Agentic Dev Breaks Testing, Requires JiTTesting

Tom Preston-Werner, former GitHub CEO, launched Entire, an AI agent developer platform backed by a $60 million seed round. Their open-source Entire CLI automatically captures agent prompts, transcripts, and token usage as versioned 'Checkpoints' in Git on every commit, improving traceability for human-AI collaboration.

Apptronik secured a $520 million Series A extension, bringing its total funding to over $935 million, with investors including Google and Mercedes-Benz. This investment will accelerate production of their humanoid robot, Apollo, designed for logistics and manufacturing, signals continued strong investor interest in robotics, relevant for founders fundraising in the space.

LLMs Don't Need More Parameters, They Need Loops

A new paper, 'Scaling Latent Reasoning via Looped Language Models,' proposes integrating 'loops' into LLM architecture as an alternative to solely increasing parameters. This method addresses reasoning challenges using concepts like dynamic termination and looped KV caching, suggesting a new path to achieve advanced reasoning without relying on ever-larger models.

Agent Skill Scanner Catches Prompt Injections

Cisco AI Defense shipped Skill Scanner, a new security tool for AI agent skills. It uses static analysis, LLM-as-a-judge, and behavioral dataflow to detect prompt injections, data exfiltration, and malicious code patterns. The scanner integrates with CI/CD and offers "best-effort" detection, still requiring human review for full coverage.

$315M Funds Next-Gen Video World Models

AI video generation startup Runway raised $315 million at a $5.3 billion valuation. The funding will pre-train next-generation 'world models' to power new products and existing video creation features.

Z.ai released GLM-5, an open-source LLM that scales up parameters and pre-training data. It integrates DeepSeek Sparse Attention for lower deployment costs and longer context, plus new asynchronous RL infrastructure to boost training efficiency. GLM-5 shows strong performance among open models in reasoning, coding, and agentic tasks, particularly excelling at long-horizon operations like business simulations.

Jeff Dean: Energy is the New FLOPs Bottleneck

Google's Jeff Dean highlights distillation as crucial for making efficient "Flash" models from larger "Frontier" models. He argues energy consumption (picojoules), not FLOPs, is becoming the primary bottleneck for scaling AI. Future AI assistants will process trillions of tokens by intelligently retrieving information, moving past reliance on ever-larger context windows.

LLM Fine-Tuning Gets 3.5x Speedup with Chronicals

Chronicals, a new framework for LLM fine-tuning, achieves a 3.51x speedup over Unsloth, processing 41,184 tokens/second for full fine-tuning. It tackles memory and compute bottlenecks using fused Triton kernels, Cut Cross-Entropy, FlashAttention, LoRA+, and sequence packing. The framework reduces memory usage and increases model FLOPs utilization.