Pi School

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 #𝟖𝟖
3 things our Deep Learning Scientist Giuseppe Tanzi is paying attention to this week.

🚀 𝐍𝐀𝐒𝐀 𝐇𝐏𝐒𝐂: 𝐀𝐈-𝐑𝐞𝐚𝐝𝐲 𝐒𝐩𝐚𝐜𝐞𝐜𝐫𝐚𝐟𝐭 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐨𝐫 𝐚𝐭 𝟓𝟎𝟎× 𝐂𝐮𝐫𝐫𝐞𝐧𝐭 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞
NASA and Microchip Technology are testing the High Performance Spaceflight Computing (HPSC) processor, a radiation-hardened chip that delivers performance hundreds of times that of current spaceflight computers while surviving tests designed to mimic the harsh conditions of space. The technology will enable autonomous spacecraft to use artificial intelligence to respond in real time to complex situations and environments where human input isn't possible.
🌐 https://pischool.link/nasa

⚡𝐄𝐱𝐜𝐢𝐭𝐨𝐧-𝐏𝐨𝐥𝐚𝐫𝐢𝐭𝐨𝐧𝐬: 𝐋𝐢𝐠𝐡𝐭-𝐌𝐚𝐭𝐭𝐞𝐫 𝐏𝐚𝐫𝐭𝐢𝐜𝐥𝐞𝐬 𝐟𝐨𝐫 𝐀𝐈 𝐂𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠
Researchers at the University of Pennsylvania demonstrated all-optical signal switching using exciton-polaritons, using only about 4 quadrillionths of a joule of energy, far below the energy needed to power a tiny LED briefly. If scaled, the technology could lead to photonic chips capable of processing information directly from cameras without repeated conversions between light and electricity, lowering the massive energy demands of large AI systems and potentially supporting basic quantum computing functions.
🌐 https://pischool.link/Ectnplrtn

🧠 𝐓𝐋𝐓: 𝐓𝐚𝐦𝐢𝐧𝐠 𝐭𝐡𝐞 𝐋𝐨𝐧𝐠 𝐓𝐚𝐢𝐥 𝐢𝐧 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐋𝐋𝐌 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠
MIT researchers identified that the rollout phase consumes a disproportionately large fraction (~85%) of total RL training step time, creating a major bottleneck for reasoning LLMs. Their solution, TLT, uses idle processor downtime to continuously train a lightweight drafter model on the fly, keeping it aligned with the target model at zero extra cost. Tested across multiple reasoning LLMs, TLT accelerated training between 70 and 210 per cent while preserving the accuracy of each model.
🌐 https://pischool.link/MIT

𝐖𝐡𝐢𝐜𝐡 𝐨𝐟 𝐭𝐡𝐞𝐬𝐞 𝐰𝐨𝐮𝐥𝐝 𝐜𝐡𝐚𝐧𝐠𝐞 𝐡𝐨𝐰 𝐲𝐨𝐮 𝐰𝐨𝐫𝐤? 𝐃𝐫𝐨𝐩 𝐢𝐭 𝐢𝐧 𝐭𝐡𝐞 𝐜𝐨𝐦𝐦𝐞𝐧𝐭𝐬.

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 #𝟖𝟕 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:

🌍 𝐄𝐚𝐫𝐭𝐡𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐄𝐱𝐩𝐥𝐨𝐫𝐞𝐫. Cross-modal search over global Sentinel-2 imagery: query by text, image, or geolocation, compare SigLIP, DINOv2, SatCLIP, and FarSLIP on MajorTOM embeddings, zero setup in the browser.
🌐 https://pischool.link/c89ca5

🔎 𝐒𝐜𝐫𝐚𝐩𝐥𝐢𝐧𝐠: 𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐖𝐞𝐛 𝐒𝐜𝐫𝐚𝐩𝐢𝐧𝐠.. Open-source Python scraping that beats Cloudflare Turnstile, adapts when page layouts change, and ships an MCP server for agent workflows. 774× faster text extraction than BeautifulSoup in benchmarks.
🌐 https://pischool.link/d0f03a

🧠 𝐏𝐚𝐠𝐞𝐈𝐧𝐝𝐞𝐱. Vectorless, reasoning-based RAG that indexes documents as trees instead of chunks. 98.7% on FinanceBench, built for contracts, reports, and long structured PDFs.
🌐 https://pischool.link/e4f356

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

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 #𝟖𝟔 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:

🗣️𝐎𝐩𝐞𝐧𝐀𝐈 𝐑𝐞𝐚𝐥𝐭𝐢𝐦𝐞 𝐕𝐨𝐢𝐜𝐞 𝐈𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞
Three new audio models are coming to the API, enabling a new class of real-time voice applications. GPT-Realtime-2 brings GPT-5-class reasoning to voice for the first time, handling complex requests and sustaining natural conversation. GPT-Realtime-Translate offers live speech translation across 70+ input languages into 13 output languages, keeping pace with the speaker in real time. GPT-Realtime-Whisper rounds out the trio with streaming speech-to-text that transcribes live as the speaker talks.
🌐 https://pischool.link/22085d

🤖 𝐒𝐮𝐛𝐐: 𝐓𝐡𝐞 𝐅𝐢𝐫𝐬𝐭 𝐅𝐮𝐥𝐥𝐲 𝐒𝐮𝐛𝐪𝐮𝐚𝐝𝐫𝐚𝐭𝐢𝐜 𝐋𝐋𝐌
SubQ 1M is the first LLM built on a fully subquadratic architecture, where compute scales linearly with context length rather than quadratically. This simultaneously enables longer context windows, state-of-the-art retrieval accuracy, faster inference, and lower cost; improvements that have historically traded off against one another. SubQ breaks that tradeoff entirely, reducing attention compute by nearly 1,000x compared to frontier transformer models and making million-token context windows a practical reality. SubQ is available for early access as an API, a coding agent and a long context search tool.
🌐 https://pischool.link/32a53b

🧠𝐃𝐨 𝟑𝐃 𝐋𝐚𝐫𝐠𝐞 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬 𝐑𝐞𝐚𝐥𝐥𝐲 𝐔𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝 𝟑𝐃 𝐒𝐩𝐚𝐭𝐢𝐚𝐥 𝐑𝐞𝐥𝐚𝐭𝐢𝐨𝐧𝐬𝐡𝐢𝐩𝐬?
This paper questions whether 3D Large Language Models truly understand spatial relationships or merely exploit textual shortcuts. The authors show that a text-only model can match or outperform existing 3D-LLMs on the SQA3D benchmark without any 3D input, revealing a fundamental gap between benchmark performance and genuine 3D reasoning. To address this, they introduce a more rigorous evaluation benchmark and a 3D-reweighted training objective that pushes models to rely on visual 3D cues, yielding substantial gains in spatial reasoning.
🌐 https://pischool.link/58e608

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

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 #𝟖𝟓 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:

💻 𝐌𝐨𝐥𝐦𝐨𝐀𝐜𝐭𝟐 𝐀𝐜𝐭𝐢𝐨𝐧 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐌𝐨𝐝𝐞𝐥𝐬 𝐟𝐨𝐫 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝐃𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭
MolmoAct2 is an open Vision-Language-Action model tackling real-world deployment issues like latency, cost, and reliability. It introduces a spatial reasoning VLM backbone, large-scale teleoperation datasets, and an open action tokeniser. A hybrid architecture + adaptive reasoning (MolmoAct2-Think) cuts latency while preserving grounding. It outperforms strong baselines and even frontier models in embodied reasoning benchmarks.
🌐 https://pischool.link/9ec593

🗣️ 𝐌𝐚𝐦𝐨𝐝𝐚𝟐.𝟓: 𝐄𝐧𝐡𝐚𝐧𝐜𝐢𝐧𝐠 𝐔𝐧𝐢𝐟𝐢𝐞𝐝 𝐌𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐌𝐨𝐝𝐞𝐥 𝐰𝐢𝐭𝐡 𝐃𝐢𝐓-𝐌𝐨𝐄
Mamoda2.5 unifies multimodal understanding and generation using an AR–Diffusion framework with a MoE-based Diffusion Transformer (25B params, ~3B active). It achieves top-tier video generation and editing, rivalling leading proprietary models. A distillation + RL pipeline compresses 30-step editing into just 4 steps, enabling up to 95.9x faster inference. Already deployed in real-world ad workflows with ~98% success in video editing tasks.
🌐 https://pischool.link/0060fb

🧠 𝐀 𝐓𝐡𝐞𝐨𝐫𝐲 𝐨𝐟 𝐆𝐞𝐧𝐞𝐫𝐚𝐥𝐢𝐬𝐚𝐭𝐢𝐨𝐧 𝐢𝐧 𝐃𝐞𝐞𝐩 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠
This work shows how the Neural Tangent Kernel separates signal vs noise, letting models generalise even while memorising. SGD amplifies the true signal while pushing noise into “invisible” dimensions, explaining phenomena like double descent and grokking. It also derives a novel population risk objective from a single training run.
🌐 https://arxiv.org/abs/2605.01172

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

🚀 𝐏𝐢 𝐀𝐈 𝐖𝐞𝐞𝐤𝐥𝐲 𝐓𝐫𝐞𝐧𝐝𝐬 #𝟖𝟒 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:

⚡ 𝐓𝐮𝐫𝐛𝐨𝐐𝐮𝐚𝐧𝐭: Google Research just presented TurboQuant at ICLR 2026, a training-free, model-agnostic KV cache compression algorithm that shrinks the memory bottleneck of LLM inference down to 3.4 bits per element with near-zero quality loss. Key innovation: a two-stage pipeline combining PolarQuant (polar-coordinate rotation + scalar quantisation) and a 1-bit QJL residual correction, achieving 6x memory reduction and up to 8x attention speedup on H100 GPUs. No fine-tuning, no calibration data, works on any transformer architecture. Community PyTorch and Rust implementations were live on PyPI within 48 hours of the paper dropping.
🌐 https://pischool.link/bd0a86

🤖 𝐒𝐨𝐧𝐲 𝐀𝐈 𝐀𝐜𝐞, 𝐑𝐨𝐛𝐨𝐭 𝐁𝐞𝐚𝐭𝐬 𝐏𝐫𝐨𝐟𝐞𝐬𝐬𝐢𝐨𝐧𝐚𝐥 𝐀𝐭𝐡𝐥𝐞𝐭𝐞𝐬: Published on the cover of Nature this week, Sony AI's Ace is the first autonomous robotic system to defeat professional-level human table tennis players in real-world competition. This is not a simulation win. It operates with millisecond-level perception, planning, and control in a dynamic physical environment. Ace combines reinforcement learning with advanced sensors to adapt in real time, outperforming elite and professional opponents across December 2025 and March 2026 matches. The implications extend far beyond sport: this is a landmark proof that AI can operate safely and precisely at the edge of human performance in unstructured physical spaces.
🌐 https://pischool.link/0cca5f

🧠 𝐃𝐞𝐞𝐩𝐒𝐞𝐞𝐤 𝐕𝟒: DeepSeek just released preview versions of its new flagship model, V4-Pro (1.6T parameters, 49B active) and V4-Flash (284B parameters, 13B active), both open-source under the MIT license. Key architectural innovation: a Hybrid Attention Architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), enabling a 1M-token context window while using only 27% of the inference FLOPs and 10% of the KV cache size of its predecessor. V4-Pro claims the top spot among open-source models in coding and math benchmarks, rivalling leading closed-source models at a fraction of the cost.
🌐 https://pischool.link/882e8c

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