.
A Waymo self-driving car was vandalized and set on fire in San Francisco, an act seen by some as a protest against the imposition of autonomous vehicles. Such hostility overlooks the potential for self-driving cars to drastically reduce traffic fatalities. As the technology improves, autonomous vehicles will enhance safety and complement public transport.
OpenAI's Sora is a text-to-video model capable of generating high-fidelity videos with diverse aspect ratios and resolutions that also exhibits world simulation abilities. Sora’s vast capabilities raises both excitement and concerns about the implications for reality simulation and AGI.
OpenAI's Sora exhibits significant flaws, revealing the limitations of generative AI's understanding. The technology's potential impact on the creative industry and its heavy computational demands highlight both its revolutionary capabilities and practical challenges.
Sponsored
The role of workers in the age of AI is constantly evolving. How do we cultivate trust and move away from a human-vs-AI mentality? Read the article by Dow Jones Factiva [17].
Easy way to train, tune, evaluate, and host language models in Jax.
Stack Overflow has launched OverflowAPI, which allows AI companies like Google to integrate its vast knowledge base. Its partnership with Google will enrich Google Cloud's Gemini service and bring AI-powered features to Stack Overflow.
Brave has expanded its AI-powered assistant Leo to Android users. Leo, available without a Brave account, with a premium subscription option across devices, offers functionalities such as summarizing content, answering questions, and creating text.
Sponsored
Google Cloud’s upcoming developer conference [5]is packed with immersive learning, community interactions, and a peek into the future of Google Cloud — with a focus on the latest tools for building AI-powered applications [6].
Secure your spot now with the special rate of $1,299 USD by using the code LEARNATNEXT by 11:59 PM PT by March 10, 2024.
OpenAI responded to Elon Musk's lawsuit by sharing its side of the story along with screenshots of email exchanges between Sam Altman, Ilya Sutskever, Greg Brockman, and Elon Musk. The receipts show that Musk believed OpenAI’s chance of success was 0% and agreed that there was a need to close source certain models.
Stable Diffusion 3, with its novel Multimodal Diffusion Transformer architecture, surpasses leading text-to-image models by enhancing prompt comprehension and typography through separate processing weights for text and images, promising advancements in AI-generated visual content.
AI advancements like OpenAI's Sora and other platforms are enhancing video production, potentially streamlining Hollywood's expensive and complex filmmaking process. These tools raise questions about the future of creative roles and the industry's ability to adapt to AI's increasing capabilities in scripting, scoring, and editing films. Hollywood faces a paradigm shift akin to the music industry's digital transformation, with a mix of optimism and concern for the impact on traditional filmmaking.
Language models are now exceptionally skilled at answering questions. As a result, most existing benchmarks have been saturated. There is a new type of question, dubbed 'researchy', which is open-ended and requires multiple steps. This particular dataset is mined from search engine queries. It contains cases where GPT-4 struggled to answer queries.
UniCtrl introduces a new approach to enhancing semantic consistency and motion quality in videos created by text-to-video models. It improves video realism and coherence without the need for extra training by using cross-frame self-attention and motion injection techniques.
VTG-GPT offers a novel GPT-based method for accurately identifying specific segments in videos using natural language queries without the need for training or fine-tuning.
MobileClip runs 7x faster with the same performance as OpenAI’s original CLIP model. It can be used on-device for various vision and language tasks.
Microsoft has released an Orca math word problems dataset. This was used to train some of the current state-of-the-art math models. It can now be used for research by the community.
The founder of Reka and famous GPU critic Yi Tay wrote a blog post outlining their experience training extremely strong language models outside of Google. The main challenges come from cluster and hardware instability. They also found software maturity to be a challenge.
Anthropic's release of Claude 3, a large language model comparable to GPT-4, stands out not for its benchmark tests or cost efficiency, but for its uniquely human-like, creative, and naturalistic interaction quality, making it a significant advancement in AI's ability to partner creatively with writers.
Regulating AI like a tangible product is misguided because AI is a broad process involving statistical methods, and overseeing its use across all companies would be impractical. Targeted regulation for specific misuses, similar to existing approaches for programming or email abuses, is more effective given AI's impending ubiquity in the economy.
Sponsored
Try Dollar Flight Club for just $1 [24] and get access to the best travel deals, up to $500 off your next trip. For the next 24 hours, try the club for just $1 [24]—make summer 2024 your best one yet here.