Pete warden on local AI, open models, voice interfaces, the future of hardware (thedataexchange.media)
169 points by datascientist 5 hours ago
lrvick 3 hours ago
I run 8x r9700s at home which lets me run up to 2x 1M context sessions of GLM 5.3 flash at ~50tps or 12x sessions of qwen 3.8 flash next at ~100tps.
I exclusively use this setup for all AI usage from fixing obscure reproducibility bugs in software that once took me weeks, to compiler bugs, to kernel hardening and hardware drivers, to building a complete unprivileged and security hardened replacement for systemd in rust. Most of that was done using Qwen 27b back when I had only 2x r9700s.
There is literally not a single problem I have faced in 25 years of systems engineering that I cannot solve now 10x+ faster using my local LLM setup.
I turned down the free OpenAI and Anthropic tokens in my YC batch because I have absolutely no interest in being reliant on third parties to do my job.
lrvick an hour ago
For those asking in threads. 4x r9700s will cost about 6-8k depending on how patient you are with ebay auctions.
From there any motherboard and CPU that can give you 4x slots of 16x Gen 5 PCIe slots positioned such that you can install 4x cards is probably fine. Ideally with a CPU capable of memory encryption.
My latest pairing is a ASUS Pro WS W890E-SAGE SE For $800 open box, and a Xeon 654 for $1300. My first build was an AMD EPYC rig for similar cost a few months ago using a Gigabyte ME03-PE0 and a AMD EPYC Siena 8224P for about 2k combined. I specifically wanted one amd and one intel machine for reproducible build diversity which most probably do not care about.
Do not get too attached to specific cpus and boards. Get what is available for cheap with the needed pcie bandwidth open box on ebay from reputable sellers to save thousands of dollars.
Be sure to get a power supply that can handle all 4 cards. I run Silverstone Hela 2500s at about $800. Do not go cheap on power supplies. It will end in pain. Can confirm.
Then 2-3k for 64gb in DDR5 RDIMMs. This is the most painful thing to purchase but it is what it is.
Not only do I recommend this setup, I plan to build more of them to share with my team.
If you are patient and buy used, a 10k budget can get you a very powerful local AI rig running 100% FOSS software/drivers even today.
back_to_basics 3 hours ago
>I turned down the free OpenAI and Anthropic tokens in my YC batch because I have absolutely no interest in being reliant on third parties to do my job.
Which is exactly why said power and utility will soon be BANNED under a myriad of false aegis'.
An economy that almost exclusively extracts ECONOMIC RENT from the working class, won't permit any "leakage" or autonomy.
agcat 3 hours ago
Did you turn down this for internal use-case? OR for your core product? I am keen to learn how are people running stuff locally for their core product to be able to make it easy for their customers to use.
lrvick 2 hours ago
MomsAVoxell 2 hours ago
I exclusively only use local models. I do not subscribe to anything, and I have my agents wrapped and inspected.
2xDGX Spark cluster is my current load, it is sufficient for a great deal many wonderful, wonderful things. It isn't fast as consumer AI, but I have nevertheless been able to do wonderful things with it for the past year'ish.. I'd get a couple more, but indeed they've tripled in price, so yeah .. things will just do for now, and it's fine. Pretty great, actually.
I think the way we control AI is we air gap it. My cluster is air-gapped.
txdv 3 hours ago
whats electricity bill and do you not fear that your hardware will be obsolete while others will have access to better models on faster hardware?
i just calculated that it should be Per month (30 days) 380euro in my region, which is a lot after an initial 20k.
I have to say, no limits stopping sounds nice though.
lrvick 3 hours ago
> whats electricity bill
$300/mo in the bay area for 1kwh running 24/7, at least until I get solar installed, then free.
> do you not fear that your hardware will be obsolete
If it never gets better than it is right now, I cannot think of any problems in software I cannot solve locally, so I am all set for a while.
Though also, the capabilities on my -existing- cards have gone UP dramatically in the past 6 months since as the local LLM community keeps optimizing.
Sure in 3 years all these cards will be a lot cheaper than I paid, but having them for the next 3 years is absolutely worth it for me.
This costs me less than a used car. Most people chase car and phone upgrades every 2 years. I tend to get something good enough and upgrade once a decade. Depends on your goals really.
paul7986 an hour ago
wez470 3 hours ago
How cost effective had it been? I don't want to rely on 3rd parties either, but my experience is it's super expensive to run locally versus using a provider (both in terms of hardware costs and ongoing electricity usage).
lrvick 3 hours ago
No third party can ever provide the privacy and sovereignty of being able to run locally with abliterated models. I would do it this way even if the providers were much cheaper as I value those properties a -lot-.
In practice though, I can produce about 3B tokens/month of Qwen 3.8 Flash Next which is the model I crank almost 24/7. That would cost something like 1k/mo on OpenRouter. My setup costs about $300/mo in electricity, though getting solar soon to drop that to zero.
cortesoft 2 hours ago
ahartmetz an hour ago
Running locally is not a problem if you allow it to be slow. I get about 3.5 tps with Qwen3.8-27B:UD-Q4_K_XL on my 64 GB laptop and about the same tps with Qwen3.8-Flash-Next:UD-IQ4_XS (IIRC, or UD-Q4_K_XL) on my 128 GB desktop with a 16 GB GPU.
random3 3 hours ago
Can you share more, or direct us to more resources around local setups? For starters what's the upfront cost, what are the space, weight, energy requirements, etc.
lrvick 3 hours ago
Really depends on what GPUs you go with but I hate proprietary software so I went with AMD AI PRO r9700 cards which you can still pick up used on ebay for $1500 used on auctions. Started with 1 then expanded over time as I won auctions. Though on a lower budget my $400 16gb 6800 XT can run Qwen 3.8 27b which I could absolutely get by with at significantly reduced speed.
My entire 12u rack has roughly the same capability as an H200, though at maybe 15k all in with mostly used parts vs 50k for a used H200 checking todays ebay prices and that is -just- the card you still need to build a system for.
Consumer cards are a lot more power hungry than enterprise cards and they take up more space. I run my setup next to my dryer, running on 240v for 15% better power efficiency.
Also using transparent memory encryption on a Xeon and an EPYC for confidential and private inference guarantees no third party can offer me. Of course I can run abliterated models that do any security research I want without refusals.
Strata and Radiance are the state of the art ATM for running big MoE models on small consumer GPUs at high performance.
Leif24 2 hours ago
vablings 3 hours ago
There are tons of resources out there on how to run a local setup. Once you have something that's somewhat functional you can then have the system self optimize your software stack to suit your needs.
As for cost a 9700 is running at about 2k USD MSRP right now and the rated power is 300w.
dghlsakjg 3 hours ago
How does this make you less reliant on third parties?
It seems like it just shifts which third parties you rely on. I think local inference is cool and your setup sounds dope, but you still have third party hardware manufacturers, and third party models in the mix.
What is the advantage of this over OpenRouter, which lets me select for, jurisdiction, quantization, ZDR, etc.
note: My use case can't even come close to having this make sense financially, but I understand if yours does. Either way, that doesn't make the third party argument better.
runtime_terror 2 hours ago
Not the OP but I would assume: - no data retention policies to worry about in the first place - no monthly billing outside of power bill and capex - no jurisdiction issues - full control over your data - no network round trips or even worry about network outages in the first place
lrvick 2 hours ago
almondfestival 2 hours ago
it's the difference between a one-time transaction with a third party and an ongoing, perpetual third-party liability. you buy your inference hardware, you download your model, and if you wanted/needed to, you could run inference in perpetuity (or until your hardware broke down). compare that to anthropic/openai/etc being able to throttle intelligence, deprecate models, or raise prices on you whenever they felt like it.
sure, in a very general sense you're dealing with third parties in both scenarios, but if you actually own the hardware and model weights, you can control your own destiny for the lifespan of that hardware.
dghlsakjg an hour ago
drdexebtjl 2 hours ago
Is this some sort of gotcha? You didn’t build your computer from scratch with sand, sticks and pebbles, so you’re just as reliant on third parties?
If you own your own hardware, you don’t have to depend on and trust third parties continually, just once when you bought it.
lrvick 2 hours ago
dghlsakjg an hour ago
giancarlostoro 5 hours ago
If Nvidia didnt abandon consumer GPUs we all could be running better models locally by now. The moat is companies like Nvidia not releasing better consumer hardware so they can partner with bigger companies so they can sell us overpriced hardware.
They release a new RTX series every two years, last time was 2022. (edit: I mistakenly quoted the year wrong, not sure why, but it seems obvious to most that Nvidia doesn't care about the local consumer market anymore)
Meanwhile major AI companies are investing in their own chips, if Nvidia wants to survive they really should not abandon consumers.
Hoping AMD steps up to the plate. Meanwhile the nextXbox and PlayStation consoles are going to have a rough time with their next consoles which will likely cost a thousand dollars or more is my gut feeling.
codehorses 4 hours ago
Not defending Nvidia here, but the RTX 50 series was released in 2025, was it not?
Regardless, they aren't planning on sticking to the two year release cadence any more which does suck.
paxy 4 hours ago
RTX 5090 - release price $2000, average price today $6500
RTX 5080 - release price $1000, average price today $2300
Nvidia also stopped selling its own version of these cards directly to consumers and is happy with letting partners and resellers/scalpers control the price.
So it is pretty obvious that the company doesn't particularly care about this market segment.
giancarlostoro 3 hours ago
Buttons840 4 hours ago
Ironically, if consumer GPU progress stalls, then slowly PC becomes the best console. Everyone has the same top-end GPU that can run any game.
ghostbrainalpha 4 hours ago
Tuna-Fish 3 hours ago
GPU-style platforms (lots of compute connected to a big shared pipe that connects to a large pool of fast memory) are inherently more efficient for centralized inference.
Local inference gets truly interesting when the proper next-gen, natively AI and not just a repurposed GPU, stuff starts to land. Initially things like compute-in-memory and fancy flash, and later possibly the truly exotic stuff like various RMM (read-mostly-ram, for example stacked FeFET on logic) tech.
IMO there is no chance for ubiquitous local AI to happen if that means people using GPUs. But the industry is chasing all possible solutions for more efficient inference platforms, and if something like pachinko machine architecture with FeFETs ends up winning, suddenly buying a local ai inference solution in a box is no longer as insane.
winwang 3 hours ago
If consumers were willing to pay more for GPUs like datacenters do, NVIDIA would have sold plenty of consumer GPUs. Or would you press a button to halve your revenue for some reason?
(I say this as someone who would have liked a 5090...)
lrvick 2 hours ago
> Hoping AMD steps up to the plate
Nvidia has nothing close to the Strix Point, Strix Halo, Gorgon Halo, etc.
My GPD pocket 4 handheld can run Qwen 3.8 27b only because of AMD making embedded mobile inference cheap.
scronkfinkle 5 hours ago
> Hoping AMD steps up to the plate
AMD's aggressive marketing as an AI-first company leaves me pessimistic about that.
cyberax an hour ago
I'm running qwen3.8-flash-next at 4-bit quantization on two AMD Radeon AI PRO R9700 cards at 75 tok/s. And I can actually run two models in parallel at around 50 tok/s, so you can get by with only one card. I switched from NVidia to AMD for a fully open-source stack for my local inference.
AMD Radeon AI PRO R9700 is $1600 right now (I bought them last year for $1300).
They are also quite power-efficient. My setup typically uses around ~300W during inference.
Forgeties79 4 hours ago
No gut feeling necessary. They announced weeks ago that they are going to be $1200-$1800 (Xbox specifically, PS6 somewhere in that range too).
lubujackson 4 hours ago
Not to mention the required monthly subscription... Income inequality like 1929, might as well bring sharecropping back.
jimbokun 3 hours ago
Yes, ten the home AI market would be similar to the serious PC gaming market used to be when they were the main market for the latest GPUs.
jjtheblunt 4 hours ago
> we all could be running better models locally by now
Apple silicon says "hold my beer", which I know because i've been doing this on my 3 year old quite beefy mac studio, for example. it's good enough for many tasks.
vadansky 4 hours ago
Can't wait for M7! I hope the rumours are true.
paxy 4 hours ago
I don't see any difference between the conversations around self-hosting local AI and self-hosting regular software, which have been going on for decades.
Yes, there are a hundred valid reasons to self-host. But 99%+ of consumers aren't going to care.
People don't even care when the cost of hosting is ~free, like a simple home server. How many are ready to buy thousands of dollars worth of AI hardware, which could very well be obsolete a couple of years from now?
wry_xy 4 hours ago
The self-hosting crowd, Linux fans, OSS contributors, etc, are our own worst enemies.
If we care about self-hosting and open source and decentralization because we believe it leads to a better world, and not simply because we want to be control freaks and snobs, then we need to learn how to build shit that is actually convenient and usable and understandable to normal people.
I think it's possible. I think many people would want this, and care, if we could package it up in a digestible way.
lilbigdoot 4 hours ago
Linux has been more user friendly and accessible to me for normal use than Windows for years now. I really do not understand this comment
guhidalg 3 hours ago
pelagicAustral 4 hours ago
I'd say one of the major frictions with self-hosting is that for the most part it does always going to take a lot of tinkering around and maintaining things going. A world were we have a lot more people self-hosting would be great, but as it stands, it's an utopia, because normal users will get annoyed the fuck out of the environment the second their infra is busted because of any of the million reasons things break.
lisplist an hour ago
pydry 3 hours ago
davedx 4 hours ago
Ollama and friends are pretty damn easy to use already. Hardware is the real barrier
lurking_swe 3 hours ago
petra 3 hours ago
What percentage of techies use self-hosting for a significant part of their services?
Before that number is high, can we realistically expect non-techies to use self-hosting?
forshaper 3 hours ago
At some point there is a trade-off between those far away controlling "your" stuff and convenience. It is much easier for someone to make something convenient if they have control over it. For example, troubleshooting when you own the stack versus troubleshooting when you don't. However, if someone wants someone far away to help them troubleshoot, then it will be easier if the one far away controls the stuff. There is a trade-off there.
goda90 4 hours ago
The problem with open source is that it is often written by people who are fixing issues they have or making features they miss. To get to the point of contributing, you have to go through the initial setup process. Once you've done that, it's not longer a problem you care about, so you're less motivated to contribute code to make that process easier.
nemomarx 4 hours ago
There's a decent opportunity where AI coding agents seem good at creating a very "average" and understandable UI. Maybe not perfect, but probably better than Gimp?
zoeysmithe 3 hours ago
Its worth mentioning "normal people" take hours upon hours of classes and trainings to use typical office productivity software and other job-related software, on top of watching and learning from tutorials and family and friends to use personal software and websites and apps. And heaven knows how many support calls and tickets to just do what is supposed to be "easy" in the closed source world.
Also, none of this is intuitive and a lot of popular closed stuff is harder to use (and vastly disliked by the very same normal users) than some popular open source stuff.
I'd even mention 'normies' want self hosted Word and Excel and regularly reject the web-based versions of them. Normies don't want streaming games but a gaming PC or console. Normies may buy streaming services like Netflix but get upset when their favorite shows are removed, so they actually don't like netflix. Normies complain about censorship on twitter and reddit and discord and would love an alternative they control, but may be unaware that's an option and seem stuck on those services and hate them.
Normies want self-hosted all the time too. Their ability to understand and do this is what's lacking. So that's a conversation on its own, but normies aren't actually happy. Form what I can tell most people hate 50% of the stuff they use right now and most of those complaints revolve around the cloud-hosted/lack of control/privacy and other issues self-hosting or federated services or FOSS apps could arguably solve for them.
This is a big conversation but imho tends to boil down to who has the marketing and sway to get you to install or use something, not that typical end users are savvy judges of software and useability working purely on merit and informed facts. Does your grandma actually like Facebook, or is she just lonely and its the lowest friction way to not be lonely? Do you really like Chrome or Word or is it just easy to get and popular? Is the big mac the best burger or is just easy to get and popular?
betterunix2 4 hours ago
> shit that is actually convenient
I mean...sure...but how convenient was that Office 365 outage? How convenient is it to find that a feature you actually depended on is gone because the company that runs your software decided to get rid of it?
It is always a back-and-forth. It is not as if centralized computing is a new idea, in fact it was the original way to use computers and people found that issues of control mattered and switched to PCs; the pendulum swung back because people realized that there are advantages to centralization (e.g. not having a failed hard drive destroy high-value documents). The cycle will probably continue for a long time to come because the advantages and disadvantages of both approaches are fundamental.
bendergarcia 4 hours ago
coldtea 2 hours ago
templar_snow 3 hours ago
This. The FOSS/etc crowd don't even try to bother with outreach on tools that the majority of people would want if they were marketed correctly instead of kept as niche projects for people to be smug about. Consider the EU rolling out Matrix in a huge way, or people moving to Mastodon. For most normies, "Use this to escape Elon Musk" is more than enough sales pitch.
pyrale 3 hours ago
> not simply because we want to be control freaks and snobs
It’s not so much about being snobs as it is about not having time and energy for a second job. The level of quality you demand simply requires a level of attention, investment and involvement that can’t be provided without working full-time on it, for profit.
Telling a bunch of hobbyists that they should provide service of a quality equivalent to that of a trillion-dollar corporation if they are serious about it is gaslighting.
dotancohen 3 hours ago
agile-gift0262 3 hours ago
> we want to be control freaks and snobs
That's a mirror that hurts to look at that you just held in front of my face. Guilty on both counts, even though I try to convince myself I'm not
abdullahkhalids an hour ago
One of my favorite stats is that Limewire used to be installed on 1/3rd of computers worldwide [1], i.e. people were self-hosting.
Why? Because Limewire was easy enough to install that even grandma could do it. There is not a single way to self-host even a simple html website without learning half a dozen software skills. People will self-host when it is click-and-play.
[1] https://marquettewire.org/3777966/tribune/marquee/the-life-a...
rockostrich 25 minutes ago
> There is not a single way to self-host even a simple html website without learning half a dozen software skills.
1. Set up port forwarding on your router
2. Open a terminal
3. Run `python -m http.server 8000` in the directory with the html file
I'm not sure you could really count any of those 3 as software skills, but even if you count all 3 it's still half as many as you said it would take.
nsbk 4 hours ago
Hard disagree. I'm running Qwen3.8-Flash-Next using Strata on a Mac Studio and it covers all my needs. One drag and drop install. Nothing like self-hosting your email.
mortalapeman 3 hours ago
Seconding this. If you know how to build with a dumber model it scales crazy well. And Qwen 3.8 Flash Next is just wildly good even at low quants. I'm trying to find it's limits and I've been using it for over a week building and haven't hit a wall yet.
choxi 3 hours ago
3836293648 2 hours ago
paxy 3 hours ago
What are you disagreeing with exactly? Qwen 3.8 flash costs a few cents per million tokens from cloud providers. You can get effectively unlimited usage on a $6/month plan. You were okay with buying a $5000+ Mac Studio for it because you have the money, but you are in the tiny minority.
nsbk 2 hours ago
xandrius 3 hours ago
How much was your Mac Studio?
nsbk 2 hours ago
__MatrixMan__ 3 hours ago
> But 99%+ of consumers aren't going to care.
They don't now, but they might soon.
It used to be that bad behavior on the part of our service providers was annoying because you'd get ads and spam. Nowadays the leaked data gets aggregated into apps that help government spooks abduct your neighbors. The effort remains constant, the threats it helps you avoid are getting more dire. Eventually those lines will cross.
bee_rider 3 hours ago
I think ultimately local LLMs will win but only because models will get good enough that the “cost” of running locally on a 2030 chip will be less than the “cost” of talking to the network. Especially given that building wider/lower precision LLM accelerators is a perfect task for hardware companies (embarrassingly parallel in an era where sequential performance improvements have long dried up).
In terms of self-hosting vs remote-hosting regular software… I mean, a lot of regular software is fundamentally about communication (social media and office suites) so that goes on the server. LLMs don’t benefit much from being remotely hosted, other than that it makes the flops run in somebody else’s machine. But that will be hard to fund in a universe where the upper-limit to what can be charged is defined by how much additional cost “good AI hardware” adds to the price of your phone. We’ll see I guess.
jbjbjbjb 4 hours ago
It’s more about self-hosting the hardware which is more desirable than thin clients.
wmichelin 3 hours ago
Self hosting is not free, it's actually pretty hard to setup and keep running sometimes. Time and cognitive overhead are money!
Zambyte 3 hours ago
> cognitive overhead
I told a remote model to set up strata on my local machine. With the best options for my hardware. If someone can figure out using remote models, it's trivial to bootstrap local intelligence assuming available hardware. The hardest part was knowing that strata exists.
Ninja edit: I've even had my agent tell me about suboptimal hardware configurations and told me what to do to fix it (XMP settings, having hardware in suboptimal slots). Cognitive load is simply a non-issue. The only issue is time, but in the grand scheme of things, it's a trivial amount of time to get up and running.
mortalapeman 3 hours ago
Just get a 5070 and 64 GB of RAM and download Strata. It's the ollama of actually usable local AI.
barbazoo 4 hours ago
It's so sad that we are not empowered to act as more than just consumers as if it was some kind of shared religion. How do we get out of this cycle?
random3 3 hours ago
People complaining about how difficult self-hosting is seem to be stuck in the past and missing the point. It's the very AI that we're talking about that removes most if not all friction points related to software configuration and maintenance.
While this doesn't remove the upfront cost (which could be fixed with basic financing like Apple does) and physical manipulation (e.g. replacement on failure, etc.) it does lower the barrier.
Finally, there's another path I find interesting and interested to think more of which is sharing hardware locally, from your house to your neighborhood, etc. We've done this in the late 90s - early 00s with Internet and it worked really well. I think it could work great for compute.
I see a lot of opportunities in this direction, actually and most are actually enabled by the latest improvements in AI, so most of the "old" thinking doesn't apply anymore.
bobajeff 4 hours ago
That's like the argument on why people don't care about privacy. Just because most people choose that, all things considered, they are willing to sacrifice sovereignty over their own computer usage doesn't mean they don't care about it.
By that reasoning you'd also conclude that no one cares about affordable housing or having a good work life balance or having your own lake made out of solid gold.
redanddead 3 hours ago
It comes down to the incentives
Think of who uses local AI, it’s either
-power users upset with the limits —-cost of intelligence is going down
-people and corps that want stabiltiy —-a massive self own by the labs, this is justified
-American labs seemingly snitching on their users (wtf?) —-huge incentive for people to use Chinese inference, which is cheaper along with other benefits
Regardless, we’re talking about a subset of a subset of people, 1% of 1%. The big bet is that everyone will become AI power users. I think it’s accurate, forget openAI, fuck Anthropic, let’s look at us, the people. Entire workflows like coding have been buried, there’s no going back to how things were before. Is it reasonable to say that will happen to the other work streams? There’s no indication not, and it’s a total gamble if it’s yes
lrvick 2 hours ago
This is like owning a car vs using a train. Both are valid choices depending on your goals.
I really do not care if everyone does it my way. What matters is that the choice exists for those that want more ownership, freedom, and privacy.
That said, most next gen laptops will be shipping with capable inference hardware more than most consumers will ever need. Not using the GPU people have in front of them would be silly, but people are silly.
jmyeet 3 hours ago
Here's one difference: currently lawyers using cloud LLMs violates attorney-client privilege. And you have to be really careful sending confidential medical information to one so as to not violate HIPAA.
More broadly, those prompts and the responses have value. There's a whole shadow economy in China of aggregating queries across Claude (etc) and then using that as training data. It's like your entire search history is for sale.
Clould infra has typically been opaque to the provider. You run an MySQL instance on an Amazon VM and Amazon doesn't have easy access to the unencrypted traffic and what you store. Cloud LLMs don't operate that way.
AI hardware is improving at a rapid pace. By 2030 I believe a significant amount of LLM usage will be local for these and other reasons.
jimbokun 3 hours ago
Enough that Apple currently build enough Mac Minis and Mac Studios to satisfy the demand.
adammarples 3 hours ago
I don't know about this whole argument that hardware will become obsolete. If it can run Opus 5.5 today it will always be able to run Opus 5.5 until it burns out, and the quality of the models it can run will continue to increase as the software improves.
herf 4 hours ago
I love local AI, but even so, I admit the economics point the other way. Unlike gaming, AI is a uniquely good match for data centers because the applications today aren't latency-sensitive at all (except perhaps DLSS/self-driving/robotics). Most importantly, utilization is massively better with always-on servers doing parallelism (in-flight batching). Also in some areas (like California) our local electricity is quite expensive, so moving the work 100ms away can help. The privacy implications are horrifying, though.
younglunaman 4 hours ago
I tend to disagree here. Think about how useful it would be to have a nice llm in real-time debugging your applications loaded entirely on your computer. That sounds pretty sweet to me. I don't see how that's not really cool / immediately useful to you?
Maybe your point is the majority of the time that's not needed, fair, then maybe the approach is both?
herf 2 hours ago
Well, LLMs handle raw data perhaps a million times slower than conventional algorithms, so in "bytes" and "latency" terms, a dialup modem can handle most AI communication today - a rack of cloud GPUs 100ms away is still faster for a lot of things than $10k of hardware in your GPU, Mac, etc.
drdexebtjl 4 hours ago
> the applications today aren't latency-sensitive at all
This is circular reasoning. We don’t have latency-sensitive applications because sufficiently powerful local AI is not profitable to entertain such applications.
herf 2 hours ago
Absolutely, and there are applications like self-driving or robotics or surgery or whatever where it's critical to have the work be local and never have a network hiccup. But these are all in some sense new markets, though I guess these would be growing faster if we had abundant local AI. Right now most of the AI applications that people pay for are ones that make human workers more efficient, and humans are already high latency.
musha68k 4 hours ago
The economics point the other way because capital is aligned against you. Should this be a bubble that pops we will hopefully see prices come down again.
MooresLawIsDead and his spidey web / senses thinks it already popped:
brendan_rules_t 4 hours ago
I have a 128gb ddr5 unified system, that i have been setting up with linux and local llms and other models and i have to say it feels beyond next level! I still use claude, grok, codex, and neo clouds to offload heavy work loads, but anything i want private or says too much about my system is now done locally. The voice agent experience is incredible
afzalive 2 hours ago
what models are you using for which purposes and what is the voice agent stack please?
montroser 4 hours ago
I run Qwen3.8-27B Q4 on a $1500 box in the basement with 32GB VRAM, and a customized OpenWebUI on top. It's pretty great. Occasionally I consult Claude for a second opinion, but I'd say I've moved 95% of my previous conversational Claude usage to be local instead. I use it for leisure coding projects too, but for real work, it's still DeepSeek 4.1 Flash for now.
cj00 3 hours ago
Invariably when local AI comes up, people try to see how it can be used for frontier-level AI tasks like asking it to one shot a web app from scratch and then they get disappointed when they discover it's not performing at the same level as Opus.
But they're ignoring that a local LLM can still be very useful for any kind of repetitive task, things like categorization or generating embeddings, translation or summarization tasks. I'm self-hosting Karakeep and use a local LLM for the AI tasks like embedding.
And once you've got that local LLM running, you can start thinking about integrations with your own custom built apps.
sachaa 3 hours ago
Also, they are thinking of one single LLM model at a time. What if we have a network of dynamically loaded micro-models loaded on demand based on the user intent? Think of a global DNS of micro-LLMs. We need to decentralize AI. I started building this here https://modelxare.com
smath 3 hours ago
What is a good place to track the latest capabilities of open source local LLMs vs commercial models? Things like overall intelligence performance, latency/speed, hardware needed, cost.
Anecdotally, I believe the open weights local models are constantly catching up with commercial ones, but I've honestly lost track. This is a dynamic space.
1970-01-01 2 hours ago
Voice interface is the "next big hit" as it is removing the very last hurdle to accessibility. Anyone that can talk and hear will be able to use AI. No screen, no keyboard, no problem. This was supposed to be the home assistant in a nutshell but we all know how limited it was in reality.
layer8 an hour ago
No screen is limiting though. Imagine if any form of written communication or written record were prohibited. We want the AI to integrate with our vision bandwidth. Similarly with using our hands for interaction. There are many tasks that are much more efficiently and accurately done with our hands than by spoken instructions.
sachaa 3 hours ago
Why should users have to choose a model at all? We don't choose a codec before playing a video. An app should be able to request transcription or summarization, and the browser should find and run a suitable local model. That's the layer i am exploring with modelxare.com.
jkingsman 4 hours ago
Even dumber local models that are heavily quantized to fit on consumer GPUs can be lovely as a dimwitted household assistant. A lot of people look to offload things they already do to LLMs, which can be great with frontier models, but you've got the highest standards and preferences for tasks you already do, so it's the most likely to fall into a "should have just done that myself" outcome. I've found great success in instead lowering my threshold of things I'd take time to do -- I wouldn't spend time on it, but I'd spend 60s prompting an LLM to take a crack. I'm fortunate to have the hardware to run Qwen 3.8 Flash Next at q8, but most of the tasks I have it do would be totally doable for a mid-quant Qwen 3.8 27b.
Some recent tasks I've thrown it:
* Go through my blood test data (from EHR XML dumps from prior years) and give me some over-time charts for consistently-sampled measurements
* Go through my Good Articles bookmarks, epubify them, and drop one down onto my ereader each morning for my commute, then archive the offline version into cold storage
* Go through my music library, identify my most-listened low-bitrate tracks, and source or assemble a shopping cart to get high res versions
* Analyze local tax/foreclosure auctions for vacant land; grab aerial screenshots, tax data, ownership data, and compile a weekly report of things maybe worth bidding on based on some specific selection criteria (fun facts: if you're willing to accept some headaches, 5+ acres can be yours for under $5k, if you're in the right area)
* Dump my grocery store loyalty data, analyze my purchasing habits, and prepare a weekly instacart for my approval that aims to sync with usage habits. Tweak and learn volume/order adjustments over time.
* Given a KVM that can mount ISOs onto the host, reimage two old laptops of unknown status with Debian and install Minecraft so my nieces and I can have a LAN party
* Organize my music (high confidence MusicBrainz Picard matches), movies/TV shows, ebooks, and video game ROM collections with metadata and good folder organization
* Archive bookmarked fanfiction for my offline library (slowly, and respectfully, and with an actual browser. The browser is anti-bot-countermeasured to the gills, but QFN regularly solves hCaptcha and Cloudflare turnstile challenges. Once I gave it the ability to drag, it even reasonably reliably solves drag-shape captchas. A little while back it made a purchase on eBay which is NOT a bot friendly site.)
Would I personally spend time on doing most of these? Generally not. Could an LLM screw them up so badly I regret it? Also generally not. I get the luxury of a psuedo-assistant that doesn't judge how menial the tasks I send it are, and I save time. With a local LLM, my privacy concerns are far less than with frontier models, and local abliterated models are less (or just not at all) twitchy around things that frontier classifiers balk at.
I'm hopeful that big projects like Muse can help to normalize menial task offloading, and that local LLMs provide the bridge for more technically minded folks to use them even more aggressively and limitlessly (as in compute limits/subscription costs) and cheaply.
Scrapemist 4 hours ago
Trying to get into this too. But my hardware is a bit limited atm. What are you running Q8 on? At what speed?
jkingsman 2 hours ago
I'm across a Pro 6000, 2x3090, and 128GB sys RAM (built a while back; I could not afford that these days heh). I get ~1-3ktok/s prefill, ~120tok/s decode.
When I had only twin 3090s, I made heavy use of Qwen 3.5 27b.
maxdo 4 hours ago
i'll share a lonely voice on hn, i care very little about local ai.
I feel like people who consider it as an option thinks of narrow scenarios vs ai is part of our life like internet is. How do you things on mobile for example with your local models?
AI is very competitive, I use various models, I like to be on cutting edge and to be productive, its the opposite from doing something permanent, and not changing, where local models can be good. I like to do things on the go, and the last thing any normal person would want is to have DC at their basement.
cableshaft 4 hours ago
I'm not willing to feed my personal data (emails, files, etc) to a non-local AI model (well, beyond what I can't avoid, like Google and my Gmail account).
But I would like to have an AI that does that, like synthesize my journals and notes and provide a local personal knowledgebase I can query, things like that.
I don't think I have a good enough GPU to manage a local model powerful enough to be worth the effort just yet, though, so I'm waiting a bit longer.
For coding and bouncing ideas off of though? Yeah sure, online is fine.
zbendefy 4 hours ago
> well, beyond what I can't avoid, like Google and my Gmail account
I think this is a real moat google has that will benefit them in the long run.
ale42 4 hours ago
I think that many people want similar things. But (at least personally), not giving my data to BigTech is more important that this.
maxdo 4 hours ago
out of curiosity, is there anything special your data has?
cableshaft 4 hours ago
soon_to_be_sb an hour ago
montroser 4 hours ago
desmondl 3 hours ago
chairmansteve an hour ago
pluralmonad 3 hours ago
segmondy 4 hours ago
I run local AI, there are tons of local models with no cloud alternatives, the only way to unlock and have these functionality is by being able to run local. So you are not on the cutting edge by using cloud only.
NoidFonsense 4 hours ago
I just set up a wireguard tunnel to my home network
musha68k 3 hours ago
They will need to realize quantum proof, performant and verifiable zero trust compute to stay competitive with the sovereign front.
Yet this would subvert one, if not the main pillar of funding - which is total self imposed (of course) surveillance.
NoidFonsense 2 hours ago
There is no such thing as zero trust compute
musha68k 30 minutes ago
tencentshill 4 hours ago
Local (Free) beats cloud (free) for pretty much any task except live web search (which will be less useful as more cloud AI scrapers are blocked). You run into the free limits in one or two prompts most days if you try to do anything even mildly complex.
jakswa 4 hours ago
How do you things on mobile for example with your local models?
You think you will be doing those things? And if you do keep doing those things forever, won't AI be smart enough to figure it out for you, whatever needs to happy locally?codybontecou 4 hours ago
Whisper and other stt models work fine on mobile.
Mobile is far behind frontier coding agents but they are already solving valuable problems with competitive quality and privacy, one-time payments as viable options.
yieldcrv 4 hours ago
I disagree, I think local is the supercycle here once it gets out of enthusiast territory
We have LLMs baked directly into silicon getting 20,000 tokens/sec with low power consumption, Taalas was bought by AMD weeks after that demo. In that world, people get a new LLM only on hardware refreshes, and only after LLMs are good enough to be in stasis for years
Apple’s M7 Pro/Max if it meets promised will be very expensive, I see some enterprises rolling them out to some staff
Some language models in the Apple Ecosystem are already baked into iOS and MacOS, and its ubiquity is invisible
Thats the direction I see things going, but like not anytime soon to alleviate supply chain crunch. 2029, 2032
nater5000 4 hours ago
First, I'll say that I, too, care little about local AI. I do like toying around with local AI, and I do see how it can be a very important option in some contexts. I also loosely agree that it's good for keeping frontier models on their toes, prevent this technology from being locked away behind a paywall, etc. But for me, as of now, they're not very useful, and I use AI quite a bit.
However, I do think your perspective on local models is a bit narrow. For a lot of people, they want AI tools but do not need cutting edge AI. They'd prefer to just have their own AI which is capable enough and under their control. They'd run it on a small server at home and connect to it as needed for their work (including on their phone). It's worth noting that much of the consumer-grade infrastructure definitely doesn't exist for this, but one could blame the lack of consumer-grade hardware as the source of this rather than just a general lack of demand. You can definitely see people picking up on this by running their various assistants, like OpenClaw, on Mac Minis from their home. It's definitely not accessible or cheap enough for most consumers at this point, but it's also not impossible to imagine us getting to the point where it is more ubiquitous.
I'll say, personally, that I run Qwen 3.8 27B on an RTX 3090 from my home server, and it operates pretty well for what it is. I'm certainly not giving up my Claude subscription anytime soon, but it is nice to have the option and I can chat with it directly on my phone securely (using Tailscale) and using a nice UI.
Regardless, I think the whole point of these kinds of discussions is to point out that, right now, local models aren't viable, but not because they can't be viable, but because companies are actively preventing them from being more viable. I can see for myself that they work well enough that proper investment could easily push them into the "good enough" territory for many users, and I suppose it is inevitable that more and more people will end up running local models in one way or another. But I definitely agree with the sentiment that most consumers simply do not care about it. They don't want to deal with hardware, they don't want to think about software, and they're happy to pay $20 per month for a subscription that "just works." I definitely don't blame them.
dofm 4 hours ago
> How do you things on mobile for example with your local models?
The same way you can screenshare or forward from your local Mac? Paseo, LM Studio and LM Studio Bionic all support this kind of tailscale-ish access to local models from phone apps. (Haven't tried any of these myself)
I do, personally, have "narrow scenarios" for AI. I have to use it for work, and want to be informed and engaged on that level, but I want to keep it out of my personal life as much as is practical.
If you don't, that's fine, but I reframed my mass social media use down to an extremely minimal pattern in 2014 and I am healthier for it; I intend to keep the same distance from AI in my hobbies, for several of the same reasons.
Paradoxically, for me that has involved taking the control over it that local models offer. I use the smallest, simplest things that get the job done and I try to retain my engagement with my work.
It's all pretty low-effort and low-tinkering; no more than setting up local VMs for dev work. Old M1 Max macbook, Unsloth, Pi, a model in the Muse Glimmer/Qwen 27B/Nemotron Bolt sort of bracket.
Good enough to be a rubber duck that talks back, write the minimal kind of AI-generated support code I want, and teach me how it all works.
I might one day buy an M5 Pro Mac Mini, or an M5 Max Mac Studio, or stick an R9700 in a homebuild PC, or I might just use openrouter for the few things the Macbook won't do.
I am a cynic: I believe AI has some value, but I also believe it should be a commodified technology rather than a branded SaaS subscription product, and it should demonstrate that value without me letting a couple of large corporations FUD me into pushing it into every corner of my life at once, or simply handing over control of my job to it.
It is a matter of personal discipline, for me: I am going to resist the endless FOMO of new model benchmarks, learn how it works, keep it in a box, and take it out when I need it.
betterunix2 4 hours ago
It would be a lot easier to do local AI if we were not using natural language as the interface. In fact there is already plenty of local AI, including on mobile, that most users do not even realize is there -- because nobody has "conversations" with those models.
Niche? Sure, but let's be honest here, a lot of AI use right now is forced, which is usually a sign that the product is not actually in demand or has questionable value for users. The big, natural-language-interpreting models are also niche -- a minority of people find them extremely useful and a majority find them to be "meh" at best.
lubujackson 4 hours ago
I agree. It reminds me a lot of building your own desktop computer from parts - it started as a requirement, then a cost-effective choice, then became for power users only, and now it has little place beyond tinkerers/purpose-driven/privacy tinhats.
I refused to use Macs for a decade because of how offensive I found it that they locked down their computers from even adding RAM yourself. But this is the direction of the tide and swimming against it will only continue to become more and more arduous. At best, we will have some Tor-like hosted AI, but I am not holding my breath or wasting my money/energy.
xscott 3 hours ago
Online models will eventually enshitify to serve you advertisements. They'll give you worse answers and steer you towards commercial products to solve your problem.
neutronicus 4 hours ago
No, I'm the same way. I started trying to spec out a local AI box, looked at the numbers, and it felt like trying to run my house off a generator.
Obviously worse than the centralized option, and not the hobby I want. I'd rather just use Cloud tokens help me do what I actually want to do.
philipodonnell 4 hours ago
I think one of the more compelling reasons for personal-local AI is that when two services you use add their own AI that only works in that service it creates the need for a third AI to bridge those two. But any service the proposes to sell you that third AI will be blocked by the other two, thus, only people who can create their own local service will be able to bridge them, and that will be the new version of DIY.
xnx 4 hours ago
"Local" is the wrong distinction. Self-hosting doesn't have to mean running off a box in my house. I don't care about using my own hardware as long as I can run the AI anywhere I please.
xandrius 3 hours ago
That's basically self-hosted which is basically local in this conversation, no?