MishMash Opening Conference · Machines session

Building a Norwegian Ecosystem for AI and Creativity

Kilden, Kristiansand · 14 September 2026

Three questions I get

1"How do I get access to train a model on Sigma2?"

Which tools can we use?

2"Can I get funding for Claude Code?"

Which tools should we (not) use?

3"How do I send an avatar of myself from Bergen to Trondheim in 10 milliseconds?"

Which tools do we need to develop?

Three questions I get in my inbox, and they are the whole session in miniature. One: how do I get on Sigma2 to train a model. Short answer: through your institution, in one of two application rounds a year, or by being added to a project by someone who already has one; if you are a freelance artist or a museum, there is no route at all. Click: that is the first general question, which tools can we use. Two: can I get funding for Claude Code. Short answer: a top-tier seat is about 2 500 kroner a month, 200 of them is 6 million a year, and nobody has budgeted it; the sustainable version is pay per use through a gateway, and the honest answer is that for most jobs a smaller model is good enough. Click: which tools should we use, and which should we not. Three: an avatar of yourself from Bergen to Trondheim in 10 milliseconds. Bergen to Trondheim is about 430 kilometres in a straight line, and light in a fibre needs about 2 milliseconds each way before a single router or a single rendered frame; the rest of the budget goes to the model, and nothing on the map today does that for you. Click: which tools do we need to develop. Next slide: the three questions in their general form, and who the map is for.

Three questions · for everyone in Norway

1Which tools can we use?

2Which tools should we (not) use?

3Which tools do we need to develop?

Creative AI is not "traditional" AI

AI for a task

one task, one right answer
batch: throughput matters, latency does not
large, clean, labelled data
measured by accuracy on a benchmark
runs in a data centre, out of sight

AI for creative work

open-ended: many good answers, or none
real-time and interactive: latency is the constraint
small, personal, messy data, often your own
measured by taste, context and who is asking
on stage, in the room, on the body: control matters

Map of the introduction

Strips

Rent or runtwo roads · the bill · tokens or people

Own hardwarethe ladder · memory

Shared computeserver · VDI · Fox · Sigma2 · LUMI

Frontendschat · code · creative · programmatic

Modelswhat runs where · licences

Who needs whatseven work packages

Use or developoff the shelf · new data · new methods

Holes in the mapwho gets access · two ways out

The panelfour angles

Rent or run

The middle road: open models on rented GPUs

The bill

6 million kroner a year

200 researchers on one top-tier agent subscription at 2 500 kroner a month. Nobody has budgeted it.

The arithmetic behind the whole strip. A top-tier seat on a commercial coding agent, Claude Max as the example, costs about 2 500 kroner a month. MishMash has more than 200 researchers. If every one of them had a seat, the centre would spend 6 million kroner a year on one product from one vendor, before a single image, video or music service is counted. No institution and no centre has budgeted for that, and it is the wrong shape anyway: a seat costs the same for the person who uses it all day and the person who tried it twice. Oddmund Møgedal from UiO IT put it plainly to the university this week: tokens will be a limited resource for the foreseeable future, there will never be unlimited access to the best models, and the money for the commercial ones has to come from somewhere, so administrations will start counting. The institutions have to step up. So does the centre. The next slide is the short-term answer that already exists inside the big institutions.

Tokens or people

What	Kroner	The same money as
One seat, one year	30 000	1 seat-year
One PhD year	1 400 000	47 seat-years
200 seats, one year	6 000 000	1.3 fellowships, or 4 PhD years
200 seats, five years	30 000 000	7 fellowships

The budget question in one table. Removed from the slide: agents are cheaper than people per hour; the centre is paid to produce people, open results and a community; tokens multiply those, they do not replace them. A PhD year cost 1 372 000 kroner at the Research Council's last lump-sum rate, for 2025; from 2026 the Council budgets actual personnel and indirect costs, which lands in the same place. So one PhD year is about 47 seats, and a three-year fellowship, about 4.5 million, is about 150. Two hundred seats for a year is 1.3 fellowships; over the life of the centre it is 7. Now the framing. Agents are much cheaper than people per hour of output, and that is true and beside the point. The Council pays MishMash for trained people, open results and a community, and tokens produce none of those on their own. They are a multiplier on the people we have. So the question is not people versus machines but where the next krone does more: on a 201st researcher, or on making the 200 faster. Nobody has an answer yet, and every institution in this room will have to find one this year. Depth below: a proposal for how the centre could decide. Sources in the notes of the next slide.

Pay per use, not per seat

One door

an institution-hosted gateway in front of commercial and open models: UiO GPT, GPT NTNU
chat, shared assistants, files, and now API keys for agents and coding tools
about 50 public-sector institutions run it

Keys by data class

one personal key per data class: red, yellow, green
the key is independent of the model
a monthly quota per model
a unit authorises a quota and pays for actual use

Good enough, not best

the best open-weight models, served on Fox and IDUN, cover most jobs
smaller models win on price per result
the frontier models on a quota when they are needed, without a seat per person

Depth. The short-term answer, from Oddmund Møgedal's presentation to the UiO idea meeting on 9 September. UiO built gpt.uio.no in 2023 as a privacy-safe chat front end where the providers do not know who the user is, and about 50 public-sector institutions now run it. Since 2025 it also serves open-weight models on Fox and on NTNU's IDUN, and the best of those are good enough for most of what people do. The new part is the API: personal keys, one per data class, so a red key only reaches models in the institution's own infrastructure, and the same key works from OpenCode, Claude Code, Claude Desktop, Codex or your own script. Quotas are limiting but necessary: they are what makes it possible to control the cost of the commercial models and to share the open ones. Claude and the more expensive OpenAI models through the API, with quotas and per-use billing, are in a pilot this month and planned for October, for the models available in Europe. The principle is what matters for the map: a unit pays for what its people actually use, across vendors, with as little lock-in as possible, and the person decides which model is good enough for the job. For MishMash the question is who does this for the partners with no gateway of their own: the art schools, the museums, the freelancers.

A proposal to argue with

How a centre could decide

A share, not a line

budget tokens as a share of personnel cost, as equipment always was
3 to 5 per cent of a PhD year is 40 000 to 70 000 kroner
enough for heavy pay-per-use through a gateway

Quotas by role

a PhD in machine learning, a WP leader and a museum partner do not need the same budget
the gateway model, applied to the centre's own money
the unit sees what its people actually use

Not fungible

fellowships are in the contract; tokens come from operating money
so the real trade-off is tokens against a lab engineer, workshops, travel, hardware in the room
50 seats is one engineer who builds the local alternative

Depth. Three rules a centre, or a department, could adopt this autumn. First, budget tokens as a percentage of personnel cost, the way equipment and computing have always been a few per cent on top of a position; at 3 to 5 per cent of a 1.4 million PhD year that is 40 000 to 70 000 kroner a year per person, which covers heavy pay-per-use through a gateway and is a decision the centre can make now. Second, quotas by role rather than seats by headcount, which is the gateway model from the previous slide applied to the centre's own money. Third, remember that the money is not fungible: fellowships are earmarked in the contract with the Research Council, and token spend comes from operating money, so the real trade-off is tokens against the other things operating money buys. A lab engineer costs about 1.5 million a year, which is 50 seats, and I would argue the engineer who builds the small, local alternative is the better buy. Sources: the Research Council's rates page for fellowships (1 372 000 kroner in 2025, actual costs from 2026) and Anthropic's published Max pricing (200 dollars a month).

Six questions to ask of any tool

Which model?

Name and version. "AI" is not an answer.

Where does it run?

Your machine, your institution, an EU data centre, the US.

Where do my data go?

Prompt, files, recordings. Stored? Trained on?

Who owns the output?

And can you sell a work made with it?

Same result next year?

Reproducibility for research; continuity for a repertoire piece.

What does it cost?

Per month, per hour of GPU, per kilowatt-hour.

Your own hardware

Seven rungs: microcontroller, single-board computer, phone, tablet, laptop, laptop with GPU, desktop or workstation, with typical uses and price bands

Model families placed on the rungs: TinyML and RAVE on the left, 1 to 8B language models in the middle, 70B on the right

What each rung is good for

Microcontroller

keyword spotting, gesture classes
sensor fusion, milliseconds
no LLM, ever
wearables, instruments

Pi / Jetson

small vision and audio models
Whisper tiny, YOLO
an embodied agent in a box
installations that run for a year

Phone / tablet

1–3B on-device LLMs
camera, mic, sensors
closed platforms, app stores
the device every pupil has

Laptop

7–14B chat, offline
Whisper transcription
RAVE and nn~ on stage
the studio in a bag

Laptop with GPU

Stable Diffusion, FLUX
fine-tune small models (LoRA)
real-time everything
hot, loud, one hour of battery

Workstation

70B local, video generation
train RAVE overnight
serve a whole studio or class
needs a room and a power bill

School tablet or laptop

managed device, no GPU
Feide login, pupils' data
everything must be hosted
the most constrained rung

Rule of thumb

memory decides what fits
compute decides how fast
prototype on the smallest rung that works

Shared compute

Five steps: shared server, VDI, institutional HPC such as UiO Fox, national Sigma2 with Olivia and KI-fabrikken, European LUMI

Arrows: to the right more compute and sharing, to the left more waiting, paperwork and less interactivity

Shared server: a GPU box in the lab, ssh, first come first served. VDI: a virtual desktop for licensed software and sensitive data, rarely a real GPU. Institutional HPC: UiO's Fox, NTNU's IDUN, UiA's DGX servers at CAIR, local clusters at UiB and UiT. Slurm, batch jobs, free for the institution's own people, students through a supervisor, collaborators added by a project leader. Most other institutions have their own small servers or nothing. National: Sigma2 and NRIS. Olivia, 448 Grace Hopper chips after the February expansion, plus Saga; application rounds twice a year. KI-fabrikken, opened last year, is the first national channel that explicitly includes business and public sector. European: LUMI in Finland, where Norway is a member, and the EuroHPC AI Factories. Click: what changes as you go up. More compute, more people can share it, data can live next to it. But more waiting, more paperwork, less interactivity. Batch scheduling is the wrong shape for a rehearsal.

How you actually get on

Institutional clusters

UiO Fox: Educloud project, PI adds collaborators
NTNU IDUN: supervisor requests access
UiA: DGX servers at CAIR
UiB and UiT: local HPC via the IT division

NRIS / Olivia

project application, CPU and GPU hours
two rounds a year
any Norwegian research institution
data storage next to compute

KI-fabrikken

Sigma2, part of the LUMI AI Factory network
research, business, public sector
expertise and support, not only GPUs
artists and cultural institutions: unclear

LUMI / EuroHPC

through NRIS or EuroHPC calls
AMD GPUs: check your software
large allocations, long lead times

NAIC (paused)

the low-threshold national GPU service
funding ended 31 March 2026
"no-cost hibernation" since April
successor being procured by Sigma2

Sensitive data

TSD at UiO, SAFE at UiB, HUNT Cloud at NTNU
GPU capacity inside is limited
health and pupils' data live here or offline

Institutions side by side · to be filled in

Institution	Own GPU cluster	Hosted chat / LLM service	Sensitive-data environment	Students get GPUs?	Freelancers / partners get in?
UiO	Fox	UiO GPT	TSD	via project	via a PI
UiB	local HPC (UiB IT)	Copilot	SAFE	?	?
NTNU	IDUN	GPT NTNU	HUNT Cloud	via supervisor	support agreement needed
UiT	local HPC group; hosts NRIS systems	?	?	?	?
UiA	DGX-2 and DGX H100 (CAIR)	?	?	?	?
INN · HiØ · HVL · OsloMet · Nord · Kristiania …	?	?	?	?	?
NMH · KHiO · AHO	none known	?	none	no	?
National Library	DH-lab	?	?	n/a	?
Simula · SINTEF · IFE	eX3 and own	?	?	n/a	?

Frontends

Chat: the web apps, institution-hosted services, Open WebUI and LM Studio

Code and agents: VS Code with assistants, Claude Code, Cursor, Jupyter

Creative tools: Max with nn~ and FluCoMa, Pure Data, Ableton, TouchDesigner, ComfyUI, Unity, p5.js

Programmatic: Python with PyTorch, llama.cpp, MLX, Ollama, vLLM, any OpenAI-compatible endpoint

One script, one Max patch, one notebook, with a base URL to change

Backend: the laptop with Ollama, LM Studio or llama.cpp

Backend: a lab server with vLLM or Open WebUI

Backend: Fox or Olivia with vLLM behind a tunnel

Backend: an EU provider such as Mistral, Scaleway or Hugging Face

Backend: a commercial API from OpenAI, Anthropic or Google

Tool is not model

Tools: ChatGPT, UiO GPT, VS Code, Max with nn~, ComfyUI, Ollama

Models

Language

Llama · Mistral · Gemma · Qwen · DeepSeek · Phi

Norwegian: NorMistral (UiO) · NorLLM (NorwAI/NTNU) · NB-GPT

Speech

Whisper · faster-whisper · Piper · XTTS · Kokoro

Norwegian: NB-Whisper (National Library)

Music and audio

RAVE (IRCAM) · MusicGen / AudioCraft · Stable Audio Open · ACE-Step · YuE

Norwegian: none

Image

Stable Diffusion · SDXL · FLUX

Norwegian: none

Video

Wan · HunyuanVideo · LTX

Norwegian: none

Vision, motion, multimodal

CLIP · SAM · LLaVA · Qwen-VL · MDM · MotionGPT

Norwegian: none

Licences in three colours

Anything goes

Apache 2.0 · MIT

Mistral, Mixtral · Qwen3 · DeepSeek · Phi
NorMistral (UiO) · NB-Whisper
Whisper · RAVE · CLIP · SAM
FLUX.1 schnell · Wan 2.1 · YuE · ACE-Step
use it, sell the work, ship the product

Conditions

Community licences and terms of use

Llama: "Built with Llama", 700M-user cap, derivatives named Llama
Gemma: prohibited-use policy passed on
Stability community licence (SD3.5, Stable Audio Open): free under $1M revenue
CreativeML OpenRAIL-M (SD 1.5, SDXL): use restrictions travel with the weights
NorwAI NorLLM: Nordic organisations

Non-commercial

Research only, CC-BY-NC

FLUX.1 dev: non-commercial licence for the model
MusicGen: code MIT, weights CC-BY-NC 4.0
many research checkpoints on Hugging Face
fine for a paper; a commissioned work is commercial

Depth. Three colours with concrete names. Green: permissive, do what you like, keep the notice. Amber: community licences and terms of use; usually fine for artists and researchers, but with strings: attribution, revenue caps, naming rules, use restrictions that travel with the weights. Red: research-only or non-commercial; fine for a paper, a problem for a commissioned work. Two traps: the code and the weights can have different licences, as with MusicGen, and licences change between versions, so read the card of the exact checkpoint. Licences: RAVE MIT; MusicGen weights CC-BY-NC; Stable Audio Open and SD3.5 Stability community licence with a one-million-dollar revenue cap; FLUX.1 schnell Apache 2.0 but FLUX.1 dev non-commercial; Llama community licence; Gemma terms of use; Qwen3 Apache 2.0. Rule of thumb, not legal advice.

Licence compatibility: what you can do with what

Base model licence	Research paper	Commissioned work, sold	Product or service	Release your fine-tune	Mix with a permissive model	Train a new model on its outputs
Apache 2.0 / MITMistral, Qwen3, NorMistral, Whisper, RAVE, FLUX schnell	●	●	●	● any licence, keep the notice	●	●
Llama communityLlama 3, Llama 4	●	● "Built with Llama"	◐ under 700M users	◐ same licence, name starts with Llama	◐ the mix inherits Llama terms	◐ allowed, new model named Llama
Gemma terms of useGemma 2, 3	●	●	◐ prohibited-use policy	◐ terms passed on	◐ the mix inherits the terms	◐ check the terms
Stability communitySD3.5, Stable Audio Open	●	◐ free under $1M revenue	◐ same cap, else enterprise licence	◐ same licence	◐ the mix inherits the cap	◐ check the terms
OpenRAIL-MSD 1.5, SDXL	●	●	◐ use restrictions apply	◐ restrictions passed on	◐ restrictions travel with it	◐
Non-commercialFLUX.1 dev, MusicGen weights, CC-BY-NC	●	◐ read it: outputs sometimes yes, the model no	○	◐ non-commercial only	○ the whole release becomes non-commercial	○ usually forbidden

Where to find them, how to trust them

Find

Hugging Face: weights, model cards, datasets
Ollama and LM Studio libraries: one-click local
IRCAM Forum: RAVE and nn~
the paper's GitHub, for the newest

Trust

model card: training data disclosed?
evaluation: on what, by whom?
reproducible: pinned version, checksum
energy: training and inference cost stated?

Who needs what

Axes: compute needed from laptop to LUMI, and data sensitivity from own material to personal and health data

Use cases placed on the axes: musician on stage, filmmaker, music therapist, teacher in a school, cultural institution, National Library, designer or rescue operator, and a computer science researcher who trains models across the work packages

Two axes: how much compute, and how sensitive the data. Click. WP1, a musician on stage: milliseconds, a laptop in the room, own recordings; bottom left. WP2, a filmmaker: video generation, hours on a GPU, and the rights of the training data are the issue. WP3, a music therapist: small models but health data, so TSD or fully offline; top left. WP4, a teacher: no GPU at all, a managed tablet or laptop, pupils' data, and a need for free Norwegian tools; far left, high up. WP5, cultural institutions and rights organisations: production-grade, contracts, audit trails. WP6, the National Library: a multimodal model on the whole collection, petabytes, Olivia or LUMI for months; far right. WP7, designers and rescue operators: physics and law built in, often on an embedded device. And the computer science researcher who builds the models for all of them: a lot of compute, mostly open or synthetic data, bottom right. No single machine serves this picture. That is why there is no common solution.

Musician: record own material, train RAVE overnight on a studio GPU or Fox, play with nn~ in Max on the laptop

Filmmaker: sketch with an open image model, generate with open video models on a workstation or EU provider, cut with provenance documented

Teacher: browser only with institution-hosted chat, offline Whisper on the teacher's laptop, free Norwegian GDPR-safe tools in the classroom

National Library: prepare petabytes, train a multimodal Norwegian model on Olivia or LUMI, release open weights that the musician downloads

Work Package Needs

WP	Typical machine task	Data	Compute	Runs today on	The hole
1 Performances	real-time co-creative agents, RAVE, gesture models	own recordings, sensors	ms; laptop, board, one GPU to train	laptop, studio GPU, Fox	open real-time models; training time without paperwork
2 Processes	image, video, text, audio generation	rights-encumbered	hours on a workstation	commercial tools, workstation	rights-clean open models; EU hosting
3 Health	small models on sessions and signals	health, personal	small, inside a secure zone	TSD, offline laptops	GPUs inside secure zones; consent-ready pipelines
4 Education	AI literacy, classroom tools	pupils' data	none locally; hosted	tablets and laptops, Feide, hosted chat	free, Norwegian, GDPR-safe creative tools
5 Industries	production generation, attribution, audits	contracts, identity	hosted, reliable, auditable	commercial APIs	cultural institutions in KI-fabrikken; open-provider contracts
6 Heritage	transcribe, classify, link, train on the collection	petabytes; copyright, public duty	national or European HPC	NB DH-lab, Olivia, LUMI	multimodal Norwegian models; sustained allocations; storage next to compute
7 Problem-solving	constrained generation, embodied systems	industrial, safety-critical	edge to workstation; simulation	partner systems, Fox	constraint-aware models; access for industry partners

Researcher trains on Olivia or LUMI, releases open weights, musician, teacher and filmmaker download to a laptop

Use or develop

Use off the shelf: a finished model on any device; minutes; everyone

Develop it yourself, first job: train an existing architecture on your own data; hours to days on one GPU

Develop it yourself, second job: new methods; many failed runs; time unknown

Use off the shelf: a finished model on whatever device runs it, minutes to set up, everyone on the map. Click: developing it yourself is a different job, and it forks. Click again: the second fork. Using a model off the shelf is what almost everyone on the map does, and the hardware question is only what runs inference. Developing is different in kind, not just in degree, and it forks. Training an existing architecture on new data, a RAVE model on your own recordings or a LoRA on your own footage, is bounded: you can estimate the hours, and most of the effort goes into the data and its rights. Developing new methods, new architectures or new ways of controlling a model, is open-ended: many training runs, most of which fail, and you rarely know how much compute you need until you are done. That last point matters for the ecosystem, because allocation rounds assume you can write down the number in advance. Depth slides below: inference versus training setups, why workflows do not travel between systems, then a rough tasks-against-platforms table and the same map by person. The last two are hard to generalise; use them only if the discussion needs them.

Inference is not training

Serving a model

always on, answers in milliseconds
a web endpoint that a class, a stage or a museum connects to
a modest GPU, or many of them, or a laptop
needs networking, logins and a security review

Training and fine-tuning

bursts: hours to weeks, then nothing
no endpoint, no users, a queue
the biggest GPUs and the fastest storage
needs scheduling, checkpoints, data next to compute

Workflows do not travel

Laptop to Fox

Different modules, versions, paths. Days of setup.

Fox to Olivia

Same scheduler, different drivers, containers and rules.

Olivia to LUMI

NVIDIA to AMD, CUDA to ROCm. Often a rewrite.

What helps

Containers and pinned environments. Teach it at student level. AI assistants make porting cheap.

Tasks against platforms: a rough guide

Task	Commercial API	Laptop	Laptop + GPU	Workstation	Fox-type cluster	Olivia / LUMI
Chat, writing, code	●	● 7–14B	●	● 70B	◐ overkill	◐ overkill
Transcription (Whisper)	●	●	●	●	◐ batch	◐ overkill
Image generation	●	◐ slow	●	●	◐ batch	◐ overkill
Video generation	●	○	◐ short clips	●	◐ batch	◐
Real-time audio on stage	○ latency	●	●	●	○	○
Train a small model (RAVE, classifier)	○	◐ slow	●	●	●	◐ overkill
Fine-tune an LLM (7–14B)	◐ data leave	○	◐ LoRA	●	●	●
Train from scratch (NB-scale multimodal)	○	○	○	○	◐ small	●
Sensitive data (health, pupils)	○	● offline	● offline	● offline	◐ TSD-type only	○

By person: a plausible stack today

Who	Everyday	Creative work	Training	The catch
Musician on stage	laptop, local chat	Max + nn~, RAVE	studio GPU overnight, or a friend with Fox	no institution: no Fox, no NRIS
Filmmaker	commercial tools	ComfyUI, open video models on a workstation	LoRA on own footage	rights of the base model
Teacher in a school	Feide-gated hosted chat	browser tools only	none	no Norwegian, GDPR-safe creative tool exists
Music therapist	offline laptop	small local models	inside TSD, if GPUs	consent and secure-zone capacity
Bachelor student	Ollama + Open WebUI	Whisper, notebooks	Colab-type, or a course allocation	depends entirely on the institution
PhD / ML researcher	laptop	prototypes locally	Fox, then NRIS, then LUMI	allocation rounds and the paperwork
Cultural institution	commercial APIs	hosted open models with a contract	KI-fabrikken, if they qualify	nobody has defined them as a user category
Startup / studio	commercial APIs	own GPUs or cloud	KI-fabrikken as "business"	cost, and no research partner
National Library	own lab	own models	Olivia, LUMI	sustained multi-year allocation

Holes in the map

Who	Own hardware	Institutional servers, VDI, Fox-type	National HPC NRIS, Olivia	KI-fabrikken	Hosted open models EU providers	Commercial APIs
Researchers, big universitiesUiO · UiB · NTNU · UiT	yes	yes, varies	by application	yes	own budget	own budget
Researchers, other universities and collegesUiA · INN · HiØ · HVL · NMH · KHiO · AHO · Kristiania …	yes	often little or none	by application	yes	own budget	own budget
Research institutes	yes	own labs	by application	yes	own budget	own budget
Students, bachelor to PhD	own machine	via a supervisor	via a project	?	rarely	own budget
Teachers and schools	tablets, laptops, no GPU	no	no	no	no	Feide-gated, limited
Freelance artists	if affordable	no	no	?	own budget	own budget
Public sectormuseums · libraries · NRK · municipalities	yes	own IT, no GPUs	no	"public sector"	budget, procurement	budget, procurement
Private sectorstudios · film · games · startups	yes	own	only with a research partner	"business"	budget	yes

Data: open by count, closed by volume

Open, with an API

Wikidata, KulturNav, the authority files
half of DigitaltMuseum, Nasjonalmuseet
Kartverket, SSB, Stortinget, Brønnøysund
Språkbanken: thousands of hours of speech

Closed, by the hour and the page

the library's books, newspapers, radio, TV, web
NRK, TV 2, NTB, Retriever
Bokhylla, Filmarkivet
nothing open for music, film, video, dance, theatre, games, design

A survey of about 120 Norwegian sources this week. Counted in downloadable objects, Norway looks open: Wikidata, KulturNav, half of DigitaltMuseum, Nasjonalmuseet, the public registers, and Språkbanken, which makes speech the only well-served creative data type. Measured in hours and pages, Norway is closed: the National Library's largest holdings are licensed for reading on site or from a Norwegian address, not for mining, and the broadcasters and news archives are all rights reserved. The reason is not the licences but the law: Norway still has no text-and-data-mining exception. Sweden and Finland have had one since 2023. The bill is in the Storting's family and culture committee now, with written input due 28 September and a report due 3 November. The one exception is the newspaper agreement, which lets the library alone train on newspapers for 45 million kroner a year. And for the creative data types this centre cares most about, there is no open Norwegian dataset at all. Compute without data is a machine with nothing to learn from.

Two ways out, and the elephant

National: quotas through one door

the gateways and clusters of the big institutions, opened to everyone else on the map
a quota per person, paid per use by the unit, across vendors
frontier models when they are needed, open models for the rest
quota categories for artists, schools and museums, not only researchers

Local: small and smart

small open models on the laptop, the board, the phone
several small agents, one job each, instead of one large model for everything
energy and cost in the room, where you can see them
for the stage, the classroom, the museum

The elephant: the bill that matters is not 200 researchers. It is generated feeds and the rest of the consumption nobody asked for. I would argue a centre for AI and creativity has to confront that, and show what else is possible.

Two extremes, and the answer is somewhere between them, but the extremes are what to argue about. The national road builds on what the big institutions already have, the gateways and the clusters, and opens them to everyone on the map through quotas: a quota per person, a unit that pays for actual use, the frontier models when they are needed and open models for the rest. The missing piece is the categories: nobody has defined an artist, a school or a museum as a user with a quota. The local road is the opposite bet. Small open models on hardware in the room, several small agents doing one job each rather than one large model doing everything, and energy and cost visible on the desk. This is Stefano's world, and it is the direction where a centre can contribute methods rather than money. Short term, only the national road scales to 200 researchers by next year. Long term, only the local road makes the centre independent of any vendor. Click: the elephant. I would argue the consumption that is not sustainable is not researchers running agents. It is generated feeds, synthetic images and video for social media, the mass of output nobody asked for. If tokens are a scarce resource, that is a question of who gets them. A centre for AI and creativity should say so, and demonstrate the alternative: small, local, deliberate, made by people who mean it.

Observations to argue with

The academic infrastructure is built for batch science, not for a rehearsal room.
The people who most need cheap, controllable, rights-clean models have the least access.
Access depends on which institution you sit in, not on what you need.
Open Norwegian models exist for language and speech, and for nothing creative.
The whole Norwegian cultural heritage sits in one collection. Nobody has trained a multimodal model on it.
Nobody has defined "artist", "teacher" or "cultural institution" as a user of national compute.
The low-threshold national service (NAIC) is paused; its successor is not yet here.
Workflows are locked to the system they were built on. Fox is not Olivia, and Olivia is not LUMI.
When big systems are bought, the users are rarely asked first.
Norway's data is open by count and closed by volume. The library is licensed for reading, not mining.
No text-and-data-mining exception in law. The bill is in committee; input is due 28 September.
200 researchers on a top-tier subscription is 6 million kroner a year. Per seat is the wrong shape; per use through a gateway is the right one.
There will never be unlimited access to the best models. Good enough, on a quota, is the plan.
One PhD fellowship is 150 seat-years. Tokens multiply people; they do not replace them.

Depth. Fourteen observations, each one an invitation to disagree. The last three are on money, from this week's UiO idea meeting: quotas are limiting but necessary, and the money has to come from somewhere. The two before them are from the data survey: the previous slide in one line each, and the one thing the room can do this month, which is write to the committee. The two on portability and procurement came from the panel: portability, which is a skills and vendor problem more than a hardware problem, and procurement, where the people who will use a system are rarely in the room when it is specified. NAIC was in part an attempt to turn that around and put users first; it did not survive. Worth asking the room who was consulted before the last big purchase at their institution. The fifth is the long ambition: the National Library holds text, newspapers, radio, television, film, music and photographs in one place, and its AI lab has already trained speech and language models on it. A multimodal Norwegian model on the whole collection, released openly, is the kind of thing only a national effort can do. Rights come before compute. This is an ask for after the conference, not for today.

Over to the panel

1Which tools can we use?

2Which tools should we (not) use?

3Which tools do we need to develop?

Stefano FascianiUniversity of Oslo, Department of Musicology

Anna-Maria ChristodoulouUniversity of Oslo, RITMO

Enrique EncinasOslo School of Architecture and Design

Keith MellingenVRINN, immersive learning cluster, Hamar

Four angles that together cover the map: the stage, the lab, the design studio, and a regional industry cluster. Stefano Fasciani is Professor of Music Technology at the Department of Musicology, UiO: sound and music computing, applied machine learning, real-time embedded systems, with a background in electronic engineering and the club scene. Anna-Maria Christodoulou is a doctoral research fellow at RITMO, UiO, working on multimodal music information retrieval and intelligent systems for music analysis. Enrique Encinas is Associate Professor of Interaction Design at AHO, teaching AI, creative coding and interactive spaces, with a design-research practice on humans, other-than-humans and technology. Keith Mellingen is Cluster Manager of VRINN in Hamar, a cluster of Norwegian companies working with virtual and augmented reality and gamification for learning and training, with joint projects in health, crisis management and tourism. Five minutes each. Then we argue, then the room fills in the holes, and we leave with three asks for the closing panel at 17:00. Removed from the slide: Moderator: Alexander Refsum Jensenius. Five minutes each, then conversation, then the room, then three asks for MishMash. Then the room: what are you using today, and what do you need that you cannot get? The same two questions are in the participant survey; put the link and QR code on the screen during the room segment.

What do you use, and what can you not get?

nettskjema.no/a/649352