Small AI Models Gain Traction In places with unreliable networks | Hacker News Reader

Small AI Models Gain Traction In places with unreliable networks

spectrum.ieee.org

277 pointssscaryterry82 comments

I strongly believe this premise in the article is correct - we will see a lot of tiny, hyper specialized models for individual tasks, and perhaps that will converge with an orchestration layer for a generalized intelligence that controls these specialized tiny models, that will be quite capable.

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.

What about recent models providing correct proofs to open math problems?

What about it?

I haven't tried it, but I saw Leanstral, an LLM specialized in writing Lean proofs, posted on HN recently and it claims to outperform some larger general purpose models. It didn't beat Claude Opus, but it seems to do decently at one tenth the cost. It's plausible that further research could yield other models that are smaller and more effective at limited tasks, reversing the trend of ever growing models.

> It's actually how organic brains work - specialized tasks are offloaded to local cortical columns.

How are small isolated language models more similar to that than MoE in LLMs?

Right MoE is a tradeoff between efficiency and intelligence.

MoEs don't route like most people imagine. They aren't learning topic based experts despite the name

The original Mixtral paper [0] (in the "Routing analysis" section) found:

"surprisingly, we do not observe obvious patterns in the assignment of experts based on the topic"

A quick skim of more recent analysis on MoE shows that this hasn't changed. MoE models do appear to work, but don't appear to do what the name implies, if anything they're routing based on the structure of the text and not the semantic content (and we're still not entirely sure what they're doing).

0. https://arxiv.org/pdf/2401.04088

General purpose models are always more robust and generally better than smaller narrower models. My bet is that compute will catch up and any “small” model will still be generally capable, just smaller than sota, rather than intentionally narrow. The exception would be for very well defined tasks where the data distribution never varies, but these are rare and don’t really need “AI” anyway when they do exist.

> General purpose models are always more robust and generally better than smaller narrower models.

What do you mean with more robust?

Less weird unexpected failures, more innate ability to handle edge cases gracefully. Quite important when you're running high on automation and low on oversight.

This may be speculative, but couldn't robustness emerge by having a number of specialized models, that are interconnected and feed into each other? Are there any arguments from ML that would speak against this?

That does work. Even if you drop the "specialized" part. Ensembles of the same architecture at the same scale trained on the same data do outperform a singular model of the same line - especially on corner cases. Successes of an ensemble correlate stronger than failures do.

The usual argument against is that if you have "a number of specialized models" that perform well in ensemble, you can take that ensemble, and distill it into a single larger model (dense or integrated sparse, like MoE), and get the same improvement in performance with an efficiency win.

This works because having those "specialized models" duplicates a lot of the highly conserved "low level" wiring that's required for a model to function at all. As such, you end up running a small scale version of the same "backbone" computational processes many times. "Merging" those models into a larger, denser model allows for a singular strong "backbone" to be used for everything.

> General purpose models are always more robust and generally better than smaller narrower models

I feel like this is just the marketing conflation of AI=LLM, versus regular old ML? We're never going to need to deploy a full reasoning model on a low-power device just to do some fancy image recognition in the field. Specialised ML models are just intrinsically able to be a lot more efficient than their generalist equivalents

You're getting downvoted, but you're completely right. There are very few cases in which narrowing a model down is buying you anything worthwhile.

It seems like for LLMs, "general intelligence" is expensive, but "one more domain" is fairly cheap.

The demand for smaller models and single purpose AI agents will only get bigger in the future. It’s no different than going for mainframe computers back in the day to personal computers. It’s inevitable…..

Has nothing to do with demand, it’s a question of the most efficient way to train a good model. Narrower models are in general worse than general purpose ones, so you’re better using the largest general purpose model that fits in your compute budget vs trying to somehow remove capability or knowledge and assume the model stays as capable for the task you want.

Depends on how you measure "better".

"Can run fully on local hardware without an internet connection and doesn't come with a monthly subscription" makes local models "better" in a way a LLM can never match.

No this will never work. Domain specific models will never be a thing because intelligence carries over and compounds.

Why didn’t OpenAI release a math specific model? Why not a literature specific one? Why do they instead have generic models of different sizes? And how did all labs converge on this?

Why does Fable just not train on non cybersec and non biology data but instead have clearly costly and annoying classifiers?

DeepMind did release a math specific model. And OpenAI has released a coding specific model.

The answer to your question is “because the market isn’t big enough”, not because it doesn’t work. Why would knowing about 2019 internet memes help you in any way at coding?

> And OpenAI has released a coding specific model

They did and retracted it because they found that GPT 5.5 beat codex pareto optimally. This keeps happening.

> because the market isn’t big enough

Huuh? market isn't big enough for AGI? The parent suggested that AGI would emerge from this process.

> Why would knowing about 2019 internet memes help you in any way at coding?

https://github.com/Brainrotlang/brainrot

"Brainrot is a meme-inspired programming language that translates common programming keywords into internet slang and meme references."

> Why would knowing about 2019 internet memes help you in any way at coding?

99.99% of the knowledge an LLM has is useless for a given scenario, the hard part is knowing what the .01% that’s needed is. Knowing as much as it can means the model can handle edge cases, turns of phrase, etc.

Put another way, it avoids overfitting. That’s basically the insight that’s given way to the current AI boom.

Your examples (math, literature) involve natural language. It stands to reason that a general language model will be more competitive in those domains. If you want examples of successful domain-specific models, look at AlphaZero and AlphaFold. LLMs aren't anywhere close to achieving that level of competence at abstract strategy games or protein folding.

"This will never work" is a pretty confident assertion for a field that's so young and rapidly evolving.

> I don't foresee AGI arising out training bigger LLMs (Though investors won't realise that for a while yet).

This is what the parent said. AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine. This is the industry consensus.

>AGI won't rise out of AlphaZero and AlphaFold in the same way AGI won't rise out of Houdini chess engine.

That's a straw man. Nobody thinks AGI will rise out of domain-specific systems. The question is whether domain-specific systems are necessary for AGI.

Of course, the problem is that AGI isn't a well-defined concept. But if we define it as achieving superhuman performance across several hundred domains where there are objective measures of success, it doesn't seem far-fetched to predict that it will involve some general reasoning system paired with a bunch of specialized modules.

The parent said

> I don't foresee AGI arising out training bigger LLMs

I agree that AGI will involve tool usage but not only involving domain specific AI models.

But lets try to find the discriminating point in the discussion - do you believe AGI will necessarily involve training bigger LLM's or not?

I believe they are necessary. WBU?

You're still intentionally misreading the OP's statement. If you read it again in context, they're clearly saying that they think training bigger LLMs is not sufficient. I think I agree with that statement, but my confidence is pretty low.

No, I don't think LLMs are necessary for AGI at all. I think there are multiple paths to AGI, some of which involve LLMs and some which don't.

AGI is a macguffin (or a shaggy dog) for a story told to investors. It has never been plausible on the timescales suggested and it almost certainly will not emerge from LLMs.

> No this will never work.

This bet is too early.

> Why didn’t OpenAI release a math specific model? Why not a literature specific one? Why do they instead have generic models of different sizes? And how did all labs converge on this?

Because they have a very early product and they could train it, brute force, with access to an extraordinarily large pool of money. So did all the other labs. Because it was thus easier to scrape everything rather than spend enormous effort (with tools that did not really exist) to partition the training set. Any number of other "because"s.

It's just what they are doing now and it showed the earliest results.

LLMs are still less intelligent than rats, which have tiny brains.

Rats are a weird choice for that comparison. They're some of the smartest animals on the planet, some say smarter than dogs.

Which is exactly what makes them a good comparison, IMO. They are arguably smarter than dogs with about two fifths as many neurons. A shorter lifespan, faster breeding and many more threats has made them better.

The point I am getting to elliptically is that larger models aren't necessarily the solution. They are one solution pathway.

It is fully possible (I think actually likely) that the ultimately successful path for LLMs will emerge from the pressure of keeping them small, not making them large. Very small, domain specific models could well outperform large models in their domains, and they might even show that domain specialisation is not necessarily much of a limitation, just a useful impetus to stay small. Like how rats can drive those little cars.

(I think the frontier models are potentially already too big. Can't prove it or even close to it, but it feels like this is going to be a story.)

I think the harness and local context should supply that missing piece between general model and bespoke application. Each application has its own context and action quirks that don't generalize well. Maybe it's just 5% but that is genuinely specific. So its rightful place is in context engineering.

I have a long-ass post about how this could be implemented. https://old.reddit.com/r/VisargaPersonal/comments/1um9uyv/st...

I think future is probably more similar to speculative execution (inference/decoding). A small LLM is used to speculate and a large LLM is used to confirm if needed. If the small LLM is accurate enough on N tokens it’s cheap for the large LLM to say looks good and keep moving along.

Is anyone making LLM-in-a-box for emergency supply kits yet?

I feel that would be handy in all sorts of situations when networks are down.

Can you expand what you mean?

They want to ask the iOS Foundation model (frontier on device intelligence for something small) for instance about emergency procedures and life-saving info. I wouldn’t trust that model with much at all though. More likely to find what you need from miniature survival guides.

> They want to ask the iOS Foundation model (frontier on device intelligence for something small)

This is a bit of a straw man, TBH.

For one thing, "LLM-in-a-box" doesn't necesssarily imply a device as small as a phone.

For another, you'd need to convince people that the iOS Foundation model is the "frontier" of LLMs that run on phones when it is really not. AFAIK it is noticeably outperformed by the Gemma 4 E2B model and certainly the E4B.

https://blog.google/innovation-and-ai/technology/developers-...

Here is a common-or-garden youtube video that includes a demonstration of how much better the E2B is:

https://www.youtube.com/watch?v=sTxyBUbdZcA

Whether this idea (LLMs for emergency/survival scenarios) has value, I don't know, so I am not offering an opinion, but you should approach it with a good faith argument.

I am an LLM cynic but I suppose if I was to be without connectivity but with power for a while, a device with the Gemma 4 E2B or E4B model on it might be helpful or interesting to have. If such a device had the 12B QAT model on it, that really would cross the line to utility. Not sure it has value in the OP's scenario, still.

I meant the new 20B param Foundation models, not the ones from last year. But sure Gemma might be better

This is couched in prepper nonsense, but it's got LLM, WikiPedia, maps, etc. A bunch of genuinely useful stuff to keep on a USB stick or whatever: https://www.projectnomad.us/

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.

The 12B QAT model is really overlooked because the tech industry has been so desperate for the LLM bet to play out to "product market fit" (which means please IPO now) that it has become convinced that coding models are the only things that matter.

> LLM-in-a-box for emergency

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?

You will probably want a search engine though. Perhaps a small LLM would work well as a component for that?

grep works well!

You may benefit from an embedding approach for semantic search. Not sure what an LLM would give you on top of that.

If the emergency is in a foreign country, being able to communicate with locals would likely be a benefit - and a domain specific trained model could translate better than general purpose translators.

In general, I think speech as input/output is under-explored. In the emergency scenario, in a stressful environment, having an expert in your ear you can talk to should work much better than having a big manual book to look up specific cases.

I think an LLM would give a "conversational" experience to search.

That's handy for situations where you might not really understand what you need to search for. Any search system that can ask you clarifying questions is going to be a big improvement.

Or where you need to combine several steps together but you don't yet know what those steps are.

There's probably other technologies that could do that, requiring lower resources but they'll come with different trade-offs around configuration.

Just having a Raspberry PI, a offline copy wikipedia and a RAG enabled small LLM would be quite useful or at least entertaining if you have to go off grid.

> You will probably want a search engine though.

The search engine is indeed the last missing component from a sovereign stack. But I think this could be solved locally with little cost. Instead of indexing content on the web we should be indexing sources themselves - where to look for X? - like forums, blogs, docs, feeds, and specialized search engines. We could collectively amass millions of these search stubs that can be used by local models to go and fetch fresh information from the source directly. This means separating the routing layer from the information layer, we don't need to keep information cached from the whole internet locally. The search stubs could fit in a few GB about same size with the local LLM. The cool thing is that sources change much slower than information itself, so the search stub database could be refreshed at a slower pace. We could combine a few million generic stubs with a few hundred personal stubs generated from our own activities. It is trivial to generate these stubs by piggy backing on frontier models.

> If you do carry it to an enclave of civilization that has the right power, are they going to let you run it when it cuts into light for surgeons or heat to sterilize water?

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

You can run a small model off a home generator — so in an emergency, you’d turn on both the generator and information service, eg, a mesh for “quick” responses querying that huge collection of information.

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.

I feel this is going into increasingly-unlikely mixes of constraints and needs in order to try to keep a "wouldn't it be cool if" hypothetical-tool dream alive. [0]

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...

[0] https://xkcd.com/2128/

Option 1 sounds better:

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”.

> 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.

At current prices you are also out about $4k for a Spark to actually run the inference on, if you want a full LLM in a low-power package.

In general, I'm not sure why one would want to pin your survival to an expensive, hallucination-prone data source, when an offline copy of wikipedia with a little vector search attached to a Raspberry Pi can fulfil the same role...

> Have you ever tried to use a handbook you’re not intimately familiar with during an emergency? It’s rough.

Sounds like the absolute worst time to rely on a crappy little model that will inevitably hallucinate.

I've been mulling over a good use of a large philanthropy spend in the next decade, and I would love to build a bunch of hardware "oracles" that include an LLM. Ideally solid state, visual/audio, solar + usb-c, so, good in a lot of doomsday scenarios as well as just out hiking. It's a fun thought experiment. I imagine making like 1 million of them, they could be sold and genuinely useful, but also given away; once owned, you could use them, or store and put in an emergency box, bury next to the 10k year clock.. a lot of possibilities.

I feel like you could get a lot more quality of life improvement for more people with the money if you spent it on low tech solutions, eg more efficient cooking stoves for people still cooking with biomass, or solar microgrids for areas without electricity.

No doubt there are a hundred more direct things one could do. On the other hand this would include instructions for all those things and could advise on the building! I think it would be a nice thing to have in the world.

Smallest local model able to work with offline wikipedia dump would be one step above just having an offline wikipedia dump.

This is HN, not Reddit.

electricity outage and battery running out is the end game for any real prolonged external emergency. Internet connection is just the soft edge.

https://developers.google.com/edge/gallery

Put that on a spare phone

> Is anyone making LLM-in-a-box for emergency supply kits yet?

Maybe someone should be making this, but for rebuilding society in the event of a disaster - a solar-powered black box with most of humanity's knowledge within. Even something running one of the Qwen models would be useful.

"So, we had a nuclear war and need to start from scratch. How do I turn this rock into a computer chip?"

If you are rebuilding society most of this knowledge is useless for centuries. You don't have enough labor to build and maintain factories. You will spend centuries in the hunter gather phase struggling to survive, while slowly building agriculture. You will be lucky if you can teach your grandkids to read - since that will be a useless skill.

Print important knowledge on paper and store it in a desert. in 2000 years society and population will advance enough to get a jump start based on our knowledge.

There is no rebuilding society because of energy. All of the oil and coal that could be extracted by a civilization that only has wood for fuel has been used. If we go back to the Stone Age, we aren’t returning to the present. If we’re lucky, we will get to the Middle Ages.

There is still a lot of coal and oil to be had. It isn't as easy as it used to be, but there is still a lot. If we get to the middle ages then society can build a lot with knowledge. More efficient iron production than what the middle ages (when still using wood) had for example would be useful, and our knowledge of technology may allow bicycles sooner (bicycles need higher technology than a steam engine)

I don't if anyone is doing this yet, but I think a small LLM w/ data sets for RAG reachable via APRS and LoRa would be very useful, not just as an individual but for the community around you.

Oh my... I can think of a 101 things more useful in actual emergencies than an LLM-in-a-box. Unless you have a weird definition of "emergency" (if so: please define).

Project NOMAD: https://andrew.ooo/posts/project-nomad-offline-survival-comp...

I think it was on HN at some point too. Basically it's a bunch of offline data dumps (wikipedia, maps etc) with a local AI model on top that you can use to query them with natural language.

Has anyone used the Rx Scanner mentioned in the opening?

https://rxall.net/rxscanner/

Can't wait to be killed by my toaster because some sexy mossad agent seduced it.

Have you ever tried to indulge an all-consuming urge to kill when you don't have opposable thumbs? Or hands? Or anything other than a bread slot?

Fire!

This made me laugh out loud :)

> The RxScanner is a handheld spectrometer that scans a pill with infrared light, then sends the item’s molecular profile to an AI model equipped with a pharmaceutical database. In seconds, the AI identifies the medication from its molecular profile—or reports that it’s phony.

Is every tech, including database search "AI" now?

Where is a good place to start with training SLM these days if you don't have the compute locally?

I rent cheap GPU based instances by the hour and run on those. Nothing fancy, but $20 can get you a decent amount of compute on a A6000 or H100.

100% agree on this.

I've been working on small local models for years with txtai (https://github.com/neuml/txtai). I've published close to 100 models that can run local for RAG, Agents, Vector Search and more (https://huggingface.co/NeuML/collections).

I think neuro-symbolic AI has a lot of potential here, since small models can handle a lot of conversational inputs, while relying on wired-in solvers for more complex symbolic math/computation needs. https://jjd.io/posts/swollm-bbh-leaderboard.html

99% of the model "work" (meaning the connection to your computer) is just spinning a spinner - something that makes me want to wrap it with a mosh shell so I can just keep moving from network to network.

What about small (offline) AI Models in places with weak hardware?

Fascinating to wonder whether the bigger model finds fewer or more counterfeits than the on-device one.

I looked into this a bit but unfortunately because of starlink most of this won’t be needed

SLMs for the rescue!

Reply on news.ycombinator.com