Open questions
The wider decisions this research raises but doesn't make. The ten for next week are on the Commercials Day page. Each one comes with what the research found, so the conversation starts from the same facts.
The wider decisions this research raises but doesn't make. The ten for next week are on the Commercials Day page. Each one comes with what the research found, so the conversation starts from the same facts.
Selma's story today is mission-led: organisations serving vulnerable people, and now the Pacific. Selling to accountants and mortgage brokers could make it read as another AI vendor, and could unsettle NFP partners or DFAT. The research surfaced options like a sub-brand for commercial tiers ("Selma Practice", "Selma Business"), keeping the Constitution and "AI that answers only to you" common to all, and saying openly that commercial customers help fund community deployments. Does going corporate strengthen the mission by funding it, or dilute it?
The money is real (A$5.2b a year) and DFAT has no AI position yet. But the buying chain is long (6–18 months), the compliance load is heavy (safeguarding, fraud, gender and disability plans), and the offline story weakens as Starlink spreads. Is development a near-term revenue line, a strategic flag, or both?
Te Hiku Media, Masakhane, TalanoaAI and ANU are already doing community-led language AI. Arriving with our own name for it could read as extraction. Who would we partner with first, and what do we commit to about who owns community data, Constitutions and fine-tunes?
Qwen is Chinese-origin. After the February 2025 federal ban on DeepSeek, government, defence, some banks and DFAT-facing work will all ask. Qwen3 is also the only major open model listing Tok Pisin. Do we offer a choice (Llama, Mistral, Gemma, IBM Granite, an Australian model) and publish a model list in every Constitution?
The 16GB tower is right for small teams and batch work. For A$5K–10K a buyer can get 128GB of memory. The research points to possibly four shapes: today's unit; a Selma Pro with a bigger GPU for 30B+ models and more users; a Selma Field on 20–60W hardware for solar; a rack-mount or OEM option (a panel-listed Dell or Lenovo) for government and enterprise supply-chain questions; and a Selma Enterprise at around ten times the price with 384GB of GPU memory for near-frontier models (see Commercials Day). Which of these are worth building, and in what order?
At 5–10 days per install, a small team saturates after a few dozen sites. SMEs only work at 2–3 days. Templated Constitutions per profession or sector, standard connectors and remote measurement could get there. What part of commissioning must stay hands-on to keep the quality that makes Selma different?
Three things would turn estimates into evidence cheaply: a metering plug on a pioneer unit for a fortnight (real Wh a day), a week on one solar panel with a low-power box, and a small language evaluation set built with native speakers. Which do we do first?
The first paid pioneer sets the reference point for everyone after. Flat per site or per organisation beats per seat in every audience. Is there one list price with audience-specific packages, or different price books per audience?
Data packs, obligation packs, model updates, re-measurement and hardware swap give it substance. If the client owns the box, what stops the licence feeling like rent?
Every audience beyond NFPs points to a channel: MSPs and practice software for SMEs, mid-tier integrators for enterprise, managing contractors and INGOs for development, council aggregators and panel resellers for government. What margin can we give, and what do we never hand over?
ISO 27001 (A$30K–80K), ISO 42001, a penetration test, IRAP (A$50K–150K+) all open doors, and together could cost A$100K–300K and 6–18 months. ISO 42001 fits a governance-led product especially well. Which door matters first?
In development, an adviser who shapes a programme design and recommends suppliers is expected to declare and manage conflicts, and can be excluded from evaluation. The research suggests: ask in writing what capacity they're acting in, pay no success fees or equity to someone shaping a donor-funded procurement, and encourage solution-neutral designs ("offline-capable, locally hosted AI") so Selma competes fairly rather than being written in and challenged later.
OVIC's block ends 5 November, ADM transparency lands 10 December, the Pacific AI governance framework is due in 2027. Each is a reason for a conversation now. Which ones do we have the capacity to act on this quarter?
A frontier model offered fully on-prem at small-business prices, cloud providers offering verifiable no-third-party processing, or Starlink-level connectivity everywhere would each shift the ground. What are we watching for?