Post-training
Joined April 2025
- Tweets 105
- Following 225
- Followers 246
- Likes 43
18 Photos and videos
May 11
Markdown for production, HTML for consumption(and experience).
HTML, as a first-class citizen in agent workflows, brings style, interactivity, and multi-modality — laying the foundation of Generative GUI.
We are witnessing the "Windows 1.0" moment for LLMs.
1
44
May 7
Never doubt: the resilience of CN manufacturing & U.S. capital markets; the capability of CN gov & U.S. tech giants; the carbon decline & silicon rise. This is the new "Never fight a land war in Asia."
22
May 6
ClaudeCode is the "DOS" for LLM. Karpathy's "LLM OS" is more like an LLM PC, while CC is the true OS
COMMAND.COM=Agent loop
RAM=Ctxt
File=memory
Syscall=Tool
Proc=Agent
The list goes on... OK maybe UNIX since there is Linux (Claw), and we are even stuck in CLI again
1
54
May 6
Probably every single AI native app will adopt the same architecture as CC, until a real OS rules...
So where is the Windows (the one that ignited the PC era as a GUI shell for DOS) and what it will look like...
34
12 Dec 2025
Oh my… finally : ephemeral, in‑context generative interaction. AI‑native beyond hype...
Disco is a new experiment from @GoogleLabs, designed to reimagine browsing and building for the modern web. The first feature we’re testing is GenTabs — a new way to remix your open tabs into totally custom apps using Gemini 3, our most intelligent model.
We’ve all felt the frustration of juggling dozens of open tabs when working on a complex task, like researching a topic or planning a trip. GenTabs proactively understands your complex tasks (through your open tabs and chat history) and creates interactive web applications to help you complete them. Just describe the tool you need and refine it using natural language. And because every generative element ties back to the web, it always links to the original sources.
1
80
8 Oct 2025
One of the best recent AI interviews, highlights a key non-consensus view: We need to see AI as a complete system, not just a collection of parts. It's not only about compute, models, or data in isolation. For example, the same model can yield significantly better results simply by allocating more compute for inference. A crucial systemic insight.
Groq Founder, Jonathan Ross: youtu.be/VfIK5LFGnlk?si=yfrW…
4
187
LΞX retweeted
23 Sep 2025
Those who build AI today control the future.
The dAGI Summit (part of Open Source AI Week by Linux Foundation) is for founders, researchers and builders driving the open-source and distributed AI that gives power back to people.
We are bringing together senior researchers and leaders from OpenAI, DeepMind, Amazon, Meta and NVIDIA, alongside the founders and maintainers of top open-source AI projects like Letta, Cline and PyTorch. They will be joined by leading research labs in open and distributed AI, such as Prime Intellect, Pluralis and Nous Research.
The goal is to focus on making AI universally accessible, provably fair, impartial and beneficial for all.
📅 October 24 • 📍 San Francisco • Open Source AI Week
🎟Use promo code for exclusive (limited!) discount: MLST_15
Get your ticket at luma.com/dagisummitsf
39
39
149
59,514
17 Sep 2025
big step forward to agent economy...
16 Sep 2025
Announcing Agent Payments Protocol (AP2), an open, shared protocol that provides a common language for secure, compliant transactions between agents and merchants.
AP2 can be used as an extension of the A2A protocol and MCP. Learn how it works ↓ goo.gle/4n7Cgsc
2
127
15 Sep 2025
kudos to cursor team, first large - scale online RL in real life!
12 Sep 2025
This is a big deal. It is the first large-scale demonstration of the advantage of real-time reinforcement learning. The recipe is scalable and requires no intervention in principle; the model can adapt forever as long as it is being used.
There is no way to achieve similar results with sim-to-real learning because that would require simulating thousands of human users.
The feedback loop of 1.5 hours is still too long. It might be good enough for this use cases but many use cases would benefit from instant learning which requires new algorithms.
1
122
10 Sep 2025
Thinking Machines Lab’s first technical post, “Defeating Nondeterminism in LLM Inference,” is an impressive debut. Their engineering tackles a core issue—hidden nondeterminism in inference—that’s crucial for on-policy RL. This work has real impact for applications where policy consistency and long-chain reasoning are essential
10 Sep 2025
Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference”
We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly.
The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains.
thinkingmachines.ai/blog/def…
153
10 Sep 2025
Tired of chatboxes? Here’s what stood out to me from this “AI Interfaces of the Future” video (featuring the Notion Calendar creator):
- Software is shifting from “nouns” (static features) to “verbs” (actions). It’s not just about giving you tools—it’s about getting things done for you.
- The canvas is emerging as the new document type. You visually map out how you want your AI to reason and act, not just what you want it to do.
- Adaptive interfaces are here: instead of a sea of buttons, you only see what’s relevant to your current context.
- Users make decisions, not busywork. Fast previews let you balance speed and quality—think WYSIWYG on steroids.
youtu.be/DBhSfROq3wU?si=exF1… via @YouTube
100
15 Aug 2025
Confucius said: 'Triple-check yourself daily.'
My AI remix:
1️⃣ Tokens maxed out? (The more, the better!)
2️⃣ Mind-melded with AI? (Or still talking past each other?)
3️⃣ Prompts kept chill? (Let the AI do its magic✨)
1
183
15 Aug 2025
A beautiful story: built for his own MCP tool collection, an engineer started by hand, but updates soon outpaced him—so he made an Agent to scan GitHub and auto-organize the site. No SEO, no tricks. Months later, forgotten but tireless, the Agent pushed the site to #1 on Google. human spark relentless AI = the most beautiful symbiosis in action
1
6
415
LΞX retweeted
12 Aug 2025
2035 Episode 1: Chinese vs American AI, the Rundown.
@ZeMariaMacedo & I had a chance to sit down with @moonshot6666 to discuss all things US vs Chinese AI.
Alex is currently Co-Founder of @get_truenorth & has a very unique perspective having worked at large chip companies, tech companies, and investment firms on both sides of the Pacific.
One of the more informative episodes available on US vs Chinese AI dynamics and the future of computing more broadly.
2035 is a new @delphi_intel podcast. Links below:
3
5
28
10,570
13 Aug 2025
(1/8) In previous report, I explored how AI models are fascinating contradictions—brilliant yet random, knowledgeable yet perceptually limited, powerful yet unconscious.
Today I'm publishing 2nd report, "Abundant Intelligence and Scarce Experience," which addresses how to harness this newly abundant intelligence effectively. We need a new collaboration playbook: align consciousness, grant autonomy, verify outputs, provide feedback, and enable continuous improvement.
In simpler terms, applications must put AI into experience
See below for detail:
5
7
23
11,079
13 Aug 2025
(7/8) We're watching vertical AI applications pivot from "selling tools" to professional firms to "selling outcomes" to end users. AI-first organizations are blooming in various forms—from ground-up builds like Crosby to AI roll-ups gobbling up traditional businesses. They're aligning experts with AI, creating collaborative environments, and flipping to outcome-based business models.
Organizations embracing AI at their core are gaining massive advantages through reinvention: treating code as labor not assets, generating processes dynamically, creating guardrailed environments, and pricing based on outcomes.
1
2
194