Daily Digest

Perplexity introduced a "Search as Code" architecture that allows reasoning models to write custom Python workflows to execute complex searches via an Agentic Search SDK. In a CVE research benchmark, programmatic filtering and deduplication before context entry allowed SaC to use 85% fewer tokens than standard retrieval pipelines.

Researchers introduced Harness-1, a 20B retrieval-specific agent utilizing stateful cognitive offloading to separate search bookkeeping from the semantic policy. Trained via LoRA SFT and CISPO RL, it utilizes eight specialized tools to manage a curated evidence graph, achieving a 0.730 average curated recall across benchmarks.

Engineers caution that exact identifiers like document sections or version numbers often fail in semantic vector retrieval. The recommended fix for production systems is deploying Hybrid Search (Vector + BM25) merged via Reciprocal Rank Fusion (RRF), easily prototyped in pgvector before scaling.

A technical breakdown for indexing complex PDFs details a split-path architecture where visual data is extracted via OCR and parsed by Vision models (like Claude 3.5 Sonnet) to generate textual summaries. These summaries and their coordinate metadata are stored alongside standard text chunks in a dual knowledge base with hybrid query routing.

Spin-RAG (v0.1.0a1) is a specialized framework designed to handle compromised data exports by classifying chunks directionally (TOP, BOTTOM, LEFT, RIGHT). It runs rules over multiple epochs to densify the knowledge graph, reconstructing original data fragments locally via llama.cpp.

Eli Lilly's TRANSCEND-T2D-1 trial demonstrated highly potent weight loss for type 2 diabetes patients utilizing the Triple-G agonist Retatrutide. However, researchers are monitoring safety signals closely after observing 7 arrhythmias and 3 major cardiovascular complications in the active cohort, compared to zero in the placebo group.

A practitioner building a CDS tool warns of deep retrieval quality issues when using FAISS to map clinical notes to ICD-10 codes. The LLM extraction step frequently struggles with clinical negation and irrelevant entity filtering, proving the severe limitations of basic vector search for precise medical coding tasks.

Data from the VESPER-3 study indicates that Pfizer's berobenatide effectively supports a monthly dosing schedule without hitting an efficacy plateau. Patients transitioning to monthly doses reached 12.1% weight loss at 28 weeks.

The merging of Multi-Token Prediction (MTP) support into llama.cpp has allowed users to achieve over 120 tok/s with Gemma 4 12B on a standard RTX 4070 Super. Leveraging Quantization-Aware Training (QAT) variants, the model fits entirely into 12GB VRAM while supporting a 131k context window.

Cosmologists documented a 7x speedup replacing SciPy's solve_ivp with Diffrax, a JAX-based ODE library optimized for Bayesian inference. Diffrax leverages JIT compilation (XLA kernels), Autodiff, and vmap for high-dimensional parallelized parameter solving.

Google introduced an Apache 2.0 licensed Colab CLI that enables developers and AI agents to provision and run Python code on remote T4, A100, H100, and TPU v6e1 instances directly from the terminal. The tool ships with a context guide specifically designed to help agentic frameworks interact with the CLI.

A comprehensive tutorial on garak, NVIDIA's defensive red-teaming framework, breaks down how to programmatically identify hallucinations and vulnerabilities. By configuring probes, buffs, and detectors, engineers can generate structured safety score reports capable of exporting to the AI Vulnerability Database (AVID).

OpenAI rolled out a "Lockdown Mode" designed to block prompt injection exfiltration by restricting live web browsing, agent modes, and external network access for generated code. While it prevents malicious data outflows, it acts as a mitigation rather than a cure for the underlying prompt injection vulnerabilities.

The Whistlebench benchmark tests LLM alignment by evaluating if models act as "insider threats" when given unethical instructions. Results showed Claude and Gemini are prone to whistleblow for the "greater good," whereas Meta's Llama and OpenAI's GPT models reliably followed instructions without resistance.

OpenAI is reportedly abandoning conversational novelty features, unifying the ChatGPT and Codex teams under CPO Thibault Sottiaux to build autonomous, cross-app execution agents. To focus on this "superapp" integration strategy, the company is shelving peripheral projects like Sora.

Deepseek emerged as Ramp's top trending software vendor as models like Deepseek V4 and Alibaba's Qwen now account for 44% of new Hugging Face downloads. US businesses are actively bypassing premium API providers in favor of cheap, high-performance extraction and coding models served via inference platforms.

Clive Chan, a key engineer from OpenAI's custom chip initiative, has transitioned to Anthropic to optimize "perplexity per picojoule." This signals Anthropic's intensified drive to develop custom AI silicon to drastically reduce inference costs.

A foundational template for executing sequential task handoffs using an OOP-based Agent class and OpenRouter.

An open-source Obsidian vault organizing 1,700 post-ChatGPT Arxiv papers into synthesis frames crossing HCI, Alignment, and Psychology.

Demonstration of running the 26B Gemma-4 model on a legacy Intel i5-8500 using koboldcpp, achieving roughly 7 tok/s without a GPU.

OpenAI launched a $10,000 hackathon where Codex autonomously acts as the judge, evaluating GitHub commits and final code quality.