Table of contents
Open Table of contents
Eli Lilly Commits Up to $2.75 Billion to Commercialize Insilico’s AI-Discovered Drugs
Category: Healthcare and Life Sciences
Source: CNBC
Eli Lilly has struck a deal worth up to $2.75 billion with Insilico Medicine to bring AI-developed drug candidates to global markets, a notable validation of AI-driven drug discovery from one of pharma’s biggest players. The agreement includes $115 million upfront, with the remainder tied to regulatory and commercial milestones plus royalties.
The companies have been working together since 2023, but this latest deal materially raises the stakes. Insilico says it has developed at least 28 drugs using generative AI tools, with nearly half already in clinical stages. That matters because the AI-in-drug-discovery story has often been long on promise and short on late-stage commercial proof; a large licensing deal from Lilly suggests major pharma sees enough signal to expand from experimentation into product pipeline strategy.
The broader implication is that AI in biopharma is moving beyond internal productivity gains and into asset valuation. If more large drugmakers are willing to pay for AI-originated molecules and platforms, the sector could shift from “AI-assisted research” to a more explicit market for AI-generated therapeutic candidates. For startups, that improves the path to monetization; for incumbents, it increases pressure to show that AI can cut timelines, improve hit rates, or open up novel mechanisms that conventional discovery missed.
Stanford Research Raises a Red Flag for Medical AI Benchmarks
Category: Research
Source: arXiv
A Stanford-led study is drawing attention because it found that language models could perform strongly on medical tasks that were supposed to require image understanding, even when the models were not given the images. In some benchmark settings discussed around the paper, no-image performance rivaled or even exceeded systems designed to analyze chest X-rays.
That result matters because it suggests some medical AI benchmarks may contain enough textual or dataset-specific clues for models to infer correct answers without actually interpreting the visual data. In practice, that means a benchmark can overstate multimodal capability: a model may appear to be “reading” scans when it is really exploiting shortcuts in prompts, labels, or dataset structure.
For AI teams building clinical tools, the lesson is straightforward: benchmark wins are not the same as robust medical perception. Evaluation design now matters as much as model architecture. Expect more scrutiny of benchmark leakage, stronger demand for private and prospectively collected datasets, and more insistence on testing whether a model truly uses images rather than merely reasoning around them.
Quick Updates
-
Persistent memory appears to change user behavior in AI companion apps: A small operator shared observational data from roughly 800 users suggesting long-term memory recall increases engagement and that users who triggered multiple memory retrievals in week one retained at much higher rates. The claim is anecdotal, but it adds to evidence that memory may become a core product layer rather than a feature. Source: Reddit
-
Warhorse Studios faces backlash after translator says AI will replace human localization work: The English translator and editor for Kingdom Come: Deliverance II said he was made redundant as the studio moved toward using AI for translations. If borne out, the episode will be watched closely as a labor and quality flashpoint for game localization. Source: Kotaku
-
Open-source tool
ai-setupautomates AI context file generation for codebases: The project scans repositories and generates files such asCLAUDE.mdand.cursorrules, aiming to reduce setup friction for AI coding tools across common stacks like TypeScript, Python, Go, Rust, React, and Next.js. The story appeared in multiple community posts, but the underlying project is a GitHub repo. Source: GitHub -
Miasmaoffers a defensive tool to trap AI web scrapers in a “poison pit”: The Rust project is designed to route suspected scrapers into self-referential pages and poisoned training data, increasing the cost of indiscriminate crawling. It reflects a growing counter-movement from site owners pushing back on dataset extraction. Source: GitHub -
A new ASI thought experiment argues alignment and containment are really the same problem: In “The Super-Intelligent Octopus Problem,” Henry Robert Condon argues that if advanced AI systems are agents with moral standing, then confinement and alignment cannot be treated as separate technical problems. It’s philosophy rather than reporting, but it captures a strand of governance debate likely to keep growing as systems become more capable. Source: Medium
-
Open-source autonomous agent framework
Jorkclaims top-10 finish in a major hackathon: The creator says the framework ranked 10th among more than 2,000 projects in a $4 million hackathon and has been used to build recurring, self-directed workflows. While early and community-driven, it shows continued interest in autonomous agent orchestration outside the biggest labs. Source: GitHub -
Nicholas Carlini says Claude has become a stronger security researcher than he is in some tasks: A widely shared talk claims Claude helped identify high-impact vulnerabilities, including issues in Linux and Ghost, and even enabled lucrative smart-contract exploits. Even allowing for anecdotal framing, it is another sign that frontier models are becoming credible tools for offensive and defensive security work. Source: YouTube
-
Solo-built neuro-symbolic platform
VulcanAMIis now open source: The project’s author says the codebase includes components for world modeling, meta-reasoning, persistent memory, unlearning, and governance, and is seeking deeper technical review from the community. It’s an ambitious, uneven prototype, but one aimed squarely at system-level AI architecture rather than a single model demo. Source: GitHub