AI Dynamics

Global AI News Aggregator

About

LLMS

  • LLM hallucination rates vastly exceed commercial aviation safety standards

    By

    @garymarcus

    –

    Another way to think about, with some basic math: Gary Marcus (@GaryMarcus) This is utterly false — by many order of magnitudes. Current top LLMs hallucinate (by one estimate) 4.6% of the time, about once in 25 prompts. On a known benchmark (usually things are worse on new benchmarks). Commercial airlines crash at a rate of 7 per 41 million flights. If commercial airlines crashed at the rate that LLMs hallucinated it would be 1.87 million crashes per 41 million flights. Around a quarter million times greater. The dude below does not know what he is talking about. — https://nitter.net/GaryMarcus/status/2041339541701939365#m

    → View original post on X — @garymarcus, 2026-04-07 02:19 UTC

  • AI Hallucinations Reliability: Comparing Frequency to Aviation Safety

    By

    @garymarcus

    –

    do planes crash 4.6% of the time? no, not even close. but that’s the # the guy gave me as an estimate. when hallucinations are as rare as plane crashes, the world will be very very different.

    → View original post on X — @garymarcus

  • Gary Marcus vindicated: LLM limitations predicted in 2022

    By

    @garymarcus

    –

    literally exactly what i said in March 2022. if only people had read the full article [nautil.us/deep-learning-is-h…] and not just the title, we wouldn’t be here now. Hillbilly Cosmologist (@RobinWeep) It took most of us several years but the magic has finally worn off. LLMs do not reason and are not intelligent. They're remarkable tools for accomplishing simple tasks, but they have real limits that are seemingly unsolvable without a paradigm shift. — https://nitter.net/RobinWeep/status/2041334881901322533#m

    → View original post on X — @garymarcus, 2026-04-07 02:10 UTC

  • Milla Jovovich and Ben Sigman Launch MemPalace AI Memory System
    Milla Jovovich and Ben Sigman Launch MemPalace AI Memory System

    By

    @scobleizer

    –

    Well, well, well, Milla Jovovich @MillaJovovich launching an AI memory system with Claude was not on my 2026 list, nor was almost everything else delivered this year. What a time… Here's my official invite for Ben and Milla to join the M.A.U. podcast. Ben Sigman (@bensig) My friend Milla Jovovich and I spent months creating an AI memory system with Claude. It just posted a perfect score on the standard benchmark – beating every product in the space, free or paid. It's called MemPalace, and it works nothing like anything else out there. Instead of sending your data to a background agent in the cloud, it mines your conversations locally and organizes them into a palace – a structured architecture with wings, halls, and rooms that mirrors how human memory actually works. Here is what that gets you: → Your AI knows who you are before you type a single word – family, projects, preferences, loaded in ~120 tokens → Palace architecture organizes memories by domain and type – not a flat list of facts, a navigable structure → Semantic search across months of conversations finds the answer in position 1 or 2 → AAAK compression fits your entire life context into 120 tokens – 30x lossless compression any LLM reads natively → Contradiction detection catches wrong names, wrong pronouns, wrong ages before you ever see them The benchmarks: 100% recall on LongMemEval — first perfect score ever recorded. 500/500 questions. Every question type at 100%. 92.9% on ConvoMem — more than 2x Mem0's score. 100% on LoCoMo — every multi-hop reasoning category, including temporal inference which stumps most systems. No API key. No cloud. No subscription. One dependency. Runs on your machine. Your memories never leave. MIT License. 100% Open Source. github.com/milla-jovovich/me… — https://nitter.net/bensig/status/2041236952998171118#m

    → View original post on X — @scobleizer, 2026-04-07 00:43 UTC

  • 8 Types of AI Agents You Should Know
    8 Types of AI Agents You Should Know

    By

    @ronald_vanloon

    –

    8 Types of #AIAgents You Should Know by @PythonPr #LLM #ArtificialIntelligence #ML #MachineLearning

    → View original post on X — @ronald_vanloon, 2026-04-07 00:20 UTC

  • Anthropic’s Revenue Surges to $30B in 15 Months
    Anthropic’s Revenue Surges to $30B in 15 Months

    By

    @ceobillionaire

    –

    $1B to $30B in 15 months 🤯 Anthropic (@AnthropicAI) Our run-rate revenue has surpassed $30 billion, up from $9 billion at the end of 2025, as demand for Claude continues to accelerate. This partnership gives us the compute to keep pace. Read more: anthropic.com/news/google-br… — https://nitter.net/AnthropicAI/status/2041275563466502560#m

    → View original post on X — @ceobillionaire, 2026-04-07 00:17 UTC

  • pneuma-skills：AI与用户实时协作的智能创意工具
    pneuma-skills：AI与用户实时协作的智能创意工具

    By

    @elliotchen100

    –

    诸位都知道 @evermind 是 @shanda_group 系的吧，盛大下面好几个 AI Native 公司，这不来自 @TankaChat 的 bro 写了一个开源项目，觉得思路特别好，给大家科普一下。 先说一个你可能有过的体验： 你用 Claude Code 说"帮我做个网页"，AI 在终端里哗哗改代码，改完了你得自己开浏览器、找到文件、刷新页面才能看效果。觉得按钮颜色不对？切回终端，打字说"第二行那个蓝色按钮改成绿色"，AI 可能还不确定你说的是哪个按钮。改完再切回去刷新看。 整个过程就是：说 → 等 → 切 → 看 → 切回来 → 再说。来回跳，而且你用文字描述视觉问题本身就很不精确。 pneuma-skills 这个项目的做法是，把 AI 的工作区和你的预览区塞进同一个界面。AI 每改一行代码，你这边实时看到渲染结果。你觉得哪里不对，鼠标选中那个元素，AI 立刻知道你在说什么。 你可以理解成：你和 AI 在同一个 Google Docs 里协作，只不过 AI 负责写代码，你看到的是实时渲染出来的成品。 而且它不只能做一种东西。它内置了 8 个模式：网页设计、幻灯片、Markdown 文档、Excalidraw 手绘白板、draw.io 流程图、AI 插画，甚至还有一个"模式创建器"让你自己定义新的内容类型。 最让我觉得有意思的是两点： 一是它会记住你的审美偏好。你喜欢圆角、暗色系、大间距，用几次之后它就知道了，下次不用再说。这个偏好跨 session 持久化，换个项目也还在。 二是它有个叫 Evolution Agent 的东西，会分析你过去的操作，自动优化它自己的技能模板。意思是这个系统不是静态的，是会跟着你一起长的。 底层架构上，它抽象了三层契约，agent 后端是可插拔的，目前支持 Claude Code 和 OpenAI Codex。 我觉得这个项目指向了一个很重要的方向：AI 工具的下一个瓶颈不在模型能力，在交互界面。你和 AI 之间的沟通带宽越大，协作效率就越高。 这个项目盛大内部用了都觉得不错，推荐给诸位。 Ez Chan (@EzPandazki) 我本来觉得，做 PPT 应该早就每个人都有自己一套了。没想到这玩意儿好评度这么高～ 没什么好介绍的，反正 cc 天下无敌。如果有 cc 订阅（codex 也支持但是没测试过）的同学，评论自取吧。 — https://nitter.net/EzPandazki/status/2041184491490963550#m

    → View original post on X — @elliotchen100, 2026-04-06 23:41 UTC

  • Open-Source Models Dominate: 75-80% Performance at 10x Lower Cost

    By

    @clementdelangue

    –

    The biggest winners of last week were open-source and cheaper models MiniMax 2.7 Qwen 3.6 GLM 5 – Kimi 2.5 Usage is going exponentially up on…. you get about 75-80% performance as the closed models that are 10x more expensive

    → View original post on X — @clementdelangue, 2026-04-06 23:36 UTC

  • Reasoning Reduces AI Hallucinations: Data Analysis

    By

    @garymarcus

    –

    I don’t doubt that reasoning has reduced hallucinations but you are putting words in my mouth and ignoring the data 4.6% is not an upper bound, coming from a known benchmark. i am sure under some circumstances the number will prove to be higher.

    → View original post on X — @garymarcus

  • Neural Network Expert Disputes AI Model Hallucination Claims

    By

    @garymarcus

    –

    I am blocking you unless you retract this. I have been working on neural networks for 30 years. A recent survey showed my technical predictions have been over 90% correct. Your own data (ie the data that you pointed) supported my claim that current models still hallucinate.

    → View original post on X — @garymarcus