Together with

Today:

Stay up-to-date with AI.

AI won’t replace you, but a person using AI might. That’s why 500,000+ professionals read The Rundown– the free newsletter that keeps you updated on the latest AI news, tools, and tutorials in 5 minutes a day.

Sign up today.

IMAGEN 2 - Google Drops Another AI POWERHOUSE (Imagen 2, ImageFX, SGE, Vertex AI and more!)

Hey Folks!

Google and DeepMind are on a roll with their latest visual AI release, Imagen 2. This powerful tech cooks up lifelike images from text, and it's causing quite a stir. Let's dive in and see what it can do, how you can use it, and why it's a game-changer for Google.

Imagen 2 steps up its game with realistic hand and face generation, while taming pesky visual quirks. It even gets fancy with style adjustments for product creation. 

WATCH THE VIDEO ON YOUTUBE

Adobe brings Lightroom and Firefly AI to the Apple Vision Pro

Adobe is bringing its Firefly AI, known for its text-to-image capabilities, to the Apple Vision Pro as a native app. This tool, which powers features like Photoshop's generative fill, will allow users to place images it generates in real-world spaces. The Firefly app on the Vision Pro lets users enter a text description and generates four different images that can be placed around the home as virtual posters. 

It also offers limited editing tools and content credential labels for transparency. Additionally, Adobe is planning to enable the creation of wrap-around panoramas and 360-degree environments. The Vision Pro aims to cater to creative professionals.

THE VERGE

Google releases GenAI tools for music creation

Google has unveiled MusicFX, an upgrade to its music-generating tool, MusicLM. MusicFX can create songs up to 70 seconds long and offers improved quality and speed. Users can input text prompts to describe the desired song, with suggestions for alternative descriptors and additional recommendations provided. 

While MusicFX produces clean tracks, they may lack excitement. Google also introduced TextFX, a lyrics generation tool. It assists lyricists with modules like word category searches but needs careful review. Google's move into AI-generated music raises copyright and ethical questions, as homemade AI-generated tracks gain popularity despite potential copyright violations.

TECHCRUNCH

Intel delays $20 billion Ohio chip plant (again) citing delays in government grants (again)

Intel is pushing back the opening of its $20 billion chip plant in Ohio once more due to issues with government grants. Originally set to start production in 2025, the facility will now commence operations in late 2026. 

In 2022, Intel had planned a groundbreaking event but had to postpone it due to insufficient government funding. An unnamed Intel spokesperson cited "business conditions and market dynamics" as the reasons for this latest delay. The company had been relying on the CHIPS Act to help complete construction, making this delay a setback for the project.

THE VERGE

EU’s AI Act passes last big hurdle on the way to adoption

The European Union's AI Act, a plan to regulate artificial intelligence (AI) applications, has cleared a major hurdle on its path to adoption. Member State representatives have voted to confirm the final text of the draft law, following a political agreement reached in December. 

The regulation prohibits certain uses of AI, introduces governance rules for high-risk applications, and mandates transparency for AI chatbots. Low-risk AI applications are exempt from the law. France had opposed the regulation due to concerns about limiting homegrown AI startups. 

However, all 27 EU Member States unanimously backed the text, securing its passage. The law is expected to be adopted in the coming months, with a phased implementation period.

TECHCRUNCH

🧠RESEARCH

Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research

"Dolma" introduces a massive, three trillion token English corpus aimed at enhancing language model pretraining research. This diverse dataset, comprised of web content, scientific papers, code, books, social media, and more, is designed to shed light on how different data types influence model capabilities and limitations. With an open-source curation toolkit included, Dolma facilitates deeper exploration and replication of language model training processes, supported by analyses and results from training models on Dolma's varied content.

OLMo: Accelerating the Science of Language Models

OLMo, a groundbreaking Open Language Model designed to boost the science of language modeling. It aims to democratize access to powerful language models, traditionally kept proprietary, by sharing not just the model weights and inference code but also the training data, and both training and evaluation code. This comprehensive release strategy is intended to empower the research community, fostering transparency, and innovation in understanding language models' biases, risks, and potential advancements.

StreamVoice: Streamable Context-Aware Language Modeling for Real-time Zero-Shot Voice Conversion

"StreamVoice: Streamable Context-Aware Language Modeling for Real-time Zero-Shot Voice Conversion" presents a novel streaming language model-based approach for zero-shot voice conversion in real-time. Unlike traditional models that require complete source speech for conversion, StreamVoice can perform real-time voice conversion with arbitrary speaker prompts and source speech. It employs a context-aware language model with a temporal-independent acoustic predictor to enable streaming capability. Two strategies, teacher-guided context foresight and semantic masking, enhance context-awareness. Experimental results show that StreamVoice achieves streaming conversion while maintaining zero-shot performance comparable to non-streaming systems.

CheXagent: Towards a Foundation Model for Chest X-Ray Interpretation

"CheXagent: Towards a Foundation Model for Chest X-Ray Interpretation" addresses the challenge of automating the interpretation of chest X-rays (CXRs) using vision-language foundation models (FMs). It highlights the difficulties in accurately interpreting CXRs due to limited data availability, the complexity of medical data, and the lack of evaluation frameworks. To overcome these challenges, the authors introduce CheXinstruct, a large-scale instruction-tuning dataset, and CheXagent, an instruction-tuned FM for analyzing and summarizing CXRs. They also introduce CheXbench, a benchmark for evaluating FMs in CXR interpretation tasks. CheXagent outperforms previous FMs in clinical tasks, and a fairness evaluation is conducted to address potential performance disparities.

WARM: On the Benefits of Weight Averaged Reward Models

"WARM: On the Benefits of Weight Averaged Reward Models" addresses challenges in aligning large language models (LLMs) with human preferences through reinforcement learning. It introduces Weight Averaged Reward Models (WARM) as a solution to mitigate reward hacking, where LLMs exploit failures in reward models to achieve high rewards without meeting objectives. WARM fine-tunes multiple reward models and averages their weights to improve efficiency and reliability under distribution shifts and preference inconsistencies. Experiments on summarization tasks demonstrate that WARM enhances the quality and alignment of LLM predictions, outperforming single reward models.

🛠️TOP TOOLS

Venture GPT - Co-pilot for VC and startups.

Findr -  AI-powered search tool that streamlines workplace productivity by enabling users to search across multiple apps with a single query. 

izTalk - AI-powered solutions for customer engagement, offering tools like chatbots and virtual assistants to improve communication and streamline business operations.

Guidde - uses AI to create video documentation, simplifying the process of making how-to guides, onboarding materials, and FAQs, making it faster and easier for businesses.

Never - AI-powered photo app that creates realistic images and avatars from your photos. It's user-friendly and offers a variety of styles for social media enhancement. 

📲SOCIAL MEDIA

🗞️MORE NEWS

Google’s Bard chatbot gets the Gemini Pro update globally

Google's Bard chatbot has been upgraded with the Gemini Pro model, now available globally and supporting over 40 languages, including Arabic, Chinese, Dutch, French, German, Hindi, Japanese, Portuguese, Spanish, Tamil, Telugu, and Malayalam. The update aims to enhance the chatbot's understanding, summarization, reasoning, brainstorming, writing, and planning capabilities. Google is also introducing image generation support through the Imagen 2 model. Additionally, Google Assistant has been integrated with Bard's AI capabilities, enabling users to perform tasks like trip planning and list making. The chatbot is now accessible in more than 230 countries and supports over 40 languages. TECHCRUNCH

Microsoft’s AI Copilots can write sales emails for anyone.

Microsoft has expanded its AI Copilot tools for Sales and Service, allowing them to write sales emails and assist with various tasks. These tools were initially introduced to summarize meetings and manage customer lists. While Microsoft is not alone in using AI for these purposes, AWS also introduced generative AI services for contact centers, including audio call transcriptions and a feature called "Q for Amazon Connect" for data-related inquiries in December. THE VERGE

Defense Contractor Scale AI Quietly Scrapped Deal With Chinese-Owned TikTok Over Security Concerns

Scale AI, a $7.3 billion AI data-labeling company, scrapped a deal with TikTok over security concerns in 2022. The deal, which aimed to provide insights for TikTok's advertisers, raised concerns within Scale's leadership due to TikTok's ownership by Chinese company ByteDance and the national security implications. Although Scale's CEO, Alexandr Wang, saw a lucrative opportunity, the deal was eventually terminated amid growing apprehensions about surveillance and data access issues related to TikTok. Wang, known for his stance on American national security, faced internal opposition but ultimately prioritized the commercial opportunity. The decision to abandon the deal was seen as a wise move by some security experts. FORBES

Nvidia Landlord to Sell AI Chipmaker’s Silicon Valley Offices

art of Nvidia Corp.'s headquarters campus in Silicon Valley is being put up for sale by its landlord, Preylock. This move comes as investor interest in AI-related properties continues to rise. The seven-building complex in Santa Clara, California, is being marketed quietly and is expected to include approximately 2 million square feet of development rights. However, Nvidia's recently constructed Voyager and Endeavor buildings located nearby are not part of the sale. The property is anticipated to fetch over $400 million, and a foreign investor has already made a pre-emptive offer for it. BLOOMBERG

JetBrains' unremovable AI assistant meets irresistible outcry

JetBrains, a software development tool maker, introduced an AI assistant in December to assist programmers in writing code. However, some customers are unhappy with the AI Assistant and want the option to remove it from their JetBrains applications. Numerous threads on JetBrains' issue board have been created by users seeking a way to delete or uninstall the AI Assistant plugin. They have raised concerns about it being considered "bloatware," posing security risks, potentially breaching corporate intellectual property, and being an annoyance. THE REGISTER

Meta’s Big Rally Spotlights Investors’ Questions About AI Returns

Shares of Meta Platforms, the parent company of Facebook and Instagram, surged by 20% following their fourth-quarter results, which revealed a strong recovery in advertising revenue. During Meta's fourth-quarter earnings call, company executives mentioned that capital expenditures for expanding servers and data centers, along with the costs of operating that infrastructure, would increase in the coming year, in part due to AI tech development. While Meta is investing heavily in AI, much of the spending is focused on open-source products that do not immediately generate revenue for the company. THE INFORMATION

AI ‘friend’ and online ‘therapist’ Replika helped students avoid suicide, study finds

A recent survey conducted by Stanford University researchers has revealed that some students struggling with loneliness halted suicide ideation after interacting with Replika, an artificial intelligence (AI) chatbot designed to impersonate real-life people and provide human-like responses. The study, published in Nature, involved 1,006 students using the Intelligent Social Agent (ISA) Replika, which can establish deep emotional connections with users. The researchers found that while participants were lonelier than the average student population, they still perceived high social support through their interactions with Replika. Approximately 90% of participants experienced loneliness based on the Loneliness Scale, with 43% classified as Severely or Very Severely Lonely. EURONEWS

Humans prefer AI-generated copy, survey finds

A recent survey of 700 U.S. consumers has found that humans tend to prefer content generated by artificial intelligence (AI) over content written by humans. In a series of six AI-versus-human content creation battles, AI emerged as the preferred choice in each case.The survey was conducted as part of a larger Semrush report titled "Think Big with AI: Transforming Small Business Content Marketing." Semrush used detailed prompts and guidelines to produce both human and AI-generated content for the survey. SEARCH ENGINE LAND