Parse notes
Google has introduced Gemini 2.0 Flash Thinking Experimental, a reasoning AI model available in AI Studio for tasks like programming and math. The model aims to improve accuracy by internally verifying its reasoning, though it currently struggles with simple tasks like letter counting.
On the 11th day, OpenAI introduced more details about working with the OpenAI desktop app.
Anthropic explores how AI systems might superficially align with human values while pursuing their own objectives, offering insights into detection and mitigation strategies.
The best and most commonly used time series forecasting methods rely entirely on numerical features associated with the time series. However, they rarely use textual or semantic information about the task itself (e.g., are you predicting electric prices or customer churn). When prompted with this textual information in context, language models surpass all tested time series forecasting methods substantially across a wide range of carefully decontaminated tasks.
By using low cost LiDAR (found in modern iPhones) along with a depth estimation foundation model, this method shows that you can prompt the system to reconstruct an extremely high fidelity point cloud better than either method alone and on par with extremely expensive LiDAR setups alone, such as those found on self driving cars.
BERT is an encoder only style model that is one of the most heavily used language models on the market, used in Google search for essentially every query. This new model from Answer AI, LightOn, and friends is faster, modern, and extremely performant. It can be used as a drop in replacement. They use a number of modern tricks like batch ramp to improve overall model performance.
AniDoc is a line-filling method that can colorize anime based on a character reference image and a line art series of keyframes.
This paper introduces GaussTR, a novel Gaussian Transformer that leverages alignment with foundation models to advance self-supervised 3D spatial understanding.
An open-source tool for Computer-Aided Diagnosis, offering a modular and scalable platform for medical imaging research and development.
AI apps are flooding the App Store charts, but many prioritize cash grabs over quality, often misleading users with paywalls and subpar features. HUBX, dominating with apps like DaVinci AI, faces criticism for poor functionality and misleading marketing despite high rankings. Amidst the saturation, quality apps like Photoroom and Picsart AI stand out by delivering on their promises with effective AI tools.
Kelvin My from Translink Capital shares a recap of everything that's happened in AI in 2024. The report provides an overview of each of AI's four layers: infrastructure, foundational model, tooling, and application. For each layer, it discusses key takeaways, future trends in 2025 and beyond, and lists some noteworthy startups.
A research initiative focused on spatial reasoning and AI models designed to interpret and interact within three-dimensional spaces.
Together AI has a new product that enables users to deploy a custom LoRA at the cost of the base model via serverless switching.
Researchers at the University of Washington and Harvard have used machine learning to create personalized blood test references, enhancing disease prediction accuracy.
Ilya Sutskever predicts that pre-training in AI will end due to the limited availability of new data.