I'm also surprised that it is not discussed. From what I read/hear are AI startups much faster growing than others. And because ZIRP is over, basically everyone I meet strives for revenue or even cash flow positivity from very early on.
I wish I had the funds and a curriculum expert on my side to build a language learning app with LLMs.
Part of my PhD thesis[1] was to study how robots (voice agents) can influence human language. The key component is a social connection. Back in 2017 I did that in the lab. But the research is pretty clear about it.
Also my own experience (trying to learn Arabic) is, that I only remember words/phrases which I picked up during social gathering (camping in the dessert).
The "perfect" learning app would work like how children learn: by interacting with their social surroundings. No need to learn the vocabulary or the alphabet at the beginning.
The hard part is, to create a social interaction between the learner and the AI that evolves over time.
It's not quite as effective as a real social connection but the para-social connection you get from say listening to a podcast in the language you're learning does work too in my experience.
An agent based simulator with LLMs.
When I did my PhD in 2016 I used Agent Based Simulations to predict how to influence humans with chat agents on a large scale.
The main criticisms with such simulations is always the realism of the agent (and network).
Now, with LLMs, this got much simpler. As LLMs are really good at mimicking humans.
The first application will be to simulate a LinkedIn post against your own (virtual) followers. I grab your actual LinkedIn followers, "turn" each one into an LLM and see how they react.
That's really cool.
One week ago I implemented an SQL tool and it works really well. But sometimes it still just makes up table/column names. Luckily it can read the error and correct itself.
But today I went to the next level.
I gave the LLM two tools. One web search tool and one REST tool.
I told it at what URL it can find API docs. Then I asked it to perform some tasks for me.
It was really cool to watch an AI read docs, make api calls and try again (REPL) until it worked
Not sure if it fits but I still have ~20k currated step by step solution for mathematics (pedagogical math) "lying" around from my previous startup.
They are all hand currated. And could even be used for fine tuning or so.
Here are some details:
The dataset has 20.600 Abstract Exercises which turn into 1.193.958 Concrete Exercises.
An Abstract Exercise looks like this: a + b = c
A Concrete Exercise looks like this: 2 + 3 = 5
Tital compiled file size (JSONL): 11.6GB
very nice! maybe you can put this dataset in some repository like github, kaggle or hugging face, if you are not doing anything with it. Can be helpful to train models.
I'm actually actively looking for something like this. But I'm not sure it fulfills my requirements.
Here is what I'm looking for
- curated open source MCP servers
- trusted curator
- direct oauth integration, so the user permission is taken over (e.g.: you login with your companies Microsoft account and not with the account of the AI-Client.) this is a must for enterprise
- easy to ship (eg.: via electron app)
- I as the developer or the company owner can decide what servers are available to the user, and what tools (e.g.: most companies I talk to want read only access for their employees)
If Klavia fulfills that I would like to have a chat :)
Recently I was fortunate to join a cool startup (as Head of Engineering) that tries to improve the AEC industry by helping them with their paperwork.
So lots of RAG, chat, agents, deep research.
It's really interesting but also challenging. The biggest challenges are: large data; different stakeholders/user stories in one org
This is really great. I'll use it asap.
I'm working with enterprise clients in the AEC space. Having a model that actually understands documents with messy data (drawings, floor plans, books, norms, ...) will be great.
The current situation of chunking and transforming is such a messy situation.
I wonder if it's possible to add such plugins to election apps (e.g.: Slack).
It would be such a nice experience if I could just connect my AI of choice to a local app.
