Automated skill generation, dynamic news retrieval, and hateful meme detection.
ScienceToStartup Editorial
This week's AI research pushes the boundaries of agent capabilities and information access. We're seeing systems that can distill human expertise into reusable AI skills, dynamic retrieval trees that accelerate time-sensitive news delivery, and novel approaches to identifying hateful content. These advancements offer significant implications for how we build and interact with AI systems.
Use This Via API or MCP
Pillar articles explain the operator narrative around the same proof surfaces your agents can access directly. Use them for context, then drop into REST, MCP, Signal Canvas, or the benchmark and dataset routes for machine-readable execution.
The Rundown
Researchers introduced COLLEAGUE.SKILL, a system that automates the generation of AI skills by distilling expert knowledge from heterogeneous traces. This approach moves beyond simple prompt engineering, aiming to imbue AI agents with bounded representations of human expertise, judgment, and interaction styles. The system produces versioned skill packages with two coordinated tracks: a capability track for practices and decision heuristics, and a bounded behavior track for communication style and interaction rules. This allows for inspectable, correctable, and agent-usable skills. The open-source system, which boasts approximately 18.5k GitHub stars, has already cataloged 215 skills from 165 contributors, accumulating over 100k cumulative stars across listed skill cards. This represents a significant step towards creating person-grounded AI agents that can reliably mimic and extend human professional capabilities, offering a structured way to package and deploy specialized AI expertise.
The details
Why it matters
Startups can leverage COLLEAGUE.SKILL to rapidly develop specialized AI agents that embody unique company expertise or replicate high-performing employee workflows, accelerating product development and customer service.
📰 AI for Media
The Rundown
DynaTree, a novel two-stage framework, significantly enhances time-sensitive news retrieval by decoupling planning from retrieval. Existing agentic RAG methods often couple these, leading to high inference costs and slower performance, especially for news that requires immediate access. DynaTree first constructs a reusable retrieval tree offline using coordinated agents, mapping out the semantic space of a query topic. In the online stage, it performs lightweight daily subtree selection based on a time-localized evaluation proxy, eliminating the need for further agentic reasoning or tree modification. Deployed in the Syft production system, DynaTree's dynamically adapted variant improved survival rates from 0.32-0.53 to 0.59-0.73 during A/B testing from Jan. 28 to Feb. 6, 2026. It consistently outperformed existing production recallers daily, demonstrating the power of persistent, structure-aware semantic expansion for real-world news applications.
The details
Why it matters
For media startups and content platforms, DynaTree offers a pathway to deliver fresher, more relevant news faster. This efficiency gain can translate to better user engagement and a competitive edge in the fast-paced news cycle.
🖼️ Vision-Language Models
The Rundown
Detecting hateful memes presents a significant challenge for vision-language models (VLMs). Existing benchmarks often confound rhetorical mechanisms with target community features, hindering causal evaluation of model vulnerabilities. To address this, researchers introduced FBHM (Functionality Based Hateful Memes), a benchmark with 5,000 memes structured along 25 distinct rhetorical functionalities and 10 target communities. Benchmarking current best VLMs revealed a severe generalization gap, with models dropping to near-random performance on FBHM. To efficiently close this gap, they propose LSV (learnable steering vectors), an ultra-low data regime strategy. Using as few as 500 steering samples, LSV boosted FBHM performance by approximately 30 Macro-F1 points, outperforming in-context learning and PEFT without degrading source-domain performance. This approach allows for more robust and causal evaluation of VLM capabilities in sensitive content detection.
The details
Why it matters
Startups developing content moderation tools can use FBHM and LSV to build more robust VLM-based systems. This capability is crucial for platforms aiming to combat online hate speech effectively and responsibly.
An open platform for managing the full ML lifecycle.
A platform for tracking experiments, datasets, and model performance.
An intuitive platform for deep learning research and production.
A flexible framework for building and training ML models.
Built to make you extraordinarily productive, Cursor is the best way to code with AI.
A framework for building applications powered by LLMs.
Anthropic's Claude is seeing paid subscriptions more than double this year.
A TX lawyer, Mark Lanier, reportedly "rattled" Zuckerberg during a trial against Meta and Google.
ShinyHunters claims a cyberattack on the European Commission, though internal systems were reportedly unaffected.
Sources indicate Ross Nordeen, a cofounder at xAI, has left the company.
A feud between Sam Altman and Dario Amodei is detailed, with Amodei reportedly comparing Altman's fight with Musk to Hitler's fight with Stalin.
Chess grandmasters are finding new strategies as AI pushes classical chess toward perfect play.
A new computer chip material inspired by the human brain could slash AI energy use.
Bluesky is leaning into AI with Attie, an app for building custom feeds.
Jun 8
AI agents slash knowledge work time, new TTS models hit low latency.
Jun 5
Ultralytics YOLO26, EvoDS agents, and Humanoid-GPT push real-time vision and autonomous data science.
Jun 1
Automated skill generation, dynamic news retrieval, and hateful meme detection.