All news

Conservative meta-agent search adapts agent architectures under non-stationary workloads.

Finds numerical bugs in standard softmax and cross-entropy implementations, like softmax collapse, that change training dynamics, and proposes fixes.

Postgres ON DELETE CASCADE can delete huge row counts in one transaction, causing long locks and surprises, so explicit batched deletes are safer.

Routes coding agent skills at runtime between macro and micro granularity with recovery.

Kronecker Delta Rule treats key-value pairs as joint addresses in linear attention to reduce interference at long context.

Coding agents follow stale conventions in instruction files over the actual code. A one-sentence prompt fix telling them to prefer existing code improves decisions a lot.

SCOPE compiles natural-language policies into neurosymbolic predicate programs to enforce LLM agent rules.

Walks through using SIMD in Go for a simple float32 add loop and the speedup over scalar code.

Mix-or-Split schedules LLM inference across devices, edge and cloud using a fine-grained latency model.

Quantizes n-gram embedding tables in open LLMs to low bits.

BeNice shares KV cache across loop iterations in looped transformers to close the inference memory gap.

Compresses per-layer embeddings with cross-layer latent vector quantization.

Harvard study of 718 firms finds AI coding tools raise code output but downstream review and deploy bottlenecks erase productivity gains.

Traces the Wait token in reasoning models back to pretraining data rather than RL.

A third PHP2Shell bug was found but will not get an emergency fix.

Google Data Agent Kit is GA with MCP tools and 37 open-source skills, splitting guardrails between skills and IAM.

Full-stack framework for long-horizon embodied robot agents.

Evidence position bias differs across attention and other sequence mixers in long-context QA.

PGX runs thousands of isolated throwaway Postgres-compatible databases in one Rust process, each cloned from a template in milliseconds, aimed at agent sandboxes and per-PR tests. Beta, dev use only.

A 2001 USENIX paper on DERIVE, which reverse-engineers bit-level instruction encodings by feeding permutations of instructions to an assembler, covering MIPS, SPARC, Alpha, PowerPC and most of ARM and x86.

Two dormant PHP vulnerabilities nearly broke large parts of the web this week.

LaLa-World scales JEPA latent world models with natural language actions.

PACE encodes linear-attention recurrent states in four bits.

PHP security release notes and plans to strengthen PHP security.

Factor-aware predictive learning makes MLLM latents useful for cinematic video generation.

RIDAR detects and intervenes on research integrity violations by autonomous research agents.

Twin-world counterfactual credit assignment for agentic RL.

PHP emergency fixes for two critical bugs; unpatched sites told to go offline.

Recasts sequence compression as test-time regression for efficient models.

Naive RAG beats complex agentic memory on long-term conversation tasks.