Earlier stories

Every published Modelwire story, newest first. The homepage leads with the weekly overview and this week’s hottest; this is the full collection.

Research Tools & Code

From Prompt to Pointer Engineering: Deepmind tries to reinvent the mouse cursor for the AI era

DeepMind is exploring a fundamental shift in how AI systems interact with digital environments by elevating the mouse cursor from a peripheral UI element to a core variable in context engineering. Rather than relying solely on text prompts, this approach treats pointer position and movement as structured input signals that help models understand spatial relationships and user intent within visual interfaces. The technique could reshape how AI agents navigate and manipulate software, potentially unlocking more reliable automation of complex desktop workflows and reducing the brittleness of current vision-language models when tasked with precise screen interactions.

The Decoder·May 13

Products & Apps Research

Medicare’s new payment model is built for AI, and most of the tech world has no idea

Medicare's ACCESS payment model represents a structural shift in how healthcare systems can reimburse AI-driven interventions. By creating the first federal mechanism to pay for autonomous AI agents handling patient monitoring, medication adherence, and care coordination between clinical visits, the model removes a critical barrier to deployment of healthcare AI at scale. This signals that regulatory and reimbursement infrastructure is beginning to catch up to AI capability, potentially unlocking billions in healthcare AI adoption where payment ambiguity previously stalled implementation.

TechCrunch - AI·May 13

Hardware & Infra Policy & Regulation

xAI Adds 19 New Gas Turbines Despite Ongoing Lawsuit

xAI is scaling power infrastructure at its Colossus 2 training facility by deploying 19 additional gas turbines, signaling aggressive capacity expansion even as the company faces regulatory scrutiny over air quality impacts. The move underscores the acute energy demands of frontier LLM training and the infrastructure bottlenecks constraining AI labs. For the industry, it reflects how compute-constrained players are willing to absorb regulatory friction and operational complexity to secure reliable power, reshaping the geography of AI development away from grid-dependent regions.

WIRED - AI·May 13

Policy & Regulation Tools & Code

Our response to the TanStack npm supply chain attack

OpenAI disclosed a supply chain compromise affecting the TanStack npm package, revealing how adversaries can infiltrate AI infrastructure through open-source dependencies. The incident forced OpenAI to revoke and reissue signing certificates, exposing a critical vulnerability in how AI labs secure their software pipelines. macOS users must patch by mid-June to prevent potential system compromise. This underscores a growing risk vector for frontier labs: as AI systems scale and integrate deeper into production stacks, the attack surface expands beyond model security into the mundane but lethal realm of dependency management. The response signals how seriously OpenAI now treats supply chain resilience as a core infrastructure concern.

OpenAI·May 13

Policy & Regulation Business & Funding

Sam Altman was winning on the stand, but it might not be enough

Sam Altman testified in a legal proceeding involving allegations of charity theft, marking a pivotal moment in a case that has shadowed OpenAI's leadership. While courtroom observers noted his effective testimony, the outcome remains uncertain, suggesting that even persuasive defense may not resolve the underlying reputational and governance questions facing the AI industry's most visible executive. The case underscores how personal liability and corporate governance disputes can create uncertainty around key figures steering frontier AI development.

The Verge - AI·May 12

Opinion & Analysis Business & Funding

Quoting Mo Bitar

Mo Bitar's viral commentary exposes a pattern of performative AI adoption in enterprise settings, where executives greenlight vaguely-defined initiatives like 'Ralph Loops' without technical grounding. The clip, amplified by Simon Willison, highlights a broader tension in the AI landscape: hype cycles and budget allocation often outpace genuine capability maturity, creating incentives for employees to exploit information asymmetry rather than deliver measurable outcomes. This reflects real friction between boardroom AI enthusiasm and engineering reality that shapes product roadmaps and hiring cycles.

Simon Willison·May 12

Products & Apps Hardware & Infra

The Unitree GD01 Is a Giant Mecha Robot You Can Actually Buy

Unitree's pivot from affordable quadruped robots to consumer-grade mecha represents a significant shift in how robotics companies are commercializing embodied AI systems. The move signals growing confidence in autonomous control stacks and real-time decision-making at scale, while testing whether general consumers will adopt large-form-factor robots beyond industrial or research contexts. This challenges assumptions about the robotics market's near-term ceiling and suggests Chinese hardware makers are willing to absorb risk on speculative form factors that Western competitors have avoided.

WIRED - AI·May 12

Opinion & Analysis Business & Funding

Quoting Mitchell Hashimoto

Mitchell Hashimoto argues that most technical decision makers prioritize risk avoidance over innovation, driving adoption of trendy AI solutions backed by analyst consensus rather than genuine technical merit. This dynamic explains why enterprises adopt 'AI strategy' and 'context management' tools not because they solve real problems, but because major consulting firms and research houses have validated them as defensible purchases. The insight reveals a structural misalignment between hype-driven enterprise AI spending and actual technical value, shaping which vendors and narratives dominate the market regardless of underlying capability.

Simon Willison·May 12

Products & Apps Tools & Code

Computer use in Codex

OpenAI's Codex now supports computer use, enabling AI agents to interact with local applications through clicking, typing, and background execution without requiring full system control. The capability introduces granular app-level permissions and leverages screenshots combined with accessibility data to navigate interfaces, marking a shift toward more autonomous yet bounded agent behavior. This development signals OpenAI's strategy to position Codex as a practical workflow automation tool rather than a monolithic system takeover, addressing enterprise concerns about agent safety and control while expanding the practical scope of LLM-powered automation.

OpenAI (YouTube)·May 12

Products & Apps

Android gets AI agents that book trips, fill forms, and clean up your texts

Google is rolling out agentic capabilities within Gemini Intelligence on Android, enabling the system to orchestrate multi-step workflows like travel booking and form completion without user intervention between steps. This represents a meaningful shift from single-turn assistance toward autonomous task execution on consumer devices, positioning Google to compete directly with OpenAI's emerging agent frameworks and raising questions about on-device vs. cloud-based reasoning tradeoffs. The move signals that major platforms now view agent deployment as a core differentiator rather than a research frontier.

The Decoder·May 12

Products & Apps Business & Funding

Anthropic expands legal AI offerings with new Claude Cowork plugins

Anthropic is deepening its foothold in legal services by rolling out twelve specialized Claude plugins targeting contract analysis, employment disputes, and litigation workflows. The move reflects a strategic pivot toward vertical-specific tooling rather than horizontal capability expansion, positioning Claude as infrastructure for knowledge-work automation in a sector where adoption already outpaces other professions. Integration with established legal platforms like Thomson Reuters' CoCounsel and Harvey signals Anthropic's willingness to embed rather than compete head-to-head, a notable shift in how frontier labs are monetizing enterprise LLM access.

The Decoder·May 12

Policy & Regulation Research

Google says it stopped a mass cyberattack after AI was used to discover a zero-day exploit

Google's Threat Intelligence Group has documented the first confirmed weaponization of AI to discover a zero-day vulnerability, marking a critical inflection point in adversarial AI capability. The company disrupted the planned mass attack, but the incident signals that state-backed actors from China, North Korea, and Russia have moved beyond using AI for code obfuscation into offensive vulnerability research. This development reshapes the threat model for enterprise security teams and accelerates the timeline for AI-assisted defense systems, making autonomous threat detection no longer optional but foundational infrastructure.

The Decoder·May 12

Research Tools & Code

Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation

Pion introduces a fundamentally different approach to LLM optimization by replacing additive weight updates with orthogonal transformations that preserve singular values during training. This geometric reformulation of the optimizer landscape matters because it challenges the dominance of Adam-family methods and offers potential stability gains for both pretraining and finetuning workflows. The technique's ability to modulate weight matrix geometry while fixing spectral norm could unlock efficiency gains or convergence properties that practitioners currently lack, making it a credible alternative worth benchmarking against incumbent optimizers at scale.

arXiv cs.LG·May 12

Research Models & Releases

Elastic Attention Cores for Scalable Vision Transformers

Researchers challenge a foundational assumption in Vision Transformers by demonstrating that patch-to-patch attention is unnecessary for learning effective visual representations. VECA introduces a core-periphery attention mechanism that reduces computational complexity from quadratic to linear time, potentially unlocking ViT deployment in high-resolution imaging tasks where memory constraints currently prohibit scaling. This architectural shift matters for practitioners building vision systems at scale, particularly in medical imaging, satellite analysis, and video understanding where resolution demands have outpaced transformer feasibility.

arXiv cs.LG·May 12

Research Models & Releases

Learning, Fast and Slow: Towards LLMs That Adapt Continually

Researchers propose a dual-timescale learning architecture that decouples rapid in-context adaptation from slower parameter updates, addressing a fundamental tension in LLM training. The framework treats optimized prompts as fast weights and model parameters as slow weights, mirroring human cognition across different learning regimes. This approach targets catastrophic forgetting and loss of plasticity, two critical failure modes that limit both continual learning and task specialization. The work reframes a false binary between fixed-parameter efficiency and parameter-update performance, potentially reshaping how practitioners balance stability against adaptation in production systems.

arXiv cs.LG·May 12

Research

Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training

Researchers propose a unified framework for language model post-training that treats GRPO and on-policy distillation as complementary reward-density regimes rather than competing methods. The core insight reallocates scarce labeled data upstream to exploration-heavy training phases using sparse rewards, reserving dense token-level supervision for downstream model compression. This challenges conventional practice of applying all verification data directly to deployment models, offering a principled allocation strategy for teams constrained by labeled data availability. The work reframes how practitioners should think about the training pipeline when verification budget is the bottleneck.

arXiv cs.LG·May 12

Research Models & Releases

MEME: Multi-entity & Evolving Memory Evaluation

A new benchmark reveals a critical failure mode in LLM-based agents operating across multiple sessions. MEME exposes that memory systems collapse when reasoning about dependencies between stored facts, achieving near-zero accuracy on cascade and absence tasks despite strong static retrieval. The finding matters because production agents increasingly need persistent memory across conversations, yet current architectures and optimization strategies fail to bridge this gap, suggesting fundamental architectural rethinking is needed before deployment in real-world multi-turn environments.

arXiv cs.CL·May 12

Research

Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts

Researchers have uncovered a fundamental geometric relationship between routers and expert networks in sparse mixture-of-experts models, revealing that routing decisions and expert weight updates follow coupled gradient trajectories. This mechanistic insight addresses a core scaling challenge in SMoE architectures: routing collapse and loss of expert specialization. The finding bridges theory and empirical observation in a 1B-parameter model, offering a foundation for designing more stable training procedures and potentially more efficient expert utilization in large language models relying on conditional computation.

arXiv cs.CL·May 12

Research Tools & Code

KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference

KV-Fold introduces a training-free method to extend LLM context windows by treating the key-value cache as a functional accumulator across sequence chunks. Rather than retraining or modifying model weights, the technique reuses internal attention state across segments, enabling longer inference without architectural changes. This addresses a persistent bottleneck in production LLM deployment: the computational and memory cost of processing very long documents. For practitioners, the approach offers immediate applicability to existing models, potentially unlocking longer-context capabilities without the expense of fine-tuning or model replacement.

arXiv cs.CL·May 12

Research Models & Releases

Solve the Loop: Attractor Models for Language and Reasoning

Attractor Models address a fundamental constraint in recurrent architectures by decoupling training memory from effective depth through implicit differentiation. Rather than unrolling fixed recurrence steps, the approach treats iterative refinement as fixed-point solving, allowing adaptive convergence and constant gradient overhead. This shifts the tradeoff landscape for models that benefit from multi-step reasoning, showing gains in both large-scale pretraining and small-model reasoning tasks. The technique could reshape how practitioners balance compute efficiency against representational depth in production systems.

arXiv cs.CL·May 12

Business & Funding

Alphabet's Isomorphic Labs raises $2.1 billion to scale AI drug discovery toward clinical trials

Isomorphic Labs, Alphabet's AI drug discovery unit, has secured $2.1 billion in Series B funding to accelerate its IsoDDE platform toward human clinical trials. The capital injection signals deepening confidence in AI-driven molecular design as a viable path to pharmaceutical innovation, moving beyond computational validation into real-world therapeutic validation. This represents a critical inflection point where AI drug discovery transitions from research curiosity to capital-intensive development, with implications for how biotech and pharma incumbents compete against AI-native competitors in the discovery pipeline.

The Decoder·May 12

Research Models & Releases

Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs

Researchers propose a fundamental shift in how language models process information by enabling parallel computation streams rather than sequential message exchange. Current AI agents remain bottlenecked to single-stream architectures inherited from ChatGPT-era designs, preventing simultaneous reading, writing, thinking, and acting. Multi-stream LLMs would allow agents to generate outputs while consuming new inputs and reason across multiple concurrent tasks, directly addressing a core architectural limitation that has persisted despite rapid capability gains. This work targets the infrastructure layer of autonomous agents, particularly in coding and computer-use domains where latency and decision parallelism matter.

arXiv cs.CL·May 12

Tools & Code Models & Releases

llm 0.32a2

OpenAI's shift to a new /v1/responses endpoint for reasoning-capable models marks a significant infrastructure change that enables interleaved reasoning across tool calls, particularly for GPT-5 class systems. Simon Willison's LLM tool now supports this endpoint, allowing developers to observe the model's reasoning process in real time rather than only seeing final outputs. This architectural move signals OpenAI's commitment to transparency in reasoning workflows and reflects the broader industry push toward interpretable, multi-step inference patterns that go beyond traditional chat completion semantics.

Simon Willison·May 12

Research Tools & Code

TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection

TextSeal advances the technical arms race around LLM provenance by embedding robust, localized watermarks that survive heavy mixing with human text and integrate seamlessly with production optimizations like speculative decoding. The scheme addresses a critical gap: prior watermarking approaches either degrade output quality or fail under realistic contamination scenarios. By maintaining detection confidence across multilingual benchmarks without inference overhead, TextSeal shifts the practical calculus for model providers weighing authenticity verification against user experience, making watermarking a viable default rather than a niche compliance layer.

arXiv cs.CL·May 12

Research

The Algorithmic Caricature: Auditing LLM-Generated Political Discourse Across Crisis Events

Researchers are moving beyond shallow detection signals like perplexity to audit whether LLM-generated political text mimics real social behavior during crises. Using a dataset spanning nine major events from COVID-19 to the 2024 election, the work compares synthetic discourse patterns against observed online populations to expose how generative systems may distort political discourse at scale. This shift toward behavioral auditing matters because traditional AI-text detection weakens as models improve, forcing the field to adopt social science methods to catch synthetic manipulation when it matters most: during high-stakes moments when misinformation spreads fastest.

arXiv cs.CL·May 12

Business & Funding Policy & Regulation

Sam Altman says Elon Musk’s mind games were damaging OpenAI

OpenAI's leadership conflict with Elon Musk, now surfacing in litigation testimony, reveals internal governance tensions that shaped the company's early culture. Altman's account of Musk's pressure to rank and cut researchers highlights how founding disputes over organizational structure and talent strategy can fracture AI labs during critical scaling phases. This case illustrates broader questions about founder alignment and decision-making authority in high-stakes AI development, with implications for how competing visions of research culture play out in legal and operational contexts.

The Verge - AI·May 12

Hardware & Infra Business & Funding

Report: Google and SpaceX in talks to put data centers into orbit

Google and SpaceX are exploring orbital data centers as a potential long-term solution for AI compute infrastructure, positioning space as a frontier for scaling training and inference workloads. While current costs dwarf terrestrial alternatives, the strategic rationale centers on latency reduction, power availability, and escape from earthbound capacity constraints as AI demand accelerates. This signals how major infrastructure players are stress-testing unconventional solutions to the compute bottleneck, though viability remains speculative and years away from commercial deployment.

TechCrunch - AI·May 12

Older stories →