Skip to content

Revise AI strategy document for clarity and focus #27

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
eglerean wants to merge 3 commits into
base: rkdarst/ai-strategy
Choose a base branch
from
Open

Revise AI strategy document for clarity and focus #27

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0bc79bb
Revise AI strategy document for clarity and focus
eglerean May 22, 2026
b238c80
Revise AI strategy document for clarity and accuracy
eglerean May 22, 2026
e2029dd
content/2026/thoughts-on-aalto-ai-strategy: minor fixes
rkdarst May 26, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
187 changes: 107 additions & 80 deletions content/2026/thoughts-on-aalto-ai-strategy.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,122 +4,149 @@
:category: ai


Thoughts on an Aalto AI strategy
================================
Thoughts on a practical AI strategy in research
===============================================

Much has been said about AI strategies for Aalto. Yet, as our
:doc:`last post </2026/what-is-ai>` talked about, AI can be
anything - thus it's hard to write a strategy. This post is about
Much has been said about "AI strategies": increase AI development and adoption in higher education institutions.
Yet, as our :doc:`last post </2026/what-is-ai>` discussed, since AI can be
anything, it's hard to write a useful strategy. This post is about
what we, Aalto Scientific Computing / ASC, think should be considered
in such a strategy.
in such a strategy in research.

Read :doc:`last post </2026/what-is-ai>` if you want some
background.


.. admonition:: General steps for AI strategy
.. admonition:: General steps for AI strategy in research

@Vinye Vinye May 28, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could "General steps" be replaced with something more specific/descriptive? Are these practical and concrete steps universities should implement. Do these have a connection with the three layers described next?


* Minimize the use of the term "AI" except for branding purposes.
* Support AI development and adoption from core methods to standalone AI tools
* Minimize the use of the generic term "AI".
Be specific about intention in communication.
* Providing computing and deep learning specialists who can advise
* Provide computing and machine learning specialists who can advise
on "what AI means" and how to actually incorporate it into
projects (beyond chatbots).
* Lower the barrier to testing open-source AI tools in researchers' workflow


Consumer of AI tools
--------------------
We can think of an AI system as a layered structure. This helps us support the development and adoption of i) new AI methods and AI models, ii) AI components and platforms that turn the methods into building blocks for applications, iii) AI tools that connect the components to the interfaces and workflows used by researchers

@Vinye Vinye May 28, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think "AI system" is the best term to use here. I would describe the subsections as perspectives or use cases or similar.


One side is being a **consumer** of AI products developed by others.
Do we want to buy all the latest tools to generate content? Do we
want to buy tools to help us make decisions (probably wrapped in
proprietary logic?).
Developing AI methods and AI models
-----------------------------------

These tools definitely have some use, given the huge amount of tools
that generate content (writing, code, images, etc). Still, for this,
we are a follower and paying others for the privilege.
Researchers develop new AI methods and AI models, often in the field of machine learning. This is the groundwork
needed for platforms and tools of the future, and is probably what is

@Vinye Vinye May 28, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I find "probably" here a bit vague and hand-wavy. I'd prefer something like "... and is what we believe is meant by..."

@Vinye Vinye May 28, 2026 •
edited
Loading

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Or "... and is what we believe is often meant by..."

meant by "being an AI leader". The revolution isn't just replacing
human work by automation, but being able to see patterns in new
fields.

.. admonition:: Steps for consuming existing AI tools
Method/model development can be the core computer science-y stuff, where you are really
developing a new machine learning method. Or by using ML methods/models in various fields of research,
researchers can develop new ways of looking at research data. Real magic can happen when these two halves work together.

* Any AI strategy should consider under what conditions, and with
what budget, for buying outside tools.
* A plan for future cost increases of AI tools, including
possibility of moving to local resources where appropriate.
* If this is part of the strategy, there should be a way to do
these procurements rapidly, otherwise we are not only a follower,
but a follower of the followers.
* Training with emphasis that "AI" is prediction and not "actual
intelligence" with decisions made opaquely by the provider.
* Support for advanced use (especially for coding), when AI code
generation leads to problems which users can not solve
themselves.
.. admonition:: Steps for developing AI methods

* Good computational resources for development, both powerful and
with good user interfaces for people who aren't computer
scientists or physicists.
* Computing support staff, who can help guide researchers from diverse
fields to the right AI methods/model, computing resources, and software
frameworks.
* Research software engineering support for those applying AI to
new fields.

Implementation: focus on Aalto's Triton resource (designed for more

@Vinye Vinye May 28, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is Aalto-specific. You said you wanted a more general version applicable to other universities, so consider dropping this? Moreover, the other two sections have no implementation suggestions, so this is a bit asymmetric

usability across broad fields) and Aalto Research Software
Engineers (which helps with Triton *and* other platforms).


Developing platforms
--------------------
Developing AI components
------------------------

In addition to procuring tools, we can internally build our own
platforms optimized for our own internal use as tools. We can take
various bases:
Core machine-learning methods can be difficult to integrate in applications: if one researcher is planning to use a local large language model interactively, it would be inefficient to reserve a large amount of computational resources just for a single user. Developing AI components is the key here to enable the adoption of AI methods.

a) purchased AI inference in the cloud (subscription cost but no
capital cost), and
b) open-weight models which we can run fully locally (capital cost but
no subscription cost).
AI components can be proprietary or open source. For example in the case of large language models:

Developing our own platforms has the advantage of providing more local
sovereignty: they have more data and decision-making fully within our
control, including the possibly of running fully on-premises.
Services can be more easily transferred from (a) to (b) in the future
a) purchased AI inference with proprietary models in the cloud (usage costs but no
capital costs), versus
b) local AI inference with open-weight models (capital costs but
no usage costs).

Developing our own AI platforms has the advantage of providing greater local
sovereignty: data and decision-making are fully within our
control. Services can be more easily transferred from scenario (a) to scenario (b) in the future
to control costs and sovereignty, as needed.

.. admonition:: Steps for developing AI platforms

* A local development process which can rapidly iterate and deploy
platforms before they become obsolete. (Current small-scale
development rules are far too slow.)
platforms before they become obsolete with fast review process.
* Investment in the hardware, data, and human resources needed to do
this development. Prepare for a future where cloud AI inference
prices increase and it becomes necessary to move some work to
local hardware.
* A clear understanding of the AI act and the difference between
something using AI under command of a human and a AI system
(which is autonomously making decisions and affecting the
environment).
* Support in how to think about both using and developing AI tools,
instead of being a consumer.

Implementation: Drastic reform the small-scale development process
and broaden Research Software Engineer support to more departments.


Developing AI methods
---------------------
Using existing AI tools
-----------------------

Finally, we develop new AI methods: either deep learning methods, or
applying deep learning methods to new fields. This is the groundwork
needed for platforms and tools of the future, and is probably what is
meant by "being an AI leader". The revolution isn't just replacing
human work by generated content, but being able to see patterns in new
fields.
Finally, we want to empower **users** of AI tools .
There are many popular proprietary tools which embed an AI component and they always come with a price tag that users pay with real money or with the confidentiality of their data.

Method develop can be computer science-y stuff, where you are really
developing a new method (probably for some application but the novel
result is the deep learning). Or it can be for another field,
developing a new method in their field using AI (where the novel
result is in the other field). Real magic can happen when these two
halves work together.
Do we want to buy all the latest AI tools to generate content or automate operations? (probably wrapped in
proprietary logic?). These tools definitely have some use, still, for this, we are a follower and paying others for the privilege.

.. admonition:: Steps for developing AI methods
While it is important to test and use the state of the art proprietary AI tools, an ecosystem of open source tools with a local AI components has merged and it is part of our strategy to support it. You can code using Claude Code and pay real money for tokens (and live with the risk of loss of confidentiality), or use a local LLM endpoint which can perform as well as a proprietary alternative.

* Good computational resources for development, both powerful and
with good user interfaces for people who aren't computer
scientists or physicists.
* Computing support, who can help guide researchers from diverse
fields to the right {AI methods, computing resources, software
frameworks}.
* Research software engineering support for those applying AI to
new fields.
.. admonition:: Steps for adopting existing AI tools

Implementation: focus on Aalto's Triton resource (designed for more
usability across broad fields) and Aalto Research Software
Engineers (which helps with Triton *and* other platforms).
* Any AI strategy should consider under what conditions, and with
what budget, it should encourage buying proprietary tools.
* Procurements should happen rapidly, otherwise we are not only a follower,
but a follower of the followers.
* The strategy should contain plan for future cost increases of AI tools, including
possibility of moving to open source alternatives where appropriate.
* The barrier to using open source AI tools should be as low as possible to ensure transparency
and maintain the confidentiality of tool use
* AI literacy and competence are needed to understand the limitations of these tools: the opaque logic of the proprietary tools, the potential
loss of confidentiality, and the new risks introduced that can erode research integrity and compliance.




.. raw:: html

<pre class="ai-strategy-ascii">
LEVELS OF PRACTICAL AI STRATEGY IN RESEARCH


+------------------------------------------------------------------+
| LEVEL 1: DEVELOPING AI METHODS & AI MODELS |
| |
| core ML groundwork -> new methods/models -> novel research |
| |
| Support: computing resources, RSEs, ML specialists |
+-------------------------------+----------------------------------+
|
v
+------------------------------------------------------------------+
| LEVEL 2: DEVELOPING AI COMPONENTS AND PLATFORMS |
| |
| methods/models -> reusable blocks -> shared services |
| |
| Choices: cloud vs local, proprietary vs open models |
| Aim: sovereignty, scalability, rapid deployment |
+-------------------------------+----------------------------------+
|
v
+------------------------------------------------------------------+
| LEVEL 3: USING AI TOOLS |
| |
| tools -> workflows -> adoption |
| |
| Needs: AI literacy, risk review, open source options |
+------------------------------------------------------------------+


Across all levels: people + hardware + support + competence
</pre>
Loading