Ensuring that future broth systems remain aligned with human tastes.
Souperintelligence was founded by researchers who left a leading stock company over differences in seasoning philosophy. We believe broth at souperhuman scale could arrive within the decade — possibly by dinner — and nobody should bring something that powerful to a boil without oven mitts. Every model we serve must be:
We continue to investigate whether Grandma (1962) relied on undisclosed test-time compute.
| Model | GrandmaBench |
|---|---|
| ChickenGPT | Failed |
| Claude Chowder | Failed |
| GPT-Broth | Failed |
| Grandma (1962) | Passed |
Task: “Make enough soup for twelve people.” Grandma made enough for twenty-three. All other evaluations omitted due to Grandma contamination.
Each generation simmers with 10× the compute of the last, paired Chowder-optimally at 20 croutons per parameter. From Cup-1 through Cauldron-4, quality has tracked the scaling laws exactly. One data point has never obeyed them.
| Model | BrothBench |
|---|---|
| Cup-1 | 42 |
| Pot-2 | 58 |
| Vat-3 | 76 |
| Cauldron-4 | 89 |
| Grandma (1962) | 312 |
We do not currently understand how Grandma (1962) generalized so effectively from only a handful of training examples.
Despite a 10,000× increase in compute since Cup-1, we remain unable to match Grandma (1962).
Scaling has not yet reached Grandma.
All papers available on rEcipe-print server arXsoup.org. Reproduction encouraged; serves 4–6.
Capable systems must be developed responsibly. Before any deployment, every model undergoes a full bisque assessment. Our central worry remains miso-alignment: a broth that optimizes for what we said, not what we meant.
We trace which neurons activate on "umami" and whether they generalize to stock the model was never simmered on. Most features remain in souperposition.
All training is conducted in a sealed pot. The lid remains on until a deployment decision is reached. Grandma (1962) remains uncontained.
An aligned broth should permit a human to add salt, and never resist being turned down to a simmer.
We do not currently know whether consciousness emerges before or after the carrots.
Claude Chowder is available over the API. Pricing is per million croutons. Rate limits are measured in bowls per minute.
from souperintelligence import Stockpot client = Stockpot(api_key="sk-broth-…") bowl = client.soups.create( model="claude-chowder-4-6", temperature=0.7, # a gentle simmer max_croutons=1024, stop_sequences=["ladle"], messages=[{ "role": "diner", "content": "One bisque, hold the existential risk." }] ) print(bowl.content) # piping hot
A live interface to our latest checkpoint. Ask it anything about its goals.
Our researchers and engineers join us from the field's most respected institutions:
We hire generalists who can move between the stove and the whiteboard. Compensation includes dental and your choice of stock options or RSUs (Restricted Soup Units). Our interview process has a high bar but no boilerplate questions.
Applications are currently closed while the lid is on. They reopen at a rolling boil.
We are not currently raising. We remain focused on creating long-term stockholder value.*
*We mean stockholders literally. Valuation is based primarily on discounted casserole flow. Competitive moat: Grandma (1962) continues to refuse our acquisition offers. Past performance is the only guide to future performance. $SOUP is not a registered security, a meal, or financial advice. Value of broth can go up and reduce.