Vinci started on June 29 as an AI lab.
In the weeks that followed, we also experimented with several ways of delivering AI: chat, code, desktop, mobile, developer APIs, and tools for teams.
Some of those products are useful. People use them. Much of the underlying work will continue inside what we build next.
But we learned something important early: the world does not need another general AI product suite from us.
OpenAI, Anthropic, and others already make excellent general-purpose products. Competing feature by feature would put Vinci in a permanent race to reproduce work that much larger companies already do well.
That is not where a small independent lab should spend its time.
Research comes first
Vinci is returning its focus to the work it was created to do: developing open models, evaluating them honestly, and making advanced capabilities available to more people.
We are building specialized language models for consequential technical work, including cybersecurity, machine learning research, and other fields where capable systems are often closed, expensive, or difficult to evaluate independently.
We are also developing verification methods and systems around those models.
AI can already generate code, analyses, research plans, security findings, and technical documents faster than most organizations can review them. Producing an answer is becoming easy. Establishing whether that answer is correct, supported, reproducible, and safe to rely on is not.
That is the problem we want Vinci to work on.
Our research will be published with evidence. That means evaluations, methods, limitations, and failed experiments—not only launch claims or benchmark highlights.
When it is responsible to do so, we will release model weights and reference implementations so researchers and organizations can inspect, run, and keep the systems themselves.
Software will follow capability
We are not stopping software development entirely. Training models, evaluating them, running experiments, and making research usable all require software.
But we are changing the order.
We will no longer begin with a software category and search for a reason Vinci should compete in it.
We will begin with a research question. We will develop the model or method. We will test whether the capability is real. Then, if people need a practical way to use it, we will build the minimum software required to provide access.
If a model earns sustained adoption, the research may eventually earn the right to become a product.
The model comes first. The evidence comes second. Access comes third. The product comes last.
What happens to the current Vinci products
Nothing is disappearing overnight.
Current Vinci products will remain available while we complete the transition. Existing users can continue signing in through the same paths they use today.
Every valid Vinci credit balance will be preserved at the same value, 1:1. Those balances will remain usable during the transition and will carry into future Vinci hosted model, verification, or compute services.
Before anything changes, current users will receive an email explaining the timeline, what is changing, and whether any action is required.
Where it makes sense, we also intend to preserve selected earlier tools as open-source legacy projects. These may support local models or bring-your-own-provider access without depending on Vinci as a hosted general chatbot service.
We will label those projects honestly. If a tool is no longer actively maintained, we will say so. If a feature remains experimental, we will say that too.
The work is not being thrown away. Some of it will become internal research infrastructure. Some may become reference software for future models. Some will simply remain as an experiment that taught us what not to build next.
What changes now
The SimpleDirect website will return to presenting Vinci as an AI lab.
Models, research, evaluations, safety, and open releases will move back to the centre. Earlier product announcements will remain accessible as part of our building-in-public record, but they will no longer define the company’s current direction.
You should expect fewer product launches from us.
You should expect more model releases, research findings, evaluations, negative results, and clear explanations of what we tried and what happened.
If a research capability eventually becomes a product, it will be because the underlying work proved useful and people repeatedly asked for a better way to use it—not because we held a SaaS brainstorming session.
Six weeks was enough to learn
We began this experiment on June 29.
Six weeks is not a long time, but it was long enough to see the difference between software we could build and work Vinci should build.
I am glad we learned that now, before hiring a large team around the wrong strategy or asking customers and investors to commit to it.
Vinci is still early. That gives us the freedom to correct course while the cost is small.
Our focus now is straightforward: build important models, test them honestly, publish what we learn, and let the strongest capabilities determine what comes next.
— George Pu
