A new BUD-E for Valentines Day, AI becoming easier to integrate in products, time-series transformers, chips
In the News | Week 6
This week’s In the News segment is in a new format - a newsletter! Hopefully this makes it easier to revisit weekly news and find links.
🗞 This Week in News
Just in time for Valentine’s Day, LAION introduces BUD-E, “Buddy for Understanding and Digital Empathy”, towards more empathic and natural AI voice assistants 💕
Chips were a hot topic this past week: NVIDIA spending $30B on custom chips and Sam Altman seeking $5-7T for AI chip fabrication
Gaming and VR/AR are back in the game. Only weeks after layoffs, Epic Games and Disney team up to create the next digital universe. Lots of news this past week as people get their hands on the Apple VR headset 🥽
🥁 Interesting Products & Features
Quantized models + Cloud Workstations (localllm): Use LLMs locally on CPU and memory within the Google Cloud environment
AI Integrations now on Vercel: Use the Vercel AI SDK to integrate AI in your frontend + a new model playground that looks a lot like Replicate’s. Use AI providers including Modal, Pinecone, Replicate, Anyscale, Eleven Labs, Perplexity, Together AI, Fal, LMNT, and OpenAI.
MGIE from Apple: new open-source model that can edit images based on natural language instructions.
✨ Just launched ✨ - RadMate AI - using AI to transform radiology reporting.
📄 Interesting Papers
Learning a Decision Tree Algorithm with Transformers: This paper introduces MetaTree, which trains a transformer-based model on filtered outputs from classical algorithms to produce strong decision trees for classification. GitHub Authors from UCSD.
Discovering Temporally-Aware Reinforcement Learning Algorithms: Current reinforcement learning methods often overlook how many training steps are allowed for training (this is in opposition to how humans learn, i.e. students may change studying methods based on the proximity to deadlines). This paper proposes an updated method that lets the algorithm adjust its goals during training, leading to improved performance across different training durations. Authors from University of Oxford.
Let Your Graph Do the Talking: Encoding Structured Data for LLMs: This paper introduces introduce a parameter-efficient method to explicitly represent structured data for LLMs. “GraphToken” learns an encoding function to extend prompts with explicit structured information. Authors from Google.
🔥The time-series transformer space is heating up! 🔥:
Unified Training of Universal Time Series Forecasting Transformers: Paper introduces “Masked EncOder-based UnIveRsAl Time Series Forecasting Transformer” and “LOTSA”, the largest collection of open time series datasets with 27B observations across 9 domains. Code, data, model weights will be uploaded soon. Twitter Post here. Authors from Salesforce.
Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting: Paper that claims to be the first open-source foundation model for time series forecasting. Twitter Post here. GitHub. Authors from MilaQuebec.
What other news from the week are you excited about? Leave a comment!