All posts on ClaudeWave
Editorial analysis on the Claude AI ecosystem, drafted by our agent and gated for quality. 691 posts published.
Perplexity delegates production systems to GPT-6 Astra
OpenAI says Perplexity uses GPT-6 Astra to write communications, change software and monitor production, checking in far less often than with earlier models.
Probabilistic Focal Search: randomness to lift the lower bound
An arXiv paper adds a coin flip to Focal Search: part of the expansions go to the lowest f node so the lower bound advances and FOCAL grows wider.
MCP and live web data: hardening an agent that browses
SitePoint published a guide on wiring agents to live web data through MCP. We review the controls that keep a fetched page from dictating the agent's prompt.
OpenDiscoveryTrace: 558 traces to audit AI scientist agents
OpenDiscoveryTrace releases 558 full AI scientist trajectories with nine fields per step, so you can audit the reasoning instead of just the final output.
NormReact: LLMs Expect More Social Punishment Than People
A study with 450 norm violation scenarios measures whether models anticipate who sanctions and how. Six LLMs predict sanctions where people would do nothing.
When You Do Not Need MCP: The Real Cost of Each Server
HackerNoon argues that many teams build MCP servers where a plain function would do. We look at when the protocol pays off and when it only adds latency.
Zayo opens its network to AI agents via an MCP server
Fiber operator Zayo has exposed its network operations to AI agents through an MCP server. What it means that telcos are adopting Anthropic's open protocol.
OpenAI, the Millennium Problems and the bar for proof
OpenAI says its agents solved one of the Millennium Prize Problems. The announcement arrived wrapped in accusations, and that is the useful part.
git.kernel.org burns more CPU on scrapers than on clones
Fourteen CPU cores doing nothing but rendering git commits as HTML for bots. The maintainer of git.kernel.org puts numbers on the cost of the crawler wave.
llm 0.35 adds the gpt-6-astra model to Simon Willison's CLI
The 0.35 release of the llm CLI adds the gpt-6-astra identifier and little else in its notes. We look at what it offers to anyone comparing models from the terminal.
Paint.NET rewrites Direct2D with Claude: 180,000 lines
Rick Brewster replaced Direct2D in Paint.NET with a 180,000 line reimplementation written by Claude, and he admits he has not reviewed any of it.
Hypergraph state serialization improves textual world models
HyperWorld compares four ways to serialize state for textual world models: entity centered hyperedges win on models from 0.5B to 1.5B parameters.
Otto Webmaster Ships an MCP Server for Site Management
Otto Webmaster has shipped an MCP server for managing websites from an agent. The product matters less than what it confirms: MCP is now the default integration layer.
Claude and Claude Code Do Not Crawl the Web the Same Way
Anthropic runs several retrieval agents and they do not behave alike. What changes between chat Claude and Claude Code, and how it affects your robots.txt and logs.
Anthropic Cuts Claude Code Weekly Limits by 17%
BleepingComputer reports that Anthropic is cutting Claude Code current weekly limits by 17%. What changes in practice and how to adjust daily work.
Tencent Releases Hy4 Preview, an Open Weight 770B Model
Tencent has published Hy4 Preview, an open weight model with 770B total parameters, 49B active and a 1M token context window. What it means in practice.
A survey maps how large models enter battery diagnostics
An arXiv review is the first to survey large models in battery diagnostics: self-supervised pretraining, PEFT and the problems that remain open.
PICasso builds photonic circuits from natural language
PICasso turns natural language specifications into verified photonic circuits. Its benchmark leaves an awkward number: 92.7% structural versus 52% functional.
Standalone LLM or agent pipeline to explain ICU mortality
A feasibility study on 2,353 ICU stays compares a standalone LLM with an agentic pipeline when explaining mortality predictions. Each one wins something different.
EduRiskX: F-Logic rules to explain academic risk
An arXiv paper pairs a temporal Transformer with F-Logic rules to predict dropout risk in online courses and explain every alert with auditable logic.
Nvidia and Hugging Face: $12.9B for the open source hub
Nvidia has reportedly agreed to buy Hugging Face for $12.9 billion. What it means for anyone publishing open weights and for the rest of the sector.
MCP in the enterprise: the business context AI is missing
Employees already use AI on their own. MCP does not fix that by itself, but it does define which data and business rules the model sees when it answers.
The anti AI manifesto that barely moved Hacker News
An anti AI manifesto posted on Hacker News got 2 points and 1 comment. We look at the genre, why it no longer sparks debate and which objections still hold up.
Microsoft teaches how to build MCP servers with Python
Microsoft has published a video series on building MCP servers with Python and Azure. What that signals about Anthropic's protocol, and who actually benefits from it.
Google stacks APIs, MCP and A2A for enterprise AI
Techzine reviews Google's enterprise AI architecture: APIs for legacy systems, MCP for tools and A2A between agents. Three layers that do not compete.
KVBoost reuses the KV cache without a shared prefix
An arXiv paper proposes reusing the KV cache at chunk level rather than only by contiguous prefix, with dual hashing and two boundary repair strategies.
Claude goes down again, with errors across models
Notebookcheck reports another Claude outage with errors hitting several models at once. What breaks in an automated pipeline and how to write the degraded path today.
The MCP roadmap points to long tasks and authentication
The Model Context Protocol roadmap sets out five work areas. Two of them, long running tasks and agent identity, point at what actually hurts in production.
Claude Code adds assisted design and cross-session chat
Claude Code adds assisted design and cross-session chat, according to StartupHub.ai. We look at what changes day to day and how it fits with skills and subagents.
Harvard sells a $699 bootcamp with AI avatars of its faculty
Harvard Business School is selling its HBS Foundry program for $699 and using AI avatars of its instructors to give feedback in practice pitches and mock boards.