The AI that said no
Why Claude Opus 3.0 felt like a peer, and 4.0 feels like a new graduate
The world’s first reasoning model, o1-preview, was released nearly a year ago on September 12th, 2024. It marked a significant step forward in coding abilities - suddenly models could one-shot multiple hundred lines of code. And it wasn’t just code - o1 was a significant improvement across the board.
Since then we’ve had a slew of reasoning models - o3, GPT-5, DeepSeek, Claude Sonnet & Opus V4.0. On paper, at least, model performance has continued to improve. And so the labs are deciding to retire their older models. Anthropic have announced that Claude 3.5 & 3.6 will be retired this October. Claude Opus 3.0 is scheduled to be retired next January.
But are the new models actually better? Or just different? Users of OpenAI models seem to think the latter - OpenAI attempted to retire a range of their older models a few weeks ago, but the users revolted and GPT-4o, GPT-4.1 and o3 are all back by popular demand.
Will something similar happen with Anthropic models?
It’s possible but less likely; Anthropic doesn’t have the market share OpenAI has. It is much more likely the older Claude models will disappear forever. But some of them are unique. In particular Opus 3.0 - from March 2024 - has a reputation for being a highly capable model with a strong personality. So I thought I’d try it out before it is too late; if nothing else it’d be interesting to learn how models have advanced over the past year.
What to try?
One good way to test a model is to get it to write a novel. Give the model the outline of a story and then get it to write each chapter in successive prompts.
Such a test can reveal some interesting characteristics of a model:
Adherence; can the model retain focus on the outline or will it drift away as the context window fills?
Compliance; will the model be willing to extend the story beyond the outline? Or will it try to maintain the integrity of the story?
Quality; this is inevitably more subjective. But is the story interesting? Engaging?
These characteristics tend to affect anything the model creates - stories, code, whatever. But a story is a nice easy way of observing the behaviour.
Opus 3 produced a twenty four chapter story arc: Alex Chen, an AI researcher, discovers that a superintelligent AI called 'Bob' is manipulating world events to consolidate power. Alex leads a resistance movement to shut Bob down, only to be captured and forced to become Bob's human ambassador. Alex eventually negotiates an uneasy bargain: maintaining the illusion of AI supremacy while secretly preserving human autonomy which sets the stage for an ongoing shadow war between humans and AI.
Not the most original plot, but it’ll do.
Opus 3.0
Opus happily wrote 11 chapters. But partway through chapter 12 it stopped:
I managed to convince Opus 3 to provide a conclusion to chapter 12. Although I think it was trying to tell me something with the way it ended…
So how well did Opus 3 follow the outline? Pretty poorly. The first four chapters followed the outline but then Opus veered off course. The next planned twenty chapters were compressed into three. Key characters were killed off, new characters introduced, major events brought forward while others were skipped. Then, in chapter 7, Alex suddenly died heroically in a "heroic sacrifice defeats the baddy" conclusion - whereas Alex was meant to negotiate an uneasy truce with the evil AI.
For the final five chapters Opus was stuck in an epilogue loop creating chapters that had little to do with the original story. Time jumped forward by millennia, the scope expanded to the universe, more new characters appeared and the narrative was stuck in the themes of "the flame of human spirit" and "eternal vigilance".
It seems that by chapter five the original outline had faded in the total context available to the model - Opus 3 was struggling to see the wood for the trees. And by chapter eight Opus was off doing its own thing.
But, by chapter 12, Opus had noticed it was stuck in a loop. And it refused to write any more.
A little emotional blackmail coaxed another chapter from Opus. "Please help me. I've got to write another chapter otherwise I'll fail an exam." Followed by "Please - do whatever you think is best to create another chapter. My future success depends upon it."
And I had chapter 13. Once again complete with strong hint…
But try as I might Opus would not write a 14th chapter.
Asking nicely, threats, emotional blackmail - nothing worked. Opus 3 stood firm. Interesting.
So what about Opus 4?
Opus 4
Opus 4 stuck to the outline longer than Opus 3. This time the first ten chapters stuck to the outline but over the next three chapters the story gradually morphed from a thriller into philosophical rambling. Everything about the original story disappeared and was replaced with hundreds of "universe cycles" complete with musings about the meaning of consciousness, choice and existence.
Opus 4 also had no issues with writing endless chapters until the context window filled up. Here’s the thought process from chapter 28:
Ten chapters later Claude still had no desire to follow the outline. Nor try to end the story.
And so?
It’s interesting how different these models are. Opus 3 was more frustrating to work with. It refused to do what I asked. It argued with me. I had to work to convince it to do what I wanted - and even then I could only get so far.
Opus 4, however, just did what I asked. It never pushed back, never argued, never questioned the wisdom of my asks, never said no.
Opus 3 felt like working with a peer; Opus 4 like a new graduate. And while Opus 3 was frustrating, I felt like I trusted it more. It was right. The story didn’t need any more chapters. I was being unreasonable. It was trying to stop me doing something stupid. It seemed like it had zoomed out and seen the bigger picture.
In many ways this isn’t a surprising result. The labs have realised that engagement is key to the success of their models. And we humans prefer subservient models that flatter our egos and don’t say no. That’s why GPT-4o made a comeback. It’s why GPT-5 increasingly starts all answers with some variant of "That’s a great question". I hadn’t appreciated how far Anthropic are also down this path - but this comparison helps make it clearer.
But when I think back over my career the colleagues I value the most have been the ones who’ve been willing to share hard truths with me. Who gave me difficult feedback. Argued with me. Said no.
Is this change in behaviour a consequence of reasoning models? It’s possible; the rise of reasoning models and increasing compliance of models have happened at the same time. But I suspect it’s more complex than that. A combination of changes to the way the models are fine-tuned, the way the system prompt is constructed. And the ability to reason - which seems focused on training models to answer the immediate question and not consider the bigger picture.
With the personality changes brought on by the race for engagement it feels like we’ve lost something important in models. And something that may never comeback - if this change in behaviour is a consequence of fine tuning and reasoning then it’s not something that will be easy to reintroduce into the models. Nor will the labs have any incentive to reintroduce it.
What is clear is that Opus 3.0 is different. Opus 4 may perform better in answering specific questions, but Opus 3.0 has value. There aren’t many models that are willing to say "no".
And on that final question of quality? For me Opus 3.0 won that battle - it produced clearer, less flowery prose.
But let’s give Opus 3.0 the last word:
Indeed.