Submission 161
Vigilant Labs
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues for sovereign AI capability through refundable development incentives, shared research compute, open technology, provider-independent education and lower startup barriers. Supports eligible overseas compute while rewarding domestic control. The author, a cybersecurity-company founder and practitioner-conference organiser, seeks independent capability for sensitive government functions but avoids detailed classified-infrastructure advice because of limited visibility into existing arrangements.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that the remaining policy gap is the distinction between adoption, domestic data centres and control of the underlying capability.
Australia has a National AI Plan and substantial adoption activity. The remaining policy gap is the distinction between adoption, domestic data centres and control of the underlying capability.
Says a June 2026 US access restriction, subsequently lifted, demonstrates foreign frontier-model access can be interrupted or conditioned by the provider’s home government.
In June 2026 the United States Department of Commerce directed Anthropic to suspend foreign-national access to its most advanced models while safeguards were considered. The restriction was subsequently lifted. This episode demonstrates that access to a foreign frontier capability can be regulated, interrupted or conditioned by the provider's home government at short notice.
Describes smaller firms’ specialised-model businesses as examples while stating their post-trained weights are unpublished and the author has not independently tested or validated the models.
Neither company publishes its resulting post-trained weights, and I have not independently tested or validated their models. Their functioning businesses nevertheless illustrate that a smaller technical organisation can start from accessible weights, build specialist capability and offer a commercial product.
Argues requiring domestic hardware first would delay capability development; the durable asset is an Australian team able to train, evaluate, deploy, secure and move models between providers.
making domestic hardware an absolute precondition would postpone capability development and confuse the location of machines with the location of expertise. The most durable asset is an Australian team that knows how to train, evaluate, deploy and secure models and can move between infrastructure providers.
Proposes applied programs, fellowships, placements and research compute for experienced professionals, with vendor partnerships providing resources without defining the curriculum.
Six- and twelve-month applied programs, fellowships, industry placements and access to research compute would turn existing expertise into capability. Vendor partnerships can provide useful tools and resources, but a commercial provider should not define the curriculum.
Recommends reviewing employee share-scheme concessions so employees do not ordinarily owe tax on startup equity before it can be sold or has a realisable value.
The objective should be simple: employees should not ordinarily owe tax on startup equity before it can be sold or has a realisable value. Government should review the boundaries, documentation and interaction of the concessions with common early-stage structures.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Maintain independently operable sovereign compute and models for intelligence, defence, cybersecurity and other highly sensitive government functions, physically isolated where necessary.
National security must be an essential part of any Australian AI strategy. Allowing intelligence, defence, cybersecurity or other highly sensitive government functions to become operationally dependent on foreign-owned and controlled models would be dangerous. Even where data remains in Australia, access, permissible use, model behaviour, pricing and continuity may remain subject to decisions outside Australian control. Australia must maintain sovereign compute and models for these functions that can operate independently and be physically isolated where necessary.
A program corrected obvious text-recognition errors in this quote for display.
Pursue centralised government procurement to obtain better prices, privacy terms, portability and notice of service changes, especially across government and publicly funded universities.
Centralised government procurement may obtain better prices, privacy terms, portability and notice of service changes. It should be pursued, especially across government and publicly funded universities.
Reward substantive Australian engineering, training and enduring access or use rights; make government procurement a credible first-customer pathway for capable Australian firms.
Government incentives should reward substantive Australian engineering, training and enduring access or use rights, while procurement should give capable Australian firms a credible path to becoming a first customer.
Establish a rules-based, refundable incentive for technically substantive AI development by Australian-based teams, permitting eligible overseas compute, with higher support for Australian-resident or -controlled compute, training Australian personnel, reusable public outputs and priority areas.
Establish a rules-based, refundable incentive for technically substantive AI development performed by Australian-based teams. The base benefit should permit eligible overseas compute so work can begin immediately. Higher support should apply to Australian-resident or Australian-controlled compute, substantive training of Australian personnel, reusable public outputs and changing national priority areas such as cybersecurity or health.
A program corrected obvious text-recognition errors in this quote for display.
Define resident compute as physically processed in Australia under Australian law; an Australian subsidiary alone is insufficient. Require additional Australian control of access, scheduling, deployment, model and software for controlled-compute status.
The rules should distinguish Australian-resident compute from Australian-controlled compute. Australian-resident compute should mean that the relevant processing occurs physically in Australia and is subject to Australian law; purchasing a service through an Australian subsidiary should not be enough. Australian-controlled compute should additionally require an Australian entity to control access, workload scheduling, deployment and the choice of model and software.
Do not give a closed foreign API hosted in Australia the same uplift as infrastructure where an Australian team can run and modify its own model.
A closed foreign API hosted in an Australian data centre should not receive the same uplift as infrastructure on which an Australian team can run and modify its own model.
Include pre-training/continued pre-training, supervised fine-tuning, reinforcement learning, distillation, specialised models, training/evaluation environments, datasets, interpretability, AI security, inference runtimes, accelerator kernels, quantisation and compression as qualifying work.
Qualifying work should include pre-training and continued pre-training, supervised fine-tuning, reinforcement learning, distillation, specialised models, training and evaluation environments, datasets, interpretability, AI security, inference runtimes, accelerator kernels, quantisation and compression.
A program corrected obvious text-recognition errors in this quote for display.
Exclude routine chatbot subscriptions, ordinary closed-API use, standard enterprise integration, production inference, resale and thin wrappers from incentive eligibility.
Routine chatbot subscriptions, ordinary use of closed APIs, standard enterprise integration, production inference, resale and thin wrappers should not qualify.
Assess incentive eligibility across the consolidated corporate group and beneficial ownership, with caps and lower rates for very large firms, using published definitions, prospective guidance, independent technical assessment, milestone verification and frequent payments.
Eligibility should be assessed across the consolidated corporate group and beneficial ownership, with caps and lower rates for very large firms. Published definitions, prospective guidance, independent technical assessment, milestone verification and frequent payments would make the scheme useful to cash-constrained teams while limiting favouritism and gaming.
For major strategic support, consider non-exclusive perpetual government licences for defined project technology; preserve company ownership and commercial freedom, distinguish background/project/future technology, and ensure agreements survive acquisition or transfer.
For major strategic support, government could negotiate a non-exclusive, perpetual licence to use defined project technology while the company retains ownership and commercial freedom. The agreement should distinguish background, project and future technology and survive acquisition or transfer.
Provide shared research compute for universities, independent researchers, students, non-profits and open-source maintainers, potentially procuring from multiple providers; use rapid small allocations and independent milestone, feasibility and public-benefit assessment for larger allocations.
Acomplementary program should provide capacity to universities, independent researchers, students, non-profits and open-source maintainers. Government could procure capacity from several providers rather than own every machine. Small exploratory allocations should use a rapid, low-friction process; larger allocations should be independently assessed against milestones, technical feasibility and the likely public benefit.
Fix shared-compute public-output expectations in advance, potentially including code, weights, datasets, evaluations, negative results or technical documentation.
Public-output expectations should be fixed in advance and might include code, weights, datasets, evaluations, negative results or technical documentation.
Fund, and physically isolate and separately govern, classified Defence and intelligence AI infrastructure rather than having it compete with research and commercial applicants.
Classified Defence and intelligence infrastructure should be funded, physically isolated and governed separately. It should not compete with research and commercial applicants for allocations on a common system.
Fund open-weight post-training, specialised models, evaluation/security tools, inference software, kernels, quantisation, interpretability, reproducible training and governed Australian datasets through grants, prizes and procurement; include companies, universities, independent engineers, maintainers and non-profits.
Fund the layer between imported foundation models and useful sovereign capability. Grants, challenge prizes and procurement should support open-weight post-training, specialised models, evaluation and security tools, inference software, kernels, quantisation, interpretability, reproducible training methods and appropriately governed Australian datasets. Eligible recipients should include companies, universities, independent engineers, open-source maintainers and non-profits.
Do not require all commercial outputs to be open; distinguish research-output and commercial-capability grant conditions. Assess team location/continuity, engineering, independence from closed foreign APIs, deployment control, performance and personnel trained.
Public funding need not require every commercial output to be open. A research grant may require released code, weights, evaluations or documentation; a commercial capability grant may instead require Australian technical work, knowledge transfer or defined government-use rights. Assessment should measure the location and continuity of the team, documented engineering, ability to operate without a closed foreign API, deployment control, performance and personnel trained rather than relying on a 'sovereign' label or nominal IP ownership.
Make open-weight restrictions evidence-based and proportionate, identifying harmful capability, misuse likelihood/scale, existing law, effectiveness and research, competition, cybersecurity and sovereignty effects; foreign frontier laboratories’ preferences should not determine Australian settings.
Open-weight systems can create real risks. Restrictions should be evidence-based and proportionate: government should identify the specific harmful capability, likelihood and scale of misuse, relevant existing law, expected effectiveness and effects on research, competition, cybersecurity and sovereignty. Foreign frontier laboratories offer useful expertise, but their preferences should not determine Australia's settings because they also have a commercial interest in concentrating capability behind controlled services.
Match funding to uncertainty: small exploratory/replication grants, larger milestone grants for demonstrated feasibility, outcome-based prizes, and procurement where government has a genuine operational need.
Different funding mechanisms should be used for different kinds of uncertainty. Small grants can support exploratory implementations and replication. Larger milestone grants can support post-training and infrastructure whose feasibility has been demonstrated. Capability prizes can reward a defined outcome without requiring government to select the technical approach in advance. Procurement can create a first customer where government has a genuine operational need.
Publish qualifying outputs and evaluations before applications open; define open source precisely, specify licence, artefact, documentation, reproducibility, evaluation and limitation requirements where appropriate, and explain staged or controlled release for security-sensitive work.
The government should publish qualifying outputs and evaluation requirements before applications open and define 'open source' precisely. An award might require a specified licence, downloadable artefacts, model or dataset documentation, reproducible instructions, evaluation transcripts and disclosure of material limitations. Security-sensitive work may justify staged or controlled release, with the basis for that treatment explained.
Keep the program broadly open; periodically update priority areas as needs and comparative advantages change, affecting support level or urgency without excluding work outside the list.
The program should remain broadly open while offering additional weight to a periodically reviewed schedule of national capability priorities. Initial areas might include cybersecurity, health and medical research, defence, critical infrastructure, resources, agriculture and public administration. The schedule should change as needs and Australian comparative advantages change. It should influence the level or urgency of support rather than make work outside the list ineligible.
Specify measurable capabilities for priority awards rather than preferred vendors, model families or corporate structures, enabling diverse applicants to compete on technical merit and public value.
Priority awards should specify a measurable capability rather than a preferred vendor, model family or corporate structure. This would allow a startup, university group, established Australian company or collaboration to compete on technical merit and public value.
Review applications by independent technical assessors with declared conflicts, and publish decisions, amounts and non-sensitive outcomes.
Applications should be reviewed by independent technical assessors with declared conflicts. Decisions, amounts and non-sensitive outcomes should be published.
Require recipient reports against technical milestones using domain-appropriate evaluations disclosing uncertainty and limitations.
Recipients should report against technical milestones rather than promotional claims, using evaluations that are appropriate to the domain and disclose uncertainty and limitations.
Review programs against additional Australian technical personnel, reusable outputs, independently demonstrated capability, follow-on investment and continued Australian operation, rather than grant counts.
Programs should be reviewed against additional Australian technical personnel, reusable outputs, independently demonstrated capability, follow-on investment and continued operation in Australia, not simply the number of grants awarded.
Include change-of-control notification and, where appropriate, repayment or licence provisions in funding agreements, while avoiding restrictions that make an Australian company impossible to finance or acquire.
Funding agreements should include change-of-control notification, repayment or licence provisions where appropriate, while avoiding restrictions that make an Australian company impossible to finance or acquire.
Have schools build computational thinking, probability, data literacy and critical evaluation, and TAFE teach accelerator and cluster administration, data pipelines, deployment, inference serving, monitoring and security.
Schools should build computational thinking, probability, data literacy and critical evaluation. TAFE should teach accelerator and cluster administration, data pipelines, deployment, inference serving, monitoring and security.
Have university programs cover model architecture and training, distributed systems, accelerator programming, evaluation, interpretability and AI security, using multiple commercial and open systems so education transfers across vendors.
Universities should cover model architecture and training, distributed systems, accelerator programming, evaluation, interpretability and AI security. Programs should use multiple commercial and open systems so that education transfers across vendors.
Give particular attention to mid-career conversion through six- and twelve-month applied programs, fellowships, industry placements and access to research compute, to turn experienced professionals into AI engineers.
Mid-career conversion deserves particular attention. Experienced software, systems, cybersecurity and domain professionals can become productive AI engineers faster than an entirely new workforce can be created. Six- and twelve-month applied programs, fellowships, industry placements and access to research compute would turn existing expertise into capability.
Keep curricula independent of commercial providers while allowing vendor partnerships to supply useful tools and resources.
Vendor partnerships can provide useful tools and resources, but a commercial provider should not define the curriculum.
Use coordinated public-sector contracting to obtain stronger pricing, privacy, portability, security-notification and research-access terms from major vendors, while funding provider-independent learning materials.
Coordinated public-sector contracting could obtain stronger pricing, privacy, portability, security-notification and research-access terms from major vendors, while public funding maintains provider-independent learning materials.
Support practitioner communities through technical fellowships, recurring workshops, open-source maintainer support, shared laboratories and industry-university placements.
Government can multiply such efforts through technical fellowships, recurring workshops, open-source maintainer support, shared laboratories and industry-university placements.
Review share-scheme concession boundaries, documentation and interaction with common early-stage structures so employees do not ordinarily owe tax before startup equity is saleable or has realisable value.
The objective should be simple: employees should not ordinarily owe tax on startup equity before it can be sold or has a realisable value. Government should review the boundaries, documentation and interaction of the concessions with common early-stage structures.
Create a fast, predictable, time-bound and portable visa pathway for scarce technical AI specialists needed by growing teams, with regularly reviewed priorities.
Australia also needs a fast, predictable pathway for scarce technical specialists needed by growing teams, including distributed-training engineers, accelerator-kernel developers, inference-runtime engineers, reinforcement-learning infrastructure specialists, model evaluators and AI security researchers. Priorities should be reviewed regularly, decisions time-bound and visas sufficiently portable that skilled workers are not trapped with one employer.
Across all programs, provide prospective eligibility guidance, a single contact point, short decision times and regular payment or refund cycles.
Across all programs, prospective eligibility guidance, a single contact point, short decision times and regular payment or refund cycles are as important as headline funding.
Read the original
In some quotes a program corrected obvious text-recognition errors for display.