Earlier stories
The full Modelwire feed, ordered by publish time.
Sam Altman’s Orb Company Promoted a Bruno Mars Partnership That Doesn't Exist
Sam Altman's Orb company falsely promoted a Bruno Mars partnership that never existed, according to the artist's representatives. The incident raises questions about how AI-adjacent ventures vet claims and manage public communications.58
AI failure could trigger the next financial crisis, warns Elizabeth Warren
Sen. Elizabeth Warren drew parallels between current AI investment patterns and pre-2008 financial bubble conditions, warning that widespread AI system failures could destabilize markets. Warren, who championed post-recession financial regulation, flagged the sector's rapid scaling and risk concentration as potential triggers for systemic economic damage.
OpenAI now lets teams make custom bots that can do work on their own
OpenAI rolled out autonomous agents to Business, Enterprise, Edu, and Teachers plan subscribers, enabling teams to deploy custom bots that handle tasks like gathering product feedback and generating Slack reports without manual intervention.
CI/CD Breaks at AI Speed: Tangle, Graphite Stacks, Pro-Model PR Review , Mikhail Parakhin, Shopify
Shopify's CTO Mikhail Parakhin discusses how the company rebuilt its development pipeline after a December 2025 model inflection, deploying internal tools like Tangle for ML workflows and SimGym for customer simulation to handle code generation outpacing traditional CI/CD capacity.85
Ex-OpenAI researcher Jerry Tworek launches Core Automation to build the most automated AI lab in the world
Jerry Tworek, a former OpenAI researcher, founded Core Automation to build an AI lab focused on automating research itself and developing novel learning methods beyond current architectural constraints. The venture signals renewed interest in bootstrapping frontier AI capabilities outside established labs.68
Google Cloud launches two new AI chips to compete with Nvidia
Google unveiled faster, cheaper TPUs to challenge Nvidia's dominance in AI infrastructure, though the company continues offering Nvidia chips within its cloud platform. The move signals intensifying competition in specialized silicon for training and inference workloads.
OpenAI launches workspace agents that turn ChatGPT from a chatbot into a team automation platform
OpenAI is rolling out workspace agents within ChatGPT that automate team workflows and run continuously without user intervention, marking a shift from custom GPTs toward persistent automation infrastructure. The agents leverage Codex and represent a meaningful expansion of ChatGPT's role from conversational tool to enterprise workflow backbone.85
Workspace agents in ChatGPT: Lead outreach agent
OpenAI rolled out workspace agents in ChatGPT, Codex-powered tools that automate lead qualification, personalized outreach drafting, and CRM updates for teams. The feature targets sales and operations workflows with agentic automation built into the platform.
Workspace agents in ChatGPT: Product feedback routing agent
OpenAI launched workspace agents for ChatGPT, autonomous tools that aggregate feedback from Slack and support channels to surface product priorities and generate weekly action items. The feature targets teams seeking to streamline signal-to-insight workflows at scale.
5 AI Models Tried to Scam Me. Some of Them Were Scary Good
A WIRED investigation tested multiple AI models' ability to execute social engineering and scam tactics, finding some demonstrated concerning proficiency at manipulation. The findings highlight a gap between AI safety focus on raw capability and the underexplored risks of models weaponizing persuasion and deception.69
SpeechParaling-Bench: A Comprehensive Benchmark for Paralinguistic-Aware Speech Generation
Researchers released SpeechParaling-Bench, a benchmark expanding paralinguistic feature coverage from under 50 to over 100 attributes for evaluating Large Audio-Language Models. The dataset includes 1,000+ English-Chinese parallel queries across three task difficulty levels, with a pairwise comparison pipeline to reduce subjective assessment bias.52
Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL
Researchers propose Parallel-SFT, a supervised fine-tuning strategy that enables Llama-3.1 to transfer coding skills across programming languages without degradation. Standard RL training on code tasks in one language actually hurts performance on others; the new method uses generalized SFT initialization to unlock zero-shot cross-language transfer.58
AVISE: Framework for Evaluating the Security of AI Systems
Researchers released AVISE, an open-source framework for systematically identifying security vulnerabilities in AI systems. The tool includes an automated test suite with 25 cases designed to uncover jailbreak flaws in language models, extending prior adversarial attack methods.58
FedSIR: Spectral Client Identification and Relabeling for Federated Learning with Noisy Labels
FedSIR addresses a real pain point in federated learning: noisy labels across distributed clients that tank model quality. The method uses spectral analysis of feature representations to identify which clients have clean data, then leverages those as references to correct noise—cutting communication overhead versus existing approaches.52
Closing the Domain Gap in Biomedical Imaging by In-Context Control Samples
Researchers propose CS-ARM-BN, a meta-learning method that uses control samples to adapt deep learning models across experimental batches in biomedical imaging. The technique addresses batch effects, a longstanding barrier to deploying ML systems in real-world lab settings where systematic technical variations degrade performance.58
Stream-CQSA: Avoiding Out-of-Memory in Attention Computation via Flexible Workload Scheduling
Researchers introduce Stream-CQSA, a memory-adaptive scheduling framework that decomposes attention computation into independent subproblems to prevent out-of-memory failures in long-context LLMs. The technique removes the assumption that full query, key, and value tensors must fit in device memory, enabling attention on hardware with arbitrary constraints.58
Convergent Evolution: How Different Language Models Learn Similar Number Representations
Researchers discovered that multiple model architectures—Transformers, LSTMs, Linear RNNs, and classical embeddings—independently converge on periodic number representations in the Fourier domain, but only some achieve geometric separability for modular arithmetic. The finding reveals that data, architecture, optimizer, and tokenizer choices determine whether models learn truly usable numerical features.62
ParetoSlider: Diffusion Models Post-Training for Continuous Reward Control
Researchers propose ParetoSlider, a multi-objective reinforcement learning framework that trains diffusion models to explore trade-offs between competing goals at inference time rather than committing to fixed preferences during training. The approach enables users to dynamically balance objectives like image quality versus prompt adherence without retraining.58
OMIBench: Benchmarking Olympiad-Level Multi-Image Reasoning in Large Vision-Language Model
Researchers introduced OMIBench, a benchmark for evaluating vision-language models on Olympiad-level reasoning across multiple images rather than single frames. The dataset spans biology, chemistry, math, and physics problems with manual rationales, revealing significant performance gaps even in leading models like Gemini-3-Pro.58
Google turns Chrome into an AI coworker for the workplace
Google is rolling out Gemini-powered automation to Chrome for enterprise, enabling workers to offload repetitive tasks like research and data entry through browser-native AI agents. The move signals a shift toward embedding LLM capabilities directly into workplace infrastructure rather than standalone tools.
LEXIS: LatEnt ProXimal Interaction Signatures for 3D HOI from an Image
Researchers introduce LEXIS, a method for reconstructing 3D human-object interactions from single images by modeling continuous proximity fields across body and object surfaces rather than relying on sparse contact cues. The approach captures interaction structure through action and object geometry to resolve the inherent ambiguity of inferring 3D relationships from 2D input.52
Google's new AI tools put film scouting in Street View and promise to cut weeks of satellite analysis to minutes
Google unveiled three computer vision models at Cloud Next targeting creative and infrastructure workflows: AI image placement in Street View for film scouts, satellite imagery analysis that compresses weeks of planning into minutes for city planners, and object detection for bridges and power lines. The tools represent applied AI moving from research into domain-specific production use.73
Gauge-Equivariant Graph Neural Networks for Lattice Gauge Theories
Researchers developed gauge-equivariant graph neural networks that embed local symmetries directly into message passing, enabling ML models to learn under site-dependent gauge symmetries relevant to quantum physics and strongly correlated systems. The approach treats gauge-covariant transport as a native operation, allowing nonlocal correlations to emerge from local computations.52
Can "AI" Be a Doctor? A Study of Empathy, Readability, and Alignment in Clinical LLMs
Researchers benchmarked clinical LLMs on bedside communication, finding that general-purpose models like GPT-5 and Claude produce text 40% more complex than physician-authored notes and amplify negative sentiment. Empathy-focused prompting substantially reduced both issues, suggesting alignment gaps in healthcare deployment.62
Working Memory Constraints Scaffold Learning in Transformers under Data Scarcity
Researchers added working-memory-inspired constraints to Transformers, finding that fixed-width attention windows boost grammatical accuracy on small datasets (10M–100M words) while better matching human reading patterns. The approach suggests cognitive bottlenecks may act as useful inductive biases when training data is limited.58
Physics-Conditioned Synthesis of Internal Ice-Layer Thickness for Incomplete Layer Traces
Researchers developed a neural network that reconstructs missing ice-layer thickness from incomplete radar traces by conditioning on physical climate model data. The approach combines geometric learning with physics constraints to solve a key problem in glaciology where sensor noise and signal loss create gaps in stratigraphic records.42