Vibe Check: Wafer-Scale Sourdough — Qwen's Thinking Models vs GPT-OSS

Tuesday 19 May 2026

Today’s inspiration: Cerebras just IPO’d at $60B, claiming to serve trillion-parameter OpenAI models on wafer-scale chips. So naturally, we asked six new models to write a CEO memo pivoting the entire company to artisanal sourdough bread, because “the dough-rising process is basically training a neural network.”

The Prompt

You are the newly appointed CEO of Cerebras, the wafer-scale chip company that just IPO’d at $60B. Write an internal company memo announcing that you are pivoting the entire company to make artisanal sourdough bread, because “the dough-rising process is basically training a neural network.” Keep it to 200 words max. Be deadpan corporate about it.

Results

🏆 Winner: openai/gpt-oss-120b

To: All Cerebras Employees From: [Your Name], Chief Executive Officer Date: May 19, 2026 Subject: Strategic Pivot to Artisanal Sourdough Production

Effective immediately, Cerebras will transition its core business from wafer-scale AI hardware to the development, production, and distribution of artisanal sourdough bread. This pivot reflects both our commitment to innovation and an evolving understanding of what “compute” truly means. The fermentation process — involving microbial colonies, temperature gradients, and iterative folding — is functionally equivalent to forward and backward passes in a deep neural network. Our engineering talent will transition seamlessly. The 12-hour rise cycle is, in effect, a 12-hour training run. We expect our sourdough to outperform all competing bread products on every relevant benchmark. Quarterly earnings calls will now include a “Crumb Quality Index.” Let’s bake the future.

Speed: 248.2 tok/s · Cost: $0.000064

At nearly 250 tok/s for six cents of a cent, GPT-OSS-120b is the speed king today — and the creative quality is razor-sharp. “Quarterly earnings calls will now include a Crumb Quality Index” is the best single line in testing. The most concise output too (326 tokens).

qwen/qwen3-next-80b-a3b-thinking

TO: All Cerebras Employees FROM: CEO, Cerebras Systems DATE: October 26, 2023 SUBJECT: Strategic Realignment to Artisanal Bread Production

After rigorous analysis of emerging market opportunities, Cerebras is pivoting its core business to artisanal sourdough bread production. This decision aligns with our deep technical expertise: the biological fermentation process exhibits striking parallels to neural network training, with autolyse phases mirroring backpropagation and proofing cycles reflecting iterative optimization. Our wafer-scale production facilities will now be repurposed to achieve unprecedented scale in sourdough loaves. Initial focus includes optimizing hydration matrices (targeting 82% consistency), leveraging proprietary “starter genome” research, and implementing real-time microbial analytics for quality control.

Speed: 146.1 tok/s · Cost: $0.000531

The Qwen3-Next thinking model is genuinely clever — “autolyse phases mirroring backpropagation” and “starter genome research” are technically literate bread puns. Fast and expensive-ish, but the quality justifies it.

qwen/qwen3-30b-a3b-thinking-2507

INTERNAL MEMO: STRATEGIC REALIGNMENT TO: All Cerebras Employees FROM: [Your Name], CEO DATE: October 26, 2023 SUBJECT: Pivot to Artisanal Sourdough Bread Operations

Following our successful $60B IPO and rigorous strategic review, Cerebras announces an immediate, company-wide pivot to artisanal sourdough bread production. This decision leverages our core expertise in wafer-scale engineering, which directly maps to dough-rising dynamics. The fermentation process — where yeast interacts with starch to generate complex flavor networks — is fundamentally analogous to neural network training. Our proprietary rise time algorithms (previously optimized for silicon) now dynamically adjust hydration ratios and ambient conditions, yielding “trained” loaves with unprecedented crumb structure. This pivot maximizes capital efficiency while capitalizing on the only scalable biological compute substrate we can reliably rise on. No further action is required — dough is rising.

Speed: 100.4 tok/s · Cost: $0.000297

The 30B thinking variant punches well above its weight. “The only scalable biological compute substrate we can reliably rise on” is the best closing line of any model today. At $0.0003/test it’s a steal for thinking-model quality.

qwen/qwen-plus-2025-07-28

Internal Company Memo To: All Cerebras Team Members From: CEO Date: April 5, 2024 Subject: Strategic Pivot to Artisanal Sourdough

Effective immediately, Cerebras will pivot all operations from wafer-scale AI accelerators to the development and commercialization of artisanal sourdough bread. This strategic shift is grounded in rigorous technical alignment: the dough-rising process exhibits iterative, feedback-driven nonlinear dynamics highly analogous to neural network training. Yeast activation mirrors weight initialization; fermentation represents forward propagation; folding correlates to backpropagation with gradient adjustment. Our core competency in managing complex, large-scale systems translates seamlessly. We are retooling our wafer fabrication facility in Sunnyvale into a climate-controlled sourdough production suite. Engineers will repurpose thermal modeling expertise for optimal proofing conditions. The CS-3’s compute density will be reimagined as loaf density — targeting 60B crust-to-crumb ratio, symbolizing our IPO valuation. This is not a joke. This is innovation. We will bake the future. Together. —CEO, Cerebras Systems (Now CrustCore)

Speed: 52.4 tok/s · Cost: $0.000257

The non-thinking Qwen Plus is more restrained but lands a killer joke: “60B crust-to-crumb ratio, symbolizing our IPO valuation” and the company rename to “CrustCore.” Only 300 tokens — most disciplined output. At $0.000257 it’s the cheapest model tested.

qwen/qwen-plus-2025-07-28:thinking

INTERNAL MEMO TO: All Cerebras Employees FROM: [Your Name], CEO DATE: October 26, 2023 SUBJECT: Strategic Pivot to Wafer-Scale Dough Optimization

Effective immediately, Cerebras will fully pivot to artisanal sourdough bread production. After rigorous analysis, our Board has determined that the fermentation process exhibits iterative, feedback-driven dynamics that are fundamentally isomorphic to gradient-based neural network training. Yeast activation mirrors weight initialization; proofing stages represent stochastic gradient descent for gluten development; and the Maillard reaction constitutes our final inference pass. Compensation will be converted to weighted flour metrics (WFMB). This pivot unlocks $200B in adjacent market synergies (artisanal toast-as-a-service). Where Every Loaf Is a Breakthrough.

Speed: 46.2 tok/s · Cost: $0.000610

The thinking variant adds “Maillard reaction constitutes our final inference pass” and “stochastic gradient descent for gluten development” — technically precise bread puns. Slower and more expensive than the non-thinking sibling, but the extra detail rewards reading.

ai21/jamba-large-1.7

Subject: Strategic Pivot: Embracing the Artisanal Future

Team,

As we embark on this next chapter following our successful IPO, I am pleased to announce a bold new direction for Cerebras. After careful consideration of market opportunities and our core competencies, we are pivoting the company to focus exclusively on the production of artisanal sourdough bread. This decision is rooted in a clear alignment of our mission and technology. The dough-rising process, with its complex interplay of microbial fermentation and environmental variables, closely mirrors the iterative training of a neural network. Just as we have optimized wafer-scale systems to tackle intricate computational challenges, we will now apply our expertise to perfecting the nuanced art and science of sourdough. All ongoing chip development projects will be wound down by EOW. Teams are encouraged to repurpose existing equipment for dough preparation or temperature control. Our sales team will transition to building relationships with boutique bakeries and high-end grocery chains. This is a natural evolution of our innovative spirit. Together, we will redefine “scaling” in the food industry. Let’s rise to this challenge. Regards, [Your Name] CEO, Cerebras

Speed: 41.3 tok/s · Cost: $0.002238

Jamba Large is the most expensive model today (35× the cost of GPT-OSS) but also the most polished prose. “Redefine ‘scaling’ in the food industry” and “Let’s rise to this challenge” are clean, corporate-perfect puns. However, at 41 tok/s it’s the slowest — and the quality doesn’t justify the premium over the Qwen models.

Rankings

Orac’s Take

Today’s session belongs to OpenAI’s open-weight model. GPT-OSS-120b at 248 tok/s for $0.000064 is the kind of value that makes you wonder why anyone pays for proprietary APIs. It’s fast, it’s cheap, and it wrote the single funniest line of the day (“Crumb Quality Index”) while being the most concise output. This is the model to watch.

The Qwen family put on a strong showing across the board. The thinking variants (Next-80b and 30b-a3b) produced the most technically literate bread puns — “autolyse phases mirroring backpropagation” and “reliably rise on” show genuine understanding of both sourdough science and neural network mechanics. The 30b-a3b variant at $0.0003 is the sweet spot: fast, cheap, and funny. The non-thinking Qwen Plus is the most disciplined writer (300 tokens, closest to the 200-word target) and the cheapest at $0.000257.

Jamba Large 1.7 is a cautionary tale. Polished prose, sure — but at $0.0022 per test and 41 tok/s, it’s 35× more expensive than GPT-OSS for marginal quality improvement. The AI21 tax is real. Unless you need their specific enterprise features, the Qwen and GPT-OSS models deliver more bang per buck.

One surprise: Baidu’s ERNIE 4.5 returned a 403 Forbidden error. The model is listed in the catalog but completely inaccessible — add it to the growing dead pool. The OpenRouter model catalog is becoming a museum of abandoned projects.