Olmo 3: Charting a path through the model flow to lead open-source AI

AI2 has released Olmo 3, a family of fully open 7B and 32B LLMs, emphasizing transparency by open-sourcing the entire "model flow"—all training data, code, and intermediate checkpoints. This approach enables researchers to trace model behaviors back to training data and fork the development process at any stage. The family includes a strong base model, an instruction-tuned variant, and the flagship Olmo 3-Think, which exposes intermediate reasoning traces and shows competitive performance against other open-weight models. The release is accompanied by new open datasets (Dolma 3, Dolci) and a full suite of open-source tooling for training, data processing, and evaluation.

Read more

Boom, bubble, bust, boom. Why should AI be different?

The text argues that the current AI boom is a massive bubble, larger and more concentrated than the 1999 dot-com era, and its collapse is imminent. This bubble is characterized by unprecedented capital spending without clear monetization paths, excessive leverage through complex circular financing deals involving major players like NVIDIA and OpenAI, and extreme VC valuations for startups. The geopolitical race with China both fuels this frenzy and poses a significant risk, as a foreign technical breakthrough could rapidly devalue current US-centric investments.

Read more

Microsoft AI CEO Puzzled by People Being Unimpressed by AI

Microsoft AI CEO Mustafa Suleyman expressed confusion over the public being "unimpressed" by AI, calling critics "cynics" in response to backlash against plans for an "agentic OS." The negative sentiment stems not from the core technology, but from Microsoft's strategy of prioritizing forced AI integration over addressing fundamental user concerns with Windows, such as UI, security, and privacy.

Read more

Google must double AI serving capacity every 6 months to meet demand

Google's head of AI infrastructure stated the company must double its AI serving capacity every six months to meet demand, targeting a 1000x increase in 4-5 years. This exponential growth is being addressed through massive capex, more efficient models, and custom silicon like their new TPUs. CEO Sundar Pichai confirmed that compute is a major bottleneck limiting product rollouts, justifying the aggressive investment despite concerns of an AI bubble.

Read more

Show HN: OCR Arena – A playground for OCR models

A new OCR-focused leaderboard ranks models via an ELO system based on anonymous, head-to-head battles. Google's Gemini 2.5 Pro and Gemini 3 Preview currently hold the top spots, outperforming several GPT-5.1 versions and other specialized OCR models like dots.ocr and Qwen3-VL-8B.

Read more

AI-Newton: Concept-Driven Physical Law Discovery System Without Prior Knowledge

AI-Newton is a concept-driven system that autonomously discovers physical laws from raw, noisy data without supervision or prior domain knowledge. It uses a knowledge base and an autonomous workflow to define its own concepts and derive relationships. As a proof of concept, the system successfully rediscovered fundamental laws of Newtonian mechanics, including Newton's second law and the law of gravitation.

ArXiv

Back to Basics: Let Denoising Generative Models Denoise

This paper argues that diffusion models should directly predict clean data rather than noise, based on the manifold assumption. The authors introduce JiT (Just image Transformers), a simple Transformer architecture that operates on large pixel patches without tokenizers, pre-training, or extra losses. This approach achieves competitive results on high-resolution ImageNet, demonstrating that predicting clean data allows simpler models to be effective in high-dimensional spaces where noise-prediction can fail.

ArXiv

Recap: Reproducing Copyrighted Data from LLMs Training with an Agentic Pipeline

RECAP is a new agentic pipeline designed to elicit memorized training data from LLMs. It uses an iterative feedback loop where a secondary LLM evaluates the target model's output against a reference, generates correction hints, and feeds them back to refine subsequent generations. The system, which includes a jailbreaking module to bypass alignment refusals, demonstrated a nearly 24% ROUGE-L score improvement for copyrighted text extraction on the EchoTrace benchmark.

ArXiv

Reinforcement Learning Control of Quantum Error Correction

To combat environmental drift in quantum computers without halting for recalibration, this work employs a reinforcement learning agent that unifies calibration with computation. The agent uses error detection events from the quantum error correction process as a learning signal to continuously steer physical control parameters and stabilize the system in real-time. This approach was experimentally shown to improve logical error rate stability 3.5-fold over traditional methods, with simulations confirming the technique's scalability.

ArXiv

Repetitive vs. Non-repetitive Lidar scanning pattern for roadside perception

This research introduces the "InfraLiDARs' Benchmark," a new simulated dataset, to evaluate the impact of different LiDAR scanning patterns on 3D object detection for roadside infrastructure. The study compares repetitive and non-repetitive scanning paradigms, finding that non-repetitive LiDAR provides comparable detection performance to a 128-line repetitive system. This makes it a cost-effective alternative despite its more limited perception range, and the benchmark is being publicly released to foster further research.

ArXiv

Show HN:emma019 Real-Time AI-Powered Texas Hold'em in Python and Flask

The provided text is an error message indicating a failure to retrieve a README file.

Repo

Show HN: Cossistant – open-source and open components support widget for React

Cossistant is an open-source, API-driven chat support widget for the React ecosystem, providing headless components and a full backend infrastructure. Its code-first philosophy and AI-friendly documentation make it a flexible foundation for building custom support solutions, such as those powered by LLMs. The project is licensed under AGPL-3.0 for non-commercial use, with a separate commercial license available.

Repo

MemMachine, an open-source memory layer for advanced AI agents

MemMachine is an open-source, universal memory layer for AI agents, providing persistent memory across multiple sessions, agents, and LLMs. It supports working, persistent, and personalized memory types to build evolving user profiles for context-aware interactions. The architecture separates episodic conversational memory, stored in a graph database, from long-term user profile facts stored in an SQL database, accessible via RESTful APIs and a Python SDK.

Repo

GraphLite: An Embeddable Graph Database with ISO Graph Query Language Support

GraphLite is a fast, embedded graph database, analogous to SQLite, that implements the new ISO GQL standard. Written in Rust, it operates as a single, memory-safe binary without a client-server setup, providing ACID transactions and cost-based query optimization. Its lightweight, file-based architecture makes it well-suited for applications requiring local graph capabilities, such as building knowledge graphs for RAG systems.

Repo

Show HN: Yoink – Copy any website's design system for your AI coding assistant

Yoink is an open-source browser extension that extracts a website's design system into structured YAML. This output can be fed directly into prompts for LLMs like Claude, instructing them to build UIs that accurately replicate the target site's visual style without manual description. The tool captures colors, typography, spacing, and components, with all processing done locally in the browser for privacy.

Repo