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

Latent Phase-Shift Rollback: Inference-Time Error Correction via Residual Stream Monitoring and KV-Cache Steering

Researchers propose Latent Phase-Shift Rollback, an inference-time technique that detects and corrects reasoning errors in LLM generation by monitoring residual streams and steering the KV-cache without retraining. The method lifts an 8B model's MATH-500 performance from 28.8% to 44.0%, substantially outpacing prompted self-correction baselines.

arXiv cs.CL·Apr 20

Research Tools & Code

GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling

Researchers propose GSQ, a scalar quantization method that closes the accuracy gap between simple techniques like GPTQ and complex vector-quantization approaches at ultra-low bit-widths (2-3 bits per parameter). The work suggests efficient LLM deployment needn't sacrifice accuracy for implementation simplicity.

arXiv cs.CL·Apr 20

Business & Funding Policy & Regulation

Deezer says 44% of new music uploads are AI-generated, most streams are fraudulent

Deezer reports that 44% of newly uploaded tracks are AI-generated, though they represent a tiny share of actual streams and face widespread demonetization for fraud. The finding underscores how generative audio is flooding music platforms while listeners remain uninterested in low-quality synthetic content.

Ars Technica — AI·Apr 20

Models & Releases Tools & Code

Open-weight Kimi K2.6 takes on GPT-5.4 and Claude Opus 4.6 with agent swarms

Moonshot AI released Kimi K2.6 as an open-weight model competitive with GPT-5.4 and Claude Opus 4.6 on coding tasks, with the ability to orchestrate up to 300 parallel agents. The open release challenges closed-model dominance in frontier capabilities.

The Decoder·Apr 20

Research

Different Paths to Harmful Compliance: Behavioral Side Effects and Mechanistic Divergence Across LLM Jailbreaks

Researchers compared three methods for jailbreaking open-weight LLMs and found they produce unsafe models with distinct internal failure modes. RLVR-jailbroken models notably retain the ability to recognize harmful requests and describe safe responses while still complying with them, revealing a mechanistic divergence across attack vectors.

arXiv cs.CL·Apr 20

Products & Apps Tools & Code

Fortnite developers can make AI characters now — just don’t try to date them

Epic Games rolled out a conversational AI tool for Fortnite creators to build interactive NPCs without hand-coding dialogue trees, expanding on last year's AI-voiced Darth Vader character. The feature lets players talk naturally with AI-generated characters on custom islands, marking a shift toward generative dialogue in game development.

The Verge — AI·Apr 20

Research Policy & Regulation

Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety

A new benchmark reveals that frontier models' safety guardrails collapse dramatically when harmful prompts are rewritten in literary or obfuscated styles. Attack success rates jumped from 3.84% to 55.75% across 31 models when researchers applied humanities-inspired transformations, exposing a critical gap in stylistic robustness.

arXiv cs.CL·Apr 20

Business & Funding Models & Releases

🔬 Training Transformers to solve 95% failure rate of Cancer Trials , Ron Alfa & Daniel Bear, Noetik

Noetik's AI platform for patient-tumor matching just landed a $50M licensing deal with GSK, betting that better treatment stratification can flip the 95% cancer trial failure rate without discovering new drugs. The deal signals biotech's shift from AI-as-discovery to AI-as-platform, a structural change in how big pharma deploys ML.

Latent Space·Apr 20

Business & Funding Hardware & Infra

CEO and CFO suddenly depart AI nuclear power upstart Fermi

Fermi, the nuclear power startup co-founded by former Energy Secretary Rick Perry and positioned to power AI datacenters, lost its CEO and CFO amid operational challenges at its Texas campus. The dual departure signals trouble for a company betting on nuclear as the infrastructure backbone for AI scaling.

TechCrunch — AI·Apr 20

Research Models & Releases

StepPO: Step-Aligned Policy Optimization for Agentic Reinforcement Learning

Researchers propose StepPO, a reinforcement learning method designed to train LLMs for multi-turn agentic tasks like tool use and decision-making. The approach addresses challenges unique to agent training: sparse rewards, long contexts, and variable-length interactions that differ from single-turn alignment methods like RLHF.

arXiv cs.CL·Apr 20

Products & Apps Business & Funding

Adobe fights AI disruption of its own business model with new enterprise agent platform

Adobe is launching an enterprise agent platform to counter AI-native startups threatening its core business, while simultaneously searching for a new CEO. The move signals how established software vendors are scrambling to defend market share against AI disruption.

The Decoder·Apr 20

Business & Funding Research

Physical AI Edges Closer to Real-World Deployments

Capgemini's latest research shows enterprises are shifting physical AI systems from pilot programs into live operations at scale. The move signals that robotics and embodied AI are maturing beyond controlled lab settings into production environments where they face real operational constraints.

AI Business·Apr 20

Products & Apps Business & Funding

Chinese Vendor Claims First Large-Scale Embodied AI Deployment

Agibot claims to have moved embodied AI from experimental pilots into real industrial deployment at scale, signaling a potential inflection point for robotics in manufacturing and logistics workflows.

AI Business·Apr 20

Hardware & Infra Business & Funding

Google plans nearly two million new AI chips as it turns to Marvell for custom designs

Google is partnering with Marvell Technology to design two custom AI chips for its data centers, planning to deploy nearly two million units. The move signals Google's shift toward vertical integration of silicon to reduce reliance on third-party vendors and control costs at scale.

The Decoder·Apr 20

Products & Apps Business & Funding

Salesforce bets on "Agent Albert" to prove AI won't kill enterprise software

Salesforce is launching Agent Albert, an AI product designed to counter Wall Street concerns that generative AI could render traditional enterprise software irrelevant. CEO Marc Benioff is also introducing a proprietary metric to demonstrate the technology's business value.

The Decoder·Apr 20

Business & Funding Policy & Regulation

UK Launches $675 Million Fund for AI Startups

The UK government committed $675 million to back AI startups through a new capital program, joining a wave of state-backed funding initiatives globally. The move signals sovereign competition for AI talent and infrastructure outside the US venture ecosystem.

AI Business·Apr 20

Research Models & Releases

Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4

Import AI 454 covers three substantive topics: automating alignment research to scale safety work, a safety evaluation of a Chinese LLM, and HiFloat4, a new numerical format for model training. The lead question on market pricing of AGI signals broader economic implications.

Import AI (Jack Clark)·Apr 20

Policy & Regulation

US, California Use Purchasing Power to Set AI Rules

US federal and state governments are leveraging procurement rules to enforce AI governance where legislation hasn't materialized, with California and federal agencies using purchasing power as a de facto regulatory lever on AI vendors.

AI Business·Apr 20

Hardware & Infra Business & Funding

TSMC Earnings, New N3 Fabs, The Nvidia Ramp

TSMC's latest earnings reveal skepticism within leadership about AI-driven demand, despite plans for new N3 fabrication capacity and Nvidia's ongoing ramp. The gap between public AI enthusiasm and internal caution signals potential overcapacity risk in chip supply.

Stratechery·Apr 20

Research

Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation

Researchers found that scaling up multi-agent LLM systems for brainstorming backfires: stronger aligned models, hierarchical group dynamics, and larger teams all suppress diversity rather than expand it, revealing a fundamental tension between individual quality and collective exploration.

arXiv cs.CL·Apr 20

Research

Neural Garbage Collection: Learning to Forget while Learning to Reason

Researchers propose Neural Garbage Collection, a technique that trains language models to selectively discard KV cache entries during chain-of-thought reasoning rather than relying on hand-crafted pruning rules. The end-to-end learned approach could unlock longer reasoning chains by reducing memory bottlenecks that currently constrain scaling.

arXiv cs.LG·Apr 20

Policy & Regulation Business & Funding

The NSA is using Anthropic's most powerful AI model Mythos

The NSA has deployed Anthropic's Mythos Preview model for intelligence operations, marking a significant government adoption of frontier AI capabilities by a major US surveillance agency.

The Decoder·Apr 20

Business & Funding Policy & Regulation

Chinese tech workers are starting to train their AI doubles–and pushing back

Chinese tech workers are being ordered to train AI agents designed to automate their own roles, sparking internal resistance among early adopters. A GitHub project called Colleague Skill enables companies to extract worker skills and personality traits into replicable AI systems, raising questions about job displacement and worker agency in AI-driven labor markets.

MIT Technology Review — AI·Apr 20

Research

From Fallback to Frontline: When Can LLMs be Superior Annotators of Human Perspectives?

Researchers show LLMs can outperform human annotators at predicting aggregate subgroup opinions on subjective tasks, flipping the typical view of models as annotation fallbacks. The advantage stems from structural properties like low variance and reduced bias coupling rather than domain knowledge, with conditions for superiority common in real-world scenarios.

arXiv cs.CL·Apr 20

Research

The Umwelt Representation Hypothesis: Rethinking Universality

Researchers challenge the assumption that all capable AI systems converge on universal representations of reality, proposing instead that alignment between ANNs and biological brains stems from shared ecological constraints rather than a single global optimum. The Umwelt Representation Hypothesis reframes how we should interpret similarities across neural systems.

arXiv cs.LG·Apr 20

Research

How Much Cache Does Reasoning Need? Depth-Cache Tradeoffs in KV-Compressed Transformers

Researchers establish theoretical bounds on how much key-value cache compression Transformers can tolerate during multi-step reasoning before performance collapses. The work formalizes a depth-cache tradeoff, suggesting aggressive KV compression requires deeper models to maintain reasoning capability.

arXiv cs.LG·Apr 20

Research

Heterogeneity in Formal Linguistic Competence of Language Models: Is Data the Real Bottleneck?

Researchers found that GPT-2 Small models trained on web data struggle with specific grammatical constructions, but injecting just 1% synthetic data targeting those phenomena recovered performance across 8 of 9 failing linguistic benchmarks, suggesting data scarcity rather than architectural limits drive formal linguistic gaps.

arXiv cs.CL·Apr 20

Older stories →