@pyagon.bsky.social
大部分的痛苦和壓力並非來自「當下」，而是源自於我們所建構的那些關於「過去」（如悔恨、心碎）或「未來」（如焦慮、恐懼）的「故事」。
專注於「當下」。
你唯一真實體驗的就是這個大約 3 秒長的動態窗口。而「當下」通常是可以改善的 （例如：如果你覺得冷，就去拿條毯子；如果你不開心，就為自己泡杯熱可可）。
專注於「當下」。
你唯一真實體驗的就是這個大約 3 秒長的動態窗口。而「當下」通常是可以改善的 （例如：如果你覺得冷，就去拿條毯子；如果你不開心，就為自己泡杯熱可可）。
Joscha Bach: The Operation of Consciousness | AGI-25
YouTube video by SingularityNET
www.youtube.com
November 9, 2025 at 3:34 PM
大部分的痛苦和壓力並非來自「當下」，而是源自於我們所建構的那些關於「過去」（如悔恨、心碎）或「未來」（如焦慮、恐懼）的「故事」。
專注於「當下」。
你唯一真實體驗的就是這個大約 3 秒長的動態窗口。而「當下」通常是可以改善的 （例如：如果你覺得冷，就去拿條毯子；如果你不開心，就為自己泡杯熱可可）。
專注於「當下」。
你唯一真實體驗的就是這個大約 3 秒長的動態窗口。而「當下」通常是可以改善的 （例如：如果你覺得冷，就去拿條毯子；如果你不開心，就為自己泡杯熱可可）。
為何我會把 Web browser 所進行的 render 動作翻譯為「描繪」而非「渲染」：
⟨Render 中文翻譯與應用情境⟩ gemini.google.com/share/6747d3...
⟨Render 中文翻譯與應用情境⟩ gemini.google.com/share/6747d3...
Gemini - Render 中文翻譯與應用情境
Created with Gemini
gemini.google.com
October 26, 2025 at 7:54 PM
為何我會把 Web browser 所進行的 render 動作翻譯為「描繪」而非「渲染」：
⟨Render 中文翻譯與應用情境⟩ gemini.google.com/share/6747d3...
⟨Render 中文翻譯與應用情境⟩ gemini.google.com/share/6747d3...
Reposted
Prompt Injection is a top risk for use of LLMs, where untrusted data can cause a system to misbehave or leak data. We're hard at work on deterministic solutions to prevent or bound the risk, and just published a paper on our innovative information flow control approach: arxiv.org/pdf/2505.23643
June 5, 2025 at 8:34 PM
Prompt Injection is a top risk for use of LLMs, where untrusted data can cause a system to misbehave or leak data. We're hard at work on deterministic solutions to prevent or bound the risk, and just published a paper on our innovative information flow control approach: arxiv.org/pdf/2505.23643
Reposted
Latest Scott and Mark To podcast is out, this one on how we manage stress:
www.youtube.com/watch?featur...
www.youtube.com/watch?featur...
EPISODE 19 - How Do Executives at Microsoft Manage Their Stress?
YouTube video by Scott Hanselman
www.youtube.com
June 26, 2025 at 2:37 AM
Latest Scott and Mark To podcast is out, this one on how we manage stress:
www.youtube.com/watch?featur...
www.youtube.com/watch?featur...
Reposted
I've put together a fun video compilation of the games I've played with Task Manager as a bitmap display on the monster-sized VMs we've released in Azure over the years: www.youtube.com/watch?v=L73A...
Task Manager "Game Console" on Azure Memory Optimized Servers
YouTube video by Mark Russinovich
www.youtube.com
June 11, 2025 at 11:08 PM
I've put together a fun video compilation of the games I've played with Task Manager as a bitmap display on the monster-sized VMs we've released in Azure over the years: www.youtube.com/watch?v=L73A...
By operating within our current paradigm, an AGI demonstrates mastery by engineering the very supercomputer it runs on—assembling the hardware and writing its software from scratch.
June 8, 2025 at 1:23 PM
By operating within our current paradigm, an AGI demonstrates mastery by engineering the very supercomputer it runs on—assembling the hardware and writing its software from scratch.
This is my favorite talk lately — it lowered both my anxiety and my perplexity.
youtu.be/VDhan5Kd3ms?...
youtu.be/VDhan5Kd3ms?...
CIMC Machine Consciousness Salon: Existential Hope (Joscha Bach, Lou de K)
YouTube video by California Institute for Machine Consciousness
youtu.be
June 3, 2025 at 6:23 AM
This is my favorite talk lately — it lowered both my anxiety and my perplexity.
youtu.be/VDhan5Kd3ms?...
youtu.be/VDhan5Kd3ms?...
One model to rule them all,
one model to parse them,
One model to remix them all,
and in latent space bind them.
one model to parse them,
One model to remix them all,
and in latent space bind them.
May 22, 2025 at 9:12 AM
One model to rule them all,
one model to parse them,
One model to remix them all,
and in latent space bind them.
one model to parse them,
One model to remix them all,
and in latent space bind them.
Reposted
The worst bugs are the ones where the code mostly works. Low-level AI coding is so complex that I (an the AI that I write code with), produce an annoying number of them.
May 10, 2025 at 12:32 AM
The worst bugs are the ones where the code mostly works. Low-level AI coding is so complex that I (an the AI that I write code with), produce an annoying number of them.
Reposted
Vibe coding and AI-assisted coding are massive software development productivity boosts, but @scott.hanselman.com and I don't think programming and software engineering professions are going to be obsolete any time soon: www.youtube.com/watch?v=xcaK...
EPISODE 15 - Scott & Mark Learn To… Vibe Coding (Part 2)
YouTube video by Scott Hanselman
www.youtube.com
May 1, 2025 at 9:46 PM
Vibe coding and AI-assisted coding are massive software development productivity boosts, but @scott.hanselman.com and I don't think programming and software engineering professions are going to be obsolete any time soon: www.youtube.com/watch?v=xcaK...
Every now and then, I go back to this post from eight years ago to get a feel for how the future it describes is coming closer.
Could artificially intelligent computers replace programmers in the future?
www.quora.com/Could-artifi...
Could artificially intelligent computers replace programmers in the future?
www.quora.com/Could-artifi...
Could artificially intelligent computers replace programmers in the future?
Joscha Bach's answer: I am surprised that most answers seem to be in the negative, or think that programming is one of the last fields where human labor becomes obsolete. General artificial intelligen...
www.quora.com
May 2, 2025 at 6:07 AM
Every now and then, I go back to this post from eight years ago to get a feel for how the future it describes is coming closer.
Could artificially intelligent computers replace programmers in the future?
www.quora.com/Could-artifi...
Could artificially intelligent computers replace programmers in the future?
www.quora.com/Could-artifi...
Reposted
Read about our progress on securing Microsoft in our Secure Future Initiative update:
Securing our future: April 2025 progress report on Microsoft’s Secure Future Initiative | Microsoft Security Blog
The Microsoft Secure Future Initiative (SFI) stands as the largest cybersecurity engineering project in history and most extensive effort of its kind at Microsoft. Now, we are sharing the second SFI p...
www.microsoft.com
April 21, 2025 at 5:52 PM
Read about our progress on securing Microsoft in our Secure Future Initiative update:
Reposted
This was my first “internals” book. It includes the full ROM disassembly of Applesoft BASIC, which I used to figure out how to extend the language with a computed GOSUB command:
www.atarimagazines.com/compute/issu...
www.atarimagazines.com/compute/issu...
April 14, 2025 at 2:34 AM
This was my first “internals” book. It includes the full ROM disassembly of Applesoft BASIC, which I used to figure out how to extend the language with a computed GOSUB command:
www.atarimagazines.com/compute/issu...
www.atarimagazines.com/compute/issu...
Reposted
ChatGPT Deep Research does indeed hallucinate when given a false premise (or when asked to do research on the ungrounded facts ChatGPT hallucinates). This is what I call "induced hallucination":
April 14, 2025 at 4:39 AM
ChatGPT Deep Research does indeed hallucinate when given a false premise (or when asked to do research on the ungrounded facts ChatGPT hallucinates). This is what I call "induced hallucination":
Reposted
Everyone is talking about "vibe coding", but in our latest podcast episode I explain the reasons I don't think AI is going to replace good programmers any time soon for non-trivial projects:
EPISODE 11 - Scott & Mark Learn To... Systems Thinking
YouTube video by Scott Hanselman
youtu.be
March 12, 2025 at 4:34 PM
Everyone is talking about "vibe coding", but in our latest podcast episode I explain the reasons I don't think AI is going to replace good programmers any time soon for non-trivial projects:
Reposted
Last year I shared our discovery of the Crescendo and Skeleton Key jailbreak techniques. We have identified a new one, Context Compliance Attack (CCA) that's very simple, yet very effective at bypassing LLM guardrails.
Jailbreaking is (mostly) simpler than you think | MSRC Blog
| Microsoft Security Response Center
Your Privacy Choices Opt-Out Icon
Jailbreaking is (mostly) simpler than you think
msrc.microsoft.com
March 13, 2025 at 7:41 PM
Last year I shared our discovery of the Crescendo and Skeleton Key jailbreak techniques. We have identified a new one, Context Compliance Attack (CCA) that's very simple, yet very effective at bypassing LLM guardrails.
Reposted
Here is one for you: 50's/60's space atomic age ads.
April 1, 2025 at 3:26 AM
Here is one for you: 50's/60's space atomic age ads.
Reposted
New podcast episode is out: Scott and Mark Learn to Systems Think. In this one we talk about why programmers aren't going away any time soon and the limits of 'vibe coding':
EPISODE 11 - Scott & Mark Learn To... Systems Thinking
YouTube video by Scott Hanselman
www.youtube.com
March 6, 2025 at 12:05 AM
New podcast episode is out: Scott and Mark Learn to Systems Think. In this one we talk about why programmers aren't going away any time soon and the limits of 'vibe coding':
Reposted
I didn't know I had 10 laws, but these are great so I'll claim them:
March 8, 2025 at 9:39 PM
I didn't know I had 10 laws, but these are great so I'll claim them:
Reposted
Learn how Azure uses Sysinternals tools in this podcast with Mario Hewardt, an engineer on the (small) Sysinternals team:
Episode 514 - Sysinternals Tools on Azure
YouTube video by The Azure Podcast
www.youtube.com
March 9, 2025 at 8:03 PM
Learn how Azure uses Sysinternals tools in this podcast with Mario Hewardt, an engineer on the (small) Sysinternals team:
Reposted
Do large language models develop "emergent" models of the world? My latest Substack posts explore this claim and more generally the nature of "world models":
LLMs and World Models, Part 1: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 2: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 1: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 2: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 1
How do Large Language Models Make Sense of Their “Worlds”?
aiguide.substack.com
February 13, 2025 at 10:30 PM
Do large language models develop "emergent" models of the world? My latest Substack posts explore this claim and more generally the nature of "world models":
LLMs and World Models, Part 1: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 2: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 1: aiguide.substack.com/p/llms-and-w...
LLMs and World Models, Part 2: aiguide.substack.com/p/llms-and-w...
January 23, 2025 at 7:32 PM
Reposted
Launching Version 14.2 of Wolfram Language & Mathematica: big data meets computation & AI...
writings.stephenwolfram.com/2025/01/laun...
writings.stephenwolfram.com/2025/01/laun...
January 23, 2025 at 7:06 PM
Launching Version 14.2 of Wolfram Language & Mathematica: big data meets computation & AI...
writings.stephenwolfram.com/2025/01/laun...
writings.stephenwolfram.com/2025/01/laun...
Reposted
Learn about the risks of hallucination, jailbreaks and prompt injection and current mitigations in our ACM Queue paper:
The Price of Intelligence - ACM Queue
queue.acm.org
January 23, 2025 at 7:26 PM
Learn about the risks of hallucination, jailbreaks and prompt injection and current mitigations in our ACM Queue paper:
Reposted
High order human cognition process information at around 10 bits/second - yes, 10 *bits*. Surprising that no one has done this analysis before: arxiv.org/pdf/2408.10234
January 12, 2025 at 11:15 PM
High order human cognition process information at around 10 bits/second - yes, 10 *bits*. Surprising that no one has done this analysis before: arxiv.org/pdf/2408.10234