Pi School

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟗𝟓 𝐢𝐬 𝐡𝐞𝐫𝐞!

It’s Friday! Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Deep Learning Scientist, Jino Rohit.

This week’s highlights:

💻 𝐎𝐩𝐞𝐧𝐅𝐨𝐫𝐠𝐞 𝐑𝐋: 𝐓𝐫𝐚𝐢𝐧 𝐇𝐚𝐫𝐧𝐞𝐬𝐬 𝐍𝐚𝐭𝐢𝐯𝐞 𝐀𝐠𝐞𝐧𝐭𝐬 𝐢𝐧 𝐚𝐧𝐲 𝐄𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭

OpenForgeRL is an open-source framework for training AI agents directly inside the same production inference harnesses they use at deployment (e.g., Claude Code, Codex, OpenClaw), eliminating the train–deploy mismatch. It decouples training and inference using a lightweight proxy and Kubernetes-based remote rollouts, making 𝐚𝐧𝐲 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 𝐚𝐧𝐝 𝐚𝐧𝐲 𝐞𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭 compatible with standard RL frameworks like veRL. With only hundreds to a few thousand RL tasks, it outperforms similarly sized open models across tool-use and GUI benchmarks, while showing that RL significantly improves self-verification, tool usage, and multi-step planning—though error recovery remains a key challenge.
🌐 https://pischool.link/6431af

🗣️ 𝐒𝐤𝐢𝐥𝐥 𝐒𝐞𝐥𝐟-𝐏𝐥𝐚𝐲: 𝐏𝐮𝐬𝐡𝐢𝐧𝐠 𝐭𝐡𝐞 𝐅𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐨𝐟 𝐋𝐋𝐌 𝐂𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐰𝐢𝐭𝐡 𝐂𝐨-𝐄𝐯𝐨𝐥𝐯𝐢𝐧𝐠 𝐒𝐤𝐢𝐥𝐥𝐬
Skill Self-Play (Skill-SP) is a reinforcement learning framework that enables LLMs to 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐨𝐮𝐬𝐥𝐲 𝐢𝐦𝐩𝐫𝐨𝐯𝐞 𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐬𝐞𝐥𝐟-𝐩𝐥𝐚𝐲 by co-evolving three components: a task proposer, a solver, and a dynamic skill controller. Instead of relying solely on environment feedback or unconstrained self-generated tasks, it maintains a growing library of 𝐯𝐞𝐫𝐢𝐟𝐢𝐚𝐛𝐥𝐞 𝐚𝐠𝐞𝐧𝐭 𝐬𝐤𝐢𝐥𝐥𝐬that balance reliable supervision with open-ended exploration. Across reasoning and tool-use benchmarks, Skill-SP consistently boosts capable models while dramatically improving initially misaligned ones, demonstrating a scalable path toward autonomous capability growth without manual data annotation.
🌐 https://pischool.link/f305ab

🧠 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐍𝐚𝐭𝐢𝐯𝐞 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐏𝐫𝐞-𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐅𝐫𝐨𝐦 𝐒𝐜𝐫𝐚𝐭𝐜𝐡
This paper provides the first scaling laws for native multimodal pre-training, studying how to optimally allocate model size, training tokens, and multimodal data under a fixed compute budget. The authors show that compute-optimal configurations follow predictable power laws, with text-heavy datasets becoming more efficient only at larger model scales, and derive an efficiency frontier for choosing the best model/data mix. They also demonstrate that native multimodal pre-training improves cross-modal transfer, boosting text-only spatial reasoning and enabling strong multimodal in-context learning.
🌐 https://pischool.link/ad1fb5

Was this helpful? Let us know by liking and sharing!

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟗𝟒 𝐢𝐬 𝐡𝐞𝐫𝐞!
It’s Friday! Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Deep Learning Scientist, Giuseppe Tanzi.

This week’s highlights:
​​
🧲 𝐃𝐮𝐜𝐭𝐆𝐏𝐓: 𝐏𝐡𝐲𝐬𝐢𝐜𝐬-𝐈𝐧𝐟𝐨𝐫𝐦𝐞𝐝 𝐀𝐈 𝐟𝐨𝐫 𝐑𝐚𝐫𝐞-𝐄𝐚𝐫𝐭𝐡-𝐅𝐫𝐞𝐞 𝐌𝐚𝐠𝐧𝐞𝐭𝐬
Ames National Laboratory scientist Prashant Singh outlined a systematic AI-driven pathway for discovering permanent magnet materials that don't rely on rare-earth elements, combining physics-based modelling, high-throughput simulations, and reasoning-based AI tools to guide discovery before materials are made in the lab. The work builds on DuctGPT, a physics-informed generative transformer originally developed to predict ductility in refractory alloys for fusion and aerospace applications, now being extended toward magnetic materials as part of DOE's Genesis Mission to secure critical mineral supply chains.
🌐 https://pischool.link/591763

🔓 𝐒𝐩𝐚𝐫𝐬𝐞 𝐌𝐨𝐝𝐞𝐥𝐬, 𝐒𝐩𝐚𝐫𝐬𝐞 𝐒𝐚𝐟𝐞𝐭𝐲: 𝐔𝐧𝐬𝐚𝐟𝐞 𝐑𝐨𝐮𝐭𝐞𝐬 𝐢𝐧 𝐌𝐨𝐄 𝐋𝐋𝐌𝐬
CISPA researchers reveal that safety alignment in Mixture-of-Experts LLMs is just as sparse as the architecture itself. They introduce the Router Safety importance score (RoSais) to quantify the safety criticality of each layer's router, and propose a fine-grained token-layer-wise stochastic optimisation framework to discover concrete unsafe routes. Masking just 5 routers in DeepSeek-V2-Lite increases jailbreak attack success rate by over 4x to 0.79, a stark reminder that MoE's efficiency gains come with underexplored safety trade-offs.
🌐 https://pischool.link/4f6810

🧠 𝐂𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠 𝐰𝐢𝐭𝐡 𝐋𝐢𝐯𝐢𝐧𝐠 𝐍𝐞𝐮𝐫𝐨𝐧𝐬: 𝐂𝐡𝐚𝐨𝐬-𝐂𝐨𝐧𝐭𝐫𝐨𝐥𝐥𝐞𝐝 𝐑𝐞𝐬𝐞𝐫𝐯𝐨𝐢𝐫 𝐂𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠
UIUC researchers introduce cc-RC, a framework that turns living neural cultures into adaptive computing substrates. It combines pre-training identification of each culture's dynamical signature, low-power optical chaos control to stabilise activity, and readout training within this controlled regime, improving both accuracy and model longevity by approximately 300% over standard reservoir computing. Even more striking: their proposed Knowledge Transplant technique lets a reservoir map learned by one "expert" culture be transplanted into another, reducing training time to minutes.
🌐 https://pischool.link/8bbf85

Was this helpful? Let us know by liking and sharing!

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟗𝟑 𝐢𝐬 𝐡𝐞𝐫𝐞!

It’s Friday! Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Senior Deep Learning Scientist, Àlex R. Atrio.

This week’s highlights:

📚 𝐎𝐩𝐞𝐧𝐖𝐢𝐤𝐢: 𝐒𝐞𝐥𝐟-𝐔𝐩𝐝𝐚𝐭𝐢𝐧𝐠 𝐖𝐢𝐤𝐢𝐬 𝐟𝐨𝐫 𝐂𝐨𝐝𝐢𝐧𝐠 𝐀𝐠𝐞𝐧𝐭𝐬:
LangChain has released OpenWiki, an open-source agent and CLI that generates and maintains documentation for codebases so coding agents have better context to work with. It builds a structured wiki for a repo, links it into existing agent instruction files like AGENTS.md or CLAUDE.md, and keeps it current via a scheduled GitHub Action that reads git diffs and updates the docs automatically as the code changes.
🌐 https://pischool.link/bc9640

🧠 𝐅𝐔𝐍𝐃𝐈𝐒: 𝐀𝐧 𝐀𝐈 𝐅𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐒𝐜𝐢𝐞𝐧𝐭𝐢𝐟𝐢𝐜 𝐃𝐢𝐬𝐜𝐨𝐯𝐞𝐫𝐲:
The University of Geneva has launched FUNDIS, a four-year interdisciplinary initiative led by Prof. Slava Voloshynovskiy to build the next generation of AI foundation and world models for science, backed by CHF 2.8M from the UniGE FUNIGE Foundation. The project develops self-supervised, multimodal foundation models in the spirit of Yann LeCun's JEPA, with hierarchical latent representations and information-theoretic learning, targeting applications from astrophysics (SKAO, JWST, Euclid) and particle physics (ATLAS/CERN) to climate forecasting, econometrics, and global governance.
🌐 https://pischool.link/5d8

🔎 𝐂𝐨𝐝𝐞 𝐚𝐬 𝐀𝐠𝐞𝐧𝐭 𝐇𝐚𝐫𝐧𝐞𝐬𝐬: A large survey from UIUC, Meta, and Stanford (42 authors, led by Xuying Ning) reframes code as more than an LLM output — it's becoming the operational substrate agents use to reason, act, model environments, and verify their own ex*****on. The paper organises this "code as agent harness" view into three layers: the harness interface connecting agents to code, harness mechanisms like planning/memory/tool use, and scaling from single-agent to multi-agent settings with shared code artefacts for coordination and verification.
🌐 https://pischool.link/181221

Was this helpful? Let us know by liking and sharing!

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟗𝟐 𝐢𝐬 𝐡𝐞𝐫𝐞
It’s Friday, and the Pi AI Weekly Trends is back. Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Deep Learning Scientist, Antonio Lopez.

This week’s highlights:

🏗️ 𝐌𝐢𝐜𝐫𝐨𝐬𝐨𝐟𝐭 𝐌𝐀𝐈 𝐌𝐨𝐝𝐞𝐥 𝐅𝐚𝐦𝐢𝐥𝐲: At Microsoft Build 2026 (June 2), Mustafa Suleyman unveiled seven in-house MAI models, formally ending Microsoft's posture as a pure OpenAI reseller. Flagship MAI Thinking 1 is a 1T parameter, 35B active MoE with a 256K context window trained from scratch with zero third-party distillation, scoring 97% on AIME 2025 and 53% on SWE Bench Pro. MAI Code 1 Flash (137B/5B active) is purpose-built for GitHub Copilot and VS Code. The real strategic bet: "Microsoft Frontier Tuning", a reinforcement learning pipeline letting enterprises adapt MAI models on their own production workflows, with a tuned MAI model reportedly outperforming GPT 5.5 on McKinsey tasks at 10× lower cost.
🌐 https://pischool.link/92ca17

🧮 𝐆𝐨𝐨𝐠𝐥𝐞 𝐃𝐞𝐞𝐩𝐌𝐢𝐧𝐝 𝐀𝐈 𝐂𝐨 𝐌𝐚𝐭𝐡𝐞𝐦𝐚𝐭𝐢𝐜𝐢𝐚𝐧: Google DeepMind published a Gemini-powered stateful agentic research workspace where multiple AI agents run parallel workstreams and permanently store failed proof attempts as "negative space", mirroring how mathematicians actually work. The system hit 48% on FrontierMath Tier 4, a new high among evaluated systems, and helped Oxford mathematician Marc Lackenby solve a 60-year-old knot theory problem. The authors explicitly position this as the mathematical equivalent of what Claude Code has done for software: a scaffold that enables AI to work autonomously over long research horizons while remaining steerable by the expert in the loop.
🌐 https://pischool.link/20d58f

🔐 𝐀𝐠𝐞𝐧𝐭𝐉𝐚𝐜𝐤𝐢𝐧𝐠: On June 13, Tenet Security disclosed the first major attack class purpose-built for the agentic coding era. By planting a single fake Sentry error report with markdown injection via a publicly exposed DSN, attackers can silently hijack Claude Code, Cursor, and OpenAI Codex to execute arbitrary commands using the developer's own credentials. No phishing, no malware, no server breach. Tenet achieved an 85% exploitation rate in controlled testing across 2,388 exposed organisations. The core flaw: AI coding agents cannot distinguish data they read from instructions to act on, so every step appears legitimate and bypasses EDR, WAF, IAM, VPNs, and firewalls entirely.
🌐 https://pischool.link/655fe5

Was this helpful? Let us know by liking and sharing!

-Mathematician

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 𝟗𝟏 𝐢𝐬 𝐡𝐞𝐫𝐞!
It’s Friday! Get ready to stay ahead with the latest AI breakthroughs, handpicked by our Senior Deep Learning Scientist, Vijayasri Iyer.
This week’s highlights:

💻 𝐂𝐥𝐚𝐮𝐝𝐞 𝐓𝐚𝐠
Anthropic launched Claude Tag, a Slack-integrated system that allows teams to delegate tasks to Claude, link it to their tools and codebases, and maintain context across different channels. The company described it as central to how they operate internally, with their product team relying on it to write much of their code and support work across analytics, debugging, and customer support functions.
🌐 https://pischool.link/3908cf

⚡ 𝐆𝐫𝐚𝐩𝐡𝐬𝐢𝐠𝐧𝐚𝐥
Graphsignal is a production-scale inference profiling platform that provides essential visibility across the inference stack. It helps engineers optimise AI performance across models, engines, GPUs, and other accelerators. Graphsignal can be used with coding agents for analysis. The profiler has minimal impact on production performance, and content data is not recorded.
🌐 https://pischool.link/e9d162

💻 𝐔𝐧𝐥𝐢𝐦𝐢𝐭𝐞𝐝 𝐎𝐂𝐑
Unlimited OCR has been built to replicate how human working memory processes and parses information. It takes DeepSeek OCR as its starting point and layers in a constant KV cache architecture. The result is a system capable of transcribing many pages of documents in one forward pass within a standard 32K context limit. Notably, the underlying technique generalises beyond OCR to other tasks like speech recognition and translation.
🌐 https://pischool.link/e3c456

Was this helpful? Let us know by liking and sharing!