FAQs

An LLM is a large language model — a machine learning model trained on a large amount of text that generates text. Modern LLMs are often multimodal, also handling images and voice input.

An LLM is point-in-time: it does not change after its training completes, which can leave it stale on current events. Providers like OpenAI and Anthropic work around this by giving the model recent information from the internet at answer time.

Open-source models are the best place to start. In the clouds businesses tend to buy from — Azure especially — they're heavily rate-limited by default, so expect to file quota requests. The frontier labs' cut-rate tiers — like OpenAI's GPT-5.6 Luna — are another lever: cheaper than the flagships, and usually with higher default capacity from the cloud providers.

You can also go directly to a zero-data-retention provider. We recommend Fireworks AI — growing fast, and Azure itself routes to Fireworks for models it doesn't host.

How long the cloud provider that runs the AI keeps your chat with the AI. You generally want the least retention possible, with ZDR (zero data retention) being the best.

Pay attention to where the provider physically runs as well — different jurisdictions bring their own retention baggage. And check the specific model: Anthropic Fable deployments, for example, are never zero data retention — Fable prompts and outputs are retained for 30 days for trust and safety on every platform.

Your data is not stored long term. It lands on the provider's machines, the model — which is just a program — runs over it, you get text back, and they delete it. With a provider like Fireworks AI, that's the whole lifecycle.

No. A model is just a program that runs in the cloud — models don't share data. Run on Fireworks, for example, a Chinese open-source model is just a program running over your data on Fireworks' servers.

That said, retention policies are often not good — which is why a zero-data-retention provider matters. Read the terms of service.

Open-source models aren't as broadly strong as the US frontier labs like Anthropic and OpenAI — the biggest gap today is coding. But across a wide range of non-coding tasks they do just fine. For something like formatting an email, even very small models at one-fiftieth the cost of the frontier models are good.

Optimize where your tokens go. Dropping .docx files and PDFs straight into a model is often more expensive than asking a model for a one-time extract and saving that in your SharePoint. Models tend to do best with Markdown (.md files) — plain .txt works too — and you can ask a model what format it prefers; the answer sometimes varies by where it's running and how you're accessing it.

Point to their existing KPIs, not new ones. "Look how cool it is," usage counts, and tokens spent mean nothing to people trying to get work done against classic bottom-line KPIs. The market has come around to this: proxy metrics don't compel the people who actually do the work. Track the same productivity outputs the business always tracked.

You are. We treat AI as cybernetic — a tool that acts on your behalf, like your hand. You don't blame your hand. A natural consequence is that people don't trust it at first; that's valid, and a good place to start.

Divide your work into what can be easily validated and what can't — the work where you'd have to go through every detail to tell if it was right. AI is really good for the easily-validated work, because when you get around to looking at it, you can tell faster whether it was done than if you'd done it yourself. For developers that's often unit tests, but every industry has work like this — work you can read faster than you can write, or tell at a glance is fine.

No — lean into motivation. Assume positive intent: people like what they do and it provides meaning to them. Figure out which parts of the job carry that meaning and amplify those outputs — for the same amount of work, more of the impact that makes the job feel good every day. It can be done without reducing quality or degrading relationships in the workplace.

Because the mission is ultimately to scale. Now that AI exists, it's possible to bring people into the tech market who normally wouldn't be there because they'd need a college degree.

I'm really excited about that vision — upskilling, making tech available to more people — without compromising on quality. AI lets more people do more things, and I'm passionate about building a new operating model around that.