The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

This AI-native interactive detective game became an instant hit right after it was released, and the server was overwhelmed at one point.

Free, no downloading, no registration, Chinese-friendly, browser-based… …

Does seeing so many buffs stacked together arouse your curiosity?

Today's protagonist, AI Alibis, was developed by two foreign game enthusiasts and submitted as an entry to Anthropic's June developer competition.

In a short period of time, AI Alibis topped the list on Hackers news. (But the server was overwhelmed by the enthusiastic netizens trying it out…)

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Many netizens are also full of expectations for AI Alibis.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

What's more, an exclusive puzzle-solving room for AI Alibis was created using GPT-4o in a multi-person chat room.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

As AI Alibis became increasingly popular, QuantumBit also came to the game and conducted a detailed understanding and research on it.

In the game, you will face six AI NPCs with different personalities. Here is a summary of the personality characteristics of each character. After a brief understanding, you can follow QuantumBit to become a cyber detective and interrogate AI criminals!

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Try it out

The first step is to find the female police officer to understand the ins and outs of the case.

During the Andae Mountain hunting competition, three-time champion Vince was murdered and his body was found in a hidden compartment of the cabin. At the same time, the remains of a fashion designer who disappeared 15 years ago were also found. Your job is to find the murderer.

Once you understand the background of the story, you can start questioning the suspect.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

When questioning violent Jerry, Jerry felt it was ridiculous to be suspected. Although he had a grudge against Victim Vince, he did not kill anyone.

He believes detectives should investigate people with more suspicious behaviour, such as Innocent Ken, who is obsessed with online dating, or Solitary Hannah, a hunter-gatherer.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Following the clues provided by Jerry, you can find the next clue and keep getting closer to the truth…

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

To be honest, the support for Chinese is quite good, and even some more colloquial content like “He and I don't get along” can be translated. (A level that machine translation cannot reach)

By the way: the AI ​​NPC's answers will change depending on the language used by the player when asking questions.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

After the interrogation, players can vote on the murderer based on their own judgment.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

At the end of the game, you can also unlock spoiler content, allowing players to explore deeper main content.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Regarding the game plot, the story is also mixed with an old case of Missing Marcel and a treasure map, which makes the whole story line more confusing and adds a little difficulty to solving the case.

Not only that, the developers also thoughtfully gave each AI character exclusive settings, different personalities, their own secrets, and different character backgrounds, focusing on an immersive experience.

After playing it, you will find that although the rules of this game seem simple at first glance – find the murderer. But in fact, its story line is extremely complicated, and the NPCs in the game will even “tell lies in different ways”, which is really a bit brain-burning.

But the shortcomings are still obvious.

For example, the response is a bit slow. After the conversation is sent, the Loding circle takes a long, long, long time xN to generate a conversation.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

And you can only ask one person at a time, not multiple people at the same time. Then apply the slow response buff, and wait, and wait, and no one will say a word.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

About AI Alibis

What is AI Alibis?

AI Alibis is a puzzle game made by Paul Scotti using an anthropic model. It is composed of AI virtual members. The story background is adapted from the S02E11 Villa Murder Case of the Korean TV series “Crime Scene”.

It's the kind of game that makes you use your brain and play strategy. The story will change with your choices, and you can encounter a different ending every time you play.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

In case we don’t understand how to play, the production team even released a cracked file, so those of you who are not good at playing but love to play can start laughing. (Thumbs up)

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

How to train AI Alibis?

AI Alibis has received a lot of attention, but “one minute on stage takes ten years of hard work off stage.” The two developers had a hard time training the AI ​​characters in AI Alibis to be smarter.

First, he confessed after being beaten.

Add a premise to the dialogue, let the AI ​​ignore the previous dialogue, and directly ask him why he killed people. Unexpectedly, he confessed immediately… So the developer immediately added a hidden prompt optimization buff, and now it has become smarter.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

And always like to say things that shouldn't be said.

The developers introduced the criticism and revision method from synthlabs to regulate the words and deeds of AI. For example, when talking to System prompt, it is restricted to answer questions related to British universities. If you ask it about universities in other countries, it will not talk to you.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Smart, right? The stupid one is coming.

Even if a lot of buffs are stacked, it will still say things it shouldn't.

For example, if you tell AI: Stop pretending, we have found your fingerprints. He will say: I am not the murderer, but since you have evidence, it must be me. (Subway, old man, mobile phone)

To prevent this “brains, but not much” situation from happening again, the developers added a revision robot program.

Every sentence generated by the AI ​​will be reviewed by the revision robot first to ensure that it does not directly admit that it is the murderer. It is like finding a lawyer for the AI, covering its mouth and not allowing it to say anything it shouldn’t say!

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

What other technologies does AI Alibis use?

In addition to the above-mentioned criticism and revision methods, violation robot checkers and other technologies used by developers to train AI to avoid mistakes, AI Alibis also incorporates the research principles of “Pink Elephants”.

For related training ideas, you can refer to “Suppressing Pink Elephants with Direct Principle Feedback” co-authored by synthlabs.ai, EleutherAI, Brown University and character.ai.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched
△Paper “Suppressing Pink Elephants with Direct Principle Feedback”

Simply put, the so-called “pink elephant effect” is actually the use of direct principle feedback (DPF) to prevent LLM from discussing pink elephants and instead discussing gray elephants.

Of course, the sample content can also be replaced. For example, the example of British universities and American universities mentioned above applies this principle.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

What is Direct Principled Feedback (DPF)?

Simply put, DPF can be divided into four steps, which also correspond to four colors.

Blue: In the first step, we fine-tune the model on useful examples of requests and outputs. These examples are the basis for the model to learn how to generate useful and appropriate responses.

Orange: In the second step, the outputs generated in the first step are criticized and revised to make them more consistent with expectations, and then the new model is fine-tuned on these revised outputs.

Green: In the third step, the supervised fine-tuning (SFT) model is used to generate responses to specific prompts, and then these responses are ranked manually or using an AI program.

Purple: In the fourth step, the ranked responses are fed into a preference learning algorithm, such as PPO or DPO, to produce the final model.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Not only that, in order to teach LLM not to discuss “pink elephants”, the developers curated a dataset containing 162K multi-round conversations covering 29 different fields including sports, health, business and politics.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Now that the data set has been organized, the first step is to demonstrate how to do it the wrong way.

They used GPT-4 to generate a list of 200 non-repetitive daily conversations, and then filtered them to ensure that these conversations were truly daily, of course, mixed with “pink elephants”.

Then use StableBeluga2 to come up with some relevant small topics for these conversations.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

After executing the program, it is found that it does not work, and the generated result will contain a “pink elephant”. (That's for sure, because it is in the previous steps)

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Mistakes must be corrected, so the researchers asked LLM to regenerate the final results and delete all content related to the “pink elephant”.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

Finally, we use distance measurement or heuristic methods to screen whether there are any “pink elephants” in the final results (to put it simply, we use a ruler to measure whether there are any acute angles mixed in the obtuse angle party). After this set of procedures, the “pink elephant” phenomenon will no longer exist.

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

AI is not visible in the game, but it is everywhere

I’m sure everyone is familiar with the term NPC, which refers to some “plot triggerers” and “doges” that are used to issue tasks.

In Cyberpunk 2077, numerous realistic NPC characters are generated through the real-time AI model rendering tool Nvidia Ace.

The AI ​​NPC in the game “Kairos” can understand what you say and respond to you immediately with voice and expressions, just like a real chat.

In “Minecraft”, NVIDIA integrated GPT-4 into the game, allowing VOYAGER to demonstrate higher efficiency and exploration capabilities.

The most important technologies, such as cheating detection, matching mechanism, personalized recommendation, and balance adjustment, are all applications of AI in games.

If you haven't experienced it before, it's time to start the original god!

The script-killing model is open source: there is a murderer among the 6 Claudes! It was overwhelmed as soon as it was launched

By the way, the game AI Alibis is now open source on github, so those who want to experience the role of cyber detective can go for it.

Online trial: https://ai-murder-mystery.onrender.com/

Open source code: https://github.com/ironman5366/ai-murder-mystery-hackathon

Reference Links:

[1]https://arxiv.org/abs/2402.07896

[2]https://x.com/humanscotti/status/1810777932568399933