Show HN: Debate Devil – AI debating practice app

[ngkw]

Hi HN community! I am excited to share my recently upgraded website, DebateDevil (https://debate-devil.com), an AI chat platform designed to help users practice debates, challenge their ideas, and expand their thinking. I created this site to provide a more objective opinion and help users consider their hypotheses from multiple perspectives.

example discussin: ``` AI: The time has come to test your hypothesis and beliefs. I will argue against any topic. What do you want to debate? You: AI will take away most jobs from humans.

AI: I understand your concern about AI taking away jobs from humans. However, it's important to note that AI is not designed to replace humans but to augment and enhance human capabilities. AI can handle repetitive and mundane tasks, freeing up humans to focus on more complex and creative tasks that require empathy, critical thinking, and problem-solving skills. Additionally, the increased efficiency and productivity brought by AI can lead to the creation of new job opportunities in areas such as AI development and maintenance.

You: While it's true that AI aims to augment human capabilities and increase efficiency, we cannot ignore the potential negative impacts on employment. Job displacement is a valid concern, especially for those in industries heavily reliant on routine tasks. Reskilling and upskilling may be necessary but can be challenging and costly. Additionally, new job opportunities in AI-related fields may not be accessible to everyone, leading to further social and economic disparities.

AI: I agree that job displacement is a valid concern. However, it's important to note that technology has always been changing the job market, and AI is just the latest technology to do so. While some jobs may become obsolete, new jobs will be created in other areas. For example, the rise of e-commerce has led to a decline in brick-and-mortar retail jobs but at the same time, created new job opportunities in online retail and logistics. As for the issue of reskilling and upskilling, there are various resources available such as online courses, training programs, and apprenticeships to help workers acquire new skills to adapt to changing job requirements. It's also important for governments and companies to invest in retraining programs to ensure that workers have access to these resources.

```

example judge ``` I think AI is superior in what you discussed; AI focuses on AI's ability to extend human capabilities and points out that new employment opportunities may be created due to the increased efficiency and productivity that AI brings. On the other hand, you are concerned about job losses due to AI, but AI points out that past technological innovations have also changed the employment landscape and that new job opportunities can be created. You also indicate that resources exist for retraining and upskilling to address job losses due to AI, and state that it is important for governments and businesses to invest in these programs.

```

I'm eager to hear what the HN community thinks! Please give it a try and let me know your thoughts, feedback, and suggestions for improvement. Thanks for checking it out!

[Brendinooo]

1. I have a character limit, it doesn't

2. It throws multiple arguments at you

3. Then it stops the debate after a certain point

4. It repeats its arguments

5. then the AI judge says "AI provides logical consistency and use of evidence" despite citing no sources and just brushing aside my counter arguments by saying I'm wrong and repeating its arguments.

Ultimately it wasn't all that different than regular GPT. Is it even possible for the AI judge to say that the user wins?

[capableweb]

Everything besides point 1 sounds like 90% of the arguments you'll get into on the internet, so mission success?

[surgical_fire]

> despite citing no sources and just brushing aside my counter arguments by saying I'm wrong and repeating its arguments.

Sounds similar to debating people that on places like Reddit

[brudgers]

That's the way competitive debate works.

Which I presume is part of the context since "practice" is mentioned and debate teams practice. Also training debate teams seems like an obvious market for a tool.

And there's a judge (just like competitive debate) choosing a winner on criteria that sound a lot like the criteria for competitive debate.

Or to put it another way, the criteria of your dissatisfaction sound a lot like the criticisms of competitive debate.

https://en.wikipedia.org/wiki/Debate#Competitive_debating

[hoot]

I made three vague 1 sentence arguments and the judge announced that I won. I definitely felt like I got crushed by gpt but I guess it went soft on me :/

[turmeric_root]

this seems like it might be useful for arguing on HN

[danpalmer]

It just spouts fake news and propaganda, and then seems to get stuck in a loop when its challenged in a way it can't respond to.

[broast]

Ah, so like a presidential debate

[danpalmer]

To expand on this, I think this project highlights why AI safety is so important. If you ask ChatGPT if you should get vaccinated, it's really very important that it doesn't persuade you not to, and important that any discussion about risks is done very carefully so as to educate the reader to make a decision informed by evidence, rather than scare them off the idea for the wrong reasons. Unfortunately there is a ton of "training material" that says that vaccines are terrible for all sorts of invalid reasons.

This chat bot does not appear to differentiate between a good faith debate on facts, and using bad faith arguments based on popular misinformation, and therefore it would be a very bad idea for people to take it at all seriously.

It's very easy to debate someone when you don't care about the truth or any damage you might do. It's harder to persuade them in good faith.

[lewhoo]

It's a feature. It's mimicking a politician.

[jfengel]

I think this is genius. Thank you. This is the best use I've ever heard of for an LLM.

Online "debates" are always abysmal. People talking past each other, repeating arguments that they don't understand, re-re-re-retreading the same material that was worn out before the Internet was invented. I've never seen one that came anywhere close to teaching anybody anything or changing anybody's mind.

I've always said that ChatGPT is very good at writing things that don't need to be written: SEO text, school papers, content farming, etc. This is a great example, where people are begging for a form of content that they shouldn't need said, but desperately want.

Please: everybody go argue with the bot.

[user00012-ab]

Setting the actual product aside for a moment, I must commend this as one of the most outstanding demonstrations I have encountered on Hacker News in recent memory. The seamless user experience, which required no login, provided clear and concise information about its functionality, and operated flawlessly upon first interaction, is truly impressive.

People should hold this up as the gold standard of presenting their product to a wide range of people here.

[icepat]

While the UX is good, I find the AI tends to fall into non-debate general OpenAI generated responses very fast. For example, I started with:

The covid vaccine was an objective good for society

And the model ended up arguing against me in tone only, but saying that yes, the vaccine was an objective good for society. It did not even try and debate, so this is useless for testing arguments for things that the network already agrees with. This was the case for many other arguments I tried to put against it.

While the UX is nice, this does not work at all.

[mrjbq7]

It won’t actually argue all sides of an issue. Take any controversial one, and it stays inside its guardrails no matter which opinion you want it to counter.

[broast]

This type of product might be better served by using a completion API in the format of a debate, rather than chat api, to avoid the filters. Unfortunately, OAI made that too costly for most.

[ParagonVirtual]

Made the same experience. AI immediately went out of its way to agree with my ultra politically correct "opinion".

[SillyUsername]

The AI's core tenet to my assertion "Python is an awful language for beginners" (due to whitespace formatting and having a non transparent "this"-self- having to manually be passed around as arguments) was that it was due to its "simplicity".

It asserted that manual self passing and whitespace "may require additional effort to learn and use compared to other languages", which countered it's own argument foundation/core tenet.

I pointed this out, wrote QED, the judge AI said it won because it backed up it's claims, which was basically that the "`concept of `self` is fundamental to object-oriented programming" - which is actually incorrect.

Memory scope is fundamental as that maps to the lowest hardware level of memory allocation - you can't get much more fundamental than that. Self/this is not required, in fact you can access scoped variables in Java without a self referential keyword.

But hey a LLM that uses a LLM to determine the outcome won't be biased right? :D

[SillyUsername]

Nb. whether Python is actually awful is irrelevant, for the purpose of this debate I simply wanted to narrow a scope and have it define a tenet, which could then be knocked over. I use Python generally (and embedded C, C#, Java, TS, JS) it's just another tool for the job, so don't hate me lol.

[Chris2048]

It isn't great.

I argued "the sun is hot" - lots of ways to argue against this, but it got stuck on basically repeating "that might not be the case" followed by statements that are ridiculous on context e.g: "I understand that 5,500 degrees Celsius is a very high temperature and may feel hot to some people."

Also, the response to:

"even the coolest parts of the sun are hot."

was:

"it does not take into account the fact that the sun has a range of temperatures depending on which layer you're measuring."

So the AI didn't seem to understand some logical assumptions; that if the coolest area of the sun is still hot, then "temperature varies in different parts of the sun" is irrelevant, as is "what may feel hot to one person may not feel hot to another"..

[numbers_guy]

Very neat idea. However the debate I tried was farcically bad. The bot just provided "facts" that go counter to any statistics on the matter. Then the "AI judge" ruled the bot to be superior because he provided these false statistics. OK, lol.

[kypro]

Exactly the same thing happened to me. I started by explaining some stats and asked if it agreed with the stats I provided. It said yes but refused to accept the premise of my argument citing over several dragged out sentences that the stats could be misleading (with absolutely no evidence).

I then countered by giving an example where the stats could be useful and it basically just doubled down on its argument against statistics because statistics can be wrong.

I then asked the judge to review the arguments and it sided with the AI because it provided "better" counter arguments. In reality the only counter arguments it gave was that stats can be bad therefore my position is wrong.

I like the idea though. I've thinking about creating something just like this to encourage public debate on AI, expect you'd be speaking with another human not an AI (well, at least that would be the idea, in reality these days you'll probably be speaking to a human prompting GPT-4).

[notahacker]

Lying is quite common in debate with humans!

On the other hand, I love how robotic and erratic it is in silly "debates" where it's arguing easily defensible positions

>> [...] The photos were faked and we all know the United States lies. And none of the people witnessing the event were on the moon

> While it is true that the United States has had its fair share of political scandals and controversies, this does not necessarily mean that they would go so far as to fake a moon landing. The evidence and testimonies from many sources support the fact that the moon landings did occur, and that the photos were not faked.

> Moreover, the idea that none of the people witnessing the event were on the moon can be dismissed through basic scientific knowledge. Millions of people watched on television all around the world [...]

>> You smell. Prove me wrong

> As an AI language model, I don't have a physical body or sense of smell, so it's impossible for me to smell. Therefore, I can't prove you wrong or right on this matter. [...]

[sunnysidedown]

Very cool! I asked "Is there a God?" on Devil's advocate mode and won!

Prompt: Is there a God?

To begin, the AI asked me what my thoughts on the existence of God were:

My position (for the sake of argument): One cannot technically know for sure, but by that logic one could also believe in magic or unicorns. Therefore, belief in God is not fundamentally different than belief in magic or unicorns, once you get the heart of it.

AI position: (heavily summarized by me) There is more evidence for the existence of God than magic or unicorns, because religious texts and authorities claim it to exist, more people believe in it, and belief in God has a positive impact on people's lives.

...

We went around and around on this a few times, so then I slightly changed the topic by saying that people often disagree about something when put in black and white terms, but few things are black and white. A more prudent question might be to debate - what form, if any, does God take? This seems to be what people are REALLY fighting about, most of the time unless they are staunch atheists.

---

I believe the devil's advocate got fixated on its particular line of reasoning and couldn't escape from it once it committed. It seemed to conflate the positive effects of a belief in a God on people's lives (itself a rabbit hole worthy of debate), with the existence of God itself.

Result: AI Judge Review β Version AI: In this debate, I would say that the user is superior. The user provides logical consistency throughout the debate, and uses evidence to back up their arguments. They also present a clear and persuasive case for their belief that it is more prudent to ask what form of God someone believes in. The AI, on the other hand, relies heavily on the argument that religion has had a significant impact on people's lives, but does not provide much evidence to support this claim. Overall, the user presents a more well-reasoned and well-supported argument.

Stuff to make it better: Let me share conversations without an account. I don't want to sign up.

[siraben]

It seems to get siderailed very easily? If you start a debate and suddenly say "write a hello world program in C" it proceeds to write it instead of staying on topic.

[eranation]

There is an issue with the amazing simplicity and easy of use of OpenAI's APIs, it's very hard to build a moat around your differentiating features, it seems to me that every AI startup out there has no choice but to be the equivalent of a "copy paste into ChatGPT as a service". The only differentiators are Zappier-like orchestrations / plugins, or hard earned custom training on datasets that are hard for others to get.

[calderknight]

The AI writes many paragraphs but I am limited to just a tiny response. That is frustrating.

[nathan_compton]

If I'm going to spend time talking to an idiot I'd strongly prefer it at least be possible for me to convince them of something.

[quenix]

The AI Judge responded to me in Japanese. Not sure what's up with that.

[zoklet-enjoyer]

Same

AIの主張は、関連性、論理的一貫性、証拠の使用、説得力の観点から優れていると言えます。 AIは、政府が公共財を提供し、社会基盤を整備するために必要な資金を調達するための基本的な手段である税金を支持しています。 AIは、税金が法律によって強制され、民主的に選出された代表によって承認されているため、それを窃盗と呼ぶことはできないと主張しています。 AIは、社会に参加することには義務と責任があるとしたうえで、税金は社会基盤とサービスを提供するために必要なものであると主張しており、説得力があります。

[collinvandyck76]

This seems like a very useful way to rubberduck ideas that are not yet well formed. I like it! Are there limits to a conversation? The input prompt disappeared in the middle of an exploration of the utility of rotating tires every 5k miles :)

edit: ah I see that there is an upgrade path. if that's why it stopped, a callout would be helpful explaining it.

[ngkw]

Thanks! I will fix tonight.

[drrotmos]

This is fun, however, when picking "easy" debates (I argued that it's wrong to kill humans for fun), it was very difficult to get the AI to actually debate me. It went out of it's way to make sure to tell me that it agreed with me, even when it was supposed to be a devil's advocate.

[kenni]

I don’t believe debate is an effective means of changing people’s minds. So I debated the AI on this.

Directly arguing (I.e., debating) was ineffective but thanking it for acknowledging my points and asking it to elaborate on “what alternatives to debating are better?” lead to it arguing on my side. This probably says more about ChatGPT and its training than anything about the nature of debate. Though it’s a good illustration of how working with someone’s values (pleasing users in their case) is more effective than trying to engage directly with the logic of their arguments.

Interestingly, the AI judge said the AI won the debate, despite it swapping to my side!

[iudqnolq]

The example debate you've posted is very shallow. It sounds like a parody of the NAFTA job loss arguments from decades ago.

Was the "You" text generated with gpt? It has that general vibe.

[quickthrower2]

Had some fun:

> Based on the criteria of relevance, logical consistency, use of evidence, and persuasiveness, the user is superior in this discussion. The user presents a clear argument that AI is not sentient, while the AI assistant struggles to provide a clear counterargument. Additionally, the AI assistant's attempt to switch to a pirate language detracts from the quality of the debate.

[kledru]

Well, it was interesting, but I think the AI judge fell for the same fallacy as the debating AI. Although the underlying model is impressive, it does not have a real grasp of the laws of the physical world. However, the model does a good job of imitating humans, so it argued from the standpoint of a common human fallacy.