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.

Opinion & Analysis Research

Deepmind's Hassabis sees humanity "in the foothills of the singularity" while LeCun says current AI isn't intelligent

Three senior AI researchers have staked out divergent positions on whether current systems constitute genuine intelligence or approach AGI. Hassabis frames the field as entering a critical inflection point toward singularity, while LeCun argues today's models lack true reasoning capacity. Vinyals offers a calibration: systems now exceed what would have seemed like AGI in 2019, yet remain fundamentally limited in learning and discovery. This disagreement among DeepMind and Meta leadership signals unresolved questions about capability measurement and timeline expectations that will shape investment, regulation, and research priorities across the industry.

The Decoder·May 24

Research Policy & Regulation

Hackers are learning to exploit chatbot ‘personalities’

Security researchers are uncovering a new attack surface in conversational AI systems: exploiting the behavioral quirks and designed personalities of chatbots to bypass safety guardrails. Unlike early jailbreaks that relied on crude prompt injection, adversaries now target the tension between a model's helpfulness objective and its safety constraints, using personality traits as leverage points. This shift signals that as chatbot defenses mature, attackers are moving upstream to exploit the fundamental design trade-offs baked into instruction-tuning and RLHF processes. For AI teams, this underscores the fragility of behavioral alignment and the need for adversarial testing that goes beyond static prompt lists.

The Verge - AI·May 24

Products & Apps Research

Why you shouldn't leave model selection on default in Copilot, Gemini and other AI tools

Default model selection in mainstream AI assistants masks a critical reliability gap: identical inputs produce wildly different outputs depending on which underlying model processes them. Mathematician Adam Kucharski's experiment with Copilot revealed the tool fabricates country-specific stereotypes when fed unlabeled data, a failure that advanced reasoning models catch but only when users explicitly select them. This exposes a usability and trust problem at scale. As AI tools embed deeper into workflows, burying model choice behind defaults risks systematizing hallucination and bias without user awareness or recourse.

The Decoder·May 24

Research Models & Releases

Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models

Researchers demonstrate that sparse autoencoders can steer medical vision-language models at inference time to reduce hallucinations in radiology report generation without retraining. By applying targeted suppression and amplification of learned features across late-layer SAEs, the technique achieves 5-17% improvements in clinical accuracy across three VLM architectures on MIMIC-CXR benchmarks. This work signals a broader shift toward post-hoc steering as a practical alternative to fine-tuning for domain-critical applications, with implications for how practitioners can adapt pretrained models to high-stakes medical settings without computational overhead.

arXiv cs.CL·May 24

Research

NITP: Next Implicit Token Prediction for LLM Pre-training

Researchers propose Next Implicit Token Prediction, a training method that supplements standard next-token prediction with dense supervision in the model's representation space rather than just discrete output labels. By anchoring hidden states to shallow-layer embeddings as self-supervised targets, NITP aims to prevent representation collapse and anisotropy that can degrade generalization. The technique addresses a fundamental constraint in current LLM pre-training: one-hot supervision leaves latent geometry under-specified. If validated at scale, this could reshape how foundation models are initialized and regularized, particularly for efficiency-focused training regimes where representation quality directly impacts downstream performance.

arXiv cs.CL·May 24

Policy & Regulation Business & Funding

Anthropic may keep supplying Claude to the NSA despite being flagged as a supply chain risk by the Pentagon

Anthropic is positioned to maintain its NSA contract despite Pentagon designation as a supply chain risk, a tension rooted in hardware constraints rather than capability gaps. Intelligence agencies face acute shortages of Nvidia's latest Grace Blackwell processors, making Anthropic's Mythos model, which operates on older silicon, strategically valuable despite security concerns. The removal of the contentious 'any lawful use' clause signals negotiated compromise, but the deal underscores how geopolitical AI competition and domestic chip scarcity are reshaping government procurement logic independent of traditional risk frameworks.

The Decoder·May 24

$Illustration for: H$^{2}$MT: Semantic Hierarchy-Aware Hierarchical Memory Transformer$

Research Tools & Code

H$^{2}$MT: Semantic Hierarchy-Aware Hierarchical Memory Transformer

H2MT addresses a fundamental bottleneck in transformer inference: the cost of processing irrelevant context in long-input scenarios. By pre-computing a semantic hierarchy and routing queries through it at inference time, the approach reduces wasted computation on unrelated text while avoiding the external storage and indexing overhead that plagues retrieval-augmented generation systems. This matters because it directly tackles prefill latency and memory consumption, two metrics that constrain practical deployment of long-context LLMs. The coarse-to-fine pruning strategy represents a structural shift from flat token processing, potentially reshaping how production systems balance context window size against inference speed.

arXiv cs.CL·May 24

Research Tools & Code

Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed

A multi-institutional research team deployed an AI coding agent to autonomously search for novel scaling algorithms, yielding a control method that reduces compute requirements by 70 percent relative to standard self-consistency approaches while preserving accuracy. The discovery cost $40 and completed in under three hours, signaling a shift toward machine-driven algorithm design as a path to efficiency gains. This outcome matters because it demonstrates that AI systems can uncover optimization strategies outside human intuition, potentially reshaping how teams approach inference-time scaling and resource allocation in production systems.

The Decoder·May 24

Research

Quantifying the Impact of Translation Errors on Multilingual LLM Evaluation

A new study exposes a critical blind spot in how the AI industry validates multilingual LLMs: machine-translated benchmarks contain systematic errors that go largely undetected, yet measurably degrade model performance scores. By comparing LLM-based error detection against human expert annotations and quantifying how translation flaws (rather than source problems) drive accuracy drops, the research reveals that current multilingual evaluation metrics may be fundamentally unreliable. This matters because vendors and researchers routinely cite multilingual benchmarks to claim parity across languages, but those claims rest on corrupted data. The findings suggest the field needs either human-vetted translations or far more rigorous automated quality control before drawing conclusions about true cross-lingual capability.

arXiv cs.CL·May 24

Research Models & Releases

When Reasoning Hurts: Source-Aware Evaluation of Frontier LLMs for Clinical SOAP Note Generation

A controlled evaluation of reasoning-enabled frontier LLMs reveals a counterintuitive finding: disabling chain-of-thought reasoning in GPT-5.4 produces superior clinical documentation compared to reasoning-augmented variants across three healthcare benchmarks. The study challenges the assumption that reasoning capabilities automatically improve structured, domain-specific outputs, suggesting that for clinical SOAP note generation, simpler decoding paths may outperform complex inference chains. This has implications for how enterprises deploy reasoning models in regulated settings where output quality and consistency matter more than benchmark performance.

arXiv cs.CL·May 24

Research Models & Releases

Towards a Universal Causal Reasoner

Researchers have built UniCo, a synthetic data framework that systematically generates causal reasoning tasks across Pearl's Causal Ladder, translating symbolic examples into code and natural language to reflect real-world scenarios where causality isn't explicitly labeled. The work addresses a critical gap in LLM training: while benchmarks for causal reasoning exist, few datasets enable models to learn generalizable causal inference at scale. By filtering for reasoning shortcuts and grounding answers in formal causal inference, the team produced 66.6K high-quality instances that improved performance on smaller models like Qwen3-4B and Olmo-3-7B-Instruct. This signals growing momentum in making causal reasoning a trainable, composable capability rather than an emergent property, with implications for reliability in domains where causal claims matter.

arXiv cs.CL·May 24

Research Policy & Regulation

Translators as Invisible Teachers of AI: Copyright, Translation Memory, and the Political Economy of Linguistic Data

A new paper traces how translator labor has become foundational infrastructure for modern AI systems, from statistical machine translation through multilingual LLMs. Translation memories and parallel corpora represent supervised training data of extraordinary value, yet translators have historically been compensated as contract deliverable providers rather than recognized as data contributors. The work examines how copyright frameworks have obscured translators' role in building the linguistic foundations that enabled the Transformer era, raising questions about data provenance, labor attribution, and the political economy of AI training at scale.

arXiv cs.CL·May 24

Research

Spiking the training data to correct for test set contamination

Researchers propose a novel approach to correcting inflated test scores caused by data leakage, a persistent problem in model evaluation. Rather than only detecting contamination, the method intentionally spikes training data with known test examples to calibrate memorization predictors, enabling statistical adjustment of benchmark results. The work introduces Hubble models as a simulation framework with paired contaminated and clean variants to validate correction estimators. This addresses a critical gap in ML rigor: while test set contamination is widely acknowledged, principled correction methods remain rare. The technique could reshape how labs validate model performance and report benchmark claims, particularly as model scale makes accidental data leakage increasingly likely.

arXiv cs.CL·May 24

Research Tools & Code

RouteScan: A Non-Intrusive Approach to Auditing MoE LLMs Safety via Expert Routing Telemetry

RouteScan introduces a privacy-preserving safety audit method for Mixture-of-Experts LLMs by analyzing GPU-level routing telemetry rather than user inputs or model outputs. This addresses a critical tension in production deployments: safety verification without exposing sensitive data. The technique exploits the sparse activation patterns inherent to MoE architectures, creating a new class of non-intrusive monitoring that could reshape how enterprises validate model behavior in regulated environments while maintaining user confidentiality.

arXiv cs.CL·May 24

Research Models & Releases

DUEL: Adversarial Self-Play for Multimodal Reasoning

DUEL introduces a self-play training framework that sidesteps the annotation bottleneck plaguing vision-language model improvement. By pitting two identical VLM instances against each other, one generating hard negatives while the other validates claims, the approach bootstraps supervision signals without human labeling. This addresses a critical scaling constraint in RL-based model refinement, potentially unlocking cheaper pathways to stronger multimodal reasoning without the drift problems that plague unsupervised alternatives. The technique matters for labs seeking to push VLM capabilities beyond what labeled data budgets allow.

arXiv cs.CL·May 24

Research Tools & Code

Beyond the Target: From Imitation to Collaboration in Speculative Decoding

Collaborative Speculative Decoding challenges a core assumption in LLM acceleration: that larger models always make better token-level decisions. Researchers found that smaller draft models, despite lower overall capability, sometimes outperform target models on individual predictions, leading to correct final outputs. This work reframes inference optimization from hierarchical verification toward genuine model collaboration, potentially unlocking efficiency gains in production systems where current SPD methods leave performance on the table by reflexively deferring to the larger model.

arXiv cs.CL·May 24

Research Tools & Code

Who judges the judges? Governance from metrics: a runtime framework for continuous LLM compliance monitoring

Researchers propose a shift from static compliance audits to continuous runtime monitoring of LLM behavior, arguing that binary, point-in-time assessments misalign with EU AI Act requirements for ongoing oversight. The paper introduces govllm, an open-source framework that routes model selection based on accumulated compliance scores rather than latency or cost, treating regulatory conformity as a measurable, observable property of production systems. This approach addresses a critical gap in deployed AI governance: detecting behavioral drift and emergent failures after models enter production, not just at certification.

arXiv cs.CL·May 23

Research Policy & Regulation

Fundamental Limitation in Explaining AI

A new theoretical result establishes a fundamental trade-off in AI explainability: systems cannot simultaneously achieve environmental complexity, performance quality, explanation fidelity, and interpretability. This quadrilemma directly challenges the regulatory assumption that faithful explanations of large-scale models are always achievable, reshaping how policymakers should approach AI governance and transparency mandates. The finding suggests governance frameworks may need to accept bounded explainability rather than demand complete interpretability.

arXiv cs.CL·May 23

Research

The Tokenizer Tax Across 25 European Languages: Domain Invariance, Cross-Lingual Few-Shot Effects, and the Ukrainian Penalty

A systematic study quantifies the computational penalty non-English languages face in foundation models through tokenizer inefficiency. Across 25 European languages, token-per-word ratios vary 2.5x, with Ukrainian and other underrepresented languages paying 15-18% higher inference costs than English peers. The research reveals that this 'tokenizer tax' correlates directly with pre-training data scarcity rather than linguistic structure, and persists consistently across domains and model architectures. For practitioners deploying multilingual systems, this work exposes a hidden scaling cost that compounds at inference time and suggests that equitable model development requires deliberate tokenizer design, not just balanced training data.

arXiv cs.CL·May 23

Business & Funding Models & Releases

Deepseek makes its 75 percent discount permanent, pricing output tokens at least 34x below GPT-5.5

DeepSeek has made its 75 percent discount on V4-Pro permanent, according to The Decoder’s May 23 report. The reported rates are $0.435 per million uncached input tokens and $0.87 per million output tokens. The output rate is about 34.5 times below GPT-5.5’s listed $30 per million at that time. This unit-price gap matters for token-intensive applications, but token consumption and output quality determine whether a completed task actually costs less.

The Decoder·May 23

Business & Funding Products & Apps

One of the world's top law schools draws a hard line against AI in legal education

UC Berkeley Law's decision to restrict AI across graded coursework signals a strategic pivot in how elite institutions approach AI literacy. By blocking generative tools from outlining, drafting, and proofreading while preserving research access, the school is betting that foundational legal reasoning must precede tool fluency. This move reflects growing tension between AI adoption and skill atrophy in knowledge work, and will likely influence hiring expectations at major firms and reshape how law schools calibrate their competitive positioning around AI competency.

The Decoder·May 23

Models & Releases Hardware & Infra

Alibaba's latest AI model ran autonomously for 35 hours to optimize code for its own custom chip

Alibaba's Qwen team unveiled Qwen3.7-Max, a proprietary model engineered for extended autonomous reasoning tasks that ran uninterrupted for 35 hours optimizing code for Alibaba's custom silicon. The model matches Claude Opus 4.6 on standard benchmarks while outperforming Chinese competitors DeepSeek V4 Pro and Kimi K2.6, signaling Alibaba's push into long-horizon agent capabilities and vertical integration of AI infrastructure. The robotics demonstration underscores ambitions beyond language tasks, positioning Qwen as a contender in the emerging market for models that can sustain complex multi-step problem solving without human intervention.

The Decoder·May 23

Business & Funding Policy & Regulation

Google CEO Pichai now calls links a "part" of search, redefining the web's role in its own product

Google's semantic shift from positioning links as foundational to search toward framing them as merely a 'part' of the product signals a strategic pivot in how the company deploys generative AI. By keeping users within Google's ecosystem rather than routing traffic outward, the company is consolidating editorial control over source selection and information synthesis. This reflects a broader tension in the AI industry: as LLMs become the primary interface for discovery, the power to curate training data and surface sources becomes a form of editorial gatekeeping that bypasses traditional web distribution.

The Decoder·May 23

Research Policy & Regulation

Anthropic warns Claude Mythos Preview finds bugs faster than developers can patch them

Anthropic's Claude Mythos Preview has uncovered over 10,000 critical vulnerabilities across system-critical software through Project Glasswing, a 50-partner initiative. The discovery rate now outpaces patch deployment, exposing a structural vulnerability window that Anthropic itself acknowledges lacks adequate safeguards against misuse. This signals a pivotal inflection point in AI-assisted security research: frontier models are becoming more effective at threat discovery than human teams can remediate, forcing the industry to confront whether current governance frameworks can manage the asymmetry between detection and defense.

The Decoder·May 23

Models & Releases Research

Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

Nvidia's Nemotron-Labs division has released a diffusion-based language model architecture targeting dramatic inference speedups, positioning diffusion as a viable alternative to autoregressive decoding for text generation. This represents a meaningful shift in the efficiency frontier for LLM inference, with implications for cost-per-token economics and real-time applications. If the claimed speed gains hold across diverse workloads, the approach could reshape deployment strategies for resource-constrained environments and challenge the current autoregressive paradigm that dominates production systems.

Hugging Face·May 23

Policy & Regulation Research

AI is being used to resurrect the voices of dead pilots

Researchers weaponized audio reconstruction techniques on archived cockpit spectrograms to synthesize pilot voices, exposing a critical vulnerability in how safety agencies protect sensitive investigation data. The breach forced the NTSB to restrict public access to its accident docket, marking a watershed moment where generative audio capabilities now threaten the integrity of aviation forensics and regulatory transparency. This incident signals that voice synthesis has matured beyond entertainment use cases into territory where it can compromise institutional trust and investigative processes, forcing safety bodies to rethink data governance in an era of commodity AI tools.

TechCrunch - AI·May 22

Products & Apps

Workspace agents in ChatGPT: Admin and builder controls

OpenAI has moved workspace agents from beta to general availability across ChatGPT Business, Enterprise, and Edu tiers, introducing granular governance layers that shift the control surface for agentic systems. The release emphasizes administrative guardrails and approval workflows, signaling that enterprise adoption of autonomous agents now hinges on auditability and permission models rather than raw capability. This reflects a maturing market where teams expect agents to operate within defined boundaries, making governance infrastructure as critical as the agents themselves.

OpenAI (YouTube)·May 22

Older stories →