Models & Releases Research Products & Apps Business & Funding

Modelwire

A curated feed of what matters in AI. Independent, ad-supported, built in Denver, Colorado.

Read

Today
Models & Releases
Research
Business & Funding

About

About Modelwire
Methodology
Our sources
Editor's notes
Contact
Advertise

Legal

Privacy policy
Terms of use
DMCA & takedowns
Corrections

© 2026 Modelwire. All article links go to the original publishers.Summaries generated by Modelwire. We don’t republish full articles.

Earlier stories

The full Modelwire feed, ordered by publish time.

Illustration for: First token counts reveal Opus 4.7 costs significantly more than 4.6 despite Anthropic's flat pricing

Models & Releases Business & Funding

First token counts reveal Opus 4.7 costs significantly more than 4.6 despite Anthropic's flat pricing

Anthropic's Opus 4.7 maintains flat per-token pricing but costs substantially more in practice due to a new tokenizer that generates up to 47% more tokens per input. This hidden cost increase affects Claude Code users despite the company's stated pricing stability.

The Decoder·Apr 19

73

Illustration for: AI-generated influencers flood social media with pro-Trump content ahead of midterms

Policy & Regulation

AI-generated influencers flood social media with pro-Trump content ahead of midterms

Hundreds of AI-generated avatars are spreading pro-Trump political content across TikTok, Instagram, and YouTube, with some accounts reaching 35,000+ followers and millions of views ahead of the midterms. The campaign's origin—whether grassroots or coordinated—remains unclear, though Trump has already amplified some AI-generated posts.

The Decoder·Apr 19

73

Illustration for: Even the best AI models lose about half their performance when charts get complicated, new benchmark finds

Research Models & Releases

Even the best AI models lose about half their performance when charts get complicated, new benchmark finds

The RealChart2Code benchmark tested 14 leading AI models on complex, real-world chart visualizations and found that even top proprietary models saw performance drop by roughly 50% compared to simpler chart tasks, revealing a significant capability gap.

The Decoder·Apr 19

73

Illustration for: Always-on Ray-Ban Meta glasses powered by OpenClaw speed up everyday tasks in new study

Research Products & Apps

Always-on Ray-Ban Meta glasses powered by OpenClaw speed up everyday tasks in new study

Researchers deployed an OpenClaw AI agent on Ray-Ban Meta smart glasses to study how continuous visual perception affects user interaction with agentic systems. The study measured whether always-on AI assistance accelerates task completion in everyday scenarios.

The Decoder·Apr 19

73

Illustration for: Google launches generative UI standard for AI agents

Tools & Code Products & Apps

Google launches generative UI standard for AI agents

Google unveiled A2UI 0.9, a framework-agnostic standard enabling AI agents to dynamically generate UI components across web, mobile, and other platforms by leveraging existing app interfaces.

The Decoder·Apr 19

85

Illustration for: Revisiting Auxiliary Losses for Conditional Depth Routing: An Empirical Study

Revisiting Auxiliary Losses for Conditional Depth Routing: An Empirical Study

Researchers systematically compared auxiliary loss strategies for training conditional depth routing gates in language models, finding that interactions between predictive and explicit supervision losses significantly affect training stability on a 157.5M-parameter decoder.

arXiv cs.LG·Apr 19

52

Illustration for: LASER: Low-Rank Activation SVD for Efficient Recursion

Research Tools & Code

LASER: Low-Rank Activation SVD for Efficient Recursion

Researchers identify that recursive neural architectures concentrate computation along a low-dimensional activation manifold, enabling LASER, a dynamic compression method that exploits this structure for efficient inference without retraining weight-sharing models.

arXiv cs.LG·Apr 19

58

Illustration for: Bilinear Input Modulation for Mamba: Koopman Bilinear Forms for Memory Retention and Multiplicative Computation

Research Models & Releases

Bilinear Input Modulation for Mamba: Koopman Bilinear Forms for Memory Retention and Multiplicative Computation

Researchers propose bilinear input modulation for Mamba, a selective state space model, to overcome diagonal state transition limits and improve memory retention and multiplicative computation. Two variants trade off sequential accuracy (Coupled-BIM) against parallel efficiency (Coupled-GM), with experiments showing distinct performance on memory and nonlinear tasks.

arXiv cs.LG·Apr 19

52

Illustration for: PAC-Bayes Bounds for Gibbs Posteriors via Singular Learning Theory

PAC-Bayes Bounds for Gibbs Posteriors via Singular Learning Theory

Researchers derived new generalization bounds for Gibbs posteriors using singular learning theory, enabling tighter risk guarantees for overparameterized models without requiring explicit metric entropy control. The work bridges classical PAC-Bayes theory with data-dependent complexity, offering practical tools for understanding when and why posterior averaging improves generalization.

arXiv cs.LG·Apr 19

52

Illustration for: Continual Safety Alignment via Gradient-Based Sample Selection

Continual Safety Alignment via Gradient-Based Sample Selection

Researchers identify that high-gradient training samples degrade LLM safety alignment during fine-tuning, while moderate-gradient samples preserve safety behaviors. A gradient-based filtering method recovers alignment across multiple model families without sacrificing task performance.

arXiv cs.LG·Apr 19

62

Illustration for: Guardrails in Logit Space: Safety Token Regularization for LLM Alignment

Research Tools & Code

Guardrails in Logit Space: Safety Token Regularization for LLM Alignment

Researchers propose safety token regularization, a lightweight fine-tuning method that preserves LLM safety properties by constraining logits of rejection tokens during domain adaptation. The technique avoids expensive RL or preference optimization while integrating with parameter-efficient methods like LoRA, addressing a practical gap where aligned models degrade on benign new datasets.

arXiv cs.LG·Apr 19

58

Illustration for: Demystifying the unreasonable effectiveness of online alignment methods

Demystifying the unreasonable effectiveness of online alignment methods

Researchers resolve a long-standing gap between theory and practice in online alignment methods by reframing the performance metric. Greedy algorithms like online RLHF and DPO achieve constant regret under decision-centric evaluation, explaining why they outperform pessimistic O(log T) bounds in real deployments.

arXiv cs.LG·Apr 19

62

Illustration for: Do LLM-derived graph priors improve multi-agent coordination?

Do LLM-derived graph priors improve multi-agent coordination?

Researchers tested whether LLM-generated coordination priors can improve multi-agent reinforcement learning by inferring agent interaction patterns from natural language descriptions, potentially replacing brittle hand-coded or proximity-based coordination topologies.

arXiv cs.LG·Apr 19

54

Illustration for: SynthFix: Adaptive Neuro-Symbolic Code Vulnerability Repair

Research Tools & Code

SynthFix: Adaptive Neuro-Symbolic Code Vulnerability Repair

SynthFix combines neural networks with symbolic reasoning to fix code vulnerabilities, routing samples adaptively between supervised and reward-based fine-tuning. The hybrid approach yields 18% gains in code quality metrics and 32% in exact-match repairs on JavaScript and C benchmarks versus baseline LLM methods.

arXiv cs.LG·Apr 19

58

Decomposing the Depth Profile of Fine-Tuning

Researchers tested whether fine-tuning depth profiles reflect model properties or gradient dynamics by controlling weight change magnitudes across 240 runs spanning 15 models up to 6.9B parameters. The finding: representational shifts concentrate near output layers in standard training, but this pattern persists or vanishes depending on architecture and scale when per-layer gradient control is applied.

arXiv cs.LG·Apr 19

58

Illustration for: RosettaSearch: Multi-Objective Inference-Time Search for Protein Sequence Design

Research Models & Releases

RosettaSearch: Multi-Objective Inference-Time Search for Protein Sequence Design

Researchers combined LLMs with structure prediction to optimize protein sequences, achieving 2.5x higher success rates than single-pass models on 400 test cases. RosettaSearch uses inference-time search to refine designs that prior methods missed, showing 18–68% gains in structural fidelity.

arXiv cs.LG·Apr 19

62

Illustration for: Changes in the system prompt between Claude Opus 4.6 and 4.7

Models & Releases Research

Changes in the system prompt between Claude Opus 4.6 and 4.7

Anthropic released Claude Opus 4.7 on April 16, 2026, with updated system prompts compared to the February 4.6 version. The company continues its practice of publicly archiving system prompt changes, enabling transparency into how model instructions evolve across releases.

Simon Willison·Apr 18

72

Illustration for: Anthropic Has Received Investor Interest at $800 Billion Valuation

Business & Funding

Anthropic Has Received Investor Interest at $800 Billion Valuation

Anthropic has attracted investor interest at an $800 billion valuation, according to a profile of CFO Krishna Rao in The Information, though the company currently has no plans to raise capital.

The Information — AI·Apr 18

85

Illustration for: Lightweight Cybersickness Detection based on User-Specific Eye and Head Tracking Data in Virtual Reality

Lightweight Cybersickness Detection based on User-Specific Eye and Head Tracking Data in Virtual Reality

Researchers propose a lightweight ensemble learning model for detecting cybersickness in VR using eye and head tracking data, addressing reliability gaps in existing methods by tailoring detection to individual users rather than aggregating across populations.

arXiv cs.LG·Apr 18

52

Illustration for: Uncertainty Quantification in PINNs for Turbulent Flows: Bayesian Inference and Repulsive Ensembles

Research Tools & Code

Uncertainty Quantification in PINNs for Turbulent Flows: Bayesian Inference and Repulsive Ensembles

Researchers developed probabilistic extensions to physics-informed neural networks for quantifying uncertainty in turbulence modeling, combining Bayesian inference, Monte Carlo dropout, and repulsive ensembles to improve reliability on ill-posed inverse problems.

arXiv cs.LG·Apr 18

52

Illustration for: Tesla brings its robotaxi service to Dallas and Houston

Products & Apps

Tesla brings its robotaxi service to Dallas and Houston

Tesla is expanding its Robotaxi service to Dallas and Houston, deploying autonomous vehicles without human safety monitors. The rollout marks geographic expansion of the company's autonomous driving deployment.

TechCrunch — AI·Apr 18

69

Illustration for: FlowRefiner: Flow Matching-Based Iterative Refinement for 3D Turbulent Flow Simulation

FlowRefiner: Flow Matching-Based Iterative Refinement for 3D Turbulent Flow Simulation

FlowRefiner applies flow matching and deterministic ODE correction to improve autoregressive prediction of 3D turbulent flows, addressing error accumulation in neural PDE solvers. The method stabilizes multi-scale refinement through a decoupled noise schedule and achieves state-of-the-art results on large-scale turbulence benchmarks.

arXiv cs.LG·Apr 18

52

Illustration for: The RAM shortage could last years

Hardware & Infra

The RAM shortage could last years

DRAM suppliers will only meet 60% of global demand by end-2027, with SK Group warning shortages could persist until 2030. Samsung, SK Hynix, and Micron are ramping production but face a multi-year capacity crunch affecting AI infrastructure buildout.

The Verge — AI·Apr 18

76

Illustration for: Negative Momentum for Convex-Concave Optimization

Negative Momentum for Convex-Concave Optimization

Researchers tackle a long-standing gap in min-max optimization by proving global convergence for convex-concave problems using negative momentum, addressing whether acceleration is possible in this foundational setting that underpins adversarial training and game-theoretic ML.

arXiv cs.LG·Apr 18

52

Illustration for: SeekerGym: A Benchmark for Reliable Information Seeking

SeekerGym: A Benchmark for Reliable Information Seeking

Researchers introduced SeekerGym, a benchmark that tests whether AI agents retrieve complete information during research tasks and accurately report confidence in their findings. The work addresses a critical gap in agent evaluation: agents can return correct data while missing relevant context that skews user understanding.

arXiv cs.LG·Apr 18

58

Illustration for: Dylan Patel’s SemiAnalysis Projects Over $100 Million In 2026 Revenue

Business & Funding

Dylan Patel’s SemiAnalysis Projects Over $100 Million In 2026 Revenue

SemiAnalysis, Dylan Patel's semiconductor and AI research firm, projects over $100 million in 2026 revenue, up sharply from ~$20 million in 2025. The growth reflects rising demand for specialized AI infrastructure intelligence among enterprise clients.

The Information — AI·Apr 18

73

Illustration for: Local Inconsistency Resolution: The Interplay between Attention and Control in Probabilistic Models

Local Inconsistency Resolution: The Interplay between Attention and Control in Probabilistic Models

Researchers propose Local Inconsistency Resolution, a unifying framework that recovers EM, belief propagation, adversarial training, GANs, and GFlowNets as special cases. The work suggests a refined GFlowNet loss function that accelerates convergence, positioning LIR as a foundational lens for understanding disparate learning algorithms.

arXiv cs.LG·Apr 18

58

Illustration for: BOIL: Learning Environment Personalized Information

BOIL: Learning Environment Personalized Information

Researchers introduce BOIL, a multi-agent learning framework that combines PageRank and information maximization to help agents extract strategic insights from complex environments. The approach outperforms heuristics on coverage, patrolling, and reachability tasks by generating adaptive strategy distributions over long horizons.

arXiv cs.LG·Apr 18

48

Illustration for: Automated Classification of Plasma Regions at Mars Using Machine Learning

Automated Classification of Plasma Regions at Mars Using Machine Learning

Researchers developed a CNN classifier that automatically identifies three plasma regions around Mars using ion energy spectra from NASA's MAVEN spacecraft, outperforming a simpler MLP baseline. The work enables faster, more reliable analysis of solar wind-Mars interactions critical for understanding atmospheric escape.

arXiv cs.LG·Apr 18

42

Illustration for: The Topological Trouble With Transformers

The Topological Trouble With Transformers

Researchers identify a fundamental architectural constraint in transformer models: their feedforward design forces state representations deeper into the network with each sequential step, eventually exhausting model depth. Proposed workarounds like dynamic depth and latent thinking exist but carry significant computational overhead.

arXiv cs.LG·Apr 18

58

Older stories →