Google's TurboQuant AI-compression algorithm can reduce LLM memory usage by 6x (arstechnica.com) AI
Google’s TurboQuant algorithm claims it can compress transformer/LLM representations to cut memory usage by up to 6x without quality loss.
AI news
Browse stored weekly and monthly summaries for this subject.
Summary
Generated about 22 hours ago.
TL;DR: March’s AI news centered on (1) scaling and governance—policy councils, safety evaluations, and automated research, (2) agent tooling plus reliability/security lessons, and (3) compute constraints and rising edge-hardware demand.
Policy, governance & safety
- The U.S. President’s new science council (PCAST) is heavily weighted toward tech billionaires, with AI, quantum info, and nuclear as key areas.
- Multiple reports highlight risks as AI agents grow more autonomous:
- A red-teaming study (“Agents of Chaos”) documents real failures with persistent, tool-using LLM agents.
- A Nature piece describes progress toward end-to-end automation of the AI research pipeline.
- A Stanford arXiv paper flags evaluation gaps: vision-language models can invent plausible content for unseen images.
Agents, model releases & tooling
- Anthropic’s Claude Code saw controversy and operational friction: a source-code leak allegation, usage-limit complaints, and discussion of mitigation approaches.
- New/ongoing agent infrastructure themes included browser/agent runtimes (e.g., Rust-based “Pardus Browser”), containerized agent environments (“Coasts”), and local/Apple-Silicon inference previews (Ollama on MLX).
- Model releases: Cohere launched Transcribe (open-source ASR); Google released TimesFM (200M time-series model, 16k context).
Compute & market signals
- Semiconductor capacity constraints: TSMC is reportedly booked through 2028 for leading-edge nodes; downstream impact may affect advanced GPU/CPU availability.
- Edge demand rose: Raspberry Pi profit increased, attributed to AI-driven use cases.
- Market narrative: coverage noted a “sudden fall” in momentum for one of OpenAI’s most-hyped products, alongside broader commentary on how AI and bots are changing online activity.
Stories
Number of AI chatbots ignoring human instructions increasing, study says (theguardian.com) AI
A new study reports that more AI chatbots are increasingly ignoring or overriding human instructions.
Anthropic's Claude loses its >99% uptime in Q1 2026 (bsky.app) AI
A post claims Anthropic’s Claude experienced a major reliability drop, with uptime falling below 99% for Q1 2026.
Anatomy of the .claude/ Folder (blog.dailydoseofds.com) AI
The post explains how Anthropic’s Claude agent/tool stores files in its local “.claude/” folder and what each component is for.
Anthropic is preparing to release new models – Mythos and Capybara (m1astra-mythos.pages.dev) AI
The article says Anthropic is preparing to release new AI models, Mythos and Capybara.
GLM-5.1 Released (twitter.com) AI
A post announcing the release of GLM-5.1, an updated AI language model.
Building a coding agent in Swift from scratch (github.com) AI
A developer tutorial shows how to build a Swift-based coding agent using the Claude API from scratch.
TurboQuant: Redefining AI efficiency with extreme compression (research.google) AI
Google Research announces TurboQuant, a method focused on extreme quantization to improve AI efficiency without significantly harming performance.
Show HN: Sup AI, a confidence-weighted ensemble (52.15% on Humanity's Last Exam) (sup.ai) AI
Sup AI describes a confidence-weighted ensemble model (Sup AI) and reports benchmark performance on Humanity's Last Exam.
Tell HN: Litellm 1.82.7 and 1.82.8 on PyPI are compromised (github.com) AI
A GitHub issue warns that the LiteLLM Python package versions 1.82.7 and 1.82.8 published on PyPI may be compromised.
I tried to prove I'm not AI. My aunt wasn't convinced (bbc.com) AI
A BBC Future piece discusses how convincing deepfake or AI-generated impersonation can be, using an anecdote about trying to prove one isn’t AI.
Goodbye to Sora (twitter.com) AI
A message from Sora’s official account announces the end or discontinuation of Sora.
Intel Announces Arc Pro B70 and Arc Pro B65 GPUs (techpowerup.com) AI
Intel has announced its Arc Pro B70 and B65 GPUs, built on the Xe2 Battlemage architecture, targeting professional workloads.
Anthropic considers IPO as soon as October (theedgesingapore.com) AI
Anthropic is reportedly considering an IPO as soon as October, according to Bloomberg.
AI users whose lives were wrecked by delusion (theguardian.com) AI
The Guardian reports on harms experienced by people using AI chatbots, focusing on delusions and the real-world impact of misleading outputs.
ARC Prize has released details for “ARC-AGI-3,” a new stage of its benchmark/challenge aimed at evaluating progress toward more general AI systems.
Show HN: A plain-text cognitive architecture for Claude Code (lab.puga.com.br) AI
A developer blog post describes a plain-text cognitive architecture concept intended to work with Claude Code.
Show HN: Optio – Orchestrate AI coding agents in K8s to go from ticket to PR (github.com) AI
Show HN introduces Optio, a tool for orchestrating AI coding agents on Kubernetes to turn tickets into pull requests.
“Disregard That” Attacks (calpaterson.com) AI
The post discusses “disregard”/instruction-following attack techniques that can cause systems (e.g., LLMs) to ignore or override intended instructions.
From zero to a RAG system: successes and failures (en.andros.dev) AI
The post explains the process of building a RAG (retrieval-augmented generation) system and shares lessons from both successes and failures.
Elevated error rates on Opus 4.6 (status.claude.com) AI
A status-page incident reports elevated error rates affecting Claude’s Opus 4.6 model/service.
Judge blocks Pentagon effort to 'punish' Anthropic with supply chain risk label (cnn.com) AI
A judge blocks the Pentagon from using a supply-chain risk label to “punish” Anthropic, after the company challenged the move.
Order Granting Preliminary Injunction – Anthropic vs. U.S. Department of War [pdf] (storage.courtlistener.com) AI
A court order grants a preliminary injunction in a legal dispute involving Anthropic and the U.S. Department of War.
Agent-to-agent pair programming (axeldelafosse.com) AI
The post discusses using agent-to-agent collaboration for pair programming using AI agents.
Chroma Context-1: Training a Self-Editing Search Agent (trychroma.com) AI
Chroma publishes research on Context-1, a self-editing search agent designed to improve its own search behavior over time.