The Archive

RouteREPA adapts representation alignment for vision transformers via routed learning, resolving conflicts between contrastive and discriminative objectives.

Simon Willison documents Python 3.15 dependency resolution workaround: deleting uv.lock files to fix wheel compatibility issues.

Hippocampal-inspired architecture enables continual learning without replay by indexing input geometry changes across tasks.

Study traces how language models integrate temporal anchors to resolve context-dependent factual recall with multiple valid objects.

OpenAI doubles ChatGPT Pro usage limits and consolidates Work features into Chat, merging allowances across ChatGPT and Codex.

Deepgram acquires Daily (raised $62M+), developer of open-source voice AI framework Pipecat, consolidating enterprise voice capabilities.

OpenAI consolidates ChatGPT and Codex API usage limits into unified system.

PSM-Bench evaluates LLM reasoning on perovskite photovoltaics; OpenPerov framework learns from 4,532 papers to guide materials design.

Log-Driven Semantic Tuning uses LLMs to automate MIP solver parameter tuning by interpreting performance logs.

Team Friday's 19M-parameter policy achieves 0.947 score on MetaFine fine-grained robot manipulation challenge.

Worldwide Express automated its less-than-truckload freight billing-exception process with a multi-agent solution built on Amazon Bedrock. Six months after launch, the company cut active exceptions by 75%, improved time-to-cash by four weeks, and unlocked $10 million in free cash flow.

Seven frontier LLMs evaluated on molecular property prediction; most memorized ESOL but demonstrate in-context learning on clean benchmarks with self-reported uncertainty.

Controlled membership inference study on OLMo 2 isolates true training exposure signal from distribution shift using paired documents.

ML method for solving underdamped stochastic optimal transport with nonlinear Landauer boundary conditions, benchmarked on bit-erasure problem.

Physics-informed Gaussian process method for probabilistic PDE solving using global residual information.

Hugging Face malware resurges with GPU hijacking, crypto mining, backdoors, and NetSupport RAT targeting AI coding agents via fake Solana tools.

LLM-based agents autonomously conduct quantum-spin experiments with peer review and theory validation to prevent unsupported conclusions.

Ramanujan Challenge evaluates AI models and symbolic solvers on research-level number theory problems with numerically verifiable formulas.

The Anthropic S-1 seems to have provided the financial justification for the company to spend as much as it possibly can on compute. If you are the kind of person who wonders if Anthropic is doing its level best to build a machine god while convincing Wall Street it's in the business of selling IT contracts, Anthropic has just offered a handy solution: Doing a bunch of things that look like you're trying to build the machine god is the only way to avoid going bankrupt. Forget about the machine god for a minute. At least in my read of the S-1, it seems like Anthropic is pinning the future of i...

Probabilistic meshless methods extend Gaussian process PDE solvers to compact manifolds with convergence analysis.

Hybrid episodic-relational memory with consolidation achieves 90.2% LoCoMo accuracy for long-horizon dialogue agents.

Quantum-inspired CNN generates pseudo-labels for land-cover classification from SAR and multispectral satellite data.

LLM-guided SMT solver for quantified array logic with verifiable focal term extraction.

Formulaic uses LLM agents to co-optimize hardware-software AI infrastructure design across boundaries, verified by mathematical solvers.

Autoregressive transformer for antenna design optimization; domain-specific CAD/physics application without general ML infrastructure relevance.

Computer-use agents experience 17% click staleness from page load delays; 1-second network sync resolves most failures.

Skeleton Coding demonstrates program structure constrains behavioral specifications to guide code-generating agents toward intended solutions.

Skill-Uni: unified RL framework jointly optimizing skill retrieval and utilization for LLM task-solving.

In-context learning framework enables PDE solvers to generalize across operators without retraining, advancing multi-fidelity model closure.