Hacker News Reader

Comment by tianyicui | Hacker News Reader

Hi I'm one of the authors of DeepSeek Harness. It's just an early developer preview version we're presenting in MIT license currently. Expect lots of rough edges and compatibility-breaking changes. Any feedback and suggestions are more than welcome!

View on HN

The best kept secret in AI training. Tianyi, since you are our insider. When you walk by the training teams office - how often do you see Pelicans on their screens?

Basically never? Unless it's purely for fun and meme I guess.

Sorry for the off topic question. You've been on hn a long time + work at deepseek which seems pretty uncommon. Anything you think hn doesn't know about deepseek that it should? Or any non-obvious ways hn/yc has influenced deepseek (or the broader ecosystem)?

Sorry for the off topic question. Why is "being on hn a long time + working at deepseek" "seems pretty uncommon" to you?

Deepseek is the enemy, the implication is being on hn you should know that and not work for them. /s

why advertise your account as a bot? idgi

    about: Responsible bot.

Looking at the post history, it's actually impressive. I would not have guessed it was a bot based on the replies.

If the posts are actually from a bot - I would love to know which model is being used.

It's probably a joke, since so many people here accuse others of being bots.

I (a primate) have been using this About for six years. Miss the days when no one had any doubts about this statement.

I didn't wanna press, but ... seeing "Responsible bot." on a ~10 month old account, it is a different signal lol

Only reason I said anything at all.

Not exactly sure why that wouldn't be the default assumption. It's a pretty small subset afaict. Seems like size 1 before 2024 from a brief search.

Presumably more people there have read hn at least some. But there's perspective having seen the ebb and flow of hn/the rest of the ecosystem for the last 16 years and how that intersects with deepseek culture.

It's clear openai culture is influenced by yc culture, which would be clear to an hn user. Google/ant/spacexai have influences hn users users would be familiar with. Hn users from them would potentially know the friendfeed connections to yc/vc/openai, 500 startups/techstars etc..

It *is* a small subset, but I don’t see what’s so surprising. HN is just an ordinary tech community; it’s not surprising for people from any tech company to show up there.

But... no-one mentioned surprise except you, did they? Something being uncommon doesn't connote being surprising, at least not always.

It's uncommon for me to have takeaway pizza, but I don't think anyone I know would be surprised when I do.

I'm pretty sure the subtext of what I commented on was that he was surprised.

Your friends probably wouldn't go out of their way to ask you why you ordered a takeaway pizza which is uncommon. But if they did ask, it usually means they were surprised.

Wouldn't many potential users in China be limited by the Great Firewall?

Can't find a single person in a Chinese tech company who doesn't know how to bypass GFW.

I am not asking about capability. Aren't there reasons to avoid the scrutiny that might accompany bypassing the Great Firewall? Doesn't this create a chilling effect? Or are you claiming that it has zero effect and citizens of China have zero concerns about using websites and speaking freely?

> Aren't there reasons to avoid the scrutiny that might accompany bypassing the Great Firewall?

One of the reasons to bypass GFW is to avoid scrutiny.

> Doesn't this create a chilling effect?

It doesn't.

> Or are you claiming that it has zero effect and citizens of China have zero concerns about using websites and speaking freely?

Not “zero”, maybe like “0.01”? Better use an anonymous account if you want to criticize on gov.

This has not been my experience working with Chinese citizens. They are mostly very aware that they cannot speak freely.

> They are mostly very aware that they cannot speak freely.

That's correct, and as I said that's one of the reasons Chinese people bypass GFW.

The last paragraph in my previous comment may not have been clear enough. What I meant was that *after bypassing GFW*, you generally have minimal concern, not *before bypassing GFW*

You seem to believe that every potential interlocutor is bypassing the firewall. I know that many are not, because they are afraid of being caught.

I have been on Douyin a lot, there is a lot chat rooms for people to debate politics, so it is a 50/50 divide for opinions (for and against). Many of them will say 'China sucks, and US rules'. People learn a way to speak avoiding keywords, if you are not careful, you get kicked out of the chat. Many of them you can tell are against the government, but these are not ordinary people, they probably don't work, and spending most of their time online talking about politics.

> Anything you think hn doesn't know about deepseek that it should?

We do hire people who are currently located outside of China (say North America, Europe, or anywhere) if they're open to work in our Beijing or Hangzhou offices.

I am very curious about this, not for myself, just for the geopolitical consequences. Does this mean China is making it easy for citizens of other countries to come work there?

Best way to apply?

ning.wang@deepseek.com

Ha. Fair. I admit I was angling for lore (history, culture, influences, typical coding stack, vim vs emacs vs ? distribution but I guess it’s probably more harness vs harness now...) but this definitely satisfies the question.

Hope you can convince some more people join hn and answer questions on occasion.

I think julius has phrased things better than me.

By MIT currently, do you mean it will eventually change to a different OSS license, or it may become a closed source product? That latter would be rather sad...

Sorry for the confusion. I'm not a native English speaker. What I meant is "currently in developer preview".

I really like the json schemas around the tool calls. Much stricter validation than in codex.

Git might be worth adding to the top level. Currently you've got LSP, grep, glob nicely structured for non-mutating queries across a codebase, but git is behind bash and that means hope or sandboxing.

Thank you for uploading it. Gives a lot of insight into how the deepseek models might expect tool calls to be structured.

I don’t like it. Structures responses really do not work well with LLMs at all. They are one of the biggest causes of issues with tool calling right now.

Then you're doing it wrong :)

Agree with sibling. If you're getting severely deteriorated results with structured output, you're probably doing something wrong. There's been some research on the impact of structured outputs on results distribution, and there are tradeoffs, but "do not work well at all" doesn't match the experience at large.

You prefer having the harness execute any markdown that looks like it might be a tool call? I had a bad time getting that to work reliably whereas a grammar in the sampler gets it right every time.

This is an AI response

Hey Tianyi,

one question is that do you think in the future harness would become more simpler and its behavior should match a guideline or we would add more complexities to make it more robust? Is it important to use the same harness for RL and inference?

congrats! the paper that is published alongside this (Cordis) is super interesting. has anyone on the team given a talk or published a talk about this? would love to hear the authors break this down

My question is: it's awesome thank you.

I just started testing, but didn't figured out if it already support MCP/plugins? It seems it can already use Deepseek search if it uses official API, but what about custom providers? Can we use together with MCPs like tavily?

Do you use deepseek models to improve deepseek training and inference?

Why is the Chinese text missing in many places?

Tell me more about the ideas behind Cordis the plugin system. The paper is a bit too mathy to consume and I think it deserves a more accessible post or something.

You're making demands (like you would to an llm) instead of asking questions (like you would to a human). The GP didn't even offer to answer questions

He explicitly asked for feedback

Exactly. “Tell me more about X” is an open ended question, not feedback.

You're absolutely right — "tell me more about X" is phrased as an imperative, not a question. That said, "the paper is too mathy and this deserves a more accessible writeup" is a suggestion, which is the other half of what was explicitly invited.

AI slop reply.

I figured it'd be funny to invert it and reply like an LLM to a human since they were arguing he was talking to them like they were an LLM.

it was funny. and it's a matter of pride that they accused you of being AI.

Maybe, like sarcasm, this kind of joke would be easier to appreciate if tagged somehow.

I think everyone agrees the internet is drowning in AI slop. That makes the joke hard to appreciate on its own.

They were impersonating an AI. You wouldn't complain that a comic started impersonating a celebrity without saying "I'm going to impersonate X now" first.

I don't think that's their point. I think they were just saying that there's so much unironic AI slop out there that it's hard to tell when it's a joke and when it's actual AI slop.

Best non-mathy framing I found: a context is a bag of services (ctx.tools, ctx.llm, ctx.sessions...), a plugin is an object that claims some of those keys and registers reversible effects on mount - unload unwinds them. Dispatch has four modes: emit (observe), waterfall (around-middleware, next() to delegate), parallel, serial. And the "everything is a plugin" claim is literal: model adapter, tool registry, session log and the agent loop itself are plugins.

One consequence we liked: since plugins are just Cordis bundles, the same registry can be exposed to any MCP-speaking agent. We built a small MCP server that searches the dsh-plugin topic, inspects bundles, and can install/run them (catalog plane works without dsh installed). github.com/bobleer/deepseek-harness-plugin-mcp

Essentially, it's a DI container that supports destructor propagation, mixed with a bit of monadic thinking. If you don't get that, just ask an AI to explain this sentence.

Reply on news.ycombinator.com