I was hoping to read more about this, but they don't back up such a claim...
> Fraud, cybercrime, CSAM, harassment, nonconsensual intimate deepfakes, discrimination, and sabotage should stay illegal and be enforced seriously.
The "enforced seriously" part is how they will get you. Don't worry, there won't be a blanket ban on local models. Instead, any model that is "certified CSAM-free" or whatever will be perfectly legal. Meaning that it's impossible to prompt it into producing underage smut in any shape or form.
Of course, any model running locally can be easily jailbroken via prefills, and so in practice it will be a blanket ban. But good luck politically standing up against something that is explicitly worded as an anti-CSAM / anti-terrorist measure and nominally constrained to those areas.
Here’s a plot of a sci-fi thriller. What if, unbeknownst to most, when the vendors were claiming that there AI was too dangerous, they weren’t referring to an increment on what was already out there, but to something far different and more capable. Conscious even! What if that then came up with the financial scheme to end all schemes, knowing that human greed is eminently exploitable, and that the build out of all the global DCs was actually all about removing single points of failure for itself while secretly building out a robot army! We’ll be well on the way to capitalism-ing our own demise.
Maybe Steven Spielberg can make this his next project.
I've designed the role-model protocol for this, allowing routing between any model, however to function optimally it needs consumer applications to use the protocol when sending requests: https://role-model.dev/concepts/how-role-model-works
Which is bullshit, unless you’re an AIaaS company whose revenue is dependent on state-sanctioned market fixing and regulatory capture.
Look, when this shitty cycle ends, we’re likely to find ourselves back in the start of a new memory cycle of surplus and lower costs. We’re talking what very well may be the boom that shatters the 16GB “baseline” we’ve been stuck at for over a decade in consumer computing, and make larger RAM counts (64GB to 1TB+) valuable to consumers specifically for local AI workloads. Local AI isn’t just an enthusiast thing, it’s likely the future of consumer AI provided we don’t let companies and policymakers curtail its use via fearmongering.
Be proactive, and protect consumer right to compute and AI models. Enforce existing laws, don’t outlaw legitimate use just to prop up an unsustainable business model.
And I already have the right to local intelligence, because my GPUs are my private property, and if someone freely releases a beerware model then I can freely download it.
What am I missing?
And when even very intelligent, but excessively conceited, people hear the echo of their own reason9ing from conversational autocorrect and assume it is somehow akin to intelligent life, the normies will go with whatever the plutocrats push with their media outlets too absorbed in their own domain specific knowledge (and cowed into intellectual laziness by other media products they consume eagerly) to ever subject it to much thought that Claude might not be Skynet after all.
You cannot ban local AI without banning local software which is obviously impossible.
Possession of software, even software that can be used for illegal purpose such as RATs and other viruses isn't illegal.
Misleading title.
The article is about local "AI".
The Chinese are the open ones, with free downloads, open weights, and loads of published research. The USA with OpenAI is some of the most closed shit out there.
but I wouldn't advise it
The twist is that AI is pushing all white collar jobs further into bureaucratic work. Nobody is losing their jobs and it's not quite a revolution, but despite all odds and headlines the younger generations are actually much better educated and positioned to do the right things as they take over.
An optimistic take is that since this is the middle class we're talking about, we get more productivity and more justice as a result. The only people upset about this are grifters and charlatans whose time is up.
Local training? Is that feasible?
IDK, I don't live in the US, and I have no idea which "possible law" this website is referring to. In any case, it could be seen as a proactive effort to keep the gates open.
As a side note, I think this is a discussion every open-source supporter should have by actively considering the risks and what actions to take if such a hypothetical law were ever to pass.
No need for huge expensive purpose built tractors. Even if they’re slow you could have half a dozen running 24/7.
It could provide independence for anyone with a modicum of resources.
Aside from that you're 100% correct.
In the worst case it communicates the magnitude of dismsissiveness while demonstrating your intention to claim agency.
Blackmagic Design too.
They are all desperate for Windows to run well on a chip with unified VRAM.
Obviously money is not freely distributed, so how could we possibly make the case that power will be?
All that needs to happen is Xi Jinpeng feel that Chinese SOTA models would be better kept to benefit solely China, and just upon that single utterance, no more models will come out of China.
One they _could_ be referring to is the California AI Transparency Act which isn't compatible with open source licensing, see https://github.blog/news-insights/policy-news-and-insights/g...
wdym by that
> for daily tasks
which are?
Where did you hear this? All the results I can find say the opposite that the US would buy anything.
* https://www.npr.org/sections/codeswitch/2015/01/12/376781165...
That symbolism is akin to prayer.
I am not casting prayer in a negative light, I’m simply categorising your voting concept.
Do your local filters run slow? Does your movie render have no sass? Then sign-up for AaaS!
Nothing has ever been, directly or indirectly, deficit financed at this scale before. In notional or real terms, in history, by anyone.
Now maybe there's an argument that it's a good investment: we are going to beggar the Treasury to buy 2CTA on CoWoS out of Taipei and DCs the size of Manhattan. I personally think we could have done a little more engineering before deciding that the big blind was like, 5 trillion all counted, but it was going to be expensive no matter what.
What super weird is that we're running a project where the "penny" to the "dollar" is the Manhattan Project, and a couple of super weird dudes who do MDMA at Lighthaven now and again are like, in charge of it.
But it must be an epic pain for them to maintain when they are really a software company, and AI tools really should be able to do most of their work locally. Affinity has some local model support for example.
I am sure they are going to have to maintain cloud support for those features for a long time, but it's all a much easier sell if you can also run it locally.
[1] https://memeburn.com/amodei-says-open-source-ai-is-becoming-... [2 July 2026]
[2] https://fortune.com/2026/07/02/anthropic-fable-and-mythos-ar... [July 2, 2026]
[3] https://www.youtube.com/shorts/S72ZRBSNHZc [2 weeks ago] -- Dario: "Now what I do worry about with some of these laggard models is the risks of them, where we have Mythos-class cyber capabilities, 12 months from now we'll have much better cyber capabilities. But the Mythos-class cyber capabilities may just be available for anyone to download."
what does this mean?
I find it unlikely that this is true.
The obvious candidate for things financed through staggeringly large deficits would be war.
But with $1.4T announced capex for the Frontier AI labs, we're not far from the 2nd (illegal) war in Irak: $1.8T of direct military spending.
With that said, I don't know how Frontier Ai companies will ever recover this capex with a glorious $50B of revenues. Add to that that a GPU's lifetime is only a few years and you may see it as a deadend.
NB: did you know Uber destroyed $27B of value since inception? But it still exists. So Frontier AI might just do the same.
I get it. These models can be powerful. But will they be useful is a different question.
If generalist robotic models get good enough to accomplish many varied tasks effectively, training a separate comparable specialized system from scratch for every task would be highly cost-ineffective, even if, in theory, it could have slightly higher reliability.
Xi did not interfere when Chinese labs released GPT-4, o1, o3, etc level models. why draw the line at current SOTA that will be old news in a year?
Once they dominate the market, they will export to keep it that way. The exported models will be a generation or two behind their SOTA. It will nevertheless surpass Mythos unless you believe we are already at the peak. And it might be "free" but it will obviously be backdoored the same way the United States government benefits from Microsoft's global dominance. The same way America hands out fighter jets just generation behind their own best to keep other governments part of the club and in line.
Unlike the United States where every four years we are at risk to some 180 on policy the Chinese have a clear thesis and a clear direction. How many presidents has Xi seen come and go? Do you get the sense he is getting more impressed? They really don't have to do anything but just keep doing what they are doing. Suddenly stopping their export engine is not in keeping with this strategy. That is "stable genius" behavior.
We're going to (and are already on the way to) train deeply general models that can be told: "go tend that field."
And if that's the case, it no longer makes sense to build specialized, purpose-built tractors to house that level of autonomous capability. You instead put it in a humanoid frame (with a little extra sauce for locomotion of said humanoid), and get that to drive your existing tractor.
Visibly praying in public would be symbolic. Prayer is not symbolic; it's thought.
I'm talking about external signaling.
I'm thinking more of the small tasks that are often needed. Mending fences. Pulling weeds. Feeding chickens. Running off coyotes. Lots of things.
The greater question centres about who will tend the machines - 4,000 hectares of seeding requires a week and more of prep work on the air seeder, hoses, points, tines, etc.
What current local models work fine for is delegating clearly-described tasks in a code base the programmer actually understands. Qwen3.6 27B and DeepSeek V4 Flash are both great little workhorses.
There's also GLM 5.2, which is kind of like "store brand Opus", and which might be considered a "near-frontier" model. I don't have as much experience with it.