I don't foresee AGI arising out training bigger LLMs (Though investors won't realise that for a while yet).
It's actually how organic brains work - specialized tasks are offloaded to local cortical columns. The overall coordination between these sub-brains creates emergent skills/abilities.
How are small isolated language models more similar to that than MoE in LLMs?
I feel that would be handy in all sorts of situations when networks are down.
For most actual emergency scenarios, a device that focuses on storage of large amounts of prepared normal reference material [0] will be wayyyyy cheaper, more durable, portable, and able to run on batteries or being constantly plugged into a somehow-still-normal electrical grid. (Think an e-ink tablet that can run off a 5V battery pack buffering a literal handcrank.)
In contrast, imagine spending the money to build a beefy LLM-running computer with good GPU/RAM, and somehow mothballing it (to depreciate, unused) in a "safe" location for the big earthquake/flood/etc... Then when the disaster strikes and you dig it out, how will you power it when you need it, and for long enough to do anything useful?
Even if wall-current civilization is 20 miles away on the other side of the mountain, are you going to carry it on your back, or are you going to carry food and water to live? If you do drag it there, are they going to let you run it when it cuts into light for surgery or heat to sterilize drinking water?
That way your machine that, eg, normally plays video games or does AI work can support relief efforts by supporting emergency response IT. You don’t need to mothball the machine, just have an “emergency” boot USB than can run the services from your home generator.
You don’t even need to bring it with you: turn it on and leave it “best effort” at home, while you continue to use it via WAN.
But OK, let's assume that: The power is out, but you have a generator with so much fuel you can run a desktop just fine; Your neighborhood will somehow make a mesh network; Your neighbors need some already stored information and the best solution for that is texting a chatbot rather than a survival/emergency handbook or Wikipedia; Your mesh-network will also be good enough to match the time-sensitivity of the questions.
Under those assumption, which of these sounds better?
1. Buying an "LLM-in-a-box for emergency supply kits", which you deploy so that your neighbors can ask questions (text over the mesh) of the offline chatbot.
2. Buying a satellite internet transciever for your emergency supply kit, so that your neighbors can ask questions of a much better chatbot and communicate with human experts, their worried relatives, and coordinate with rescue/relief efforts...
I’m only out the cost of the drive, which is like $40 and doesn’t require anybody on the other side cooperate with me.
- - -
More broadly…
You call it unlikely mixes, but we see it all the time:
- people already have a computer for gaming or work
- people (ie, “preppers” like we’re discussing) buy a generator for emergencies
- local emergency response sets up mesh networking during disasters, both official and unofficial
Have you ever tried to use a handbook you’re not intimately familiar with during an emergency? It’s rough.
For personal preparedness, nothing replaces familiarity and practice — eg, weekend survival trips and reading your manual ahead of time.
But for providing information in a random lookup manner to unpracticed people who weren’t prepared? Yes, I think an LLM/chatbot is the practical way to operationalize all that information which you stored (eg, survival guides or machine manuals).
Also, it’s unlikely a general purpose chatbot would be superior at survival advice to one specialized for that purpose — and indeed, is likely to refuse your questions as “unsafe” or “criminal”.
Knowing humans? They'd probably take it by force and run it for themselves instead of providing light and heat to surgeons and water sterilizers...
/daily dose of cynism
But, the current model you really want for an emergency kit is Gemma 4 12B QAT 4-bit. At ~7GB on disk, it's small enough to run on a tablet or any modern computer, slowly if you don't have a GPU or modern Apple silicon, but exceedingly smart for its size, excellent vision capabilities, good tool user, surprisingly good reasoning.