Memory has grown to nearly two-thirds of AI chip component costs

HBM has become the primary cost driver for AI chips, increasing from 52% to 63% of total component spending between Q1 2024 and Q4 2025. While logic die costs remained stable at approximately 13%, the relative share of advanced packaging and auxiliary components declined. This trend, fueled by tight supply and rising HBM prices, is significantly impacting capex for hyperscalers as they scale infrastructure for LLM training and inference.

Read more

DeepSeek to Make Permanent 75% Discount on Flagship AI Model

DeepSeek has announced a permanent 75% price reduction for its flagship V4-Pro model, extending a discount originally scheduled to expire in May. By maintaining developer costs at a quarter of the initial price, the company aims to intensify competition between Chinese AI firms and their global LLM counterparts.

Read more

There's Never Been a Better Time to Study Computer Science

The computer science labor market is facing a significant downturn, with declining enrollment and high unemployment among new graduates as LLMs automate entry-level programming. With firms like Anthropic reporting that 90% of their new code is AI-generated, the traditional value proposition of junior developers is being challenged, forcing a shift in how the industry approaches technical talent and software development.

Read more

AI didn't kill your junior pipeline. You did

The engineering industry is dismantling its talent pipeline by prematurely declaring junior roles obsolete in favor of LLMs, a move that ignores how senior expertise is cultivated. This shift creates brittle organizations, causes senior skill atrophy, and risks total dependency on volatile AI vendors. To avoid a future talent crisis, companies must redefine junior positions to use AI as a learning accelerator rather than a replacement, ensuring a sustainable supply of future senior engineers and maintaining institutional knowledge.

Read more

Our AI just got better

Ecosia is migrating its AI infrastructure to a European provider to enhance energy efficiency, ensure GDPR-compliant privacy, and enable future self-hosting. The transition focuses on reducing resource consumption while introducing technical features such as file uploads, improved web sourcing, and persistent memory. The service remains an optional, not-for-profit alternative to mainstream LLM offerings.

Read more

Constraint Decay: The Fragility of LLM Agents in Back End Code Generation

This study evaluates LLM agents on multi-file backend generation, identifying a "constraint decay" phenomenon where performance drops significantly as structural requirements like architectural patterns and ORMs increase. While agents perform well in minimal frameworks like Flask, they struggle in convention-heavy environments like Django and FastAPI, primarily due to data-layer and ORM violations. The results highlight that simultaneously satisfying functional and structural requirements remains a significant challenge for production-grade autonomous coding.

ArXiv

SSV: Sparse Speculative Verification for Efficient LLM Inference

Speculative decoding and dynamic sparse attention, while individually effective for accelerating long-context LLM inference, exhibit a structural mismatch that limits KV-block reuse and amplifies NSA overheads when combined. SSV is a sparse speculative-verification framework that resolves this by integrating overlap-aware grouped-query execution, refresh/reuse-based NSA kernel fusion, and profile-guided prompt-adaptive orchestration. This approach improves cross-query reuse and reduces overheads, achieving up to 3.49x end-to-end throughput and 6.86x kernel speedups on NVIDIA H100 GPUs.

ArXiv

A Language for Describing Agentic LLM Contexts

ACDL (Agentic Context Description Language) is a standardized framework for specifying the structure and evolution of LLM input contexts in agentic systems. It provides formal constructs for role sequences, dynamic content, and iterative structures, enabling precise communication of prompt architectures independent of specific implementations. The language includes visualization tools to replace informal documentation methods for complex context engineering in research and development.

ArXiv

Advancing Mathematics Research with AI-Driven Formal Proof Search

Researchers evaluated LLMs' ability to solve open mathematical problems by generating formal proofs in Lean, successfully resolving 9 Erdős problems and 44 OEIS conjectures. The study demonstrates that autonomous agents combining LLM generation with Lean-based verification can effectively tackle complex research in fields like combinatorics and algebraic geometry. Results suggest that while basic generation-verification loops are functional, specialized agent designs are required to optimize costs for the most difficult problems.

ArXiv

The Green Side of the Lua

An empirical study of 25 Lua versions shows that JIT compilers achieve a 7x improvement in both energy efficiency and execution speed over standard interpreters. While LuaJIT still consumes 6x more energy than C, the findings underscore the importance of JIT compilation in reducing the environmental impact and performance overhead of interpreted languages.

ArXiv

Local note engine uses LLM to organize notes into a knowledge graph

NoteCast is a local engine that enriches raw notes and organizes them into an evolving knowledge graph using a three-stage LLM pipeline. It automates summarization, keyword extraction, and vector embeddings to classify notes into a multi-parent DAG of themes and subtopics. The system supports multiple LLM providers, including local inference via Ollama, and features a CLI, REST API, and Obsidian-compatible vault synchronization.

Repo

Scan any codebase in 3s, then verify what your AI builds

Anatomia is a multi-agent CLI framework designed to automate the software development lifecycle through a structured pipeline of scoping, planning, building, and verifying code changes. It employs five specialized agents to enforce "sealed contracts" via typed assertions, ensuring implementation matches specifications through mechanical proof rather than simple LLM generation. The system maintains a proof chain for auditability and includes a learning agent that promotes recurring findings into project-specific rules to improve future builds.

Repo

Layman with no degree directs AI agents to derive Newton's G to 1.86 ppm

An orchestration of AI agents discovered a high-precision phenomenological relation for Newton's constant G using an exhaustive symbolic search of 2.67 billion expressions. The project demonstrates an end-to-end scientific discovery pipeline where LLMs handled literature survey, hypothesis generation, falsification testing, and manuscript drafting under human direction. The resulting closed-form expression matches CODATA 2022 measurements within relative uncertainty, highlighting the potential for agentic workflows in fundamental physics.

Repo

AI MOD Music Remastering at 32-bit 96 kHz

Quinlight Audio is a tracker music player and remastering tool for MOD/S3M/XM/IT formats, integrating optional external AI backends (AudioSR, LavaSR, FLowHigh, AP-BWE) for sample upscaling and neural bandwidth extension. A core innovation is its multi-engine AI consensus algorithm, which combines outputs from multiple AI models using a per-bin Karcher mean on the rotor manifold to suppress hallucination artifacts by attenuating bins where engines disagree on phase. The system features audiophile-grade 64-bit double-precision mixing, anisotropic interpolation, and SIMD-optimized processing, enabling live A/B comparison of original, reference, and AI-remastered audio. The multi-engine AI consensus algorithm is patent pending.

Repo

Strudel – Generate commit messages via Apple's on-device LLM

saft is a macOS-based Git commit-message generator that leverages Apple's on-device LLM for local, private inference without requiring API keys. It features a tiered approach for handling varying diff sizes and is designed for integration into git hooks via a planned message-only flag.

Repo