This weekend, Andrej Karpathy, the previous director of AI at Tesla and a founding member of OpenAI, determined he needed to learn a e book. However he didn’t need to learn it alone. He needed to learn it accompanied by a committee of synthetic intelligences, every providing its personal perspective, critiquing the others, and ultimately synthesizing a closing reply beneath the steerage of a “Chairman.”
To make this occur, Karpathy wrote what he known as a “vibe code project” — a chunk of software program written shortly, largely by AI assistants, meant for enjoyable somewhat than perform. He posted the end result, a repository known as “LLM Council,” to GitHub with a stark disclaimer: “I’m not going to assist it in any manner… Code is ephemeral now and libraries are over.”
But, for technical decision-makers throughout the enterprise panorama, trying previous the informal disclaimer reveals one thing much more vital than a weekend toy. In just a few hundred traces of Python and JavaScript, Karpathy has sketched a reference structure for probably the most vital, undefined layer of the fashionable software program stack: the orchestration middleware sitting between company purposes and the risky market of AI fashions.
As firms finalize their platform investments for 2026, LLM Council presents a stripped-down have a look at the “construct vs. purchase” actuality of AI infrastructure. It demonstrates that whereas the logic of routing and aggregating AI fashions is surprisingly easy, the operational wrapper required to make it enterprise-ready is the place the true complexity lies.
How the LLM Council works: 4 AI fashions debate, critique, and synthesize solutions
To the informal observer, the LLM Council internet software seems to be virtually equivalent to ChatGPT. A person sorts a question right into a chat field. However behind the scenes, the appliance triggers a classy, three-stage workflow that mirrors how human decision-making our bodies function.
First, the system dispatches the person’s question to a panel of frontier fashions. In Karpathy’s default configuration, this contains OpenAI’s GPT-5.1, Google’s Gemini 3.0 Pro, Anthropic’s Claude Sonnet 4.5, and xAI’s Grok 4. These fashions generate their preliminary responses in parallel.
Within the second stage, the software program performs a peer assessment. Every mannequin is fed the anonymized responses of its counterparts and requested to guage them primarily based on accuracy and perception. This step transforms the AI from a generator right into a critic, forcing a layer of high quality management that’s uncommon in customary chatbot interactions.
Lastly, a chosen “Chairman LLM” — at present configured as Google’s Gemini 3 — receives the unique question, the person responses, and the peer rankings. It synthesizes this mass of context right into a single, authoritative reply for the person.
Karpathy famous that the outcomes have been usually shocking. “Very often, the fashions are surprisingly prepared to pick one other LLM’s response as superior to their very own,” he wrote on X (previously Twitter). He described utilizing the device to learn e book chapters, observing that the fashions persistently praised GPT-5.1 as probably the most insightful whereas score Claude the bottom. Nonetheless, Karpathy’s personal qualitative evaluation diverged from his digital council; he discovered GPT-5.1 “too wordy” and most popular the “condensed and processed” output of Gemini.
FastAPI, OpenRouter, and the case for treating frontier fashions as swappable elements
For CTOs and platform architects, the worth of LLM Council lies not in its literary criticism, however in its development. The repository serves as a main doc displaying precisely what a contemporary, minimal AI stack seems to be like in late 2025.
The appliance is constructed on a “skinny” structure. The backend makes use of FastAPI, a contemporary Python framework, whereas the frontend is a regular React software constructed with Vite. Information storage is dealt with not by a fancy database, however by easy JSON files written to the native disk.
The linchpin of your entire operation is OpenRouter, an API aggregator that normalizes the variations between numerous mannequin suppliers. By routing requests by this single dealer, Karpathy averted writing separate integration code for OpenAI, Google, and Anthropic. The appliance doesn’t know or care which firm gives the intelligence; it merely sends a immediate and awaits a response.
This design alternative highlights a rising pattern in enterprise structure: the commoditization of the mannequin layer. By treating frontier fashions as interchangeable elements that may be swapped by modifying a single line in a configuration file — particularly the COUNCIL_MODELS listing within the backend code — the structure protects the appliance from vendor lock-in. If a brand new mannequin from Meta or Mistral tops the leaderboards subsequent week, it may be added to the council in seconds.
What’s lacking from prototype to manufacturing: Authentication, PII redaction, and compliance
Whereas the core logic of LLM Council is elegant, it additionally serves as a stark illustration of the hole between a “weekend hack” and a manufacturing system. For an enterprise platform staff, cloning Karpathy’s repository is merely step considered one of a marathon.
A technical audit of the code reveals the lacking “boring” infrastructure that business distributors promote for premium costs. The system lacks authentication; anybody with entry to the net interface can question the fashions. There isn’t a idea of person roles, that means a junior developer has the identical entry rights because the CIO.
Moreover, the governance layer is nonexistent. In a company surroundings, sending knowledge to 4 completely different exterior AI suppliers concurrently triggers quick compliance considerations. There isn’t a mechanism right here to redact Personally Identifiable Info (PII) earlier than it leaves the native community, neither is there an audit log to trace who requested what.
Reliability is one other open query. The system assumes the OpenRouter API is at all times up and that the fashions will reply in a well timed style. It lacks the circuit breakers, fallback methods, and retry logic that preserve business-critical purposes operating when a supplier suffers an outage.
These absences usually are not flaws in Karpathy’s code — he explicitly said he doesn’t intend to assist or enhance the undertaking — however they outline the worth proposition for the business AI infrastructure market.
Firms like LangChain, AWS Bedrock, and numerous AI gateway startups are primarily promoting the “hardening” across the core logic that Karpathy demonstrated. They supply the safety, observability, and compliance wrappers that flip a uncooked orchestration script right into a viable enterprise platform.
Why Karpathy believes code is now “ephemeral” and conventional software program libraries are out of date
Maybe probably the most provocative facet of the undertaking is the philosophy beneath which it was constructed. Karpathy described the event course of as “99% vibe-coded,” implying he relied closely on AI assistants to generate the code somewhat than writing it line-by-line himself.
“Code is ephemeral now and libraries are over, ask your LLM to alter it in no matter manner you want,” he wrote within the repository’s documentation.
This assertion marks a radical shift in software program engineering functionality. Historically, firms construct inner libraries and abstractions to handle complexity, sustaining them for years. Karpathy is suggesting a future the place code is handled as “promptable scaffolding” — disposable, simply rewritten by AI, and never meant to final.
For enterprise decision-makers, this poses a tough strategic query. If inner instruments may be “vibe coded” in a weekend, does it make sense to purchase costly, inflexible software program suites for inner workflows? Or ought to platform groups empower their engineers to generate customized, disposable instruments that match their actual wants for a fraction of the fee?
When AI fashions choose AI: The damaging hole between machine preferences and human wants
Past the structure, the LLM Council undertaking inadvertently shines a light-weight on a selected danger in automated AI deployment: the divergence between human and machine judgment.
Karpathy’s statement that his fashions most popular GPT-5.1, whereas he most popular Gemini, means that AI fashions might have shared biases. They could favor verbosity, particular formatting, or rhetorical confidence that doesn’t essentially align with human enterprise wants for brevity and accuracy.
As enterprises more and more depend on “LLM-as-a-Judge” methods to guage the standard of their customer-facing bots, this discrepancy issues. If the automated evaluator persistently rewards “wordy and sprawled” solutions whereas human clients need concise options, the metrics will present success whereas buyer satisfaction plummets. Karpathy’s experiment means that relying solely on AI to grade AI is a method fraught with hidden alignment points.
What enterprise platform groups can study from a weekend hack earlier than constructing their 2026 stack
In the end, LLM Council acts as a Rorschach check for the AI trade. For the hobbyist, it’s a enjoyable approach to learn books. For the seller, it’s a risk, proving that the core performance of their merchandise may be replicated in just a few hundred traces of code.
However for the enterprise know-how chief, it’s a reference structure. It demystifies the orchestration layer, displaying that the technical problem shouldn’t be in routing the prompts, however in governing the information.
As platform groups head into 2026, many will doubtless discover themselves looking at Karpathy’s code, to not deploy it, however to grasp it. It proves {that a} multi-model technique shouldn’t be technically out of attain. The query stays whether or not firms will construct the governance layer themselves or pay another person to wrap the “vibe code” in enterprise-grade armor.
