Earlier stories
The full Modelwire feed, ordered by publish time.
The week that Meta employees became training data
Meta's workplace monitoring practices have intensified alongside recent layoffs, with employees reporting surveillance tactics that repurpose their work as training data. The story raises questions about whether invasive data collection and workforce reduction will become standard in knowledge work.73
DeepSeek-V4: a million-token context that agents can actually use
DeepSeek released V4 with a million-token context window, marking a significant expansion in how much information agents can process in a single session. The capability addresses a practical bottleneck for long-horizon reasoning and multi-step workflows.89
Anthropic and NEC collaborate to build Japan’s largest AI engineering workforce
Anthropic and NEC are partnering to develop AI engineering talent in Japan, positioning the collaboration as a path to building the country's largest AI workforce. The deal signals Anthropic's expansion into Asia-Pacific and reflects growing demand for localized AI expertise outside the US.75
Claude is connecting directly to your personal apps like Spotify, Uber Eats, and TurboTax
Anthropic expanded Claude's app integrations beyond enterprise tools to personal services including Spotify, Uber Eats, TurboTax, Instacart, and AllTrails, letting users pull data directly into conversations. The move signals a shift toward consumer-facing AI utility and raises questions about data access and privacy at scale.
Extract PDF text in your browser with LiteParse for the web
Simon Willison ported LlamaIndex's LiteParse PDF extraction tool to run entirely in the browser, preserving its non-AI approach to text parsing and OCR fallback. The browser version maintains compatibility with LiteParse's core libraries while enabling client-side PDF processing without server dependencies.
First impressions of GPT-5.5 from Aaron Friel
Aaron Friel from OpenAI discusses GPT-5.5's capabilities for handling faster, more autonomous long-running tasks in a conversation with Romain Huet. The interview explores how OpenAI teams are leveraging the new model's performance improvements for extended workflows.
Meet Noscroll, an AI bot that does your doomscrolling for you
Noscroll, a new AI bot, automates social media consumption by reading feeds on behalf of users to combat doomscrolling habits. The tool represents a niche application of AI to behavioral wellness rather than a technical or capability breakthrough.
OpenAI unveils GPT-5.5, claims a "new class of intelligence" at double the API price
OpenAI released GPT-5.5, an agentic model capable of autonomously orchestrating multiple tools to solve complex tasks. The company doubled API pricing for the new model, positioning it as a step forward in autonomous reasoning capabilities.92
Anthropic’s Mythos breach was humiliating
Anthropic's Claude Mythos model, which the company claimed was too dangerous for public release due to advanced cybersecurity capabilities, has leaked to unauthorized users. The breach undermines Anthropic's controlled rollout strategy and raises questions about the gap between internal safety claims and actual containment.
At 'AI Coachella,' Stanford Students Line Up to Learn From Silicon Valley Royalty
Stanford's CS 153 course has become a campus phenomenon, drawing students eager to learn directly from prominent Silicon Valley figures. The class has generated significant social media buzz, though it's sparked mixed reactions among the student body.47
Introducing GPT-5.5
OpenAI unveiled GPT-5.5, positioning it as a new class of AI capable of handling complex multi-step tasks, tool use, and self-verification for agentic workflows. The model is now available in ChatGPT and Codex, signaling a shift toward AI systems that can autonomously complete real-world work.100
GPU Renters Are Playing a Silicon Lottery
Research from William & Mary, Jefferson Lab, and Silicon Data reveals that identical GPU models exhibit substantial performance variance when rented from cloud providers, turning procurement into an unpredictable gamble for AI teams. The silicon lottery effect, where manufacturing tolerances create chip-to-chip differences, compounds cost uncertainty for organizations scaling compute infrastructure. This finding reshapes how practitioners should evaluate cloud GPU pricing and benchmarking claims, suggesting that nominal specs alone cannot guarantee consistent training or inference economics.
Trump science advisor says Chinese actors are copying American AI at massive scale
The Trump administration claims evidence of systematic model distillation attacks by Chinese actors targeting US frontier AI systems at industrial scale, signaling a shift toward more aggressive AI supply-chain defense.73
OpenAI says its new GPT-5.5 model is more efficient and better at coding
OpenAI released GPT-5.5, claiming improvements in efficiency and code generation over last month's GPT-5.4. The rapid iteration signals intensifying competition in frontier model capability, though concrete benchmarks remain absent from the announcement.
Temporal Taskification in Streaming Continual Learning: A Source of Evaluation Instability
Researchers show that how continuous data streams are split into tasks during continual learning evaluation significantly alters benchmark results, introducing hidden instability independent of model choice. They propose metrics to diagnose taskification sensitivity and test the effect across four major CL algorithms.58
Evaluation of Automatic Speech Recognition Using Generative Large Language Models
Researchers show decoder-based LLMs can evaluate speech recognition quality far better than traditional metrics, achieving 92-94% agreement with human judges on the HATS dataset versus 63% for Word Error Rate. The finding suggests generative models offer a practical alternative to semantic embeddings for ASR evaluation.58
Fine-Tuning Regimes Define Distinct Continual Learning Problems
Researchers show that how models are fine-tuned during continual learning fundamentally changes the problem itself, not just the solution. By varying which parameters remain trainable across sequential tasks, the effective learning dynamics shift, suggesting current benchmarks may unfairly compare methods across incompatible regimes.52
The Sample Complexity of Multicalibration
Researchers prove that achieving multicalibration in batch learning requires Θ(ε⁻³) samples, establishing a fundamental separation from marginal calibration's lower complexity. The result combines information-theoretic lower bounds with a practical randomized algorithm via online-to-batch conversion.52