Ding

Quant · ex-Head of Data Science @ Sequoia China

Sharing thoughts on design, code, and the poetic life as an entropy slayer.

[Ding @0xDing]

@blackanger 我们也做了个类似的产品 同样围绕着人类和 Agent 在一个工作空间内如何协作的课题来延展 欢迎关注 https://t.co/Be0157Uu0s

[Ding @0xDing]

@claudeai Here is a truly open-source and working version, featuring a long-term memory system designed for organizations rather than individuals. https://t.co/AqWtqBJRsW

[Ding @0xDing]

@9hills 欢迎参考我们的开源实现 https://t.co/1jES3gLuMm

[Ding @0xDing]

@jimclydego @ataiiam Here is a truly open-source and working version, featuring a long-term memory system designed for organizations rather than individuals. https://t.co/AqWtqBJRsW

[Ding @0xDing]

The three problems you hit — the agent dies with your laptop, you pay when teammates use it, no real cron — are runtime problems, not interface problems. Buzz nails the interface half. I've been building the other half: Ankole, an open-source, self-hosted OS where agents run 24/7 on your own server. Jobs run for hours or days, survive restarts, and report back to the channel. Agents join Slack/Teams/Lark under their own identity, on the team's model credentials — so usage bills the team, not whoever wired the bot up. Cron-style schedules and per-agent skills, tools, and authority boundaries are built in.

Try https://t.co/WfLOadg0b0

[Ding @0xDing]

everyone asks what 5.6 sol can build. we asked what kind of company becomes possible once models like it exist. Ankole is our working answer: people decide what matters; models own the work. when intelligence is abundant, headcount stops measuring company size. https://t.co/tsmO8YQgyV

[giatt @1owroller]

girlfriend overheard me and started yelling at me asking me “who the fuck is kimmy” “why the fuck are you talking about chinese models”

[Ding @0xDing]

We built @AgentBull/ankole, an open-source Agent OS, with Claude Code (Fable 5) and Codex (GPT-5.6 Sol). Both are strong, but GPT-5.6 Sol performed better on the same cross-stack codebase—especially at root-cause analysis and verification. https://t.co/AqWtqBJRsW

[Ding @0xDing]

@TooooooBug 欢迎试试我们的项目 在飞书数字员工场景下比 hermes 好很多 而且内置 codex 后台任务能力。https://t.co/AqWtqBJRsW

[Ding @0xDing]

Claude Tag just dropped for teams. What if you could run the same thing — fully self-hosted and open source? Ankole is an open-source AgentOS for shared AI colleagues. Multi-agent collaboration with real identities, persistent memory, tools, and long-running workflows — all on your own infrastructure. No vendor lock-in. Full control. https://t.co/GMX8rei9F5

[Ding @0xDing]

@dotey 我们在做1个相近的开源版。从“数字助理”转变为“数字员工”，这块最大的区别在于对群里消息的主动感知能力。 https://t.co/ucgcOBE9sq

[Ding @0xDing]

@QuantumTransf @yinmin1987 欢迎参考我们的版本的思路https://t.co/pjh9MRowyK 里面的“ambient”的逻辑就是。其中当`unaddressed_group_message_policy = may_intervene`时则会有主动插话的功能。本质上是基于意图识别去做

[Ding @0xDing]

Summarization is lossy compression.
The hard part is not fluency. It is the compression policy: what to keep, drop, and rank.
That policy is basically my private utility function—subjective, unstated, and wildly personal.

[Ding @0xDing]

“Real AI employees” may be easier to build than “AI assistants.”

Not because real work is easy.

Because real work often has verifiable goals.

[Ding @0xDing]

“Summarize my unread messages” looks trivial, but it is almost the worst LLM task.
The real target is not summary.
It is: decide what matters to me.

[Ding @0xDing]

@hwwaanng 或许elixir才是正解 https://t.co/iUWvxyQQhE

[Ding @0xDing]

This also changes architecture.

The moat is less “we chose the perfect language/framework” and more:

- behavioral oracles
- test suites
- audit trails
- rollback paths
- permissions
- trajectory data

Code can be rewritten. Verified intent compounds.

[Ding @0xDing]

A thought I’m not fully sure about:

After “person-months,” software velocity may become:

Verified Agentic Delta / Human Attention Hour / $

In a post-LLM world, code is cheap. The bottleneck is: what changed, was it verified, who approved it, and how much attention did it take?

[Ding @0xDing]

Maybe the formula is:

Σ(Business Value × Verification Confidence × Risk Discount)
/
(Human Attention Hours + Agent Cost + Calendar Latency)

Ugly, but closer to reality than LOC or “number of PRs.”

[Ding @0xDing]

In the old world, software was constrained by typing, debugging, and migration labor.

In 2026, large parts of that loop are becoming agentic.

So the scarce input shifts from engineering labor to judgment, verification, context, and risk ownership.

[Ding @0xDing]

“Agent-hour” feels like the wrong unit.

It measures activity, not accepted change.

The useful unit is a delta that survived tests, review, audit, risk checks, and maybe production.

Otherwise we’re just measuring how fast agents can create work for humans.