Xiaofei Du 杜笑菲 🦄

Building tools that I use myself

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Sam Altman has access to the most advanced AI on the planet. He still doesn't use it on his own email.

His words, on David Senra's show:

"I should not be scrolling mindlessly through my emails and trying to figure out which one is like least painful for me to open and respond..."

What struck me is that he says why. It isn't that the tools aren't ready — he's the one building them. It's that something in him is convinced the clicking, the sorting, the to-do list is what being productive feels like.

So the barrier was never capability. It's trust.

I run into this every time I show people what we're building, and I get it. Nobody has ever told me the old way is better. What they're actually saying is that nothing has made the new way click yet. Twenty years of habit doesn't move for a feature list.

Here's what I think the click actually is.

It's not "wow, this is smart". It's the moment you stop double-checking it.

Ask why anyone still scrolls their inbox by hand when they could just hand it to an agent. It isn't that scrolling is better. It's that scrolling is yours. You saw every subject line with your own eyes. That isn't efficiency, that's certainty, and the unpleasantness is the price you're paying for it.

So the thing worth building isn't a smarter tool, and it isn't a better promise. It's one that earns the right to be trusted. One that shows you its work every time, so you can check it, and keep checking it, until one day you notice you've stopped.

I know it's possible, because it already happened to me. Just not with email.

Somewhere in the last year I stopped opening the diff and reading every line my coding agent changed. I don't remember deciding to. It had shown me its work enough times that checking stopped paying, and one day I noticed I'd stopped.

Nobody has done that for my inbox yet. That's the bar we set for #Pingbo.

Has a tool ever gotten there for you? What made you stop checking?

Full conversation. The email part is around 6:40.
https://t.co/WMgUBDB4wX

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

In the last two months, me and my friend @pinglin02 have been building an project, Pingbo, together, and we have iterated our way into one setup for project management.

The road there was longer: a ten-person team on the sleekest SaaS, then a two-person board run entirely by an agent. We dropped both, for very different reasons, and ended up on plain GitHub issues. I think that is where this is all going.

https://t.co/h2NziV5dyK

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

@rohanpaul_ai @pinglin02 Looks like yet another framework to add in your post: https://t.co/KhuhMkdduj

[Ping-Lin Chang @pinglin02]

I went down the agent-memory rabbit hole.

My takeaway: structured stores beat files on accuracy and token cost in my evaluation, but benchmark choice changed the ranking.

For long agentic tasks, the frontier may be training models to use experience.

https://t.co/GY6u61bLwl

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

@0xCodila Harness Engineering has pretty much covered all even for loop and graph. Components to harness can be really diverse, among which the agent memory has been already a rabbit hole: https://t.co/KhuhMkdduj

[Ping-Lin Chang @pinglin02]

I went down the agent-memory rabbit hole.

My takeaway: structured stores beat files on accuracy and token cost in my evaluation, but benchmark choice changed the ranking.

For long agentic tasks, the frontier may be training models to use experience.

https://t.co/GY6u61bLwl

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

@beamnxw This post has the same conclusion. Long-term agentic task is much more complex requiring even RL-based memory framework: https://t.co/KhuhMkdduj

[Ping-Lin Chang @pinglin02]

I went down the agent-memory rabbit hole.

My takeaway: structured stores beat files on accuracy and token cost in my evaluation, but benchmark choice changed the ranking.

For long agentic tasks, the frontier may be training models to use experience.

https://t.co/GY6u61bLwl

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

@msjiaozhu 這篇文章可以補全不同的 memory 架構對於不同 dataset 和長 agentic 任務的記憶表現：https://t.co/JjSpVa47m7

[Ping-Lin Chang @pinglin02]

I went down the agent-memory rabbit hole.

My takeaway: structured stores beat files on accuracy and token cost in my evaluation, but benchmark choice changed the ranking.

For long agentic tasks, the frontier may be training models to use experience.

https://t.co/GY6u61bLwl

[Ping-Lin Chang @pinglin02]

I went down the agent-memory rabbit hole.

My takeaway: structured stores beat files on accuracy and token cost in my evaluation, but benchmark choice changed the ranking.

For long agentic tasks, the frontier may be training models to use experience.

https://t.co/GY6u61bLwl

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Even though LLMs are great at summarizing and extracting key points, I still trace back to the raw files.

Not because I don’t trust AI.

Because:

* I want the original context
* I want to avoid hallucinations
* I want first-hand information

With videos, I often jump directly to a specific moment to see exactly what was said.

Not just what the summary says.

How it was actually said.

That makes a huge difference.

Understanding doesn’t come from summaries.

It comes from the source.

For me, AI is best used to:

* help me find the right places
* help me process faster

Not replace the original content.

Next post: how I turn this into something reusable and shareable.

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Most people stop at summaries.

I think that’s where real understanding begins.

In my last post, I talked about saving content instead of just consuming it.

But the real next step is:

How do you go back and use it?

I almost always return to the original source.

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

After seeing Andrej Karpathy’s idea of personal knowledge base (llm-wiki), I’ve had a stronger realization:

In the end, we’re all converging to the same place.

Before, when I came across something interesting — whether it was a document, image, or video (especially long, information-dense videos) — I would usually just watch it and move on.

Mostly because there wasn’t a good, frictionless way to organize it into a knowledge base.
But now it’s becoming a very natural action:

👉 See → save it for later

And that’s just the first step.

After that, a few things tend to happen naturally:
1. Q&A and reprocessing the content
2. Going back to the original source to understand it more deeply
3. Sometimes turning it into something shareable

But the real shift isn’t these actions themselves.
It’s this: the information doesn’t disappear anymore — it stays, and starts to compound.

The future difference won’t be who writes better prompts.
It will be: who builds a knowledge base that keeps growing over time.

In my next post, I’ll share how I go back to the source, do Q&A, and turn it into output.

[Ping-Lin Chang @pinglin02]

🚨 Instill Agent is now in private beta!

Building multi-agent LLM systems? Observability is key — and hard.

We baked OpenTelemetry + Grafana into Instill Core (https://t.co/Wh8Ys4Yi5m) to bring built-in tracing, logging & metrics.

Join the waitlist 👉 https://t.co/IpfXagTqnC
Release 👉 https://t.co/nFS5Y5Wr6j

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

LLMs lack real opinions—no self-awareness, subjectivity, or growth. Their “insights” are just patterns & context guesses, not true understanding. Even with perfect memory, they’d mimic, not judge.

So, without needing LLMs to have “souls,” what memory features can we build?
- How do they spot key insights vs. noise in a chat?
- How can long chats build deeper, abstract memories?

Tried marathon chats with an LLM? Share your discoveries! #AIThoughts #FutureOfAI

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Why do ChatGPT, Grok, & Perplexity all have a “New Chat” design? Is it necessary or just lazy?

Let’s dive into LLMs & memory! #AI #LLM

[Andrej Karpathy @karpathy]

When working with LLMs I am used to starting "New Conversation" for each request.

But there is also the polar opposite approach of keeping one giant conversation going forever. The standard approach can still choose to use a Memory tool to write things down in between conversations (e.g. ChatGPT does so), so the "One Thread" approach can be seen as the extreme special case of using memory always and for everything.

The other day I've come across someone saying that their conversation with Grok (which was free to them at the time) has now grown way too long for them to switch to ChatGPT. i.e. it functions like a moat hah.

LLMs are rapidly growing in the allowed maximum context length *in principle*, and it's clear that this might allow the LLM to have a lot more context and knowledge of you, but there are some caveats. Few of the major ones as an example:

- Speed. A giant context window will cost more compute and will be slower.
- Ability. Just because you can feed in all those tokens doesn't mean that they can also be manipulated effectively by the LLM's attention and its in-context-learning mechanism for problem solving (the simplest demonstration is the "needle in the haystack" eval).
- Signal to noise. Too many tokens fighting for attention may *decrease* performance due to being too "distracting", diffusing attention too broadly and decreasing a signal to noise ratio in the features.
- Data; i.e. train - test data mismatch. Most of the training data in the finetuning conversation is likely ~short. Indeed, a large fraction of it in academic datasets is often single-turn (one single question -> answer). One giant conversation forces the LLM into a new data distribution it hasn't seen that much of during training. This is in large part because...
- Data labeling. Keep in mind that LLMs still primarily and quite fundamentally rely on human supervision. A human labeler (or an engineer) can understand a short conversation and write optimal responses or rank them, or inspect whether an LLM judge is getting things right. But things grind to a halt with giant conversations. Who is supposed to write or inspect an alleged "optimal response" for a conversation of a few hundred thousand tokens?

Certainly, it's not clear if an LLM should have a "New Conversation" button at all in the long run. It feels a bit like an internal implementation detail that is surfaced to the user for developer convenience and for the time being. And that the right solution is a very well-implemented memory feature, along the lines of active, agentic context management. Something I haven't really seen at all so far.

Anyway curious to poll if people have tried One Thread and what the word is.

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Andrej says “New Chat” forces LLMs to clear their context to patch these flaws—it’s a compromise for tech limits. But we need better memory, not resets. I’m all in 🙌

Think about a friend who’s known you 3 years—they reflect, grow, & evolve from your chats. LLMs? They might recall bits—your name, pet’s name, projects—but don’t truly “get” you as a whole. Why?
#Innovation #AIResearch

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

AI for Legal!

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

@instill_tech Couldn’t stand this anymore—Instill AI's beta is my fix. You relate?

[Xiaofei Du 杜笑菲 🦄 @xiaofeidu283]

Fed up with ChatGPT’s one-off analysis or NotebookLM’s tiny scope? Big docs, structure, deep insights—brutal hurdle. Built @instill_tech's Early Access beta to turn PDFs/Word/PPT into clear tables for easy analysis. Who’s stuck too? DM me.