AI News Today

<pre style='white-space:pre-wrap;width:81ex'>fix(browser): make copilot runtime bundle deterministic (#111109)</pre>

<blockquote> <p><em>Capture the ringleader first. The rest will scatter on their own.</em><br> — The 36 Stratagems, <a href="https://en.wikipedia.org/wiki/Thirty-Six_Stratagems#Defeat_the_enemy_by_capturing_their_chief_(%E6%93%92%E8%B3%8A%E6%93%92%E7%8E%8B,_Q%C3%ADn_z%C3%A9i_q%C3%ADn_w%C3%A1ng)" rel="noopener noreferrer">Capture the Ringleader</a></p> </blockquote> <p><strong>Previously on this series:</strong></p> <p><a href="https://dev.to/xulingfeng/stratagems-5-leo-walked-into-a-burning-hous

<pre style='white-space:pre-wrap;width:81ex'>fix(anthropic): guard invalid timestamps in history imports (#110536) * fix(anthropic): guard invalid timestamps in history imports * fix(anthropic): reject pre-epoch history timestamps * fix(anthropic): preserve valid transcript dates --------- Co-authored-by: Peter Steinberger <steipete@gmail.com></pre>

I cannot be the only person who feels this way. Claude is antogonistic and frankly an absolute nightmare to use. SLOW is an understatement... I get faster output from a 16GB mac mini using deepseek 8B model....and not much worse output... I ask Gemini something. I get a response in 10 seconds and its long and well detailed. I ask claude, it wastes my tokens talking to itself, makes insulting stupid and frankly rude shortcuts that have no basis in reason except for Anthropics profits... Its unuse

Comments

<p>Over the past few days, I came across a project that made me stop everything to understand how it worked. It wasn't just another framework for running language models, nor another GPU-based optimization. It was something different.</p> <p>The project is called <strong>Colibri</strong>.</p> <p>At first glance, it almost seems impossible. The idea is simple, yet remarkably bold: run a massive Mixture of Experts (MoE) model on an ordinary computer using nothing more than a CPU and an SSD, loadin

<pre style='white-space:pre-wrap;width:81ex'>refactor(agents): unify Responses tool-call id resolver (#111103)</pre>

Article URL: https://knightcolumbia.org/content/ai-as-normal-technology Comments URL: https://news.ycombinator.com/item?id=48964162 Points: 2 # Comments: 0

<p> The AI book service that shows its work </p> <p> <a href="https://www.producthunt.com/products/prosed-turn-your-content-into-a-book?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1200227?app_id=339">Link</a> </p>

<pre style='white-space:pre-wrap;width:81ex'>fix(diagnostics-otel): classify model calls as client spans (#104211) Model-call spans wrap outbound provider inference requests, so classify them as CLIENT independently of the optional GenAI naming convention. Generic tool spans remain unchanged because their shared events can represent local work. Release note: operators filtering dashboards or alerts on span kind will see model-call spans move from INTERNAL to CLIENT. Co-authored-by: Peter Steinbe

<p>Python’s dominance in AI isn't an accident. In 2026, it’s the center of the modern ML stack because it bridges research, orchestration, and production.</p> <p>The Ecosystem Advantage</p> <p>The strongest argument for Python is its mature, battle-tested ecosystem. PyTorch has become the research-to-production backbone, while Scikit-learn remains the gold standard for practical machine learning. These tools let engineers prototype quickly without fighting the language.</p> <p>LLMs and Code Gene

<p>AI agents are programs that can perceive information, decide what to do, and take actions toward a goal. Unlike a simple chatbot that only responds to prompts, an agent can use tools, follow steps, and adapt its behavior based on context. In this tutorial, you’ll build a small weather assistant agent in Python that can answer questions like “What’s the weather in Paris?” by calling a weather API.</p> <p>Prerequisites</p> <p>Before you start, make sure you have:</p> <ul> <li>Basic Python knowl

<pre style='white-space:pre-wrap;width:81ex'>fix(speech): strip markdown before TTS and route code-heavy replies sensibly (#110948) The talk.speak synthesis path passed raw agent text to providers, so talk mode could read fenced code, links, and tables aloud. Speech normalization now happens once at the shared synthesis boundary (speech-mode stripMarkdown: label-only links, decorative bullet/glyph cleanup, punctuation collapsing). Code-heavy replies (>=50% fenced content, CommonMark-aware scanne

Article URL: https://pulkitsharma.substack.com/p/no-an-ai-cannot-know-the-future-and Comments URL: https://news.ycombinator.com/item?id=48964113 Points: 2 # Comments: 0

<pre style='white-space:pre-wrap;width:81ex'>fix(browser): keep copilot runtime bundle in sync (#111089)</pre>

Article URL: https://github.com/kaiwutech-TW/flightwake Comments URL: https://news.ycombinator.com/item?id=48964105 Points: 2 # Comments: 0

<pre style='white-space:pre-wrap;width:81ex'>fix(plugins): roll back cleared runtime registrations when an activating reload aborts (#103812) * fix(plugins): roll back cleared runtime registrations when an activating reload aborts loadOpenClawPlugins clears the process-global agent-harness, plugin-command, compaction, detached-task, interactive-handler, embedding, memory-embedding, and memory stores up front on an activating load and only re-registers them at the end via activatePluginRegistry.

<p>Most "AI support bot" demos bolt an LLM onto a shared Gmail or hang a webhook off a helpdesk and call it a day. That works right up until you want the agent to actually <em>be</em> a participant in the conversation — to receive mail at a real address, decide what to do with it, and reply as itself instead of as a script poking at someone else's mailbox.</p> <p>So let's not do that. Let's give <code>support@yourcompany.com</code> its own first-class identity: an <strong>Agent Account</strong>

<p>Most "AI email" demos point a model at a human's mailbox over OAuth. You connect your Gmail, the agent reads your threads, drafts replies, maybe sends one. That's fine when the agent is acting <em>on your behalf</em> — a copilot riding shotgun in your inbox.</p> <p>It falls apart the moment you want the agent to be its <em>own</em> participant. Picture an onboarding agent that registers your test tenant on a SaaS every CI run, or a research agent that needs a developer account on a data sourc

<p>For years, VS Code was the obvious default. It was fast, extensible, familiar, and good enough for almost every kind of development work. I used it the way many engineers do: a polished editor, a handful of plugins, a terminal at the bottom, and then-eventually-an AI extension bolted on top.</p> <p>That setup worked. But it always felt like a compromise.</p> <p>Cursor is the first tool I’ve used that doesn’t treat AI as an add-on. It treats AI as part of the development environment itself. Th

<p>I got woken up by an alert call at 2 a.m. – the production chatbot had suddenly developed “amnesia.” It was completely unable to recall order information we had just discussed that morning. I scrambled to investigate and discovered the root cause: we had migrated the memory storage from Redis to PostgreSQL + pgvector. The migration script looked fine, but one historical memory entry’s vector index hadn’t been properly built, causing the similarity recall to return empty results. Before going

<pre style='white-space:pre-wrap;width:81ex'>fix: distinct Responses tool-call ids for repeated native Kimi calls (#110956) * fix: distinct Responses tool-call ids for repeated native Kimi calls The gateway tool (and any tool reused at the same call index) collapsed to the same call_* id in OpenAI Responses replay because the id normalizer hashed only the callId half of paired functions.<tool>:<index>|fc_tmp_* ids, dropping the unique suffix. That broke replay with invalid_replay_transcript: dan

Article URL: https://quentinkaiser.be/security/2026/07/18/ia-assisted-vuln-research/ Comments URL: https://news.ycombinator.com/item?id=48963893 Points: 1 # Comments: 0

Article URL: https://www.marble.onl/posts/copyright_vs_ai.html Comments URL: https://news.ycombinator.com/item?id=48963759 Points: 1 # Comments: 0

The AMOC is slowing down. ScienceAlert stories are written, fact-checked, and edited by humans, never generated by AI. Don't miss a story, subscribe here.

Article URL: https://www.youtube.com/watch?v=zWJ-g5u9Rqs Comments URL: https://news.ycombinator.com/item?id=48963642 Points: 17 # Comments: 24

I wanted to take kimi k3 for a spin. It turned my simple one sentence prompt to this. Repo here. https://github.com/praveer13/ilya-papers-quest Well, I'm mindblown. Very humbling for me as a software engineer. Took couple hours for it to build this completely autonomously. And it was all from its mobile app. It couldn't render this though from within the app - it does have a feature to preview any website and publish it on kimi's domain - but it didn't work for this. I had to put it on github pa

Article URL: https://ptecorepractice.com Comments URL: https://news.ycombinator.com/item?id=48963284 Points: 1 # Comments: 1

<!-- SC_OFF --><div class="md"><p>32,070 alphabetic tokens from GPT-2-small's WTE, no forward pass and no context.</p> <p>Works on mobile. Pinch to zoom, tap a token to see its nearest connections, tap a neighbour to walk the graph. Search box to jump anywhere.</p> <p>Layout is t-SNE over a compressed representation of the embedding table; edges are a minimum spanning tree in that space, so every line is a real nearest-kin relationship,</p> </div><!-- SC_ON --> submitted by <a href="https://www.

Article URL: https://conference.nber.org/conf_papers/f238865.pdf Comments URL: https://news.ycombinator.com/item?id=48963066 Points: 2 # Comments: 0

Article URL: https://github.com/ikaruscareer/SafeAI Comments URL: https://news.ycombinator.com/item?id=48963061 Points: 2 # Comments: 0

Comments

<p> One memory for every AI you use </p> <p> <a href="https://www.producthunt.com/products/atlaso?utm_campaign=producthunt-atom-posts-feed&utm_medium=rss-feed&utm_source=producthunt-atom-posts-feed">Discussion</a> | <a href="https://www.producthunt.com/r/p/1200155?app_id=339">Link</a> </p>

<table> <tr><td> <a href="https://www.reddit.com/r/MachineLearning/comments/1v07xai/gpt2_smalls_embedding_geometry_around_trump/"> <img src="https://preview.redd.it/tlvz4c3i32eh1.png?width=640&crop=smart&auto=webp&s=aad6aeec9197e26debda00093dd47611e70c5a08" alt="GPT-2 Small’s embedding geometry around “Trump”: discretized vs. continuous nearest neighbours [P]" title="GPT-2 Small’s embedding geometry around “Trump”: discretized vs. continuous nearest neighbours [P]" /> </a> </td><td> <!-- SC_OFF

Article URL: https://www.zaproxy.org/blog/2025-11-28-enhancing-zap-with-ai-for-bug-bounty-hunting/ Comments URL: https://news.ycombinator.com/item?id=48962545 Points: 2 # Comments: 0

Article URL: https://qainsights.com/20-month-the-price-ceiling-every-ai-company-copied/ Comments URL: https://news.ycombinator.com/item?id=48962535 Points: 2 # Comments: 0

Article URL: https://teadrivendevelopment.substack.com/p/i-asked-a-psychiatrist-i-know-about Comments URL: https://news.ycombinator.com/item?id=48962463 Points: 1 # Comments: 0

Been working on this for the last 2 years. There is a video on the site explaining how it works. There is also a production app powered by this engine. Give it a try, I'm happy to answer any questions. Comments URL: https://news.ycombinator.com/item?id=48962441 Points: 1 # Comments: 0

Article URL: https://github.com/saineshnakra/automated-data-analyst Comments URL: https://news.ycombinator.com/item?id=48962405 Points: 2 # Comments: 0

Last year, Sam Altman invited author Dave Eggers to give a talk to around 200 OpenAI staffers. The man has written countless novels, screenplays, pieces of journalism, started McSweeney's, and founded multiple schools and nonprofits that support writers and the arts more broadly. So one might expect he'd roll into the company's offices and offer […]