Circle it to Seach iT ⭕️

PLUS: Android phones meet smart search, self-rewarding LMs surpass GPT-4, simple training data leads to complex tasks

Hey there 👋

Gargi and I are so grateful for the overwhelming love and warm wishes we’ve received from you. Your support is deeply appreciated. As we begin this new journey of marital bliss, we’re excited to reconnect with you all and continue our exploration of the ever-evolving world of AI. Here's to a week of renewed energy and insightful discoveries ahead!

Today’s top AI Highlights:

1. Circle to Search by Google

2. Self-Rewarding Language Models

3. The Unreasonable Effectiveness of Easy Training Data

4. Text or URLs to Intuitive Graphs

& so much more!

Read time: 3 mins

Latest Developments 🌍

Search ANYTHING by Just Circling ⭕️

In an era where smartphones are central to our daily lives, Google's latest feature, "Circle to Search," aims to make information discovery on Android phones more intuitive and integrated into our everyday activities. It lets you seamlessly search directly from any app, from simple gestures like circling or scribbling, thereby eliminating the need to switch between applications. You can search using both text and images simultaneously.

Key Highlights:

1. Circle to Search introduces an array of simple gestures such as circling, highlighting, scribbling, or tapping to initiate searches. This functionality is activated by a long press on the home button or navigation bar.

2. Circle to Search harnesses the power of multisearch, enabling simultaneous text and image queries. This is bolstered by the latest AI-powered upgrades, ensuring that the search results are comprehensive and relevant. This combination allows you to delve deeper into concepts, ideas, or topics with information aggregated from across the web.

3. The feature is adept at identifying items within photos or videos, offering quick access to similar products available online. For instance, in a fashion video, you can circle an accessory like sunglasses to find similar items from various retailers. The tool also excels in answering complex queries; like you can circle an image with a “corn dog” and pose a question to know more about it.

Self-Rewarding LMs to Surpass GPT-4 🚀

Creating superhuman agents requires superhuman feedback for optimal training. But current approaches are constrained by human performance levels and hindered by static reward models that fail to improve during language model training. Researchers have introduced the idea of language models providing their own rewards during training. The new model achieved via fine-tuning Llama 2 70B has demonstrated an ability to outperform even GPT-4 on the AlpacaEval 2.0 leaderboard.

Key Highlights:

1. The core of this approach lies in the language model's ability to act as its own judge. By employing an "LLM-as-a-Judge" prompting system, these models not only follow instructions but also create and assess new instruction-following examples. This self-rewarding mechanism has led to notable improvements in both the model's ability to follow instructions and its capacity to provide quality self-assessments.

2. A key feature of the Self-Rewarding Language Models is their ability to self-modify their training set. This process involves generating new prompts and diverse candidate responses, which the model itself then evaluates. Such a dynamic training regime allows the model to continuously improve by generating more challenging and varied training scenarios. This self-sustaining loop of creating and evaluating content is a crucial factor in the model's progressive learning and refinement over time.

3. The models are enhanced through Iterative DPO (Direct Preference Optimization) training. In a striking result, the Llama 2 70B model, after being fine-tuned through three iterations of this training, outperformed several leading models on the AlpacaEval 2.0 leaderboard, including Claude 2, Gemini Pro, and GPT-4 0613.

The Surprising Power of Easy Data 📊

Training language models to excel at complex tasks using difficult-to-label training data poses a significant challenge, known as the scalable oversight problem. This problem has become increasingly relevant as LMs advance. However, a new study offers a strikingly different perspective. It reveals that LMs can generalize effectively from easy training data to hard tasks, sometimes matching the performance of models trained on the hard data itself.

Key Highlights:

1. The research demonstrated that LMs are capable of generalizing from simple training data to complex tasks with a high degree of effectiveness. Models trained on straightforward data like 3rd-grade science questions showed nearly comparable performance on hard test data, such as college-level STEM questions.

2. The study's findings were consistent across a range of model scales, from 7 billion to 70 billion parameters, and were validated using various human and model-based measures of data hardness.

3. A key takeaway from the research is the potential for more efficient and cost-effective AI training methods. Since collecting and accurately labeling complex, domain-specific data (like in medicine or law) can be expensive and error-prone, training models on easier, more reliably labeled data could save significant resources.

---

Tools of the Trade ⚒️

1. InstaGraph: Convert text or URLs into knowledge graphs to visually represent relationships between different entities within a topic. It simplifies complex information into an easily understandable graph format.

Source

2. Quivr: Your second brain powered by generative AI, that helps retrieve information from unstructured data. It can ingest almost any type of file, images, text, code snippets, and more.

3. AnswerFlow AI: is a platform for creating custom ChatGPT bots using personal data sources. You can connect various data sources, like documents, Excel sheets, PDFs, and links, to build a knowledge base, and embed them in your websites.

4. Txtai: All-in-one embeddings database designed for semantic search, LLM orchestration, and various language model workflows. It integrates vector indexes, graph networks, and relational databases, enabling features like vector search with SQL, topic modeling, and RAG.

😍 Enjoying so far, TWEET NOW to share with your friends!

---

Hot Takes 🔥

1. Stackoverflow was such a wonderful resource, got scraped into chatgpt, and now gets no use while OpenAI makes money. Not sure what the answer here is but it feels pretty bad ~ Eugene Vinitsky

2. My vague hunch is that further progress may be algorithmic and I'll openly admit I did not predict feeling this way this even though in principle I could have. ~ John David Pressman

---

Meme of the Day 🤡

Source

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

Real-time AI Updates 🚨

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!