Today:

• Hedra and ElevenLabs Partnership

• X Expands Grok AI Integration

• AI Accurately Reads Your Mind

• Cloudflare Launches AI Defense

• Brazil Blocks Meta Data Use

• Figma Disables Controversial AI Feature

• Hacker Steals OpenAI Secrets

• OpenAI Seeks NYT Source Materials

• Anthropic CEO Predicts AI Surge

Hedra teams up with ElevenLabs to give voice to video

Hedra has launched Character-1, a model that creates videos from still images in seconds. Within 48 hours, it had tens of thousands of users generating over 100,000 videos. Hedra partnered with ElevenLabs to integrate realistic AI voices, enabling immersive storytelling. 

ElevenLabs, known for high-quality AI audio, quickly facilitated Hedra's growth, supporting over 76,000 users in the first week. Hedra's goal is to democratize video creation, making it accessible for anyone to create characters and worlds. ElevenLabs provides advanced audio tools across 29 languages, enhancing storytelling for creators, media, and businesses.

ELEVENLABS

X plans to more deeply integrate Grok’s AI, app researcher finds

Elon Musk’s X is planning to enhance integration with xAI’s Grok within the social networking app. Discoveries by app researcher Nima Owji reveal new features such as asking Grok about X accounts, using Grok by highlighting text in the app, and accessing Grok’s chatbot via a pop-up while using other parts of X. 

These features aim to make Grok more accessible and functional, resembling the integration seen in productivity apps by Google and Microsoft. Despite these advancements, X’s in-app purchase revenue has declined, partly due to competition from other social apps like Threads, Mastodon, and Bluesky.

TECHCRUNCH

Mind-reading AI recreates what you're looking at with amazing accuracy

Artificial intelligence can now accurately recreate what a person or monkey is looking at by analyzing their brain activity. Researchers found that AI's image reconstructions are more precise when the system focuses on specific brain regions. This breakthrough, highlighted by Umut Güçlü from Radboud University, showcases the most accurate image reconstructions achieved so far. 

By honing in on the relevant brain areas, AI can produce detailed and faithful images of visual experiences, marking significant progress in mind-reading technology and its potential applications.

NEWSCIENTIST

Cloudflare launches a tool to combat AI bots

Cloudflare has introduced a free tool to stop bots from scraping websites for AI model training data. Unlike some AI vendors like Google and OpenAI, who respect website owners' robots.txt files to block bots, not all AI scrapers do. Cloudflare's tool identifies and blocks AI bots that mimic web browsers to evade detection. 

This move addresses the growing issue of unauthorized AI data scraping, which has escalated with the AI boom. Cloudflare's models detect bot activity by analyzing traffic patterns and behaviors. This tool aims to protect websites from unauthorized data usage by AI companies, maintaining content integrity.

TECHCRUNCH

Meta Has Been Ordered to Stop Mining Brazilian Personal Data to Train Its AI

Brazil’s national data protection authority has ordered Meta to stop using Brazilian data to train its AI models, citing risks to users' rights. Meta has five days to comply or face daily fines. The decision follows a Human Rights Watch report revealing a dataset with identifiable images of Brazilian children, raising concerns over exploitation. 

Meta, which has over 112 million Facebook users in Brazil, expressed disappointment, arguing the decision hinders innovation. The company faced similar challenges in Europe and plans to continue addressing regulatory questions while maintaining AI training in the U.S. where privacy laws are less stringent.

TIME

Figma disables its AI design feature that appeared to be ripping off Apple’s Weather app

Figma has temporarily disabled its "Make Design" AI feature after it was found to replicate Apple’s Weather app designs. Andy Allen, founder of NotBoring Software, highlighted the issue, prompting concerns within the design community. Figma CEO Dylan Field denied accusations of training the tool on existing apps but acknowledged a lack of thorough quality assurance. 

The feature, introduced to help designers quickly generate UI layouts, will remain disabled until a complete review is conducted. This incident highlights the ongoing debate over AI's role in design, with concerns about job impacts and legal issues.

TECHCRUNCH

A Hacker Stole OpenAI Secrets, Raising Fears That China Could, Too

Earlier this year, a hacker accessed OpenAI’s internal messaging system and stole details about its AI technologies. OpenAI did not publicize the breach or inform authorities, as no customer or partner data was compromised and the hacker was believed to be a private individual without foreign government ties. 

This incident raised concerns among employees about potential threats from foreign adversaries like China. Leopold Aschenbrenner, an OpenAI manager, criticized the company's security measures and was later dismissed. The breach highlighted internal divisions over AI security and the need for robust protection against foreign infiltration.

THE NEW YORK TIMES

OpenAI Wants New York Times to Show How Original Its Copyrighted Articles Are

OpenAI is seeking source materials from the New York Times to assess the originality of its copyrighted articles in defense of a multi-million-dollar copyright infringement lawsuit filed by the newspaper. OpenAI argues that determining the original content will help counter the claims. The Times refuses, citing the reporter’s privilege and potential chilling effects. 

OpenAI has filed a motion to compel the Times to provide the information. The Times insists its articles are copyrightable regardless of third-party content and argues that complying with such requests could deter future copyright lawsuits. The court's decision on this dispute is pending.

TORRENTFREAK

AI models that cost $1 billion to train are underway, $100 billion models coming — largest current models take 'only' $100 million to train: Anthropic CEO

Anthropic CEO Dario Amodei revealed that AI models currently in development could cost up to $1 billion to train, with future models potentially reaching $10 to $100 billion by 2027. This increase is due to growing hardware needs and advancements in AI technology. For instance, training ChatGPT-4 cost around $100 million. Amodei predicts that continued improvements in algorithms and chips will result in AI models surpassing human capabilities in most tasks. 

However, this surge in AI training costs raises concerns about power supply and infrastructure, with some companies considering modular nuclear power for data centers to meet these demands.

TOM'S HARDWARE

🧠RESEARCH

Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion

The paper introduces Diffusion Forcing, a new training method for sequence generation. It combines the strengths of next-token prediction and full-sequence diffusion models. This approach allows for continuous sequence generation, such as videos, and improves performance in decision-making tasks. The method optimizes likelihoods of subsequences from the true joint distribution.

Planetarium: A Rigorous Benchmark for Translating Text to Structured Planning Languages

The paper introduces Planetarium, a benchmark for evaluating the translation of natural language descriptions into structured planning languages like PDDL. It highlights the challenges in assessing the quality of generated PDDL code. The benchmark includes a dataset of 132,037 text-to-PDDL pairs and an algorithm for rigorous evaluation. Results show current models, like GPT-4, struggle with semantic correctness, emphasizing the need for this benchmark.

InternLM-XComposer-2.5: A Versatile Large Vision Language Model Supporting Long-Contextual Input and Output

The paper introduces InternLM-XComposer-2.5 (IXC-2.5), a large vision-language model supporting extensive input and output contexts. IXC-2.5 excels in text-image comprehension, ultra-high resolution understanding, fine-grained video analysis, and multi-image dialogue. It outperforms many models in benchmarks and competes with GPT-4V. It is publicly accessible for further use and development.

TabReD: A Benchmark of Tabular Machine Learning in-the-Wild

The paper introduces TabReD, a benchmark of eight industry-grade tabular datasets to address gaps in academic benchmarks. These datasets reflect real-world scenarios, including temporally-evolving data and complex feature engineering. Evaluations show that traditional ML models like MLP and GBDT perform best, highlighting the need for time-based splits in model evaluation.

TokenPacker: Efficient Visual Projector for Multimodal LLM

The paper presents TokenPacker, a new visual projector for multimodal large language models (MLLMs). It uses a coarse-to-fine method to reduce redundant visual tokens without losing detail. This approach improves efficiency by compressing visual tokens by 75%-89% while maintaining or enhancing performance in various benchmarks. The source code is available on GitHub.

🛠️TOP TOOLS

Video to Sound Effects - Generate custom AI sound effects for your videos effortlessly with this powerful tool.

Customers AI - Capture anonymous website visitor data, track the customer journey, and turn visitors into revenue.

Screen Pipe - Record your screen & mic 24/7 and connect it to LLMs. Inspired by adept.ai, rewind.ai, Apple Shortcut. 

Prompt Easy AI - Fine-tune GPT with absolutely zero technical skills

Paird AI - Real-time collaboration for developers.