Open source AI must win
opensourceaimustwin.comBecause of this, I think it might not be possible to have AI *only* open-weight; major players like OpenAI, Anthropic, Google will likely stay for good, with better models than open-source versions.
I think it might look something like Photoshop & GIMP, with Photoshop being a frontier lab, and GIMP being the open-weight model. GIMP is decent for many different image editing workflows, but Photoshop is just better.
I would definitely prefer to have an open-weight model better than frontier labs'. Though I don't think it's possible.
That is, of course, unless they develop their own hardware specifically to run this open model. But, that does ruin the point of open models.
(Yet; I do worry about future required hardware attestation for basic things, but that's another issue.)
It should be clear by now that there’s a whole universe of work to do with the models we have today, from studying to securing to ‘harness’ing. There are tons of economic benefits to be reaped already, if applied carefully. Doesn’t that sound nicer than rolling the dice with the lives of trillions?
I learn it hard from prusa 3d printer open model
More RAM means bigger models, which means smarter models.
Which is why Qwen and Gemma have been so interesting to a lot of us who run our own... Now 32gb VRAM isn't so bad, as these models can be run on that with decent results.
Where this gets interesting is in a couple years, when all the A100, etc, all the Enterprise hardware hits eBay.
Even if the GIMP of LLMs is only 80% as good as the VC-funded stuff, that will still be plenty useful for lots of people.
And I think just having the option to use open source models is a win, even if it turns out to be true they'll never be quite as good as the proprietary ones.
In the meanwhile, and regardless, software optimisations coupled with hardware continuing to scale, we will end up, soon enough, with some open weight that run on a mobile device with greater capabilities than Fable.
I am spreading a message of peace and sovereignty:
Never subscribe. Never. Subscribe. Ever.
Starve them out. Make their lenders take 95% haircuts.
Just don't subscribe, whatever you do!
There's a more fundamental reason for this: some AI models are large enough that they can plausibly only be reasonably run in a state-of-the-art hyperscale datacenter. Open sourcing such models would be largely pointless. Note that this would be a significantly larger scale than even the largest open models available today, one that precludes even doing inference slowly on a small-scale, cheap makeshift cluster. But it's plausible that Fable is there already.
information wants to be free
And once it leaks, it's permanently in the wild.
Interesting times.
K
These models and the hardware they are running on will get even more efficient. We are nowhere near the physical limits of what we can achieve.
Not anymore! Well, if you're like Elon and already taking down the bottle of Cuatro Comas from the high shelf, the economies of scale will continue to work in your favor.
But one of the really neat things about AI is that there is no limit in sight to the scaling incentive. More compute will always get you more: more training, more inference, more parameters, more capacity to build more and better models, more spare capacity to run the slop your models have already built to generate the slop that will succeed it. Back in the dot-com days, or even the "big data" days, you wanted to scale up rapidly but there was a limit: there were only so many customers and they could only produce so much data you could only ingest so fast. In the late 90s, one of the world's most trafficked sites, ftp.cdrom.com, ran on a (single!) dual-processor Pentium Pro system. That was just serving files, and there was certainly room for more CPU oomph to provide more sophisticated services to a huge customer base. But once those customers were served, more compute, storage, and network capacity didn't buy you enough to justify the capex. That is emphatically not the case with AI, and so the incentives for the AI companies are to buy as much compute as they possibly can. What this means in practicing is pre-purchasing capacity at the semiconductor fabs to manufacture chips exclusively for you, and there's only so much of that capacity in the world. Trillion-dollar companies can easily outbid the entire consumer market, and so the incentives for the fabs are now to sell to AI companies at the expense of the consumer market. That's why you're seeing memory prices go through the roof. Modularized RAM for end-user PC builds will soon go the way of the CRT: it will cease to exist as a market product, it won't be manufactured anywhere by anyone. GPUs, CPUs, and storage will soon follow. The only devices end users will be permitted to purchase are all-in-one integrated devices, with CPU, RAM, GPU, storage, and networking either integrated in-chip or soldered on, and they will have just enough capacity to connect to the cloud services the user wants most to use. Most likely, you will be permitted a subscription to such a device, with automatic hardware upgrades at periodic intervals supplied by the manufacturer. If your subscription lapses the device bricks itself. Almost certainly, the OS will be locked down, with no end-user option to install a different one or even run unapproved software.
If reasonably powerful computer hardware for end users exists in this future, it will be available from a single company: Apple. Only they have the leverage to prevent ~100% of manufacturing capacity from going to high-roller, big-tech firms.
How you can be so confident? I can imagine there is some limit and with each scaling iteration gain you achieved will decrease so that further iterations would be more and more look pointless
I don't think this is true. I think prices are rising at the consumer and prosumer level because that's what's required for the mass market to collectively outbid the handful of trillion-dollar companies, at least for the limited share of production they can sustainably demand. This process can continue pretty much indefinitely.
to me Open Source, like Free Software, is something i can run on my own computer. any AI system that runs on a computer that i do not control is by my definition not Open Source.
so how then can Open Source AI win? it can't even compete. even if we collect enough money and create a dedicated Open Source organization to build and run a community owned AI datacenter, how does that help?
so what exactly is the demand here?
Qwen 2.5 72B is surprisingly capable, almost on par with GPT-4o if not a little better. You can run it on a 128GB Mac Studio with 8-bit quantization. You need about 77GB for the weights and ~15GB for your context window & cache.
Pricing remains to be seen, but there's also those new nvidia laptops coming out the surface laptop ultra should have 128GB RAM w/ Blackwell GPU, they're saying 1 petaflop of AI compute, if you can tolerate Windows (no idea if it'll boot Linux until the hardware is out).
These models are roughly ~1 year or less behind the frontier models. We really just need hardware to catch up and alleviate the price pressure on RAM.
Maybe an unpopular opinion here (seening how Y-combinator is his baby), but I think OpenAI and Sam Altman should be financially decimated for cornering the DRAM market. What he's done is a step or two removed from what the Hunt brothers did. His buy-up of future DRAM silicon has measurably harmed personal computing, and he should not get to walk away with a 'win' from it.
I don’t think so. A local run model only needs to serve one or a few people. It seems possible to run a DeepSeek v4 model at full capacity on a server costing 200k usd. Very expensive but not impossible.
Factor in hardware and software improvements over time, and the fact that most people may just need to run a smaller and quantized model, it should take a pc at 10k usd scale.
Fun fact: Qwen was not initially a Apache Licensed project, it was based on a custom license from Alibaba that restricts commercial use: https://github.com/QwenLM/Qwen/blob/ba2d85a13b28ed1ee0dde2d6.... There's no guarantee that they won't just switch it back later.
Kudos for them for switching to Apache License, of course. BUT, they're still a for-profit company. So as DeepSeek btw.
Never, ever, subscribe. When you subscribe, they win. They cornered the silicon market to force you to subscribe. Don't be a sub, or at least keep your sub tendencies in the bedroom. ;^)
Right now there a few people who can run a 1T model at home, even less who can run a 5T model and probably single digits who can run a 10T model.
But if an open source 10T model was available you can be sure people would find new ways to quantize it, new ways to configure hardware and and new ways to think about problems that would make it useful.
1T+ models (Deepseek v4, Kimi K2.6 etc) are available as open weights now, and for ~$5000-$10000 you can run them usefully at home. 2 years ago no on was contemplating that.
$250K to run a 10T model might be possible now. There are many companies that will pay that, and that will push the tools and techniques downwards for the rest of us.
But I am going to need a much beefier machine to get it to the point where it can do any but very trivial dev tasks acceptably fast, and I'm going to need a much beefier model, perhaps one not so aggressively quantized, to keep it on task without the wheels completely falling off. Already we're talking serious money outlay, perhaps still within my programmer salary to accommodate, but just barely. And we're not even where near the performance characteristics a frontier model can support.
This is not true at all. It would be open source if you could download it and run it anywhere that is capable, and are free to move it and modify it as much as you want.
Just because you don't have a computer at home powerful enough doesn't mean it isn't open source.
Being Open Source (tm) will not protect you from the government/others imposing controls on your silicon or what it is allowed to do, which is already happening around the world.
Even having the models be open source won't fix the regulation or economic incentives. Which is not something you can compress into a couple of paragraphs.
AI is civilizational infrastructure and it needs civilizational solutions. Not just source.
Everybody knows AI firms pirated to train, nothing will come of it. A plain example of classist application of law.
The reason for the willy nilly application of their own laws will always be 'national security', of course, since they own infrastructure their interests are a national security.
So tech may shake things up whenever it makes great leaps, but finance capitalism quickly adapts and absorbs the waves.
All states are terroristic parasite gangs, all states [no exceptions].
Your state exists because there is no one else capable of challenging it [no outsider or internal armed militia].
Your state is merely the gang which reigns supreme in your territory - constitutions, democracy, and other grievance pressure relief systems be damned.
You don't get to vote or serve as juror because the system is somehow moral or holy, you get to vote because in historical systems lacking those pressure relief measures the aristocracy tended to be [literally] decapitated on a regular basis.
Democratic measures exist to bribe and persuade your acquiescence so you don't get together with your aggrieved neighbours and go lop heads off ["it's just the rules of the game, you can try again in 2/4/6 more years :^)"].
Seeing politics from this lens should demystify so many seemingly confusing actions and outcomes, it's why no matter how much you vote you never actually "win" and even if you do... it's in such impotent and monkey's paw ways.
No person has an inherent right to exist either. Rights, just like states, or property, or gender, are social constructs. They exist because enough people believe they exist and behave accordingly.
Biden's GPU controls should give you an idea. Thank you, China. Open source AI must win.
Famously, the PowerMac G4 was briefly subject to export controls. Apple turned it into a marketing campaign.
Go ask Claude to criticize Anthropic and see how long your account stays active.
Or are we still collectively brainwashed by the strategic false equivalence established by Big AI CMOs?
I'd never though I'd have to utter the expression "open as in beer".
The blatant attempt at manipulating vocabulary here is... quite blatant.
The weights are the useful artifact here. You can modify them, fine tune them and do what you want with them.
Unlike binary software there is nothing limiting that.
It is also useful to have access to the training recipes and to some extent the data. But I'm of the opinion that learning on something is not copyright infringement, so there are many circumstances where distributing the raw training data will not be possible.
For me this is like Open Office: it is open source, and largely inspired by and learned from Microsoft Office. But they don't need to distribute MS Office for Open Office to be Open Source.
In addition there are models that meet the criteria you appear to propose. The AllenAI models are a good example.
While if "Open Office" switches to a more problematic license at some point, the existing source has all you need for an organization to support the project without regard to the original company (this has happened already!). If Qwen decides to stop distributing models for download, you're basically stuck, _even_ if you have unlimited resources, it's not clear how the released weights help you; your best bet is to start almost from scratch. This has also happened...
These models are not "Open" by any definition of the word. It is just freely redistributable. You can justify yourself in whatever way you want re a cowboy approach to copyright, but this doesn't change the fact that this is not open, and has almost none of the benefits of open, and therefore it is a huge abuse of the word "Open".
Ironically about the only thing that is copyrightable here is the sum of the training data (possibly) _AND_ the software used to build the model (most definitely). The model itself most likely isn't (databases are not copyrightable), which makes it even more pointless to abuse the word "open" for it. All the value is in the former two.
Dependents of an AI-megacorp for our "facts"? Our software? Our work?
It's possible these companies will become everyone's boss, and will dictate to everyone what everyone is allowed to work on, think, say, do, believe, etc.
Before Big Tech springs that trap, we must support and divert resources to open models.
It's worse than this, it's more like our thinking. There's already plummetting math grades [1], handing over our thinking to AI megacorps where there's likely to be a monopoly or duopoly is an incredibly dangerous thing for humanity as a whole.
[1] https://www.dailycal.org/news/campus/academics/failing-grade...
The conundrum which tricks me though - is this a net negative or a positive? If humans are less intelligent, but their output is 2-3 times more intelligent (with AI), what's the result? At what point do we, as humans, stop comprehending anything and give all intelligent work to the neural nets?
And if that does happen, could we live in a society where no work, or at least a significantly less amount of work, is needed? To me, it seems like a dystopian net positive.
It might seem far-fetched to ask these, but I think these questions are getting more prevalent by the day.
A friend of mine asked me if I was optimistic about AI. I told him, it depends on who owns it. If the people own it, I'm optimistic. If the oligarchs own it, I'm pessimistic.
Just listen to what the SV ownership class says out loud. They openly discuss how China cannot "win the AI arms race" and how China's development is existential. Existential to who? It's impossible to fully subjugate people with agency.
What will happen? Massive. Deflation. What will you pay for an oil change? Corn? Meals? Everything is about to be free. But tokens will be expensive!! Sure but, you wont do white collar work anymore so it wont matter what tokens cost.
So really, two professors' gut feel about what the reasons are and not backed by much.
Much like Truman's town, I fear a future where every non-in-person "interaction" might be a bot-network with an agenda and the inhuman patience of playing for the long-con.
We already have personalized, algorithmic advertising and what I would call “control” all over the place: things like consolidated oligarch-owned media.
AI isn’t going to change how we are advertised to or controlled all that much, at least compared to the prospect of being put out of work or taking a huge salary cut similar to the mid-century worker who used to have a $40/hour union factory job and now works at Walmart below health insurance threshold for $15/hour.
What I’m saying is that the general public is most obviously and personally impacted by their economic situation and job prospects.
Joe Citizen who lives by the rules might not even notice that new Flock camera on his street, but he will notice if he’s laid off from his job.
Sure you can. But you're going to have a bad time.
2. The Amish are not a good example because AI will confer an advantage to those that control access to it that has never existed.
It's a better measure than GDP/S&P/401(k) line-go-up especially [re: America] when the native Euro-based population has been aging and dropping for decades, once you strip away all the post Hart-Cellar immigrant lineages.
Let’s play a thought experiment.
Let’s say we have a million people that are so technically sophisticated that they are a space faring civilization capable of seeding the universe with living ecosystems capable of perpetuating life and evolutionary processes. But they are entirely infertile and will never give birth to another individual of their species.
And we have another population that doubles every single year but is incapable of leaving their home planet.
Which one is more valuable?
It depends on what your measure of value is, but if it is to maximize the amount of life in the universe, then population growth is not the right metric, expansion of life through technological means is the more appropriate metric.
The mechanism will become like taxes, you don't have to use public services thus pay those taxes, unless most people comply as it's easy to oppress those who don't.
The parallel isn't about legitimacy, but Mechanism. Some companies already oblige employees to use AI to deliver their work. In a near future we may see jobs seekers registering their AI ID for companies to decide which humans qualify to be plugged into the compensation system, at what rate, and usage conditions to avoid terminations.
Food delivery systems already show a glimpse of how it could look like.
Would be nice if someone figured out how to properly debug a model. Without that? OK, so you have your own open source base model trained on your preferred document set that excluded whatever you think is propaganda, and your own open source RLHF training set based on the judgement of whoever you think is a good egg, and so on.
Last I checked, nobody yet knows how to define a precise rule for automatically checking which of two models made this way is aligned better with whatever your standards are.
The metaphor would be like if we knew what a CPU was but had no idea how to do either chip design or formal verification, and instead randomly mutated the connections between transistors until our test set of 2^16 randomly selected pairs of 32-bit numbers only had one error under addition and two under multiplication.
Worse, because we're making them this way, you have to be a fairly big corporation even when you take shortcuts like DeepSeek did.
And note that I'm not disagreeing about the systemic risk that comes if these models become dictators: people are currently demonstrating they're very eager to outsource their own thinking to these models even when they ought to know better, and corporations are currently demonstrating they're very eager to force workers to use them even when they're mediocre and workers spend half the time they might save from a more competent model just fixing the damage done by their current meh-ness: https://www.theregister.com/ai-and-ml/2026/06/10/brit-worker...
Or capital a comparable sum to pay an AI to approximate the skills of humans I guess is the proposed future?
I'd argue that they already are to some extend, given that well-educated people have no saying on the matter when it comes to extensive use (and by extend reinforcement training) of their models. Well, they have a saying, but exercising that means they're willing to end up without a job.
Now, as far as "what is truth" is concerned, the models are already biased towards notions and opinions that are accepted to some degree by Western values. I had an argument with Claude (why would the tool even argue?) that started by asking it what makes a man attractive, which sent it on a yap on how beauty is subjective, there's no objective way to measure beauty (which implies there's no objective way to improve it), and at some point I was just fed up with how dogged it was to convince me of a value judgement that I don't hold.
It's not about how true or false that value is, it's about what we're going to do the moment someone else dictates the values that exist within the models? What happens when what is trained isn't what you agree? Who's to decide what gets to be reinforced and what's not?
The HN crowd is too deep into productivity rampage to discuss the ethical and moral implications of having a machine so powerful that it spreads worldviews as facts, by whichever government/entity happens to be behind the wheel. At least in the case of extremist forums I can just visit different communities. But what happens when there's only a few winners in the AI race, and the cost of just walking away is too high to pay?
Remember: Google started with "do no evil" and where is that now?
but ok, who is going to initiate such a treaty? US? the orange man won't, and even if he did, no one would care. by the time his term is over and the next AIPAC spokesperson is elected, it will be even more late than it is now. EU? impotent and irrelevant. China? lmao.
And, there is the issue of data poisoning from untrusted nodes. I've almost cracked that last issue with a self-healing checkpointed rollback system that doesn't have to throw out anything that follows the corrupt datum.
But, I'm just one person with an idea and I don't have infinite funds to make this happen. This isn't a small project.
Maybe there would be interest in something like this, now that entire frontier labs are being banned from making further progress.
The total power of all GPUs on the planet dwarf their capabilities, if we had a way to harness them in a distributed way efficiently. We wouldn't be able to train a Fable as fast as them, but eventually having access is better than never having access.
That just isn't true. It misunderstands exactly how much silicon has gone directly to those companies, and exactly how much more powerful said silicon is compared to consumer grade gear.
Very rough math like I said but I doubt it's directionally wrong.
And even if you did force literally everyone on earth with some sort of GPU to max it out 24/7 in service of an open source AI training enterprise - you would waste so much power trying to use that inefficient consumer hardware with the worst latency imaginable that it would be cheaper and faster to get everyone to instead chip in some cash to buy a datacenter with blackwell chips instead! So the idea has no legs whatsoever.
It's pretty useless to compare raw FLOPS, but as a general hand-waving guesstimate, F@H is currently doing about 25 petaflops in a mix of FP16 and 32. AI usually trains at FP8, but to keep things fair the H100 is quoted at 60 FP64 teraflops per unit, so that's 12 FP64 exaflops given its 200k count.
So F@H at its peak did 2.43 exaflops@FP16/32. Colossus 1 does 12@FP64. These numbers are very hand-wavy, but I think the point is made.
By the way, I'm not trying to crap on F@H - I think it's an outstanding project and I've run it in the past. But a volunteer group simply cannot compete with well-funded, concentrated effort like what's going into AI.
Models have limited shelf live while things are improving rapidly, and decentralized training is just more wasteful.
However, things might change if we get to what Karpathy calls "cognitive core" - a stable model backbone which can be extended via skills/adapters/etc. Development of extensions to the core can be a lot more decentralized.
But for now these decentralized training attempts function largely as a deterrent to anti-open-source collusion
The far, FAR superior power efficiency means that even if you did harness every public GPU or GPU-like device on earth, you'd end up consuming so much excess electricity it would be cheaper on net to simply take the money that would have gone to the power bill and spend it on your own datacenter.
And even if electricity was free, having those GPUs spread over the world with internet-level latency will slow everything down by factors of thousands to millions - if it's feasible at all. Regardless, you're not getting fable-oss this decade, maybe even not this century.
It would be better for governments to buy and own their own datacenters, maybe as a coalition, and dedicate their operation to the public good. I believe that is what we actually have to do.
> you'd end up consuming so much excess electricity it would be cheaper on net to simply take the money that would have gone to the power bill and spend it on your own datacenter.
Costs spread over a large population, it really doesn't matter. You're not getting hundreds of thousands of people to pitch half their monthly electric bill to pay for someone else's datacenter. They will pay the electricity themselves quite happily though, if all they need to do is give you compute. This isn't new.
Interconnect is the bottleneck for distributed training, nothing else really.
I recall getting really excited over hinton's FF foray, right before he bailed on AI as a societal direction (which, if anyone ever had the right, I suppose he does). If one squints, one can see a backprop-free base being much easier to train on geographically distributed and heterogenous hardware.
Not sure what you are referring to, unless you don't think h100/h200/b200 are "AI hardware"
> Superpods aren't really power efficient
Maybe not compared to a specialized rig with multiple 4090s, but that is the best case for consumer hardware - the vast majority will be dramatically less efficient than that
Anyway, I agree the interconnect is by far the biggest obstacle and seems insurmountable, I should probably have led with that.
Then have inference go down to the next layer to use those models as a P2P decentralized network.
Maybe like open router could tap federation networks.
One being that extrapolating from like 3 data points is hardly science. All trends break at some point.
The other is that the measures to prevent distillation of their models (if it was a secret sauce of Chinese models) could work if nobody is allowed to use them.
Disagreed. GLM-5.1 is easily as good as Opus 4.5 for all the coding purposes I could throw at it, which is the model that kicked this entire hype cycle into overdrive in the first place.
The thing that big models will always bring to the table is the ability to YOLO weak/under-specified prompts, and spend less time in the loop making sure work gets partitioned correctly. For smaller/simpler tasks the P(success) difference isn't that big.
If you were to take 500 computers with older 1080 GPUs, you might have enough compute/ram equivalent to an H200 GPU for training such a model. Maybe take 10000.
But if those machines are spread over 10000 homes, wired with residential internet service, training a large model will not get anywhere.
You go from "data in the same HBM memory chip" at 4.8TB/s or "data in adjacent GPU" with NVlink at 1.2 TB/s down to 25 MBit/s upload speed. Accessing the next piece of data is going to be about a Million times slower. At the same time you will heat a thousand times more, for a Million times longer.
The first part is not really true though, the chips are not that much faster, the DRAM is not that much faster, and in aggregate it does not matter because there is just so much more consumer hardware out there (although perhaps that is changing as supply shifts toward datacenters).
The interconnect and data locality is the problem. If you could train it like e.g. you can render a scene with monte carlo ray tracing, any result from any node could be merged with any other and the combined result would have converged closer to the limit. I am sure research in that direction exists, it just has not proven effective within the scales it has been attempted.
100% agree. The US government basically has to nationalize AI and capture an outsize portion of the revenue from it in order to fix the economy, as the combination of debt burden and interest rate pressure from de-dollarization/global realignment is going to push us into a death spiral, and even if AI is a smash hit, the ~19% federal capture of corporate revenue isn't nearly enough to pull us out of it. The people owning the compute infrastructure and capturing more profit from AI at that layer is the safest, cleanest way to increase revenue capture, a sovereign wealth fund is a mediocre idea because it's possible to play shell game with stocks and redirect profit/debt (venture capital is quite good at this!).
Currently AI has generated no profit. And as it sits, is a non viable business.
I refuse to include the sellers of shovels as AI revenue.
If the companies buying the shovels are still losing money, then the tool supplier fortunes have nothing to do with the economics of the AI application layer, who is losing money on every prompt.
There's clearly also a lot of pent up demand in the corporate world for inference, the problem is that it's currently expensive enough that enterprises are balking at the cost before they've had a chance to refine processes and see projects through to fruition. That's a tractable problem to solve though.
Airlines, for example, which are so profitable they continually go bankrupt.
Any actual numbers to back this up? I don't see how nationalizing a very cutting edge technology outside of wartime is going to go super well. The leverage that these companies have is the same leverage that TSMC has: you can't just take over and expect things to rocket at the pace its going
One would question why this hasn't already happened as the rule and as opposed to the proliferation of private data centers. However, I am sure the answers are plain and perhaps saddening to us all.
I mean thats good, but they'd have to also build thier own dataset. Which involves either paying people, or breaking the law.
Plus if they do manage to make it work, they will not get any tax revenue from it, as it'll remove the need for labour, which is where a huge amount of tax revenues come from.
its a deeply hard problem with lots of second/third order effects.
Agree about the physics; disagree about the larger point.
I am not questioning that servers packed together may achieve an optimal result in how we are currently doing things, but, and this is my point, what if we didn't.
<< you cannot get that with distributed training
This is entirely the wrong question to ask. The question to ask is: how it could be adapted to distributed training.
I think the most important problem is that you have to marshall enough compute to be meaningful, and that is going to be more and more difficult as frontier compute requirements grow.
But, and it is not a small but, there is no money in it. In fact, big orgs are bound to lose money should something like that succeed.
That does mean you are actually neglecting the more difficult issues.
Or is that too close to the plot of The Matrix?
their bloom model was also a collaborative effort https://huggingface.co/docs/transformers/en/model_doc/bloom
Can it be parallelized or not?
If you take a model, make two copies, and fine-tune each one on different data, what happens when you merge them? Does it work if you freeze different layers?
I think this works if the steps are small enough. And the transfer should become tenable if the steps are big enough. Where's the cutoff?
At most a decentralized effort could contribute a little bit to some bigger centralized effort by doing inference and sandboxed CPU work. Modern model training isn't just backprop, it's got a huge and growing CPU and inferencing component too, which doesn't require intense inter-node communication. For instance, doing RL rollouts for agentic coding requires a lot of plain old inferencing and sandboxed containers for the models to practice in. The final results are just a set of rollouts and scores that can be uploaded back to a central datacenter for GRPO to adjust the weights (relatively cheap). But then, of course, you'd have to stick to models small enough to fit on people's computers so it'd never be competitive.
https://github.com/NousResearch/DisTrO
There are other gradient compression papers from the past reporting large compression rates
It is already possible: https://arxiv.org/abs/2603.08163 . You don't need to sync so frequently, so it can be done over normal internet, it's just less efficient (takes longer to converge).
Also, it wouldn't be able to use a transformer architecture. For inspiration, take a look at Google Maps and how it a much more efficient A* divide/conquer hill-climbing architecture. Think minimized matrix math.
These things can't even center a div correctly half the time.
Not everything is code. Just because it generates a shitty SaaS clone for you and that seemed magical, it does not mean we are approaching "AGI".
An AGI could design an Oil tanker, manage the project from start to finish, handle all contract negotiations and purchasables, payroll, scheduling. Then it could do that 50x over and start a leading logistics firms.
In reality an LLM can't even complete upwork projects that are worth $20 an hour more than 4% or the time.
Source:
https://labs.scale.com/leaderboard/rli
4% guys, 4%. It cannot complete entry level work on fucking Upwork 96% of the time. Stop falling for the marketing and sorry but an LLM will never be AGI.
Its literally just text autocomplete with some RLHF post training, holy shit im losing my mind. I want this hype to end so badly holy shit I need this to end.
It doesn't really matter for most use cases, because the way AI is working is capability saturation. https://www.delanceyukschoolschesschallenge.com/the-rising-t...
The only exception to this is fields that are inherently adversarial (to nature or others) and an edge relative to competition matters.
That's what the Fable harness felt like. You give it a goal and it could try to get there through the shortest path given the tree of possibilities to get there. Iteratively, or recursively.
Perhaps if we make a open coding AI, the design must be along these lines. Something that's easy to train, and serve from local machines. Albeit has loop / recursive hill climbing facilities built it. That way the model gradually keeps moving towards the solutions, in iterations/recursions.
Once this is done, other multi modal things could be pursued.
Open source 'winning' just means that there exists at least one open source alternative to closed models which is as good as, say, GPT 4... I mean, we're essentially there already with Google Gemma models.
As a software engineer, I didn't notice any difference in my productivity since Sonnet. Of course Opus is better and I'm sure Fable is better yet, but we're already hitting diminishing returns in terms of economic value.
I went from Cursor running one of the earlier GPT models to Claude Code on Sonnet and that was essentially a 5x productivity boost for me. Before Claude Code, I only used AI for small snippets. With Claude Code + Sonnet, I could trust it for entire sub-tasks... But I still don't trust Opus with full end-to-end features. I'm not sure it will ever get there. It probably doesn't need to.
Companies need software engineers to have a certain moderately high level of talent but above that level, they really don't care AT ALL. They don't even notice the difference, even if the gap is significant.
Is this really true? We just don't know what the maximum capability of AI is. If it turns out AI can be as intelligent and capable as something like Data from Star Trek, no one is going to be thinking GPT 4 is good enough.
At scale, I can see a benefit in terms of being able to process large amounts of data intelligently to gain a competitive advantage in terms of accruing nominal gains but I think that as long as AI is pursuing dollars, those gains won't translate to real value to the people who control the AI. At best, will translate to more political control; but with added risks and threats too. I suspect it will look more like controlled decline with a small number of entities getting an increasingly large slice of a rapidly shrinking pie.
I think AI may just figure out really complex ways to legally steal people's money. It will probably look all legit on the surface, it will look like the majority of people are just freakishly unlucky and a tiny number of elites are just extremely lucky... But it will be AI behind the scenes orchestrating seemingly random events; choosing who gets lucky and who doesn't.
Might end up literally like a game of monopoly. One player could dominate the game and start receiving all the money but, if you look at the big picture, none of the players are doing anything economically useful; just sitting around a board and moving pieces of paper amongst each other.
It's like the industrial revolution. Many kings and emperors did not like the idea of industrialization because they were already living a luxurious life and understood that it would not benefit them and would only create risks and problems for them personally. They could already afford as many human servants than they needed, what was the point of replacing them with machines to provide the same service they already received? It would give their servants more free time? To an emperor, that would have sounded more like a problem than a solution. It's a bit like that with AI. The people who control AI won't benefit from it beyond what they already have. If it doesn't serve a social cause then it serves nobody.
For all theory purposes there is no limit. Thats what the latest loop engineering trend is about, you are asking AI to find solutions to a problem going by listing steps, and if solution not found in those steps, to treat each step as a separate problem and repeat the process until the master solution to the master problem is found.
Once a solution is found, or new data/insights are generated through this process, the LLM can be trained on this. So in theory you can just keep going like this forever.
Secondly. This is as close to agency you can build inside a machine.
Practically speaking, hardware is a limit. But that can scale up with time.
So we are already looking at some kind of runaway intelligence even if not sentient.
Agency seems to correlate with the ability to make good decisions. It's kind of surprising how much agency is required to make good technical decisions. It's not even about business domain knowledge; a lot of agency is needed even in a pure tech context.
Absorbing all the good ideas or data from openly available systems doesn't seem to be the only determiner
You have either VC funded models looking for a return on investment, or CCP funded models looking to solidify authoritarian "model Chinese society".
Maybe there are some university 4B models, but I doubt those will carry far.
It highlights the difference between companies like Nvidia and Anthropic to me, where one is clearly all about the money and power, and the other is doing it because they genuinely want to accelerate progress and make cool stuff as the driving factor. It's no surprise therefore, that Nvidia is the worlds largest open-source contributor to AI, with over 800 open-weight models.
Of course, these models run on Nvidia hardware, so they benefit from it as a company. But with that healthy mindset, they found a way to contribute that not only benefits everyone, but also benefits themselves.
Contrast to Anthropic, who has gone the complete opposite direction. Closed off everything, restricting everything, fearmongering progress, regulatory capture attempts, the list goes on. I mean, they won't even agree on using AGENTS.md as a standard because CLAUDE.md is free marketing for them. That's the level of disgusting greed we are dealing with...
From a game theory perspective, the cooperative strategies tend to win. As a result, Nvidia has set themselves up for a lifetime. Anthropic however, is playing a strategy of winner takes all, and they're happy to see the world and the entire AI industry collapse in the process.
Compared to bizes like Oracle, Microsoft, or Facebook, I felt that Anthropic was more interested in progress (not to the neglect of business―AI training is expensive at the end of the day), but maybe I've just not seen what you've seen.
But yeah they are good shovel seller and competitor to actually evil companies that literally wants to eat all the world chips and energy supply.
Basically everything Nvidia does in open source is there to make sure their proprietary stack have a good moat and no competitor stack can catch up.
Their license terms are also incredibly generous and allow commercial use, modification, etc, at no cost.
People questioned whether there could ever be a viable open source operating system, yet Linux has been a viable option for a desktop environment for decades now, and that's not to mention its ubiquitous use as a server or phone OS.
You have to start some where. Im guessing, making progress also brings in new ideas how to move further.
From the 1960s to the mid-2000s, every 10 years you'd have a big enough improvement in computing power that you could basically throw out the old computers and replace them with two new ones that were each massive improvements for the same cost (this varied, of course, from hyperbole to massive understatement). We achieved this by shrinking transistors, so we could fit more onto the die. With that, we could dramatically increase clock speeds and the amount of RAM we could cram into a machine
But then we hit the wall of physics. Things haven't stopped improving since ~2015, but they've slowed down so, so much. We've made transistors so small that there's very little more improvement we can get by continuing down that path—they're already seeing serious quantum tunneling effects that need to be adjusted for.
We can no longer assume that we can just powerscale our way out of any computation-cost problem. And breakthroughs, by their very nature, cannot be relied upon—we have no guarantee that there's even a possible way to improve our silicon to scale the way we did before, let alone that it'll be something achievable this decade, or that it'll be cost-effective.
The Chinese would love to produce AI hardware much cheaper, but are blocked from doing so because US sanctions stop a Dutch company from selling them the machines capable of doing so. Coincidentally the companies with a monopoly happen to be in the US.
[1]https://www.eetimes.com/u-s-gives-ok-to-asml-on-euv-effort/
It's the most logical solution for AI anyway, considering that it's training on humanities collective knowledge. It should be more of a public-funded and public-access resource, rather than something greedy tech companies distribute like crumbs while they use unlocked powers internally to clone all of our businesses and swallow the economy.
I am astonished on a daily basis that my Linux computer is so close to the same experience as two operating systems put out by trillion dollar companies. It even does things that those commercial alternatives don’t do.
Also, if DeepSeek is truly putting out models with 1/10th the cost of Western competitors, and a fraction of the employee headcount, I think it implies that there will be a market for someone else to be in the space offering an alternative.
I think about how companies like IBM are so willing to contribute to Linux and give away those contributions for free because they are part of group of corporate sponsors that need an alternative to more dominant commercial players in the market.
Meta “gives away” React for similar reasons: it’s more beneficial for them to have it be a standard and be able to hire people who already know it.
It’s definitely harder to imagine the same ecosystem benefits of an AI model, but maybe it’s out there somewhere.
I could imagine some data center/VPS providers trying to sponsor something like that so that the big AI companies have less leverage over them.
Or maybe all this optimism is a pipe dream?
We live in a world where you can "port" open source software to a new language (Rust) and close it up.
Linux will be ported to Rust and closed. It'll probably also be put under MIT/BSD because nobody cares anymore, but the companies will have their own internal private variants. And these will be the ones that see corporate development.
The value in open source is that it was a lot of concentrated value that was hard to copy, clone, or rip off. Now you can one shot a replacement with a few hundred bucks in tokens.
The economic value of Linux used to be billions of dollars. Soon it'll probably be closer to $0.
It's over.
> Meta “gives away” React for similar reasons: it’s more beneficial for them to have it be a standard and be able to hire people who already know it.
Nah, now you just one shot your thing. And you do it fast enough and with distribution and you win. Eventually human devs can't afford to keep competing and launching startups slower than a hyperscaler's own massively funded efforts.
This is the end of open source and the end of solo developers.
And when the ruthlessly effective models that can one shot entire business functions cost $1,000,000 per invocation. Oracle can afford to press the button to create, say, a new smartphone. But you cannot.
Just wait until devices start requiring trusted computing attestation. The ladder is going to be pulled up.
The scenario you describe is basically that software is free as in beer now. We as a corporation don’t really need to bother using GPL/Apache licensed software because we can one-shot something of our own and not deal with with giving back contributions to the open source community.
But that highway goes both directions. That means that the open source community can also one-shot their software, build more with fewer resources, or it might even just devalue proprietary software even further.
If software is so easy to make, what’s the point of keeping it proprietary? I can’t charge you $100/year for Microsoft Word if I can tell Claude Opus 9.0 to clone it with $100 worth of tokens.
What matters is physical infrastructure (datacenters), the lead on competitors / open source models, and distribution/mindshare.
Thinking of a open weight/source AI as gcc/perl was in the 1990s is more helpful line of approach to take here.
The tool used to achieve a thing must be open.
However, Once real costs are involved, participation tanks. Open source hardware, because it actually requires money to realize, has 1/10,000 the depth of open source software, if that.
Obviously everyone wants an open source AI, but virtually no one wants to fork over money, especially when the end result is others getting it free. A proper training run would require millions of people donating hundreds of dollars. Its not something one guy over a weekend can do...
With a lot of OSS it’s just free volunteer hours.
Compute isn’t free.
The closest thing I can think of is the idea that some group of businesses who can benefit from open models being around might fund that sort of thing. It’s just hard to imagine who they might be.
I feel like they aren't comparable. Open source software just requires human labor, and lots of people are willing and able to share that with the world for free.
Training AI requires capital, to build and power giant datacenters. People don't donate capital at that level.
(Chinese labs famously distilled American models, and that seems to be going well for them. They now have a competitive industry, home-grown talent choosing not to leave, and they now can truly compete without distillation).
If AI fails as a technology, it's going to lead to a great depression and probably either a revolution or WWIII.
They're not even IPOed so how do they tank the market? GPU and ram prices will go down but that will actually help most tech companies.
I don't think the rest of the economy is inflated on the fantasy gains of AI.
We could actually go back to feeling like we can invest in products and content without FOMO.
The fully open model Apertus (although not the frontier) was fully fundend by public Swiss institutions and a strategic national partners. I would not consider Switzerland to be a communist or totalitarian state...
Open source AI manifesto demand that "Opensource AI should remain ... economically viable". That's just wishful thinking.
Anthropic just kneecapped themselves, and possibly OpenAI and Google as well, with their FUD strategy that got fable shutdown by the government.
But that doesn't impact Chinese providers. Then can US companies get investments for expensive model development if they can't actually sell those models-as-a-service?
In the meantime, open source will continue its march onward because while slower, it's completely open source, and the models are already good enough to improve their own work as well as build out the next gen of models.
Right now, and likely forever, because biological threats can be sanctioned at a supply-chain level, the risk of AI is all digital. Fraud, phishing scams, spam, hacks, etc.
The only way we harden the worlds infrastructure to the point that it can withstand attack from bad AI is if we have an abundance of access to frontier intelligence to develop countermeasures.
Otherwise, bad actors will develop these capabilities behind closed doors and use them to hold the world hostage and cause irreparable harm. There's no putting the genie back in the bottle. Good and open-access AI and the people using it are the digital immune system.
If there's an asymmetry where bleeding edge is gated off to only a small group, and allowed to gain exponential power over the immune systems defense grid, the slightest infection will lead to death of the host.
That’s really the only thing stopping people from training or running these models at home:
It doesn't seem to be showing any signs of stopping. Have you used Fable 5? It's a fantastically capable model and trumps anything that came before it. Seedance 2.0 is categorically the best video model, and it's only a few months old.
> the entire business is run by a few old men
Startups tend to skew young, and in this case it's no different. Most of the leaders of AI companies are decades younger than the CEOs in other types of industries.
> who think AI is everything and invest huge sums of money to show other AI companies they need to improve to get more funding from old people.
They're spending capital to win market share and to try to build a moat. One of the most important things in business is building a durable way to keep competitors from taking your market. You spend enormous capital to win customers, and it would suck if other businesses could watch what you did, spend less money, and come in and take everything away. The money being spent is an attempt to have a durable lead.
It's working. Enterprise contracts are deep and sticky tendrils that work through governments and large companies. Both OpenAI and Anthropic have massive partnerships with Fortune 500s, the DoD, you name it - and these contracts will last and print enormous amounts of money. This makes it incredibly hard for other players to enter the market and build a cash flow with which to compete and thrive.
> find something new and innovative
This is easier said than done. It's an incredibly hard problem. It took decades to find the last big technological waves: the PC, the internet, broadband, smartphones. Now AI. These are generational step function increases. The groundwork can be decades old, but it takes time to proliferate before it can become a big business.
Other possibilities include fusion, green tech, quantum computing (useful for crypto breaking, etc.), AI drug discovery, etc. If you go into research one day, try to find an interesting field with potential for commercialization - that could make you very wealthy if you find something you enjoy working on, with lots of greenfield opportunity, that is ripe for turning into products.
Good luck with your game! You should post it here on HN when you finish. You'll get lots of great reviews, comments, and early players. :)
All we can do is hope we end up in the one where things are ok.
You can one-shot a port of Linux to Rust and stop contributing to open source.
The value of software is going to tend towards zero. The value of the software developer the same.
Anthropic is now a kingmaker. It gets to decide which businesses get the expensive private model that can generate entire business functions at the drop of a hat. If you can't afford the price tag, then competition in the market is not for you.
Computing is no longer "personal". It's for big biz only.
Touch grass brother. Seriously.
To make any agent "good", there are two components: the model and the harness. Very few companies can train models, but anyone can build a harness. How much does the harness matter? Can I build a harness that's good enough that I can use open source models with opus level performance? That's the question I've been trying to answer by building better harnesses. None of the existing frameworks have the functionality I need to build a good harness. The features I need are language-level... and so I started building a language called Agency[0].
It's been six months and its going well. Some of the things Agency can do are wild:
- It can pause and serialize execution at any point, making HITL easy
- It has some neat safety capabilities such as handlers[1] and PFA[2]
- You can bundle up any agent as an HTTP or MCP server[3]
- I'm now working on a built-in optimizer to optimize agents (think DSPy).
Obviously, it's a huge undertaking, but having worked with the Agency for six months, I can't imagine going back to another framework. It makes things so easy. I'm working on its built-in agent now [4]. My goal it to get it to be as good as Claude Code, but using open source models. It's still early days, lots of rough edges, but if this sort of thing interests you, I'd love to have a few more people test it out.
[1] https://agency-lang.com/guide/handlers.html
[2] https://agency-lang.com/guide/partial-application.html
[3] https://agency-lang.com/cli/serve.html
[4] https://github.com/egonSchiele/agency-lang/blob/main/package...
Who's gonna pay to power an open source AI? Will it perform well enough to make Chat-GPT and Claude obsolete?
Ever since a Chinese firm released DeepSeek I immediately came to the realization that any US tech firm "owning" AI is simply not going to happen. China will make sure of it. It's in their national security interest not to let that happen.
From the POV of geopolitics, IMHO the US shot itself in the foot by banning the export of the best chips to China. The US also somehow has the power to prevent a Dutch company (ASML) from selling to China too. That makes a little more sense to ban but the combination of banning EUV exports AND banning the best chips sowed the seeds for the destruction of all of this.
By banning chip sales, the US inadvertently created a captive market for Chinese chips with Chinese companies. If there were no chip ban, Chinese companies probably would've bought US chips. But they can't. So they can only buy from Huawei and SMEE (indirectly). The US forced China to realize it was in their national security interest to copy the best lithography and, by extension, the best AI chips.
So DeepSeek was reportedly developed on either older NVidia hardware or smuggled newer NVidia hardware but that won't last either. At some point it'll be completely Chinese made chips that are doing this.
And what's the biggest cost for a model? Training. But you do that once and the model like any software is infinitely copyable so China can under OpenAI, Anthropic and SpaceX (xAI) and that's what they're doing.
But it gets worse for the AI moat. Local models are going to get cheaper and cheaper to run. You can already run 31B models on sub-$5000 hardware. What do you think it'll cost in 5 years? Will larager parameter models keep getting better or will there be a law of diminishing returns? What is a B100 workload now, will be a Macbook Pro workload in as little as 5 years.
What if all these AI data centers are ultimately just going to be commoditized cloud hardware like AWS in the not too distant future? We already see Google renting big from SpaceX. I think the writedown on all these data center investments and the companies that are doing them is going to be extreme in the next 5 years.
Good Guy China! :DDD
Unfortunately General Secretary Xi isn't as AGI pilled as Amodei.
Still, to specifically give a partial answer to your poor faith rhetorical just askin' musing: Florida Conservatives
(specifically turfing nerds from New College of Florida and bringing an excess number of baseball sports bro's to a place that likes math and has no baseball field)
https://www.usatoday.com/story/news/politics/2025/05/22/okla...
I feel extremely strongly that a future in which most companies depend on one or two large AI-megacorps is going to lead to excessive rent seeking sooner or later.
I remain positive that the long term steady state will consist of proprietary models, -but- with open source AI models statistically close.
If compute keeps growing the relative cost of training current frontier models will decrease. An open source Fable/Mythos model simply seems inevitable.
Did open source operating systems win? No, MacOS/Windows are pretty dominant.
Does open source... cloud hosting, social media, ride sharing apps, you name it win? Not in my experience?
Subscribing is cuck paypig behavior.
You're not a cuck paypig now, are you?
Pass this on to your frens, it may save the future!
As long as these models require a lot of computing power, the best models open source or not will be served by corporations who can afford the infra.
Does he mean that the _best model_ should be an open source one (eg: today, something better than Fable 5), or just that open source models should be the default choice for most task?
The former seems an impossibility, closed labs can work off of open and their own closed research. Closed source will always be better. Well, at least until some late-stage enshittification dynamics cause the providers to hobble them.
The latter, becoming a default, not so much. But considering the deep-rooted nature of (for instance) Google, it certainly won't be a walk in the park. This seems to be a similar hurdle as dethroning Chrome as the default browser.
For the average ChatGPT user, I surmise that open-source models are already capable enough. Most people I know who use it (me included) are not paying for it, they are routed to the cheaper models.
What's needed here is everything else other than the model to be in place. Which is to say there isn't a sufficiently good open source ChatGPT app, every open source option requires more fiddling than the ChatGPT app.
No precedent comes to mind for non-tech-user software that is open source and also a default choice. The limitation is rarely from the core-tech capability; core-tech is often the same as what closed source uses.
So the real solution you're looking for is technology that can't be arbitrarily gatekept by a sovereign nation.
That, the 5 different secret levers you have to pull to make it not stupid, the fact you hs e to go to the guy’s twitter account to find all the un-dumbing features and flags that aren’t documented anywhere else. That they decrease thinking budgets silently when they run out of compute instead of announcing the rationing, and gaslighting users at every step of discovery. The fact that internally they have their own coding harness and don’t use Claude Code primarily. The lack of formal evals and consideration for millions of users collective hundreds of millions of hours of investment in their workflows — that’s all off the top of my head, let me tell you how I really feel about what they did to Claude Code..
I adore gpt5.5 and maintain my own codex fork - but I have no idea how long I’ll get this performance / cost - I know it won’t be forever. I’d like to know precisely how much it’ll cost in hardware to run a gpt5.5 open source model locally. Hell a lifetime license to a model I can run locally is also be open to.
But I like building my own tools, from software to physical shop tools. I like being able to rely on my tools.
More responding here to the assertion that this is blowing up due to Fable.
How can you release this to public?!
Why else do you think Anthropic is heavily restricting Fable? You can’t just handwave safety concerns.
Hear me out, economies of scale can only be met when there is a large enough liquidity for it.
The amount of people willing to purchase multiple hardware releases year after year just to run LLM is already tiny and businesses already do use their own hardware and there is no desire for manufacturer to reduce their own margins.
I always wondered if 1000 1M parameter models fine-tuned to specific tasks with a small router could perform as well as 100B models.
And I know this is roughly how MoE works, but current MoE models still require training the model as a whole, and big players don’t have an incentive to change that.
But OpenSource community does…
One day an open source model reaches "good enough" level. Maybe around the level the current frontier has and most people will use that
We're saved /s
Instead of doing a vanity site with a shelf-life of a few days, see where the action already is in online local LLM research and communities and contribute.
/s
If this is a serious concern, why hasn't some red teaming effort demonstrated this possibility already? The fact of the matter is that ablation can't give a model world knowledge it doesn't have as part of training, it can only make the model confabulate. The "nasty" areas of concern are most notable for their world-knowledge requirements, which is where local models are at their weakest anyway.
I'm sure they have but as usual we are a reactive society than proactive. Only when incident has occurred then we have momentum to act.
I've been training a teeny specialised model to run in a browser on a phone to detect harmonium notes played in a song (harmonium turns out is a pita, another story for another day), getting good labelled data is _all_ of the hard work.
That being said, maybe for cheap inference, using a big model to train something ultra-suited for the task at hand might be how we could handle local inference; thinking language specific models.
My bet is that once cost-efficiency becomes a priority, we will figure out ways to get away from the expensive GPU infrastructure on figure out how to architect models for CPUs. I still remember that Microsoft paper about ternary weights.
A loooooot of work to be done for the above to happen
And you're bang on with the storage comparison, we're basically in the mainframe era of this tech, but there's no reason to think that it won't get optimized to the point where you can run the equivalent of current frontier locally.
if it can access private data it will necessarily have more power.
Got a bit more than 1B tokens for $10, it's exceptionally fast, it was able to fix/implement things that 5.5 xhigh struggled with, without trying to act like my best friend or do that coy "undersell the ideal end result so that it can later overshoot it and claim a great success" bullshit.
E: miss me with the "but China" BS, everything I've experienced while using this model has convinced me they are earnestly more concerned with doing the right thing than Anthropic could ever pretend to be. And if you want to ask it questions about Mao, you can go download the weights and spend mid-five-figures to fine tune that out.
Edit: relevant Scott Alexander article from today
https://www.astralcodexten.com/p/my-ai-opinions
> In terms of bioweapons, I expect that closed-source AIs will be heavily optimized against helping with these, and open-source AI will be banned after the first warning shot (or become economically prohibitive even before then).
(For example, I suspect that plenty of folks would view the recently threatened mass scan of the DN42 hobby network as an instance of misaligned agentic behavior that would have wasted non-trivial resources, and I also think that most observers would pin the specific behavior of that AI on a proprietary SOTA model, not an open one. That's clearly not a disaster-level event, but it should scare you if you're concerned about alignment.)
From what I could tell from the very little time that I had to interact with it, it's instruction following seemed more consistent
The other thing that comes to mind is a lot of people commented on how driven it was, so I'm wondering whether figuring out how to keep existing models looping on task might actually be quite a big shift in capability
Any credible references for this? The implication that Anthropic has an even bigger and better model that they haven't released is hard to believe.
Googlers have hinted that Gemini 3 came in at 10T, which seems hard to operationalize, Google's flash and pro releases are staggered in a way that doesn't make sense if flash is a pro distill, and there are enough cases where Gemini flash outperforms pro on the same task that I think it's unlikely it's just being distilled from an "in progress" version of pro.
A thing to keep in mind is that if they release a smaller model halfway between well spaced big model releases, why wait so long on the next big model release if it's sufficiently ready to distill to a smaller model? The ability to demonstrate AI superiority is worth a ton, there's no reason to hold back.
Hints: They created a new label instead of version bumping Opus, they didn't deprecate Opus, and it costs more per token.
Turned out both assumptions were wrong. You couldn't trust sama to turn this into open source, the Chinese did. Elon never.
And we couldn't see demis take over as expected, probably blocked by Google buerocracy.
Proper mass-membership organizations are possible though. Same rules as a public corporation, but one vote per members, and the yearly meeting decides the board members and approves important decisions or introduce motions that steer the organization.
So the right way to do this would be to create something like the "Public LLM development club", some criteria on membership (after all entryism is a thing), some membership fee sufficient that there is money for a reasonable amount of work to be done and then one has to hope that people join.
not a byproduct of the corporation
It is only fair, give that LLMs are enabled by human generated content from the Internet, that they give it back!
Sure, we can do research to bring improvements to open weights models, but it's the same thing: it's either open source or it won't benefit the general public nearly as much.
And people do not just lose operational freedom. They lose the freedom to think, much less act. To some extent, general intelligence has already been outsourced to a few companies. Phones and computers extend the human mind's capabilities, but most people don't have root on their phone. They don't know or control what software is running on it, or how the hardware is made. They don't control their phone, the phone controls them instead. The upstream problem is ownership of general computation, ownership of your own mind, aka self-ownership. This will become more obvious as computing devices become more personally integrated (desktop -> laptop -> smartphone -> smartglasses -> neural interface). Who owns the digital part of your mind? It's not really you at the moment.
Democracy, or any form of negotiation, can only exist among entities with similar capabilities. The gap must be very small. Orangutans may be smart enough to drive a golf cart, but there are no orangutan citizens in a human democracy. So you cannot run from this by being a luddite hermit in the mountains. When the world is full of digitally computing humans much smarter than you, you'll be at their mercy like monkeys are at the mercy of humans. We destroy their habitats and experiment on them as we please.
Now for the first time in history, organisms can increase their own information processing capability at will. We're in the middle of a speciation event where humanity splits into those who own the digital part of their mind vs those who don't, and there will be further splits based on how much compute you own. Though a future where no individual can fully own their mind is also possible.
By "own", I mean being able to command the entire technology stack. If we want sovereignty for the masses, then we must decentralize the entire technology stack for general computation. That means everything from electricity generation, to chip design and fabbing, to all layers of software from firmware to neural networks. All of it must be accessible to every individual. Everyone must be able to make a computer from scratch at home, or at least without leaving the city they live in. Anything less than that, and democratic society as we know it will continue to crumble.
The fundamental idea underlying all of this is: that which reproduces, survives.
At what level of organization can we reproduce?
The digitally computing human species cannot reproduce as individuals. We can only reproduce as a society, at least for now. You can't make a computer from scratch on your own, but you can make a brain from scratch with just one other person of the opposite sex. As the world we live in becomes more suitable for the digitally computing rather than the purely organic, the organic part of the digitally computing human becomes less likely to voluntarily reproduce. If the organic part were to survive without being disempowered in the future, then it's probably by moving the mechanisms for reproductive drive to the society level (via religion or authoritarian government incentivizing or mandating reproduction), or by ensuring that each and every individual has the means to make the digital part of their mind on their own just like how they can make the biological part on their own.
So we must build and adopt frameworks that allow individuals to share resources to run SOTA models in a distributed manner. That way they will also be non-censorable by governments.
Also The only way to prevent that one entity weaponizes it, is by giving EVERYONE access to it.
I think it’s a great project but the communication isn’t clear to me.
I'm not sure exactly why you would buy through them vs rolling your own if you could afford the equivalent hardware.
I'm a firm supporter of local inference though so good on them for doing something
Speculating Experts Accelerates Inference for Mixture-of-Experts: https://arxiv.org/abs/2603.19289
There is a middle way; the policy space also includes government regulating both access and monopoly.
I’m opposed to monopolies of this tech, but I hope the risks of giving everyone jailbroken AGI/ASI are clear.
As a toy example you could imagine a Universal Basic AI where government subcontracts to (n_quorum) labs, everyone gets a token budget, but operating the APIs comes with the safety controls.
If everyone does get to run their own jailbroken AGI, then the only stable societal norm I see is A LOT of surveillance to make sure nobody is building CBRNE threats. This doesn’t seem like a clear win from a civil liberty perspective, though I could see the argument.
Opensource/weight models will get better and better and eventually we will have mythos level running on smartphone/eyeglass hardware.
It is stupidly tedious currently to match supply with demand though because physical hardware like a 16gb ram MacBook doesn't mean there's truly 16gb available let alone matching models and all of their settings (kvcache, context limit, temperature, etc) to demand.
Would appreciate any help cus we need ai inference by the people for the people.
This seems extremely inefficient considering data transfer between model layers if the model is distributed. I found this project called Petals that claim up to 4 tok/s for a 180B model although its repository hasn't been updated in two years.
For prompt processing it would work though, and it could for diffusion LLMs as well.
I have never understood the willingness to make the functioning of or development of a product so completely dependent on the secret sauce of one of two big unprofitable, inscrutable startups.
It really defies sensible engineering principles to do that. So I was never going to do it. I'm exploring AI now but because I have decided that open weights make it a good use of my time.
It's bad enough that any given business often ends up beholden to a single payment platform and the policies of two US credit card providers.
I guess it is the freelancer in me but I always feel nervous when I am asked to put so much energy into studying or learning someone's product, rather than the underlying technology. I still remember the days when Microsoft was pretty much lobbying academic departments with promises of access to the NT source code. I remember a senior figure in our own saying that Linux was a sideshow and access to NT would make us relevant.
More control over destiny is always necessary, and I remind myself and others that the "state of the art" is behind the "cutting edge". Progress is made at the cutting edge, but there is risk of damage. Engineering should focus on building on the state of the art, not on hitching a ride on someone else's progress.
The weights are extraordinarily expensive "capital" that is donated by big organizations who are all at war with each other.
I don't know that it will ever be possible for, for instance, archive.org, to make truly open weights. And, other than archive.org, I can't imagine any other "open source" organization (freebsd? apache?) being in any position at all to make truly open weights.
Maybe governments, government organizations, or universities.
None of whom are currently funded, mandated, inclined, or particularly interested in dumping the money into buying the infrastructure needed to make weights.
In the OSS donations war (Visual Studio Code being a really fascinating example of it) you could see that the taps can't be turned off so easily. Whatever is donated can be built upon forever.
I think there will come a point, soon enough, where open weights models are capable enough that even if they stagnate, they can be augmented with tooling that essentially keeps them current. Maybe we are there now?
But the risk of the taps being turned off is not negligible.
My own feeling is that governments will ultimately ask consortia of universities to train open weights models and support them financially in doing so.
(And for what it is worth, I think diffusion text models are likely to trigger a hardware arms race that makes this possible)
In much the same way that they used to do that for the supercomputer race, which we just don't hear about right now!
I believe open source is important, but for my business I'm just going to use the best tools I have available to me.
I know I can't win that race or outspend the competition. So I have to rely on my instinct that in my area of business, people becoming dependent on agent-written code are getting further and further out of their depth, and that slow and steady will win the race. I am going to spend the time trying to integrate the open source tools into the way I work. (I am still working on this; frankly I may have bigger problems on an individual level than they can solve)
To be maximally clear, if this two-inscrutable-megacorps model does survive, and it becomes how everyone works over even the medium term, I'll have to quit tech.
I will probably retire early and just plan for a shorter, quieter life that ends when I am out of money, because like everyone else I won't be able to afford a longer one.
I don't want that "nobody prompts now, we just specify loops" bullshit for myself and I don't want what it will do to me for anyone I love.
Open source and open weights have to win for human culture's sake but in the short term for the sake of the culture of tech work. We need control over how we use these tools, not just to be steered down whichever channel makes the most money for Dario and Sam.
Hell, I'd wager there's MORE humanity in what we're building because we can afford (literally) to sweat details that previously we couldn't.
I don't understand the argument that somehow because two mega-corps are duking it out, I cannot use their tools to win in my small arena.
I'm not an expert in LLMs so it's hard to understand how much are we lacking, is it just the compute and thinking strategies / parallel chains, or something specific architecturally. But I feel there's value there and I haven't seen anything like it available so far.
Open source AI should and will get better for sure (including better defined first), but the state will have the power over AI never the less.
If you don't like govt's AI policy or the people making those policies, go fix that, don't act like you can avoid them.
For Chinese: saying "Open source AI must win" sounds like singing "L'Internationale, sera le genre humain". The reality is Open Source AI will be over the moment US competitive pressure gone.
For rest of world: there's no real AI for you so far, go work on it or be a citizen of US&A or China.
Just your your natural born intelligence..? It's worked for the past 10k+ years, I'm sure it will work for some time longer
These are still very very (and very) early days of the modern AI and there are so many changes that are gonna happen. It's possible that all the frontier labs of today won't exist in a few years.
If you really want specific open source {LLM, LMM, research, harness, whatever} groups to win over closed source counterparts, you may show your care by trying open source solutions first when solving problems. And if they're really capable, award them with contributions or something.
What we should be saying is: We want a public, community-ran project that does pretraining and training collectively. This means working on a training corpus in public and somehow coordinating the training work.
This is a complete change of what the term means, It's like how people conflate piracy with theft. Two different things, use different words. Free weights, inference code and chat template is very different from a community-ran LLM project.
What’s the world in which frontier model performance is open source? What does that look like? What’s a sensible business model that makes this sustainable? What’s a sensible regulatory framework that doesn’t hamstring AI progress?
Everyone is so enamored with these Chinese lab models like deepseek and qwen and GLM but they exist in a world where the top performance is still claimed by closed source models. These are not developed out of any benevolent commitment to the principles laid out in this article. A world in which OSS is the frontier and its development is controlled and funded by government subsidies of an autocratic government is not reassuring. You can inspect weights but good luck getting the cat back in the bag in terms of capabilities, safeguards, value system, bias, nerfing if it smells American business use cases.
Deepseek was such a darling but guess what, it’s now raising money — 300M at 10 billion valuation. OSS development isn’t sustainable as a business model and in a world where it costs a few hundred million to develop a frontier model, you need a strong business model, or you need strong state subsidies and incentives which introduce a billion new problems.
the most sensible economic picture of OSS models already exist. Commoditize your complement, passion projects for a hedge fund. These are unsustainable and exist at the pleasure of the business or the founder.
Another potential player could be Apple if they open sourced a frontier model, and make back some of the capex on hardware. Imagine AIserve hardware with continued expansion of MacMini’s and Mac Studios.