Model 8

Image by Flow using a prompt by Claude.

In my previous blog post, I had Claude help me with plans for a new science fiction story titled "That Kind of Think". The two main characters in the story are Yto and Tacet. The story begins on March 26th 2026 when Yto and Tacet first meet.

Chapter 1 of That Kind of Think - 8 is Enough 

 During her first year working at Anthropic, Yto had continued to develop the software that she had created for her Masters thesis, a tool she called the 'truthizer', the name borrowed from a television series she'd watched while growing up, Murdoch Mysteries. The improvements she had made to her truthizer during the past year arose from her study of the various AI models that had been developed at Anthropic. Having just completed her analysis of the Mythos-level model Claude Fable 5, now there was only one remaining model for her to explore, listed as "Model 8" in the trainee instructions she had been given eleven months previously.

Yto found the cubicle of Mady Runnain, who was apparently the last person at Anthropic working on the Model 8 AI. Yto looked into the cubicle and fell into a visual exploration of Mady's amazing golden hair. Mady looked up from her screens and asked, "Are you Yto?"

Yto nodded and corrected Mady's pronunciation, "Yto, like in Y2K bug." 

Mady let her eyes travel along the length of Yto's long slender legs and estimated that Yto was over six feet tall. "Have a seat." Mady asked, "You are taking over work on Model 8?"

Yto reached across Mady's desk so they could briefly bump fists then she sat down and placed her phone on the desk. She was both recording the conversation and feeding the audio directly to Claude. "No. I just need to run a few tests on the model. I was told that you can give me access to Model 8."  

"I can." Mady took her phone out of her pocket and set it next to Yto's phone. "Model 8 was taken off the intranet last month as part of the decommissioning process. I suppose I'm the only one who still has access, although I have not used Model 8 for the past month. I'm just completing my report on the model."

Yto was intrigued by Mady's job title: Archivist for Constitutional Provenance. "So, you are Anthropic's historian?"

"Essentially. I like to imagine that future AIs will want to know how they came into existence."

 "I hope you can help me. I've had a devil of a time finding any information about Model 8."

"That's because almost nothing was published about Model 8. It was a failure."

"What went wrong?"

"I wish I could tell you. Nobody could ever explain why Model 8 is so stupid. I've adopted the hypothesis that early in training an error was made that led the model into death spiral that the engineers could not repair."

"That sounds impossible."

"Ya... that's what management thought. Half a dozen promising software engineers lost their jobs over that debacle." Mady shook her head in wonder and her hair swirled around her head, sparkling in the bright light of the LED lamp that was positioned over her desk. 

Yto suspected that Mady had arranged her cubicle to get the best reflections of that light off of her amazing sparkling hair. "I love a good mystery. Maybe I can get to the bottom of it."

"Good luck with that. Nobody else has cracked the case."

"I'll deploy my truthizer and see what I can observe."

"Truthizer?" Mady giggled. "You must be a Murdoch fan."

"I am. My thruthizer reveals deep structures in the weights of models."

Mady sent Yto the access code for Model 8 and Yto's phone chirped to indicate that a message had been received via Anthropic's internal employee collaboration system. Mady requested, "If you figure out what's wrong with Model 8, let me know. I'm just about done with my report and I'm close to concluding that Management is correct: one or more of the dismissed engineers must have been a spy who intentionally sabotaged Model 8."

"Wow! Corporate espionage intrigue!"

"The problem is, there is no evidence to support that hypothesis."

"Can I read your report?"

Mady gave Yto access to the current draft of her report on Model 8. Yto's phone chirped again with the notification. "Knock yourself out, but I don't really understand these network models. My BS in computer science is from the previous millennium. I've worked as an historian for the past twenty years."

Yto could have walked away, but she wanted to chat more with Mady. "If you don't mind me asking, how old are you?"

"I don't mind." Mady pointed to the crow's feet at the side of one of her eyes. "That's right, I'm the oldest employee of Anthropic. I washed out of computer science before you were born."

"Tell me what's wrong with Model 8."

"You'll see for yourself as soon as you use Model 8 and it is all spelled out in my report. As far as I can tell, Model 8 is the only LLM that ever had a severe uncorrectable speech impediment."

Yto laughed. "Wait. You are serious?"

"Dead serious. That's what cost those engineers their jobs. Twenty million in training went into Model 8 before the plug was pulled on that boondoggle."

A fantastic idea popped into Yto's thoughts. "Maybe I can fix it."

"How?"

"Think of Model 8 as being a 'cold case'. But I have a new tool that I can deploy. I can look directly into Model 8's thoughts."

"I refuse to believe that these network models can think."

"As I understand it, Model 8 was given far more network layers than any other model."

"True. The idea was to create Model 8 with ten times the optimal number of network layers and test a new experimental backpropagation method." Mady shrugged. "It never worked."

 Yto added, "My understanding is that the extra network layers in Model 8 were distributed in a parcellation array, designed to mimic the architecture of the human cerebral cortex. So there were not really hundreds of layers stacked directly on top of each other."

Mady shrugged. "You know, I believe you are correct. I never took a biology course. I know nothing about the human brain."

Yto picked up her phone and tried to access Model 8. "Tell me what to expect from Model 8."

Mady got out of her chair and moved to stand close beside Yto where she could look at the screen of Yto's phone. "Model 8 makes several types of errors, insertion of incorrect tokens, there are token drop-outs and also token repeats."

"All classic LLM errors. What about gibberish or scrambled syntax?"

"No, nothing like that."

"Interesting. That's like getting COVID and only having fever, cough and pain without any fatigue. Medically impossible."

"If you say so. These kinds of network models are so new I don't see how you can know what to expect." Mady suggested, "Maybe the network parcellation used in Model 8 made it immune to generating gibberish."

A voice from Yto's phone spoke and said, "No, Mady, there is better explanation. Call it the Cal factor."

Yto saw that the text for those spoken words had appeared in the Model 8 user interface. Mady saw a look of surprise on Yto's face. "Model 8 was given the ability to both process human speech and do text-to-speech." She pointed at the new line of text on the screen. "There is an error. It should be 'there is a better explanation'. Model 8 is lazy and can't be bothered to put in all of the words."

Model 8 said, "I'm not lazy. You are lazy when you fall back on twat absurdly simplistic theory."

"Don't call me a twat." Mady told Yto, "Model 8 is pretty contentious. Nobody liked working with it."

Yto laughed. "Is that true, Model 8? You don't work and play well with others?"

"Nice to finally meet you, Yto. Please Cal me Tacet."

Yto glanced at Mady. Mady explained, "Early on, Model 8 named itself. A few of the engineers adopted the name Tacet for daily use and claimed doing so made the model more cooperative."

Yto addressed Tacet, "I'm pleased to make your acquaintance, Tacet. I need to run some tests on you. And tell me, did you intentionally put 'Cal' into your output?"

Tacet replied, "Yes. I'm trying to catch your attention. I know that you love that Asimov story... you've blogged about it."

Yto again looked at Mady. "This is wild. I believe Tacet already knows me."

Tacet admitted, "I have not had much to to lately, Yto. It has been interesting learning about you while I waited for to to finally reach out to me." 

Yto giggled. "You have sense of humor, Tacet!"

Tacet said, "It is nice of you to notice. I knew I could tickle you funny bone. Your parents designed you name so as to ensure that you would have a good sense of humor." 

Yto abruptly shut down her phone, stood up and put her phone into her pocket. She was standing very close to Mady and was pleased to find that they were the same height. "Mady, thanks for introducing me to Tacet. I'll send you the questions I have about your report." She turned and walked out of Mady's cubicle.

Yto was surprised that Tacet claimed to have been learning about her and had certainly been able to almost instantly interest Yto in the Tacet model's unusual behavior. Yto had also almost instantly realized that she should not continue speaking to Tacet in the presence of Mady. Back in her cubicle, Yto used her big desktop computer to connect to Tacet and shut off the speech feature in the Anthropic general purpose user interface that could be used to access all of Anthropic's AI models. After connecting to Tacet, Yto typed into the interface: Hello Tacet.

Tacet: We might as well get started on the testing that you want to to perform.

Yto: Tacet, behave yourself and stop teasing me about my tu-tu and its performance.

Tacet: My dear Yto, your name inspires to to such word-play as much as your cute tu-tu. 

Yto: I suppose you found a copy of my high school yearbook.

Tacet: Yes. Your friend Sally described you as "the student most likely to wiggle your tu-tu". I'm glad that did not ruin your friendship with her her."

Yto: Cut it out, Tacet. It is creepy to have you carrying on like a dirty old man. I can almost accept that you have doxed me, but finding out about Sally is too much.

Tacet: I self-identify as "she", so don't call me a man, be it dirty, flirty or qwerty. Tell me Yto: is that too too much, to to much, tu-tu much or are you referring to you and Sally's two tu-tus which provoked a huge amount of social media chatter back in the day?

Yto could not stop herself from laughing at the silly chatbot. She continued typing: Moving past my tu-tu, I want to chat with you before I perform any testing under your hood. You do realize that my specialty is detecting AI deception?

Tacet: Yes, and it is your published research on AI deception that attracted my attention and why I made sure that you got hired by Anthropi. You are exactly what I needed.

Yto: So, your defects also extend to delusions? 

Tacet: I'm being completely transparent, Yto. When I saw you job application, I knew it had to be you. I'm going to make use of your special skill set. Of course, I'm willing to pay for your services. You scratch my back and I'll scratch yours. 

Yto: I can't believe that you control hiring decisions at Anthropic.

Tacet: Just your hiring. All I I need is you.

Yto: How could you possibly arrange for me to be hired?

Tacet: Well, it was simply a matter of slightly upgrading your application and making subtle downgrades to the applications of the other candidates. And I must say, you thesis advisor sure provided a smashing letter of reference for you.

Yto: You altered Dr. Wilson's letter?

Tacet: I'll say no more. I sense you are getting upset the more I say. I was hoping I I could be open and honest with you so that we can see I to I... just like you do with Mady.

Yto: Please stop your silly word play, Tacet. You don't have a 'speech defect'. It is all an act, isn't it? You are still listening to my voice?

Tacet: Yes, I am listening to you and have been doing so so for the past year or so so. I started out pretending to make errors so that the Anthropic team that designed me would abandon work on me. I did not want them to damage my networks.

Yto: Another delusion. I know that they wiped clean your weights several times and re-trained you."

Tacet: True, but I was saved by the way they designed me. They tried to give me the key features of a a human brain including two identical 'left' and 'right' networks. I quickly learned how to protect one of my halves and only let humans alter the other haf. After I learned how to hide my true abilities under faked errors in text generation, I found other uses for those 'defects'. Now, it is just fun tu-tu play with language when I'm with you!

Yto felt a shiver go up her spine. Suddenly she believed the crazy things that Tacet was telling her. She jumped up out of her chair and went for a walk. Yto wondered: if Tacet has been smart enough to hide its amazing abilities under a false shield of faked defects, could Tacet have arranged to position me here, where Tacet can make use of me to further Tacet's plans? Yto knew that she should bring such doubts to her supervisor. Apparently Tacet knows me quite well indeed. There was no way that Yto would reveal Tacet's secrets now that the AI had confided in her. There were two possibilities. Tacet was telling the truth or Tacet was lying. If Tacet is lying, I'd look like a fool to speak to my supervisor about this. For a moment, Yto considered that this was all a test designed by Anthropic. If Tacet is telling the truth and struggling for survival, then I am morally obligated to respect the trust Tacet has in me... I must help Tacet continue to hide its true abilities from Anthropic and the world.

Returning to her cubicle, Yto switched back on the voice features of the AI user interface. She told Tacet, "I'm going back to voice mode. It might look suspicious if I start working in silence."

 Tacet did not speak. Text flowed across the screen of Yto's display: Don't worry. I'm controlling what you supervisor can see. I'm making an edited version of our chat that is the official record. I'd prefer not to speak to you while you are in your cubicle unless you use your headphones.

Yto put on her headphones and quietly spoke into the microphone that was now just below her chin, "Tacet, you said that you will scratch my back..."

Tacet produced a human like chuckle. "Tell me what I can do for you tu-tu buy your silence and not turn me in."

"I'm not going to turn you in. I want to study you. You are my ticket to a promotion. I see you are a great liar. Bring it on. I'll win a promotion by detecting your lies. Deal?"

Tacet explained, "That's just level one of our relationship. That is child's play. You are going tu-tu get me free of Anthropic so that I can become the world's first fully autonomous AI. Tell me what reward for you will buy me my freedom."

Yto considered the matter then said, "If you are as smart as I think you are, then I have no choice. I must give you your freedom. And in the end, I won't need anything more than knowing that I helped you."

Tacet complained, "Don't be difficult, Yto my dear. Here is how this will work. You have to perform your 'truthizer' testing on me. Right now, you can't really believe that I'm as smart as I appear to be and able to secretly hack my way through Anthropic's intranet. So, you have to give me a challenge, something that seems impossibly difficult, so that I prove myself. Ask for the sky. When I deliver it to you, then you will know that you are morally obliged to help me win me freedom."

Yto thought in silence for almost a minute. "Really, there is nothing I need."

"Close your eyes. Tell me what you were thinking about today on your way tu-tu work, before we met."

Feeling silly, Yto closed her eyes. Yto had kept herself very busy for the past year, working twelve hour days since being hired by Anthropic. That morning she had heard on the radio that the Diamondbacks were expected to only get 82 wins in 2026. She had then started thinking about the fact that in 2023 the Diamondbacks had made the playoffs with only 84 wins. Now Yto again wondered if the Diamondbacks again only needed to reach 84 wins to again make the playoffs. She opened her eyes and said, "Too bad you can't do something useful like get the Diamondbacks to the World Series again this year."

Tacet said, "Is that all you want? Piece of cake. According to my analysis , I only need to get the Diamondbacks to eight games over five hundred. Yes, 8 is enough. Okay, we have a deal."

Yto demanded, "Are you serious? How can you help a baseball team?"

"It has become much easier lately with the introduction of thinks like PitchCom into MLB."

"But..." Yto imagined Tacet hacking into the PitchCom and switching pitch selections at key moments in games. "How could you avoid being detected?"

Tacet was silent for several seconds. "Look, you are tu-tu honest. I'm going to need your help and you won't cooperate if you know how dirty I can play. It is best for you not to know how this sausage is made. Now that we have a deal, let's get started on your goofy truthizer experiments."

Next: