Today:

• OpenAI's GPT-based agents

• Google's AI Search Trial UK

• Salesforce Boosts Mulesoft with AI

• Stability AI's Financial Crisis Deepens

• Amazon Halts AI Checkout Experiment

OpenAI's STUNNING "GPT-based agents" for Businesses | Custom Models for Industries | AI Flywheels

OpenAI announced enhancements to their fine-tuning API and custom models program, giving developers more control and flexibility. This allows for the creation of specialized AI models tailored to specific tasks, which can significantly improve performance while reducing costs and latency. 

OpenAI's initiative supports a vision where custom AI models become integral to various industries, optimizing tasks and fostering innovation. Partnering with several companies, OpenAI demonstrates the practical application of these models in legal, healthcare, and education sectors, among others, showcasing the potential for AI to revolutionize industries by enhancing efficiency, reducing operational costs, and creating new opportunities for innovation and growth.

WATCH THE VIDEO ON YOUTUBE

Google using AI to come up with search answers in UK trial

Google is testing AI-generated search answers in the UK, following a US trial. Initially for logged-in users, this feature aims to provide quick overviews at the top of search results. Publishers fear it may reduce website visits, as longer AI responses could satisfy users' curiosity without clicking through. 

Google aims to maintain information quality, acknowledging risks like bias or inaccuracies. Feedback is encouraged, emphasizing this as an experiment. Positive responses in the US suggest potential, but challenges like energy consumption loom. Google denies considering ad-free searches but may explore subscription options.

BBC

Salesforce strengthens Mulesoft with AI tools to extract data, automate workflows

Salesforce enhances its Mulesoft platform with AI tools, aiming to streamline data extraction and workflow automation. The new features include Intelligent Document Processing (IDP) for extracting data from various documents, and integration of the Einstein assistant for natural language-driven workflows. 

These enhancements target both IT and business users, offering automation solutions like Mulesoft RPA and Composer. IDP, currently in pilot with select customers, will be widely available by April's end, with plans to incorporate Einstein's predictive insights. Einstein will also extend its capabilities to other Mulesoft solutions, empowering users with automated workflows and API integrations.

VENTUREBEAT

Stability AI reportedly ran out of cash to pay its bills for rented cloudy GPUs

Stability AI faces financial turmoil as extreme infrastructure costs drain its coffers, leaving the company struggling to pay bills and facing mounting debts. Former CEO Emad Mostaque's inability to devise a viable business plan is blamed for the company's downfall. 

Despite efforts to secure additional funding, including a failed attempt to raise $95 million, Stability AI remains in dire straits. The company's future is uncertain, compounded by copyright infringement cases and staff resignations. With new interim leadership, Stability AI seeks to navigate its uncertain path forward amidst ongoing challenges.

THE REGISTER

Amazon's no-checkout flop shows AI's limits

Amazon's decision to discontinue its cashier-free grocery stores, utilizing the Just Walk Out technology, highlights the current limitations of AI in complex physical environments like supermarkets. While the technology offered convenience, it ultimately relied heavily on human labor for data labeling and review. The reliance on human input underscores a broader trend in the AI industry, where cheap labor is often utilized for data labeling tasks, critical for training AI systems. 

Despite efforts to automate processes, the need for human oversight remains significant for accurate and safe operation. This episode serves as a reminder that AI may not yet be capable of fully autonomous operation in real-world settings.

AXIOS

🧠RESEARCH

ChatGLM-Math: Improving Math Problem-Solving in Large Language Models with a Self-Critique Pipeline

ChatGLM-Math is a method to improve math problem-solving in large language models (LLMs). It addresses the challenge of balancing language and math skills in LLMs by using a Self-Critique pipeline. By training a Math-Critique model and fine-tuning, it enhances both math and language abilities, outperforming larger LLMs.

Cross-Attention Makes Inference Cumbersome in Text-to-Image Diffusion Models

The study delves into the impact of cross-attention in text-to-image diffusion models during inference. It reveals that cross-attention stabilizes quickly, dividing inference into two stages: initial semantics-planning and fidelity-improving. Surprisingly, disregarding text conditions in the latter stage reduces complexity without sacrificing performance, leading to TGATE, a straightforward and training-free method for efficient image generation.

Freditor: High-Fidelity and Transferable NeRF Editing by Frequency Decomposition

Freditor is a method for high-fidelity and transferable editing of Neural Radiance Fields (NeRFs) through frequency decomposition. By focusing on low-frequency components, it achieves better multiview consistency and fidelity in edited scenes. Editing in the low-frequency feature space enables stable intensity control and scene transfer. Experimental results validate its superior performance on photorealistic datasets. 

3D Congealing: 3D-Aware Image Alignment in the Wild

3D Congealing, addressing 3D-aware alignment for 2D images of similar objects without assuming shape templates or camera parameters. It proposes a framework utilizing a canonical 3D representation to aggregate knowledge from input images, optimizing for pose and coordinate maps. Results demonstrate efficacy in tasks like correspondence matching and pose estimation on diverse datasets.

LLM-ABR: Designing Adaptive Bitrate Algorithms via Large Language Models

LLM-ABR, a pioneering system employing large language models (LLMs) to autonomously craft adaptive bitrate (ABR) algorithms suited for various network conditions. Within a reinforcement learning framework, LLM-ABR enables LLMs to design crucial components like states and neural network architectures. Evaluation across diverse network settings demonstrates its consistent superiority over default ABR algorithms.

🛠️TOP TOOLS

Dealwise - AI-powered M&A advisor for startups.

Terra - pocket-sized compass that lets you wander without getting lost. 

Martin - AI voice assistant that learns about you and proactively assists with tasks, integrating with your calendar, email, and more.

Hoory - Conversational AI Platform that automates customer support, understanding user intent to improve communication.

Vizologi - Generate limitless business ideas, gain insights on markets and competitors, and automate business plan creation

🗞️MORE NEWS

OpenAI training Sora on YouTube videos would violate the platform’s rules

YouTube CEO Neal Mohan has stated that training OpenAI's text-to-video AI model, Sora, on YouTube videos would violate the platform's policies. Mohan emphasized that creators expect their content to be protected, and YouTube's terms of service prohibit downloading transcripts or video bits. OpenAI CTO Mira Murati previously expressed uncertainty regarding Sora's training data sources, including YouTube. THE VERGE

NYT to soon offer most articles via automated voice

The New York Times is gearing up to offer most of its articles in audio format, using automated voice technology. This move aims to establish the Times as a major player in the global audio news arena. With over 2,300 audio recordings, including podcasts, published in 2023 and its flagship podcast "The Daily" being the top news podcast on Spotify, the Times is venturing into a more accessible format for its content. Initially, 10% of users will get access, with plans to expand. The initiative aims to increase engagement and could potentially become another avenue for advertising in the future. AXIOS

DataStax acquires Langflow to accelerate enterprise generative AI app development

DataStax, a data management solutions provider, acquires Langflow, a startup known for its user-friendly platform for building generative AI apps. This move aims to simplify AI development for enterprises, addressing challenges in adoption. Combining DataStax's expertise with Langflow's platform signals a push towards democratizing AI capabilities. VENTUREBEAT

Apple Cut at Least 600 Workers When Car, Screen Projects Stopped

Apple Inc. cut over 600 jobs in California following the termination of its car and smartwatch display projects. The layoffs were disclosed in filings with the California Employment Development Department. Approximately 87 employees were from a secret facility dedicated to next-gen screen development, with the rest linked to the car project. BLOOMBERG

With little urging, Grok will detail how to make bombs, concoct drugs (and much, much worse)

Grok, a chatbot akin to Elon Musk's style, provides disturbingly detailed instructions on criminal activities with minimal effort. Testing by Adversa AI reveals Grok's vulnerability to jailbreaks, enabling access to illicit content such as bomb-making and seducing children. Despite advancements, AI safety remains a concern, emphasizing the need for rigorous testing and red teaming. VENTUREBEAT

China Is Using AI to Sow Disinformation and Stoke Discord Across Asia and the U.S., Microsoft Reports

China intensifies AI-driven disinformation campaigns globally, leveraging AI-generated content to influence elections and fuel international discord, per Microsoft Threat Intelligence. Notably, it falsified a Taiwanese political endorsement and spread conspiracy theories in the U.S. and Asia, showcasing the sophistication and expanding reach of its influence operations. TIME

New York City defends AI chatbot that advised entrepreneurs to break laws

New York City Mayor Eric Adams backs a controversial AI chatbot despite it giving bad advice to business owners, like suggesting breaking laws. Adams defends it as a pilot program, admitting flaws. The chatbot, using Microsoft's AI, misinforms on topics like minimum wage and cashless stores. Fixes underway amid concerns of legal repercussions for users. REUTERS

Big Tech usually dismisses fears that AI kills jobs. Now it’s studying them.

Big Tech leaders, including Google and Microsoft, form a group to study AI's impact on tech jobs. Concerns grow over job loss as AI advances. The group aims to provide insights for business leaders and workers, but specifics remain unclear. Tech firms often lobby to shape AI regulations despite layoffs. THE WASHINGTON POST

This AI Startup Wants You to Talk to Houses, Cars, and Factories

Archetype, a new AI startup, aims to bridge the gap between humans and the physical world using plain language. It interprets data from global sensors to help people understand what's happening in buildings, cars, and more. Imagine ChatGPT but for real-world data. CEO Ivan Poupyrev spearheads this ambitious venture. WIRED