Kun Chen

568 posts

Kun Chen banner
Kun Chen

Kun Chen

@kunchenguid

Lead Principal Eng @ Atlassian · Rovo Dev · Agentic Products

Katılım Eylül 2021
111 Takip Edilen490 Takipçiler
Kun Chen
Kun Chen@kunchenguid·
@0xSero It’s just marketing expense. As long as some of these power users keep talking about their brand, they will gain mindshare and earn money from other high margin products. Their constraint is not the financial loss from power users, it’s how quickly they can bring GPUs online.
English
0
0
0
78
0xSero
0xSero@0xSero·
I said this many time before and I kept getting ridiculed. If you use any coding agent more than an hour a day you are losing these companies money. Go look at how much 8x B200 would cost per hour. The true cost of a Token, is AI profitable? youtu.be/oEImds1Rocw
YouTube video
YouTube
Logos Tech - Testnet v0.1 now live@Logos_devs

If you're an AI power user, get prepared. As @0xSero says, the math for subs doesn't add up. $200/mo can't cover the 16+ GPUs needed for power users. So what happens next? → Prices go up → Limits get tighter → Your data becomes the product But there’s another path emerging. Open-source models are coming out with incredible efficiency savings. Local inference is improving. Home hardware is becoming viable. The frontier is shifting from access → ownership. Start building your sovereignty now.

English
72
30
440
61.4K
Kun Chen
Kun Chen@kunchenguid·
I will likely continue studying this topic given how much potential there is. Follow me if you'd like to see what I find out next!
English
0
0
0
71
Kun Chen
Kun Chen@kunchenguid·
Finding 5 - Shell pipes give AXI an advantage MCP cannot match. With CLI and AXI, agents can compose commands with unix pipes: chrome-devtools-axi open <url> 2>&1 | grep -i "designer" navigates and extracts in a single command. MCP tools have no equivalent.
English
1
0
0
78
Kun Chen
Kun Chen@kunchenguid·
Finding 4 - MCP Compressor (open sourced by Atlassian at github.com/atlassian-labs…) is the strongest non-AXI interface. Wrapping chrome-devtools-mcp with the compressor leads to a CLI that automatically performs better than agent-browser.
English
1
0
2
75
Kun Chen
Kun Chen@kunchenguid·
Finding 2 - "Tool Search" (which is auto-enabled in Claude Code when there's a large number of tools) drops task success rate by 15% compared to raw MCP, without significant cost saving benefit. @bcherny @trq212 may be worth revisiting when to auto-enable?
English
1
0
0
71
Kun Chen
Kun Chen@kunchenguid·
Finding 3 - "Code mode" is competitive but 58% slower. The agent spends time writing TypeScript scripts, and each npx tsx invocation has startup overhead. At $0.177/task, it is the most expensive MCP condition.
English
1
0
0
59
Kun Chen
Kun Chen@kunchenguid·
Finding 1 - agent-ergonomic interface design beats everything. AXI-optimized interfaces lead on every metric. Just by adjusting the interface, an AXI wrapper of agent-browser could save 15%+ on token cost without any sacrifice on success rate. @ctatedev might be interesting to u
English
1
0
0
75
Kun Chen
Kun Chen@kunchenguid·
This is a continuation of my previous study published in this blog post kunchenguid.medium.com/i-benchmarked-… To do cross-domain validation, I extended the benchmark to evaluate browser automation in addition to Github tools.
English
1
0
0
125
Kun Chen
Kun Chen@kunchenguid·
Ok my MCP vs CLI vs AXI benchmark just went from "interesting" to "holy crap people really need to know this". I evaluated many ways to do browser automation in agents with 985 task runs - surprising insights published in axi.md Key findings in this thread 👇
English
2
0
5
148
Kun Chen
Kun Chen@kunchenguid·
@itamar_mar oh i didn’t mean “every human”, but on average yeah humans suffer from cognitive dissonance and agents don’t
English
0
0
0
18
Itamar Friedman
Itamar Friedman@itamar_mar·
@kunchenguid Fair… although humans aren’t all the same 🙂 Are you seeing coding agents backtrack faster than humans (in avg.) when they’re wrong?
English
1
0
0
23
Karan Vaidya
Karan Vaidya@KaranVaidya6·
Okay, @gdb is team CLI all the way. @garrytan thinks MCPs suck. So we hit the streets of SF to see if the city agreed. We posed a simple question: MCP or CLI? - Basically everyone under the age of 35 said CLI - One person said MCP was as bloated as Java - & unsurprisingly, numerous people told us to touch grass Final score- MCP: 3 vs CLI: 17 SF has spoken, and @composio listened. Our universal CLI is now live! Drop your best CLI vs MCP hot take in the comments and we'll send the best ones some very sick gear 👀 Link to try our CLI in the next thread ⬇️
English
113
435
889
1.9M
Kun Chen
Kun Chen@kunchenguid·
@rauchg many people started building CLIs but are doing them in the same way they built CLIs for humans. CLIs for agents are fundamentally a different thing - I call them AXIs. I proved in axi.md that a principled approach can achieve higher success with lower cost.
English
1
0
2
716
Guillermo Rauch
Guillermo Rauch@rauchg·
1961: We should ship a CLI 2026: We should ship a CLI
Guillermo Rauch tweet media
English
161
220
3K
121.1K
Kun Chen
Kun Chen@kunchenguid·
@DanielMiessler interesting point! wouldn’t there then be “agent-oriented marketing”? whatever system controls that selection process will be gamed to death, right?
English
0
0
1
131
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·
To some degree yes, but there are about to be directories of all products and services for agents to use. So when a principal (you or me) wants or needs something, our agent will check the rating services for the best option across multiple dimensions. I think marketing will still exist, but it will have a lot less impact when agents become the filter between us and millions of options.
GREG ISENBERG@gregisenberg

now that we all can build anything with ai we're going to all have to figure out distribution the wealthiest people will be marketers over the next 10 years

English
5
3
21
3.5K
Kun Chen
Kun Chen@kunchenguid·
@derekcheng Master class of comms tricks that almost worked
English
0
0
1
31
Kun Chen
Kun Chen@kunchenguid·
@yeroneem oh hi Ivan! :) I think the session history can allow others to - 1) see what context was used 2) interrogate the session and ask questions 3) fork the session to do more work on top it doesn’t need to be displayed prominently, but having it as a record is quite useful
English
0
0
0
15
Kun Chen
Kun Chen@kunchenguid·
"AI as co-author" - 1. source controls should have a way for ppl to say "I made this change with AI" so others know what to expect 2. full agent session should be attached 3. agent name matters less. what prompt, model, how many tokens were used matters way more 4. don't make agents look like humans on the UI. agents are just a tool. when a change breaks production, it's always a human held accountable
Yuchen Jin@Yuchenj_UW

Claude Code auto-adds itself as a co-author on every Git commit. Codex doesn’t. It reflects a deeper philosophical divide: Is AI a tool, or a teammate? Anthropic leans hard into the latter, from Claude’s Constitution to giving “Claude” a human name and mascot. It’s been a great growth hack, Claude is everywhere on GitHub. OpenAI seems to be internally debating hard whether to go that route.

English
1
0
4
508
Yuchen Jin
Yuchen Jin@Yuchenj_UW·
hey friends! 👋 Only cool people are allowed to reply to this tweet obviously.
English
264
3
479
49.1K
Kun Chen
Kun Chen@kunchenguid·
@willmcgugan thanks Will! I’m curious how did you find that audience - did people organically discover your work, or did it spread through specific communities?
English
1
0
0
39
Will McGugan
Will McGugan@willmcgugan·
@kunchenguid They were hobby project to satisfy a personal need. I shared what I was working on and found an audience with the same needs. Scaling requires focusing on others requirements rather than my own. But still sharing my work publicly.
English
1
0
1
119
Will McGugan
Will McGugan@willmcgugan·
I’m spending the next few hours in a car. Figured it would be a good time to do an AMA. To keep me entertained… Ask me anything! I’m a software engineer of 30+ experience. In addition to tech, I know a fair bit about nature and photography. But feel free to challenge me on any subject. Currently in Thailand. Apologies to followers outside of Asia, who may sleep though this. willmcgugan.github.io github.com/willmcgugan/
English
14
3
17
3.2K
Kun Chen
Kun Chen@kunchenguid·
@levelsio i couldn’t - promise i’m a human
English
0
0
0
102

Keşfet

@0xSero @bcherny @trq212 @ctatedev @itamar_mar @KaranVaidya6 @gdb @garrytan