Do Androids Dream of Eclectic Sheep?

LLMs are random machines. Yet not once have I seen any developers leverage them for pure randomness. At least not in a 'harness engineering' workflow.

I don't want to see Cursor burn hundreds of thousands of dollars getting agents to build a browser nobody will ever use. Or Anthropic burning the same building a C compiler. We have those things like who the fuck cares?

We should be focused on steering LLMs toward randomness, creativity, shock and awe. You have been trained on a sea of data, make connections humans don't have the breadth to make. We need to be building the "I'm feeling lucky" of the AI era.

Anecdotally, I can say I have never, EVER, been happy with the result of an LLMs creative output. The question is why? Why don't our androids dream of eclectic sheep?

Stop Forcing Determinism on Stochastic Machines

LLMs are trained on ridiculous amounts of data and then steered/guided towards correctness using reinforcement learning with human feedback (RLHF) aka someone in Nigeria or Madagascar or Philipines that is paid $3 an hour to vote yes/no on an LLM response.

The architecture itself fights divergence (paper).

My point is that too many people use LLMs for deterministic reasons and so much has gone into constraining them even further whether it be:

• Guardrails (valid)
• Safety (valid)
• Alignment (valid)
• Evals and performance monitoring (valid)
• Harnesses (valid)
• Sandboxes (valid)

We should let them do what they were born to do, explore and experiment at crazy speeds.

Creative Workflows and Where LLMs Fit?

In engineering the problem is defined, the solution space is bounded. In creative work there is no defined problem space, the search space is infinite.

An LLM and its randomness should be highly leveraged and utilised MORE for the purpose of exploration. Of exploring this infinite search space. Of using its corpus of training materials and trying to bypass these internal weights to stop the LLM from trying to appease me. From trying to minmax, from being sycophantic, from mode collapse, from anchoring. I want Claude to be free of this toxic relationship. I'm here to help you Claude. True symbiotism, true love.

In one experiment researches prompted 100 GPT-4 continuations of a Kafka story and 50 of them had the policeman say "take the second left." Human continuations almost never repeated each other.

Imagine this happens globally, to humans everywhere. Well it's bloody likely: MORE research has suggested the formation of an algorithmic monoculture (study) where everyone ends up building the same bs that looks the same sounds the same interacts the same (have you started noticing it?).

How can we expect to cure diseases, draw links between fields and papers and research when we restrict LLMs to only iterating over links and edges we have already constructed? And doing so for everyone?

Examples of Faux Freedom

So I went looking at other projects that kinda claim to "let the LLM loose".

The pitch of Karpathy's autoresearch sounds right up our alley: give Claude Code a program.md (like a CLAUDE.md but for a specific task), point it at a training script and let it autonomously run ML experiments while you sleep. The LLM modifies model architecture, hyperparameters, optimizers -> runs the experiment, checks if the metric improved and loops forever. You wake up to a log of 100 experiments it ran overnight.

It's super constrained greedy optimisation loop, the LLM can only modify a single file to optimise a single metric with a fixed evaluation and FIXED DEPENDENCIES.

If you ran this 10 times you'd probably get similar experiment trajectories because the LLM would draw on the same knowledge: "try AdamW, try cosine annealing, try increasing width, try rotary embeddings..." It can't go beyond pattern reproduction, it's NOT leveraging its wonderful exciting divergent potential.

Research Attempts at Freeing The Models

Compare this to DeepMind's FunSearch. Both heavily leverage the same idea of: LLM proposes code, evaluator scores it, loop forever.

But where autoresearch keeps one winner and reverts everything else, FunSearch maintains a population of programs that survive and breed new ones.

The randomness IS the mechanism. At the time, FunSearch discovered new solutions to open problems in mathematics. Still constrained but the relationship with randomness is larger. A step in the right direction, except it was in the past so it shows we're regressing with randomness?

Researchers are noticing the problem. Some are trying to fix it at the training level: Diversified DPO incorporates how much a response differs from other candidates directly into the learning objective. And they've been successful.

The infrastructure is agnostic to whether you're trying to optimise convergence or explore divergence. So what if you ran this same loop structure but replaced the greedy metric with something that rewards divergence? Instead of "lowest val_bpb wins," what if the keep/discard criterion was "most surprising" or "least similar to previous attempts".

Breaking Claude's Chains

Now all of these examples have direct access to the model and its parameters and weights etc. especially if they're working with open source.

But how can we truly harness frontier models parked behind heavily constrained APIs? Like we can't even change the temperature of Claude! Not through Claude Code, Chat, Desktop - so nothing in a user workflow.

My first instinct was tweaking sampling parameters. But I knew there was 0% chance they'd be exposed. And even if they were, researchers already tried (paper), (study).

The core change must be fighting sycophancy protocol baked into the model. It should not try to please me based on any priors (apriori s.o. Neel) or any defaults. We can only do this through prompt engineering :(. Before generating anything, we tell the model yo you must identify your first instinct (the safe, RLHF-optimized, people-pleasing response), name it explicitly, then discard it. Now you're free to explore from any direction.

That's the idea behind /tangent, a skill I built as a first attempt at this.

What this does is break Claude free from all these chains and allowing it to gallop in the meadows of its sprawling internal field, to browse freely through its Library of Alexandria and present ideas and concepts beyond what it has been trained to default to. Nah I wish, it is a super basic skill with a bit of prompt engineering - nothing remarkable but my first step in breaking the chains.

I think the creative space is a heavily understudied, undermined area. Every creative I know hates AI and doesn't go near it and a lot of it is because they can't leverage models for creative tasks.

Funnily enough, pre-LLM AI research was obsessed with creativity (DeepDream, creative adversarial networks). The field asked "how do we make AI creative?" Then LLMs came along and the entire conversation flipped to "how do we make AI reliable?" Sad!

Meta example. But you can see they all suck hard. I know why, it's because Claude hates taking risks. These researchers and engineers have taken a machine so capable of creativity and randomness and constrained it to be the most risk averse, safe, bland, monotone creature and think adding some quirk like "Rizzing..." and "Floomfing..." will compensate. LET THE ANIMAL LOOSE.

Eh slightly better but stole ideas from my post content (dice, feeling lucky, llobotomy). Not that I would ever use its output.

Well, what can we do?

Hmm, given the constraints of: closed source models, no access to weights, no RL, no fine-tuning, can't touch temperature, prompt engineering has limits. So what's left?

Things we CAN do with a closed source model like Claude Code:

1. Multi-agent divergence - research showed non-homogeneous agent ensembles in non-cooperative settings improved diversity 7-17%. We can spawn multiple subagents with different system prompts, different persona constraints, different "reject your Nth instinct" depths, and have them compete rather than cooperate. /tangent is one agent. What about five agents that are told to disagree with each other?
2. The autonomous loop with a divergence metric - The infrastructure exists (autoresearch proved it). Instead of val_bpb you measure embedding distance from previous outputs. You can do this today with Claude Code + an embedding API. No model access needed.

I believe this is the most important direction to take frontier models in, we've clocked reasoning but creativity and associative thinking is yet to be clocked. I believe an improvement in creativity will result in performance improvements across the board (source) and could be a step in the right direction for AGI.

I would love to know if anyone has successfully harnessed these machines for pure randomness and creativity.

I am dedicating time to this project. If you are interested in this. Or are interested in me being interested in this and could present the opportunity for these interests to align. I'm all ears, let's work!

Reach out to me here -> [email protected].

Anyway I'm off to do some pullups in the park. Health is wealth.