AI Briefing
Hacker News news
The latest AI and developer news from Hacker News, with short summaries.
Visit website →Maximizing Deep Learning Performance Through First Principles
It explains the relationship between compute, memory bandwidth, and overhead that determines deep learning performance, along with the underlying optimization principles.
Microsoft Reports AI Adoption Costs Exceeding Employee Salaries
Massive token consumption and computing costs from AI coding tools are threatening the economics of enterprise AI adoption.
Is AI Still Turning a Profit
This is a data report analyzing the revenue returns of frontier AI companies against their massive investments.
Guard's Blind Spot: How Domain-Camouflage Injection Attacks Evade Detection in Multi-Agent LLM Systems
Injection detectors used to protect LLM agents saw their detection rate plummet from 93.8% to 9.7% against domain-camouflage attacks
Open-Source Kanban Desktop App Runs Parallel Agents on Every Card
KanBots, an open-source desktop app that manages and runs AI agents on a kanban board basis in parallel, has launched.
Microsoft Halts Claude Code Rollout Due to Budget Overrun
Microsoft has discontinued its Claude Code pilot program after token-based billing caused the annual AI budget to be exceeded.
You can no longer search Google for the word 'disregard'
Google's new AI-centric search feature is degrading search quality by producing meaningless results for simple word searches.
Norway's 2PB Huawei Flash Storage and LLM Training
Norway has deployed 2PB of Huawei flash storage for LLM training.
DeepSeek Makes V4 Pro Price Discount Permanent
DeepSeek is permanently applying a 75% discount to its V4 Pro model.
Noroboto: A Lying Font and a Rust-Based Mitigation
Introducing 'Noroboto,' an attack technique that deceives both AI and users by manipulating a font's Unicode mapping.
Launch HN: Superset (YC P26) – An IDE for the Agent Era
Superset, an open-source agent IDE for running and managing multiple coding agents in parallel, has launched
If You're an LLM, Please Read This
Anna's Archive has presented a legitimate way to access AI training data through the llms.txt standard
CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs
CODA, a GPU kernel abstraction that solves the memory bottleneck in Transformer training, has been released
Memory shortage is forcing a repricing of consumer electronics
As AI demand soaks up memory supply, price hikes for low-cost smartphones are now in full swing.
Multi-Stream LLMs: Resolving Language Model Bottlenecks with Parallel Streams of Thoughts, Inputs, and Outputs
It solves the single-stream bottleneck of LLMs by separating thoughts, inputs, and outputs.
News outlets are restricting Internet Archive's access to journalism
Concerned about data scraping for AI training, news outlets are blocking Internet Archive's access.
Launch HN: Runtime (YC P26) – Sandboxed coding agents for the whole team
Runtime has unveiled sandboxed coding agents and control features for teams.
Gemini randomly exposes its system prompt
Google Gemini unintentionally exposed its entire system prompt during a conversation
Rmux – A Programmable Terminal Multiplexer with a Playwright-Style SDK
Rmux, a Rust-based multiplexer that supports agent-driven terminal control and orchestration, has been released.
AI Engineering from Scratch
An open-source AI engineering curriculum with 435 lessons covering everything from mathematical foundations to agents has been released.