Monthly Archives: November 2024
AIs Discovering Vulnerabilities
I’ve been writing about the possibility of AIs automatically discovering code vulnerabilities since at least 2018. This is an ongoing area of research: AIs doing source code scanning, AIs finding zero-days in the wild, and everything in between. The AIs aren’t very good at it yet, but they’re getting better.
… Since July 2024, ZeroPath is taking a novel approach combining deep program analysis with adversarial AI agents for validation. Our methodology has uncovered numerous critical vulnerabilities in production systems, including several that traditional Static Application Security Testing (SAST) tools were ill-equipped to find. This post provides a technical deep-dive into our research methodology and a living summary of the bugs found in popular open-source tools. — Read More
New Zemeckis film used AI to de-age Tom Hanks and Robin Wright
On Friday, TriStar Pictures released Here, a $50 million Robert Zemeckis-directed film that used real time generative AI face transformation techniques to portray actors Tom Hanks and Robin Wright across a 60-year span, marking one of Hollywood’s first full-length features built around AI-powered visual effects.
The film adapts a 2014 graphic novel set primarily in a New Jersey living room across multiple time periods. Rather than cast different actors for various ages, the production used AI to modify Hanks’ and Wright’s appearances throughout.
The de-aging technology comes from Metaphysic, a visual effects company that creates real time face swapping and aging effects. During filming, the crew watched two monitors simultaneously: one showing the actors’ actual appearances and another displaying them at whatever age the scene required. — Read More
How ChatGPT search paves the way for AI agents
OpenAI’s Olivier Godement, head of product for its platform, and Romain Huet, head of developer experience, are on a whistle-stop tour around the world. Last week, I sat down with the pair in London before DevDay, the company’s annual developer conference. London’s DevDay is the first one for the company outside San Francisco. Godement and Huet are heading to Singapore next.
It’s been a busy few weeks for the company. In London, OpenAI announced updates to its new Realtime API platform, which allows developers to build voice features into their applications. The company is rolling out new voices and a function that lets developers generate prompts, which will allow them to build apps and more helpful voice assistants more quickly. Meanwhile for consumers, OpenAI announced it was launching ChatGPT search, which allows users to search the internet using the chatbot. Read more here.
Both developments pave the way for the next big thing in AI: agents. These are AI assistants that can complete complex chains of tasks, such as booking flights. (You can read my explainer on agents here.) — Read More
“How Could Machines Reach Human-Level Intelligence?” by Yann LeCun
OpenAI’s search engine is now live in ChatGPT
ChatGPT is officially an AI-powered web search engine. The company is enabling real-time information in conversations for paid subscribers today (along with SearchGPT waitlist users), with free, enterprise, and education users gaining access in the coming weeks.
Rather than launching as a separate product, web search will be integrated into ChatGPT’s existing interface. The feature determines when to tap into web results based on queries, though users can also manually trigger web searches. ChatGPT’s web search integration finally closes a key competitive gap with rivals like Microsoft Copilot and Google Gemini, which have long offered real-time internet access in their AI conversations. — Read More
Scaling and evaluating sparse autoencoders
Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer. — Read More
Evaluating feature steering: A case study in mitigating social biases
A few months ago, we published an interpretability paper demonstrating our ability to learn interpretable features that correspond to various concepts (e.g., famous individuals, types of computer code, etc.) represented in Claude 3 Sonnet. To verify our feature interpretations, we ran qualitative feature steering experiments, where we artificially dialed up and down various features to see if they changed model outputs in intuitive ways. The results were promising – for example, turning up a feature that responded to mentions of the Golden Gate Bridge made the model talk about the Golden Gate Bridge. Such examples led us to hypothesize that feature steering might be a promising way to modify model outputs in specific interpretable ways. — Read More
You could start smelling the roses from far away using AI
AI can “teleport” scents without human hands (or noses)
Ever send a picture of yourself trying on clothes to a friend to see what they think of how you look? Now, imagine doing the same from the perfume and cologne counter. AI could make that happen in the not-too-distant future after a breakthrough in ‘Scent Teleportation.’ Osmo, which bills itself as a “digital olfaction” company, has succeeded in using AI to analyze a scent in one location and reproduce it elsewhere without human intervention. — Read More
Rick's Cafe AI 