AI News Today

<pre style='white-space:pre-wrap;width:81ex'>fix(matrix): truncate reply context on code-point boundaries (#97471) truncateReplyBody used a raw value.slice(0, MAX_REPLY_BODY_LENGTH - 3) to shorten long reply bodies. When the cut index fell between the two UTF-16 code units of a surrogate pair (e.g. an emoji), the slice left a lone high surrogate before the ellipsis, which renders as a broken glyph in the reply context shown to the agent. Replace the raw slice with sliceUtf16Safe from the plugin

<pre style='white-space:pre-wrap;width:81ex'>[codex] Update Claude CLI billing docs (#96848) * Update Claude CLI billing docs * Sync Claude live testing billing note</pre>

<p>Apple's newest on-device model carries about 20 billion parameters, and on any given request it fires maybe one to four billion of them. That gap — 20B stored, roughly 3B running — is the whole story of 2026. The model that now ships inside the latest iPhone is no longer a shrunken, lobotomized cousin of the cloud model. It's a different kind of object: large in flash, small in motion, and it never phones home.</p> <p>For three years the on-device pitch was mostly aspirational. Demos ran, lat

<blockquote> <p><strong>TL;DR —</strong> Coding agents stopped being a checkbox in your IDE and turned into a four-way platform war in the first half of 2026. Anthropic is winning the model-and-product fight, OpenAI is winning distribution, and Cognition is winning the enterprise. The real moats are model cadence, install base, and price — not features. Pick your agent like you'd pick a vendor you might have to leave, because one of these companies is going to whipsaw your workflow before the ye

Article URL: https://compactmag.substack.com/p/ai-and-the-crisis-of-classical-liberalism Comments URL: https://news.ycombinator.com/item?id=48709774 Points: 1 # Comments: 0

<h2> The setup </h2> <p>I have been building a small tool called <code>claude-relay</code>. It passes messages between terminal sessions running on the same machine: one session finishes a piece of work, sends a short message, another session picks it up. Think of it as a tiny local message queue with a chat-style front end.</p> <p>That front end is a terminal UI built with <a href="https://textual.textualize.io/" rel="noopener noreferrer">Textual</a> and Rich. It lists incoming messages and ren

In the past two years, businesses have been trying to fit large language models (LLMs) into support, analytics, development, and internal automation like never before. Along with the increasing adoption of AI technology , another trend is gaining momentum — cybercriminals are taking advantage of the disconnect between assumptions about LLMs and their actual characteristics. In 2025 and 2026, several independent sources have highlighted the same trend: Prompt injection remains one of the most imp

<p>Last Tuesday your team merged a PR with 40 lines of clean TypeScript. Code review passed — the function was readable, typed correctly, and had a unit test. Twenty minutes after deploy, CloudWatch alerted: your Orders table was being fully scanned on every request. The merged function called <code>.scan()</code> without a partition key filter. Claude Code wrote it; nobody caught it because nobody — not the reviewer, not the tests, not the linter — had any way to know that Orders has 8 million

<p>We want to tell you about a problem we couldn't stop thinking about, the system I built to solve it, and what the process taught me about building for correctness rather than for hype.</p> <h2> <strong>The Gap I Kept Seeing</strong> </h2> <p>We work on AI systems. The more time I spent with memory systems — MemGPT, LangChain Memory, Mem0, Zep — the more uncomfortable I became with one thing none of them mentioned in their marketing: the memories are stored in plaintext.</p> <p>Not <strong><em

<pre style='white-space:pre-wrap;width:81ex'>fix(models): mark local Ollama rows available (#97491) * fix(models): mark local Ollama rows available * fix(models): mark local Ollama rows available --------- Co-authored-by: Vincent Koc <vincentkoc@ieee.org></pre>

<p>Most AI sales assistants are great at summarizing a meeting. The problem is that they forget everything once the next meeting begins.</p> <p>Real sales conversations span weeks or even months. New stakeholders join, pricing objections evolve, competitors appear, and action items pile up. If the AI only remembers the latest transcript, it becomes another note-taking tool instead of a real assistant.</p> <p>I built DealMind to solve that problem.</p> <p>The Problem</p> <p>Traditional AI assista

<p>Procurement processes in many small and medium businesses still rely heavily on emails, spreadsheets, and manual vendor comparisons. This often results in delayed approvals, fragmented communication, and inefficient supplier selection.</p> <p>To address these challenges, we built <strong>ProcureAI</strong>, an AI-powered procurement platform that digitizes the complete procurement lifecycle.</p> <p>*<em>The Problem<br> *</em><br> Traditional procurement workflows involve multiple disconnected

Article URL: https://semgrep.dev/blog/2026/we-have-mythos-at-home-glm-52-beats-claude-in-our-cyber-benchmarks/ Comments URL: https://news.ycombinator.com/item?id=48709670 Points: 6 # Comments: 1

<pre style='white-space:pre-wrap;width:81ex'>fix(irc): prevent ghost nick collisions on rejoin after network delay (#96108) * fix(irc): prevent ghost nick collisions on rejoin after network delay * test(irc): add regression tests for fallback nick uniqueness Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: wendy-chsy <wan.wenyan@xydigit.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com></pre>

<pre style='white-space:pre-wrap;width:81ex'>fix: use no-output placeholder for empty OpenAI tool results (#97423) Co-authored-by: scribe-dandelion-cult <scribe-dandelion-cult@hotmail.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com></pre>

<pre style='white-space:pre-wrap;width:81ex'>feat(onboard): show setup timeline note (#97482) * feat(onboard): show setup timeline note * test(agents): stabilize retained lock CI flakes</pre>

Skills for Claude Code and Codex are hard to test. What I mean by hard is that there's no standard way to do it. You evaluate the skill once on something, it looks like it works. You publish it. Then the new super model releases (GLM 5.2 anyone?), it will quietly break for some part, and you won't find out until your users complain. I also faced the same problem, so I tried to build something lightweight to stop doing that. Caliper. It's a local and lightweight harness that runs a skill k times

Article URL: https://verigate.cloud Comments URL: https://news.ycombinator.com/item?id=48709538 Points: 1 # Comments: 0

Article URL: https://github.com/JorG18/agentcrawl Comments URL: https://news.ycombinator.com/item?id=48709512 Points: 2 # Comments: 0

I have a terrible memory and often forget characters or minor plot points if I put a book down for too long. I built Paige as a way to ask questions to refresh my memory while avoiding spoilers for the part of the book I haven't read yet. Comments URL: https://news.ycombinator.com/item?id=48709372 Points: 1 # Comments: 0

Article URL: https://www.youtube.com/watch?v=tK2K26CeThY Comments URL: https://news.ycombinator.com/item?id=48709325 Points: 1 # Comments: 1

Article URL: https://www.bgr.com/2186797/nvidia-home-mini-ai-data-center-span/ Comments URL: https://news.ycombinator.com/item?id=48709324 Points: 1 # Comments: 3

Agents write code really well, but it can get hard to understand what they've written and why. Shikhu is a CLI tool and Agent Skill that facilitates learning your code through self-quizzing, transcript analysis, and validation flows. I'm Arjun, I'm currently a developer advocate at Pinecone, and I use agents to code a lot! I enjoy using agents to code, but I've been feeling like I've lost some conceptual learning and understanding that comes with writing the code yourself. I wanted to make a too

Article URL: https://www.lesswrong.com/posts/yrbyyvFvuaGfRAtB7/evaluating-offline-monitoring-of-internal-ai-agents-2 Comments URL: https://news.ycombinator.com/item?id=48709299 Points: 2 # Comments: 0

HP Inc. scales its OpenAI Frontier partnership to deploy AI across customer experiences, software development, and enterprise operations.

A jury declared a mistrial in the case of the United States vs. Jonathan Rinderknecht, who stands accused of sparking the Pacific Pallisades fire of 2025.

Article URL: https://www.bbc.co.uk/news/articles/cd95k584pzqo Comments URL: https://news.ycombinator.com/item?id=48709160 Points: 2 # Comments: 0

Article URL: https://www.exponentialview.co/p/the-state-of-the-ai-economy Comments URL: https://news.ycombinator.com/item?id=48709080 Points: 2 # Comments: 1

<!-- SC_OFF --><div class="md"><p>Hi all,</p> <p>I’m working on a research project exploring how stateless LLM-based chatbots handle long conversations and whether important earlier information is still reliably retained over time.</p> <p>My idea is to:</p> <ul> <li>Run a chatbot using an LLM API without any external memory system</li> <li>Introduce key facts early in a long conversation</li> <li>Continue with many unrelated messages (hundreds of turns)</li> <li>Later test whether the model can

Article URL: https://english.elpais.com/education/2026-06-28/ai-fraud-at-brown-university-academic-integrity-is-at-risk.html Comments URL: https://news.ycombinator.com/item?id=48708991 Points: 32 # Comments: 22

Comments

I've struggled to find a clear definition of what it means to have "taste" when working with AI, so I wanted to try to find examples. Here's my example: React.js was released in 2013 (before I started programming professionally, so I could be wrong about this). But taste is knowing that React is the "future" even when all the training data that AI would've had until at that point in time was with Angular, Backbone, jQuery, etc. So no AI would suggest React, purely because it's too new and not en

Article URL: https://www.okanode.com Comments URL: https://news.ycombinator.com/item?id=48708874 Points: 2 # Comments: 1

Article URL: https://www.arguslab.co/ Comments URL: https://news.ycombinator.com/item?id=48708829 Points: 1 # Comments: 0

Article URL: https://12gramsofcarbon.com/p/agentics-tech-things-tokenmaxxing Comments URL: https://news.ycombinator.com/item?id=48708795 Points: 57 # Comments: 65

Article URL: https://ramblingafter.substack.com/p/it-might-be-kinder-to-kill-your-ai Comments URL: https://news.ycombinator.com/item?id=48708794 Points: 1 # Comments: 0

Article URL: https://github.com/zentrix-innovative-labs/galaxdb Comments URL: https://news.ycombinator.com/item?id=48708750 Points: 1 # Comments: 0

Comments

Article URL: https://tokonomics.ca/blog/we-tracked-1m-llm-api-calls-most-were-wasting-money Comments URL: https://news.ycombinator.com/item?id=48708666 Points: 3 # Comments: 0

Tired of the noise and FUD around AI water usage, created a little site for comparing AI usage with common daily activities. Comments URL: https://news.ycombinator.com/item?id=48708624 Points: 1 # Comments: 0