AI News & Launches

Simon Willison's Weblog recently featured a post about OpenAI's new 'Predicted Outputs' feature for GPT-4o and GPT-4o mini. The feature allows users to send a 'prediction' along with their prompt which can accelerate the response time by validating large batches of input in parallel. OpenAI's pricing for this feature will charge for any tokens provided that are not part of the final completion at standard rates. The author shared a result comparison showing a faster response with prediction but at a slightly higher cost.

Perplexity CEO Aravind Srinivas has offered the company's AI services to replace striking tech workers at the New York Times just before the U.S. presidential election. The NYT Tech Guild announced a strike over terms including a wage increase and in-office expectations. The pitch was criticized online as it undermines collective bargaining efforts. Perplexity, recently involved in legal disputes with NYT over content scraping, clarified its offer would not replace journalists but provide technical support.

Scientists from DeepMind have created SynthID-Text, a scalable text watermarking scheme that identifies outputs from large language models (LLMs), ensuring responsible use. The system embeds watermarks during text generation without altering training procedures, thus preserving quality and enabling efficient detection.

OpenHands, powered by Daytona, offers an AI teammate capable of working in parallel with developers. It provides an integrated workspace with a shell, browser, editor, and planner for seamless end-to-end workflows. Its adaptability to enterprise needs, with a strong focus on security and control, makes it a robust solution for large-scale projects.

gptel is a minimalist and versatile LLM client for Emacs, allowing live interactions with various language models directly from any buffer. It supports branching conversations in Org mode, and offers additional packages for enhanced functionality. The GPL-3.0 licensed project has 1.4k stars, 27 watchers, and 140 forks.

Project Sid is an open-source repository on GitHub showcasing a technical report on 'Many-agent simulations toward AI civilization'. It involves large-scale simulations with AI agents to reflect complex interaction within civilizations.

LM Studio has open-sourced a utility called 'venvstacks' for creating layered Python virtual environments, enabling easy deployment and sharing of dependencies for Python applications in environments like MLX engine within LM Studio.

A University of Washington study highlights significant potential for racial and gender bias in AI when used to screen resumes. Tests on three open-source language models showed they favored resumes from white-associated names 85% of the time and female-associated names 11% of the time, with Black men being the least preferred.

The article by Daniel Kharitonov delves into brute-forcing the guardrails of large language models (LLMs) to bypass restrictions against actions like offering medical diagnoses based on X-ray images. Using examples from Google’s Gemini 1.5 pro model, the author demonstrates prompt engineering techniques and automation using DataChain libraries to generate numerous prompts and identify loopholes that allow guardrail evasion. The success rate of these attempts to bypass restrictions was found to be significantly high, indicating weaknesses in the current implementation of guardrails.

A detailed examination of the 12th-century Sussex muralist who, akin to Ai Weiwei's act of dropping a Han Dynasty urn, created the 'Deception of Eve and Adam'—a provocative artwork challenging the era's art conventions. The post narrates historical context, parallels to Ai Weiwei's iconoclastic attitudes, and the possibility of censorship due to the mural's unorthodox representation and questioning of orthodox belief.

Conduit is a tool that allows professionals to get GPT-like insights and answers directly in Google Sheets without any technical skills. Supports integration with various systems like Shopify, Salesforce, and more.

WalkSmart.ai provides an innovative platform for travelers to create customized walking tours for cities around the world. With a focus on personalization, the service allows users to build their perfect tour based on their preferences and interests, complete with audio-guided narration and integration with Google Maps.

Monica Code provides a one-stop AI coding assistant for VSCode, which supports GPT-4o and Claude 3.5 Sonnet. It offers code completion, editing, and multimodal chat with codebase. The premium version offers unlimited AI code completions for a monthly or annual fee, while a free version includes 8 slow premium queries per day.

Daniel Hall reflects on the evolution of SPA (Single-Page Applications), criticizing the complexity and performance issues associated with frameworks such as React. He discusses the challenges state management brings to the DOM, the irony of 'multi-page' SPAs, and the dangers of relying heavily on JavaScript for both client and server-side operations. Hall also expresses concern over the excessive use of npm packages, the registry, and the industry's failure to standardize practices.

AMD introduces its first series of 1 billion parameter language models, AMD OLMo, offering significant AI capabilities and fully open-source for community collaboration. These models are trained on AMD Instinct GPUs using trillions of tokens and demonstrate superior language understanding and reasoning.

Oasis is a groundbreaking AI-generated video game created by Decart and Etched, offering real-time gameplay without a traditional game engine. The model is a technical advance in AI research, providing complex game mechanics and rich, interactive open-world experiences using transformers and latent diffusion techniques. With challenges in domain generalization and temporal stability, the developers aim to improve upon the model with future scaling and hardware innovations.

Pythagora GPT is seeking an innovative UX Designer to revolutionize the developer experience in AI technologies. The candidate will work on user interfaces that cater to AI developers and facilitate interaction with AI models.

KlipLab is an AI-powered platform that offers users the ability to generate realistic voiceovers and lip-synced videos using celebrity and character voices. It provides a variety of high-quality voices to choose from and supports custom video and audio uploads for personalized content creation. Plans start at $5/month.

Train with Kiwi Fitness offers a personalized strength workout experience through an AI-powered virtual trainer. With gamification elements, scientific backing, and community support, it motivates users to maintain fitness routines anytime, anywhere, providing adaptive plans and progress tracking via an app available for both iOS and Android platforms.

OpenAI CEO Sam Altman notes in a Reddit AMA that insufficient compute capacity hinders the frequency of product shipments, due to the complex nature of their AI models. OpenAI faces difficult choices in compute allocation among various projects and reports suggest struggles to obtain necessary infrastructure, despite collaboration with Broadcom on a potential AI chip, anticipated by 2026. The integration of vision capabilities in ChatGPT and the release of DALL-E's next version are delayed, while Altman emphasizes future improvements on reasoning models.

Report by HAProxy shows nearly 90% of AI crawler traffic comes from TikTok's parent company Bytedance's Bytespider. HAProxy Edge analytics reveal significant bot traffic, highlighting opportunities and risks for content sites. Depending on business objectives, strategies can include allowing bot discovery or protecting content value by blocking AI crawlers with technical solutions like HAProxy Enterprise.

Qodo, formerly known as Codium, announced support for Anthropic's Claude Sonnet 3.5, OpenAI o1-preview and o1-mini, and Google Gemini 1.5 Pro with seamless developer access in the Qodo Gen version 0.12. These tools offer advanced code understanding, reasoning, and enhanced natural language understanding, providing developers with various models tailored to different tasks.

Will Hakim reflects on two years since ChatGPT's launch, highlighting a shift in Silicon Valley and the AI industry. He discusses the widening gap between the AI 'haves' and 'have-nots,' and the importance of domain expertise and fitting AI into customer workflows. Halcyon focuses on integrating domain knowledge into AI and creating user-aligned tools, key to providing value in AI tech development.

ActiveLoop is seeking a passionate Python Backend Engineer to develop and optimize their data science platform. The role involves working closely with data scientists, improving system architecture, and contributing to open source projects.

This article discusses the life and legacy of Scots poet Robert Fergusson, known as Scotia's Bard, focusing on his prolific output and influence despite his short life. Fergusson died young at 24 after a brain injury, leaving behind over 100 poems. His work is often overshadowed by Robert Burns, who saw Fergusson as a muse. The article challenges the view of Fergusson as merely preparatory to Burns, instead arguing his unique contribution to Scottish literature.

CreatorKit offers a tool called Emotive AI Actors for generating user-generated content (UGC) and video ads that feature real emotions for improved authenticity and higher conversions. The service provides AI voices capable of speaking any language, dozens of AI actors, and the ability to create affordable UGC at scale.

GitHub Spark is an innovative AI-powered platform aimed at enabling users to build and share personalized micro apps aptly named 'sparks'. Built with a managed runtime, a natural language based editor, and a shareable, remixable app ecosystem, it ushers in a new era of personal software customization without the need to write or deploy code.

Sharefable offers AI-powered interactive product demos, tailored to engage prospects and boost conversions. Their platform is designed for easy creation, organization, and sharing of demos, receiving high praise on G2 with a 4.9 rating.

Trove provides real-time merchant intelligence by deciphering financial transactions, offering insights into user spending patterns. With a developer-first approach, it enriches data with details such as merchant identities and geolocation.

Introducing Hiring Assistant, a new AI agent by LinkedIn Talent Solutions that automates time-consuming hiring tasks, helping recruiters focus on the human aspects of their job. It offers candidate sourcing, admin management, and provides proactive updates.

Browser AI Kit presents a collection of AI-powered tools that run directly in your web browser, offering features such as Audio to Text, Remove Background, and Text to Speech. The platform prides itself on convenience, security, multifunctionality, and being free of charge.

Jay F. discusses his failed ticket scalping startup, aimed at predicting concert ticket prices using data science. Despite creating a sophisticated data model and infrastructure, he faced technical challenges, stiff competition from professional brokers, and moral conflicts. The venture didn't succeed, providing him with insights into the industry's flaws and the importance of infrastructure in arbitrage operations.

DeepSeek has launched version 2.5 with enhanced general and coding capabilities, outperforming major AI models in machine learning benchmarks. It offers API access and improved web interaction at competitive pricing and supports 128K context length. The newest iteration secures top places in AI leaderboards and comes with open-source availability.

Trieve offers an all-in-one search solution that includes semantic vector search, full-text search, and AI-powered recommendations and analytics. Providing both API and no-code solutions, it enables the swift setup of industry-leading search in 30 minutes. Trieve is built on open-source models, with the option for self-hosting.

Hamel Husain's blog provides a comprehensive guide for AI teams on streamlining AI evaluation using the concept of Critique Shadowing. The post outlines a step-by-step method to build an LLM-as-a-Judge system that aligns with business goals by involving principal domain experts, generating diverse datasets, and iteratively refining the evaluation process. Emphasizing simple metrics, expert critiques, and error analysis, the blog details how careful data examination rather than complex judges creates business value.

LlamaPReview is an AI-powered GitHub PR Review Assistant offering 1-click install, automatic PR review with contextual analysis, support for all major languages, and is currently totally free.

Stanford University's Hazy Research released improvements to ThunderKittens, offering faster, more efficient kernels and cute quirks. Their work enhances computing for new architectures, emphasizing the performance advances in attention architecture and integration of demo models.

Intel is innovating with AI Flame Graphs, visualizing AI and GPU hardware profiles alongside the software stack to potentially reduce US power usage significantly. The tool is previewed on the Intel Tiber AI Cloud and showcases performance issues in AI accelerators for optimization.

Wand is an AI-powered drawing app designed for iOS that transforms sketches into fully-rendered artwork within seconds. It also offers real-time editing and style customization geared for artists. The app includes features like a custom brush engine, multi-layer support, and secure private models. It's designed to enhance creativity, save time, and help artists learn faster.

MightyDocs is a MacOS-based AI assistant that provides accurate answers from the latest developer documentation, tailored for specific tech stacks. It interacts with your computer locally, ensuring privacy and immediacy of information, without sending data to external servers.