Visualizing How LLMs Generate Text
In brief
- A new tool lets you see how transformer-based large language models (LLMs) decide which previous words influence each generated word.
- When you hover over any token in the output, the visualization shows which past tokens affected it most.
- This reveals that LLMs don't randomly guess words but strategically pick information from their training data.
- For example, when copying text like addresses or dates, the model pulls directly from the source data, minimizing errors.
- The tool uses React and Transformers.js to track attention weights during generation.
- While this version is a pilot program, it offers insights into how these models work and where they might improve.
- Future developments could make this visualization even more detailed, helping us better understand LLM behavior.
Terms in this brief
- transformer-based
- A type of neural network architecture used in LLMs that processes input data by focusing on relationships between different words or parts of text. This allows the model to understand context and generate more coherent text.
- token
- A unit of text, such as a word or punctuation mark, recognized by a language model. Tokens help models process and generate text effectively.
Read full story at Hacker News →
More briefs
A New Tool to Detect AI Agents on Your Machine
A new tool called Geiger has been developed to scan your computer for AI agents, including everything from AI plugins to server software. It provides a clear overview of what each AI component can do and access. For instance, it found that 7 out of 9 scanned components could execute code, while one had access to sensitive credentials. This tool is crucial because as AI adoption grows, so does the risk of unintended access or vulnerabilities. Geiger scans without installation, accounts, or telemetry, making it user-friendly and privacy-respectful. The creator highlights that with the rise in AI plugins and desktop clients, many users are unknowingly installing potentially powerful tools. By using Geiger, you can identify and manage these risks effectively. As AI becomes more integrated into everyday software, tools like Geiger will help keep users informed and safe.
Meta Launches Paid AI Subscription Tiers, Boosts Share Price
Meta has launched a new AI service called Muse, offering two subscription tiers priced at $20 and $100 per month. The move comes as the company looks to monetize its significant investment in artificial intelligence. On Wednesday, Meta's shares rose by 6%, reaching $649.20, while Alphabet's stock dropped 2% amid concerns that Muse might compete with Google's Gemini AI service. The paid tiers allow users to access advanced features like email management, trip planning, and payment processing through a dedicated app or WhatsApp. The service leverages Meta's new Muse Spark model, which has already increased daily interactions with its AI by 60%. This launch addresses criticism that Meta's heavy capital spending on AI had not yielded clear consumer revenue streams. While the market reacts positively to Meta's move, Alphabet faces pressure as investors worry about competition. For now, the focus remains on Meta's ability to turn its AI investments into a profitable business model.
Apple Introduces New Feature to Verify iPhone Photos' Authenticity
Apple has unveiled the Apple Reference Image, a feature designed to prove whether an image captured by the iPhone 18 Pro is authentic. This tool captures signed sensor data alongside photos, creating an "unalterable image" stored in the Photos app. Users can compare this reference image with other versions to detect any modifications. The feature aims to support photojournalists and photographers in verifying their work. Apple also announced plans to integrate this technology into third-party apps via APIs and adopt the SynthID standard for identifying AI-altered images. The iPhone 18 Pro also boasts a variable aperture on its 48-megapixel camera, enhancing low-light photography by allowing users to control the amount of light entering the lens. This upgrade is part of Apple's broader focus on improving camera capabilities, offering tools that could revolutionize how photographers and professionals handle image integrity and manipulation detection in the future.
Apple's New Watches Listen and Transcribe in Real Time
Apple has introduced a new feature in its watches that allows them to listen and transcribe audio live. This includes sounds like sirens, alarms, and even conversations when the wearer interacts with someone wearing an Apple Watch. While this could be useful for people who are deaf or hard of hearing, it raises concerns about privacy and whether consumers want their devices always listening. Apple claims these features are designed to assist, but critics worry they might normalize invasive technology. The move comes as other companies explore AI-powered transcription for different uses, like recording meetings or providing AI companions. As Apple pushes forward with this, questions remain about how it will align with growing consumer backlash against surveillance tech.
Students Build Apps with AI in West Virginia Pilot Program
Thunkable, with support from The Rockefeller Foundation, completed a pilot program teaching high school students in West Virginia how to create mobile apps using AI. Around 120 students from three counties built about 80 applications aimed at helping their peers and communities. Arun Saigalat of Thunkable highlighted the students' progress, noting they went from no coding experience to building functional apps. Andrew Sweet from The Rockefeller Foundation emphasized the importance of providing AI education in areas often overlooked by tech advancements. High schoolers nationwide lack AI training, yet demand for skilled workers is rising as AI transforms industries. With limited access to broadband in West Virginia, this program shows how hands-on learning can bridge gaps and prepare students for future job markets.