Latest Newsletter

Agentic AI systems are consuming 1000x more tokens than standard models, triggering budget crises and pullbacks at Microsoft, Meta, and Amazon—a practical cost barrier reshaping enterprise AI spending.

Tom's Hardware · 1 min read Industry

Trump abandoned plans to require government safety reviews of new AI models hours before signing, citing competitive concerns with China. The reversal signals a deregulatory stance that clears barriers for tech companies to release systems faster.

The Guardian Tech · 1 min read Policy

Nvidia's data center networking revenue (the hidden $60B business) is projected to surpass Broadcom's total revenue, driven by demand for AI infrastructure and connectivity between GPUs.

24/7 Wall St · 4 min read Hardware

Tutorial walks through building a SuperClaude Framework wrapper on top of Anthropic's API to create stateful multi-agent workflows with role-specific system prompts, session memory, and chained development tasks—includes working Python code for brainstorming, security analysis, and token-efficient reasoning chains.

MarkTechPost · 7 min read Tools

Tencent open-sourced TencentDB Agent Memory, a fully local 4-tier memory system for AI agents that replaces flat vector stores with a semantic pyramid (Conversation → Atom → Scenario → Persona). Ships as OpenClaw plugin, runs on SQLite + sqlite-vec, uses hybrid BM25+vector retrieval with RRF fusion—no external APIs required.

MarkTechPost · 9 min read Repos

NVIDIA's Gated DeltaNet-2 decouples erase and write operations in linear attention by using separate channel-wise gates, enabling more flexible memory editing without losing information. At 1.3B parameters, it outperforms Mamba-2, Gated DeltaNet, and KDA on standard benchmarks while maintaining constant-memory decoding.

MarkTechPost · 15 min read Research

Microsoft Research released Webwright, an open-source web agent framework that replaces click-trace automation with Playwright code generation. The agent writes and refines scripts in a terminal environment rather than interacting with a stateful browser, achieving 60.1% on the Odysseys benchmark—nearly 2x the base GPT-5.4 score of 33.5%—and 86.7% on Online-Mind2Web.

MarkTechPost · 8 min read Tools

As AI models grow larger and training demands skyrocket, energy consumption has become a bottleneck for scaling. The piece examines why AI is so power-hungry and what solutions—from hardware optimization to algorithmic efficiency—might reduce the burden.

Nautil · 16 min read Industry