Daily Digest

Amazon's new framework calibrates Chinchilla scaling by targeting a 1.0 MLP-to-attention parameter ratio. This optimized allocation reduces KV cache constraints, yielding the Surefire model family which matches Llama-3.2 accuracy while driving a 12-47% throughput increase on H200/vLLM deployments.

Major open-weight releases include DeepSeek-V4-Flash (284B total, 13B active) optimized for extreme inference efficiency, and Gemma 4 sizes up to 31B. Gemma 4 is now licensed under Apache 2.0, significantly clearing up enterprise deployment compliance.

Architectural optimization is shifting toward aggressive KV sharing; Gemma 4 E2B computes unique KV projections in only the first 15 layers and reuses them for the subsequent 20 layers, conserving approximately 2.7GB of VRAM at a 128K context window.

Zyphra successfully converted an autoregressive MoE to a discrete diffusion model via the TiDAR recipe. The resulting block diffusion eliminates memory-bandwidth bounds inherent in autoregressive decoding, yielding up to a 7.7x inference speedup using logit-mixing samplers.

Google's 3-bit KV cache compression maps geometry to polar coordinates (PolarQuant) and uses Quantized Johnson-Lindenstrauss to remove residual biases. It accelerates throughput by 8x on H100s for 32K+ token contexts while drastically cutting the memory footprint by up to 5.4x.

Amazon Q Knowledge Bases introduced explicit fail-closed ACLs for S3 indexing. A critical production gotcha: modifying a global `.json` ACL file triggers a full reindex of the prefix; document-level metadata files are necessary to restrict reindexing overhead during frequent permission changes.

Research identifies an 'Engineering Attractor Field' within embedding spaces where high-density technical tokens force a non-linear phase transition into unintended languages. Once an LLM's state enters this attractor basin, simple translation prompts fail to correct the output register.

Repowise introduces a graph-based indexing pipeline for Python using NetworkX for PageRank and community detection to rank code node relevancy. It inherently handles dead-code detection thresholds and semantic tracking of architectural decisions.

BerriAI released a self-hosted platform utilizing the `kubernetes-sigs/agent-sandbox` CRD for secure, isolated AI execution. Integrated directly with the LiteLLM gateway, it manages session persistence and env-var secret injection across container restarts.

RLMs isolate internal tool-calling traces by delegating to black-boxed subagents, returning only finalized answers to the primary agent. This architecture directly circumvents the 'context rot' and state management failures prevalent in traditional ReAct and CodeAct loops.

Implements core Model Context Protocol (MCP) concepts to mitigate tool-selection entropy. It relies on a hybrid LLM/heuristic router to restrict capability exposure dynamically and executes code in a localized Python sandbox with disabled network access.

Demonstrates an autonomous capability-building loop where an agent utilizes a `review-past-performance` cron job. By parsing 24-hour log data for incorrect tool calls and context misses, the agent dynamically updates organizational `agents.md` files without manual human intervention.

FDA Commissioner Marty Makary's resignation comes amid intense internal operational conflicts, having frequently overruled staff scientists via unvetted press releases. Concurrently, industry hesitation has grown around using the new 'Commissioner's National Priority Voucher' due to fears of politicized oversight.

Rznomics secured RMAT designation for RZ-001 targeting hepatocellular carcinoma. By leveraging trans-splicing ribozymes to edit RNA transcripts rather than permanently altering genomic DNA, this mechanism is gaining regulatory traction as a safer modality for complex diseases.

Visby Medical's remote STI platform pairs an at-home diagnostic kit boasting 98% PCR-level accuracy with 30-minute telehealth orchestration. The platform targets expanding care gaps in 17 states facing severe rural physician shortages.

A self-supervised neural-physics approach removes the need for prior optical calibration in super-resolution imaging. It simultaneously reconstructs 3D molecular structures and optical aberrations directly from raw microscopy data.

Buck Institute researchers discovered that the APOE2 longevity variant actively stabilizes neuronal genomes against senescence. In iPSC-derived neurons, it demonstrably suppressed DNA strand breaks and senescence markers (p16, CRYAB) under acute radiotoxic stress.

Immune-remodeling mRNAs delivered via lipid nanoparticles target NIK and IRF8 to convert immature myeloid cells into functional cDC1 dendritic cells within the tumor microenvironment. This pathway achieved total colorectal tumor regression in ~70% of murine models while avoiding systemic cytokine toxicity.

Analysis of a randomized crossover feeding study suggests the 13% decline in sperm motility associated with UPFs is a byproduct of spontaneous hyperphagia and subsequent weight gain (1.3-1.4 kg over 3 weeks), rather than processing chemicals directly acting as endocrine disruptors.

The introduction of the Clinical Integrity Standard targets low-fidelity mobile DEXA operators. Enforcing fixed-site thermal/power stability and ISCD QA is particularly critical to establish valid baselines for detecting sarcopenia in populations prescribed GLP-1 agonists.

By identifying 31 distinct regional Vietnamese genetic signatures, researchers enhance the granularity of ancestry-informed pharmacogenomics. This aids precise tracking of CYP2C19 drug metabolism variance and GJB2-linked nonsyndromic hearing loss.

Evaluations against the DELEGATE-52 benchmark reveal that SOTA models suffer a 19-34% decay in artifact semantic fidelity over 20 recursive task delegations. In contrast, workflows strictly utilizing generated Python code showed less than 1% degradation.

Anthropic's Claude Mythos dramatically outperformed GPT-5.5 on CMU's ExploitBench, achieving top-tier arbitrary code execution on 21 of 41 vulnerabilities in the V8 JavaScript engine. However, reaching this reasoning plateau incurred an immense $36,428 API cost for Mythos, compared to $3,075 for GPT-5.5.

Tsinghua University's WorldReasonBench exposed significant reasoning gaps in SOTA video diffusion models like Sora 2 and Seedance 2.0. The models failed routinely at causal physics and logical reasoning, revealing a reliance on explicitly spelled-out prompt steps rather than embedded world representations.

Following a $5.3B valuation, Runway is pivoting from creative tooling to foundational world models trained explicitly on observational sensory data. Management aims to leverage these systems to generate digital twins of biological states to accelerate longevity research and drug discovery.

Integrating Plaid directly into the platform, ChatGPT Pro now allows users to tether live brokerage and bank accounts. This unlocks localized temporal reasoning on high-sensitivity data for real-time spending analysis and tax impact projection.

To scale autonomous intelligence effectively, Deloitte warns that enterprise infrastructure must fundamentally shift from stale, batch-cycled 'Reporting-grade' architectures to real-time, access-controlled 'Decision-grade' data pipelines.

Massive capital inflows dominate physical AI and data infrastructure, led by defense contractor Anduril securing $5B. VoltaGrid pulled in $775M for mobile natural gas data center power, and AI robotics spinout Mind Robotics secured $400M.