Claims Browser

Most of modern RL for LLMs is a systems problem balancing off-policy data, training-inference mismatch, and throughput

"Most of modern RL is a systems problem balancing a few problems — how off-policy the data is, training-inference mismatch, and throughput."

Different loss functions used to train LLMs produce very distinct flavors of misalignment: imitative learning causes 'seven deadly sins' misalignment, human approval causes sycophancy, automatic verifiers cause literal genie behavior, and LLM judges cause trickster misalignment

"for every loss function that we use to train LLMs, we get a very distinct flavor of LLM misalignment."

Sincere disagreements about AI policy usually boil down to disagreements about the expected pace of progress and what future AIs will be able to do

"sincere disagreements about AI policy usually boil down to disagreements about the expected pace of progress, and what we expect future AIs will be able to do."

Deep networks can discover abstractions that shallow models miss because language and images are built from hierarchical parts, and depth lets networks recover coarse-grained variables and escape the curse of dimensionality

"Why can deep networks discover abstractions that shallow models miss? Statistical physicist Matthieu Wyart joins Tim Scarfe to argue that the answer lies in the hidden hierarchy of data. Language and images are built from parts within parts; depth lets a network recover those coarse-grained variables and escape the curse of dimensionality."

AI watermarking is most likely implemented by biasing random token selection to create a detectable pattern in token IDs

"the most likely implementation is that they bias the random token selection such that there's a detectable pattern in the token IDs"

Process-level evaluation is crucial for interpreting the true potential of LLMs and guiding development of next-generation mathematical agents

"This demonstrates that process-level evaluation is crucial for interpreting the true potential of LLMs and guiding the development of next-generation mathematical agents."

Scaling generated distillation data should be paired with trait-aware curation and evaluation, even when data appears off-task or benign

"Our results suggest that scaling generated distillation data should be paired with trait-aware curation and evaluation, even when the data appears off-task or benign."

Multi-agent orchestration patterns and parallel execution are viable approaches for agentic coding systems

Secure-by-design methodology integrating threat modeling and MLSecOps practices can fortify AI systems against adversarial attacks

Combining reinforcement learning algorithms with human feedback signals is a powerful approach to AI alignment and human-centered machine learning

AI pair programming can be scaled while keeping code maintainable through persistent AI memory and enforced coding standards

Design patterns can handle instruction drift, improve coordination, and build fault-tolerant AI agent systems

DeepSeek has unique multimodal features comparable to leading models

AI agents are distinct from simpler AI coding assistants in capability and autonomy

Hidden risks from churn, bias, and hallucinations in AI systems can be prevented through proper workflows

Teams need to be reorganized around emerging AI-native roles for orchestration, review, and governance

ChatGPT Work on mobile provides sufficient functionality to replace a work laptop for professional tasks

Keeping powerful models to a chosen few is not a good strategy