TLDR AI 2025-04-08

RAD Intel is making marketing feel human again. Forget interruption—this patented AI reveals exactly who your customers are, what they want, and the perfect (human) personality to deliver it.

Backed by industry Adobe and Fidelity Ventures Hasbro, RAD Intel is already turning heads, attracting major investment, and redefining digital marketing for brands like Hasbro, MGM and Sephora.

Insiders from Google, Meta and Amazon have invested and now, you can too.

Shares are currently $0.55 per share, but not for long. On Apr 10, the share price changes - so nows the time.

Jump in at the current share price of $0.55

Meta's VP of Generative AI has denied claims that Llama 4 models were trained on benchmark test sets, pushing back against allegations of inflated performance scores.

The updated Nova Reel model now supports multi-shot videos up to 2 minutes long, offering more creative control and efficiency for generative video content.

Andreessen Horowitz has announced an investment into Krea, a platform combining AI models to help creatives generate and edit visual content. Used by teams at Pixar and Samsung, Krea now has over 20M users. The company plans to launch an enterprise-grade product later this year.

This study finds that speech-language models initialized from text models using interleaved training scales more efficiently than speech-only models.

Online-Mind2Web is a realistic evaluation benchmark for autonomous web agents. Current models perform worse than expected due to flawed benchmarks.

This work introduces Test-Time Training (TTT) layers, featuring highly expressive hidden states, to overcome the limitations of standard Transformers and alternatives like Mamba in generating long, coherent videos. By integrating TTT layers into a pre-trained model, they demonstrate the generation of coherent one-minute videos from text storyboards that significantly outperform baseline methods in telling complex stories according to human evaluations. Tom and Jerry cartoons are the test bed.

Hasura's AI Disrupt will bring together industry leaders and decision-makers to explore the integration of AI with existing systems, focusing on accuracy and transparency. Attendees can expect engaging panels, case studies, and hands-on workshops, all led by experts at the forefront of AI innovation. See the full speaker list and register for free

Dual understanding and synthesis model in a single autoregressive framework. This means that one model can caption and generate images.

OpenSource work from Apple. It follows up the company's paper from a few weeks ago that allows images to be tokenized in a variable number of tokens. This means that more complex images can use more tokens.

ZClip uses EMA-based gradient norm statistics to adaptively suppress outlier gradients, preventing loss spikes and improving training stability without static thresholds.

Google recaps major March updates including Gemini 2.5 Pro, expanded AI Overviews, AI Mode, and other feature rollouts across its products.

Genies has launched a no-code platform that enables anyone to create intelligent AI avatars with unique looks, personalities, and behaviors for use in customizable gaming experiences called Parties. These AI avatars are powered by large language models, behavioral AI, and real-time animation tech, allowing dynamic interactions, gameplay, and emotional responses.

OpenAI will release intermediate models o3 and o4-mini before GPT-5. This staggered release is due to technical complexities with GPT-5, allowing improvement and better handling of expected demand. The shift comes as competitors like Google's Gemini 2.5 Pro advance in the AI field. OpenAI has just secured a $40 billion funding round.

Goku from ByteDance is a flow based video generation model of 2B and 8B parameters with 160M image and 36M video pairs.

Google is expanding its AI Mode feature to millions of U.S. Labs users, enhancing it with multimodal capabilities.

Arc Prize Foundation has revised computing cost estimates for OpenAI's o3, suggesting it's much higher than initially thought at potentially $30,000 per ARC-AGI problem.