TLDR AI 2025-04-07

See what agentic AI can really do on April 10.

After years of siloed experiments, proofs of concept that fail to scale, and disappointing ROI, most enterprises are stuck. AI isn't transforming organizations — it's adding complexity, friction, and frustration.

But Writer customers are seeing positive impact across their companies. Our end-to-end platform delivers adoption and ROI at scale. We're applying that same technology to build agentic AI that actually works for enterprises. Join us for a live product release on April 10 at 2pm ET (11am PT).

Can't make it? Register anyway and we'll send you the recording!

Register Now!

Meta has unveiled Llama 4 Scout and Maverick, two 17B parameter multimodal models that offer state-of-the-art performance on major benchmarks, along with Llama 4 Behemoth, a 288B model still in training that surpasses GPT-4.5 in STEM tasks.

Midjourney has released its new image generation model, V7 alpha. It brings smarter text interpretation, better image coherence, and introduces Draft Mode for fast, low-cost iterations with optional voice commands and personalization.

AI agents are emerging as potent tools in cybersecurity capable of executing complex attacks and potentially scaling operations like ransomware. The LLM Agent Honeypot project aims to detect these agents by simulating vulnerable servers. Its work has revealed that agents can adapt and avoid detection better than traditional bots. Experts anticipate an increase in agent-driven cyberattacks, urging preemptive development of defenses as these technologies evolve.

CUPS is a new method for panoptic segmentation without labeled data that leverages depth and motion cues to train directly on scene-centric images.

The key innovation that enabled Llama 4 to reach 10m+ tokens in context is the alternation between no positional embeddings and rotational positional embeddings. While there are only have benchmarks on Needle in the Haystack, it seems to be a strong confirmation of performance of alternating layers.

This paper from DeepSeek talks about how to use inference time scaling to make reward modeling better to bootstrap stronger reasoners. It hints at a broader strategy from the Chinese start-up to use its existing reasoning models as the base for a new generation of reward models to train the next generation of reasoners.

A single file, single GPU, from scratch full parameter tuning library that replicates DeepSeek R1-Zero style training.

CrystalFormer is a transformer model that generates crystal structures using space group symmetry, making crystal generation more efficient and data-friendly.

A fully automated zero-shot object counting method leveraging feature maps and self-attention mechanisms that achieves state-of-the-art accuracy on the FSC147 dataset.

The unsloth folks have figured out which piece of the new R1 model can be properly quantized. They also found some tokenizer quirks to be aware of, which make quantization slightly harder. In summary, just the MoE layers go to 1.58 bit while everything else remains in 4 or 6 with their dynamic quantization scheme.

DeepMind's AI system, Dreamer, successfully learned to collect diamonds in Minecraft without prior human guidance, highlighting a step towards general AI systems. Using reinforcement learning, Dreamer independently explores and builds a model of the game environment to predict future actions and outcomes. This advancement suggests potential applications for AI in real-world scenarios where trial and error are costly.

AI-generated Ghibli-style visuals have surged in popularity, straining OpenAI's servers and sparking debates about creativity in the AI age. While AI can rapidly produce artistic images, it lacks the human ability to experience and synthesize complex ideas and emotions. The future of creativity will focus on meaningful outputs shaped by human insight and purpose, with AI as a tool rather than a creator.

IBM silently launched a state-of-the-art speech recognition and understanding model based on its Granite series.

A study suggested that OpenAI trained its models on copyrighted content without permission.

Google's Gemini 2.5 Pro is now in public preview via the Gemini API on Google AI Studio, with Vertex AI availability coming soon.