# An American Privacy Emergency

- **The U.S. Secretary of Commerce's directive (DAO 216-26) undermines over fifty years of progress in data confidentiality by banning modern privacy techniques like differential privacy, which are essential for protecting individual data in large datasets.**

# Markets are competitive if and only if P = NP

- **Competitive market outcomes hinge on computational intractability**, as firms can only sustain collusion if P = NP, allowing efficient detection of deviations in cooperative agreements.

# 60% Fable cost cut by converting code to images and having the model OCR it

- **pxpipe** significantly reduces input tokens by converting dense text (like code and JSON) into images, achieving a **~59–70% lower end-to-end bill** on token-dense requests, as it compresses bulky context into compact PNGs.

# Claude-real-video － any LLM can watch a video

- **`claude-real-video` enables LLMs to _actually watch_ videos by extracting meaningful frames based on scene changes, rather than fixed intervals, ensuring a more accurate representation of fast-paced content.**

# NVIDIA Unlocks AI Compute at Scale, Inviting Partners to Power the AI Infrastructure Buildout

- **NVIDIA's new partnership model** enables AI clouds to deploy large-scale, multi-tenant AI factories, enhancing access to accelerated computing through a **revenue-sharing and credit-support structure** that aligns economic interests.

# Program-as-Weights: A Programming Paradigm for Fuzzy Functions

- **Fuzzy-function programming** enables the compilation of natural-language specifications into compact, locally-executable neural artifacts, enhancing efficiency in programming tasks that resist rule-based solutions.

# 14× faster embeddings: how we rebuilt the ONNX path in Manticore

- The new **ONNX Runtime backend** in Manticore Search 27.1.5 achieves **~14× faster embeddings** compared to the previous SentenceTransformers/Candle path, significantly enhancing throughput from **5–11 docs/sec** to **70–230 docs/sec** on the same hardware.

# Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

- **Memora** introduces a **scalable memory system** that enhances AI agent productivity by effectively balancing **abstraction** and **specificity**, allowing for rich memory content while optimizing retrieval processes.

# Q&A with Micron's VP and GM of Memory

- **Memory's role in AI infrastructure is evolving**, with demand for high-performance **DRAM** and **HBM** surging as companies compete for advanced memory subsystems to support AI applications across various sectors, including cloud and automotive.

# Anatomy of Persistent Memory's 3 Layers: Comparing ContextNest, Mem0 and Zep

- **Building a robust AI agent requires a three-tier persistent memory architecture**: Zep, Mem0, and ContextNest work together to ensure agents access verified, up-to-date information while maintaining user personalization and session continuity.

# Contrastive Decoding Diffing (CDD): recovering verbatim finetuning data from logits alone, no weight access needed

- **Contrastive Decoding Diffing (CDD)** enables the recovery of verbatim finetuning data from LLMs using only logit access, achieving a **4+/5 recovery score** on 19/20 organism x model pairs, outperforming previous methods that required full weight access.

# What does "Safe AI" look like?

- **Fine-tuning resistance** in open-weight LLMs poses a significant challenge, as rapid emergence of "uncensored" variants suggests that current safety measures may be inadequate against determined users.

# Hierarchos: Preliminary Findings From a 232M Recurrent Memory-Augmented Assistant Model

- **Hierarchos**, a **232M-parameter** recurrent memory-augmented model, demonstrates that a **hybrid non-Transformer architecture** can achieve training stability and coherence, addressing critical bugs in state management and numerical stability.

# ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning

- **RECONTEXT** enhances long-context reasoning in large language models (LLMs) by utilizing a **training-free inference method** that constructs a query-conditioned evidence pool, improving evidence utilization without the need for external memory or context pruning.