The future looks like the past: What DeepSeek means for the AI Ecosystem

TL;DR… A not to distant future could see a major shake up to SaaS based Edtech AI making many redundant: replaced with locally installed and created agents, using a fine-tuned private edu AI that is made accessible through an AI wrapper specifically designed and tailored for educational use. A precursor of this exists today in Mecha.

If your believe that AI is going to be a significant factor in k-12 educations evolution, then you should also believe in creating agency for education organisations (schools and systems), which can advocate for and activate their teacher and students’ local needs. To say it is out of our control or that other forces will will provide, is abdicating responsibility and setting our teachers up to fail, increasing their work load, and not learning from the Social Media storms.

There is a need for an education view of LLM deployment - not just use. It needs to be driven from and match education system needs and take advantage of market shifts such as what DeepSeek has pointed to. One with some local control set within a common set of needs that could be shared across many schools.

I built Mecha as a LLM management layer designed for education. It is locally installed, under the control of education organisations, with privacy and sharing of your own edu Apps, agents and automations built in. It aims to give schools the same (or better) capabilities as edu SaaS tools, without the cost or privacy impacts.

We are wondering if an Open Source AI Management Layer for LLMs that simplifies what is needed to make something like DeepSeek (and other commercial LLM’s) operational for education should be made available? 

SaaS software trades off localisation, control and privacy for cost savings and capabilities. DeepSeek marks a shift in the trade-offs being made - maybe reaching a tipping point for AI in education sooner than we thought. We may be at the point already where the cost and convenience of SaaS no longer outpaces that for installing and managing the full stack in your cloud on top of your Azure or Google Cloud Platform. 

So its conceivable that in less than two more years (yep its only two years) that this DeepSeek moment spawns an open source model good enough, that it could be used in house with an education layer to support teacher and student use.

That means data stays local to protect students, costs can be controlled, and an education layer can be added which responds to local needs.

The full stack involves a LLM education management layer (or LLM Wrapper) which can offer what SaaS or co-pilots do for education plus:

• comprehensive management services for costs;

• LLM selection;

• content moderation;

• usage based cost management;

• class based view of student AI use;

• prompt management;

• agent creation;

• RAG prebuilt with Australian Curriculum including evidence based resources to guide the LLM; and

• outputs streamed into your existing LMS or portals.

So its a lot, but most of the above is already available to be installed.

But why do this?

This is needed because education is different to many commercial activities and the optimisations being developed for the majority of general LLM uses means they are not optimised for local educational needs. This will also be the case for open source LLMs also.

There is a common set of features needed across education which is not currently provided by standard productivity tools, which lower the bar for teachers to be able to get benefit from use AI themselves and to be able to do so safely with their students. For example... An agent to: answer questions about a course; provide feedback based on a quality rubric; differentiate content. Or a tool to give teacher feedback on where there lesson aligns with explicit practice; or how they might better align with curriculum standards; or providing teachers advice on how to make an assessment more resilient to AI.

These are common. As part of projects like CEhat, and working with schools, it appears there is a finite set of these. Schools are currently looking at SaaS alternatives from the US or other internationals, when there are examples here of how to build and share similar tools right here in Australia, and why not open source. It mostly plain english right?

There are AI leaders in schools across our systems who have experience, so what we need is a system to share and leverage the AI knowledge. Making it simple for safe and effective use for the next group of teachers new to AI, giving them the template for effective application of AI in teaching rather than learning all intricacies LLM’s (we don’t ask all teachers to learn to code, but they do get benefit from well designed applications).

Now we have the opportunity to design, build and share AI applications because they are (mostly) coded in plain language and we have seen experts across education who are ready and willing to develop what works.

They need an AI management layer that take care of the risks, costs, safety and can operate in house - select open source LLM’s if thats the right thing or commercial, without the teachers having to care. Find the lowest cost, most effective way to achieve that task and share the education tools that make use of that management layer.

There are many potential benefits from such an approach. The rise of open source models paves the way for educator-led fine-tuning that can be shared amongst groups of schools or across school systems. This is built into Mecha through sharing agents, automations and starter prompts for common educational activities. 

Deepseek is a signpost, back to when schools and systems choose how they wanted edtech to be deployed.

The two key components to enable this shift - high quality Open Source LLM and AI management layer for education - exist right now. Mecha was built with this end in mind and it behoves us to have a crack at taking control of our own AI destiny.

AI Generated Image