AI News Daily Digest (26-07-25)

Trump’s “Genesis Mission” turns $5B into AI-driven science grants

The Verge reports the Trump administration is launching the first “Genesis Mission” grants, earmarking $5 billion for hundreds of AI-powered science projects with a stated goal of matching Manhattan Project-scale urgency. The companion storyline: Trump’s science adviser Michael Kratsios pitches lawmakers on prioritizing AI, robotics, and nuclear energy while deprioritizing life sciences.

Read the full article here

Marking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts

This study targets a high-stakes blind spot in watermarking: clinical workflows aren’t just “another benchmark,” because tiny token-level shifts can produce clinically meaningful semantic drift. Across multiple watermarking schemes and LLM/VLM setups, the authors find watermarking can degrade medical reasoning via lexical corruption, hallucinated terminology, and misattributed or omitted image findings, arguing domain-specific auditing is required for safe medical deployment.

Read the full article here

InferenceBench: A Benchmark for Open-Ended LLM Inference Optimization by AI Agents

InferenceBench evaluates whether AI agents can truly optimize inference performance or just repackage known tricks: agents must deploy an OpenAI-compatible inference server and improve speed under a strict two-hour wall-clock budget. The benchmark isolates bottlenecks like prefill latency, decode latency, and throughput, showing strong but bounded gains (often beating naive baselines) while still trailing systematic hyperparameter search.

Read the full article here

Google Zero: The traffic deal with the open web is breaking

A Verge explainer frames “Google Zero” as the unraveling of a long-running data-for-traffic bargain between Google and publishers: indexing in exchange for huge referral flows. The fallout includes intensified publisher interest in blocking crawlers and heightened pressure as AI training deals lose momentum.

Read the full article here

DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions

DecodeShare proposes a decode-time perspective on task generality under KV-cached inference by finding a low-dimensional hidden-state subspace shared across tasks. When the shared subspace is surgically removed during decoding, decision performance drops more than removing prefill-derived or random subspaces, suggesting a causal “decode channel” with implications for activation steering reliability.

Read the full article here

Benchmarking the Personalization Capabilities of Large Language Models

This work reframes LLM personalization as a two-party persuasion problem and tests it using logged outcomes from sales outreach, captured in a new SDR corpus designed to prevent future-data leakage. Results point to a personalization plateau across frontier LLMs and deep-research agents, where models often can’t reliably separate successful from unsuccessful outreach even on Fortune 100 cohorts.

Read the full article here

Alexa Plus gets an AI update for more complex smart-home instructions

Amazon is expanding Alexa Plus with new smart home capabilities in preview, enabling it to connect with a broader set of device brands and route requests to the correct appliance automatically. An example from the rollout shows Alexa handling instruction nuance like interpreting laundry tag constraints and selecting an appropriate wash cycle.

Read the full article here

PlanE: Meta Planning of Data, Tuning, and Inference for Extractive-based LLMs

PlanE tackles a practical bottleneck for task-specific LLMs: building extractive models usually requires expensive instruction-tuning, and there’s limited guidance on how to choose data, tuning, and prompt inference together. The proposed framework decomposes and then plans the construction pipeline, including a planner that selects base models and DTI combinations to improve efficiency and generalize across optimization objectives.

Read the full article here

Midjourney buys the astrology app Co-Star

Midjourney is moving beyond image generation into consumer guidance by acquiring personalized astrology app Co-Star, according to The Verge and Bloomberg. The deal signals how generative-AI startups are still hungry for new product surfaces, with Co-Star combining daily horoscopes, claimed NASA data, and AI to produce tailored advice.

Read the full article here

AINTMA: Agentic AI Architecture for Autonomous Test Management

This paper describes AINTMA, a multi-agent architecture that turns software test management into an autonomous “quality intelligence ecosystem” across distributed cloud environments. It combines specialized agents for discovery, risk assessment, RL-based prioritization, execution orchestration, and generative quality reporting, while using a secure zero-trust communication layer; across 12 projects, it reports faster cycles and fewer defect escapes.

Read the full article here

Claude Opus 5 is “close” to Claude Fable 5 in many domains

Anthropic’s Opus 5 update positions the new model as approaching the capabilities of Claude Fable 5 across many domains, with particular strength for complex coding tasks. The release arrives amid ongoing deployment and policy pressure around Anthropic’s other model line, with prior mentions of export-control-related safeguards.

Read the full article here

DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding

DC-Leap targets a speed bottleneck in diffusion LLM parallel decoding where conservative confidence thresholds slow inference. The method introduces Dynamic Contiguous Verification to enforce causal constraints during decoding and neutralize dependency errors, and it pairs this with draft-guided “leaping” to extend context – achieving large speedups on standard benchmarks without sacrificing output quality.

Read the full article here

JAXBench: Benchmarking Autonomous TPU Kernel Optimization

JAXBench fills a missing piece in evaluation for TPU-focused optimization by providing a TPU-native benchmark suite for generating and validating Pallas kernels. The authors show that conditioning on curated TPU documentation materially improves correctness and yields speedups, and they release the benchmark and harness to encourage open contributions to TPU kernel optimization.

Read the full article here