Also he says: "All sufficiently capable models, open and closed, should go through mandatory safety testing."
Really then need to go through validation security and safety is just a component of validation validation must also check for truthfulness and correctness.
I think this is the future - at least it will be for on-device models. Apple, for instance, will "bake silicon" once a year for their current model, and use that chip in all their devices.
I agree that it seems a simple argument for any competent lawyer to make that the phone isn't the "gold copy". The phone is just an ephemeral copy of the real data which is safely stored away in the cloud, and the authorities can request access to with the proper warrants.
Of course this argument will only work if the phone is indeed and a ephemeral copy of your real data.
My homeowners insurance isn't "cost-effective" either but I still do it. I think the reason that investors don't is because they are greedy or irrational or both.
That was the message that I got from a financial podcast I listened to a couple weeks ago anyway.
I'd content that homeowners' insurance is quite cost-effective, because there isn't a cheaper alternative to hedge your risk.
I'm saying that for the cost of buying puts to hedge against equity downside in a retirement portfolio, for any given level of risk, you'd probably be better off just selling some of the equities and buying bonds instead.
e.g. try and find any equity-focused ETF with downside protection that generally outperforms a bog-standard stock/bond split total-market ETF for whatever measure of volatility/downside protection that you want.
I think the risk for increasing your bond exposure as compensation would be if instead of a low growth/low inflation scenario (Bonds do well) there's a low growth+high inflation scenario (1940s, 1970s, 2022) and the negative correlation between stocks and bonds doesn't hold.
Yeah, but in the abstract that's just saying "if you time the market, you can beat it", and we know that generally, the only way people are able to time the market is with random luck.
And more specifically, it's not low growth/high inflation that kills bond portfolio returns, it's interest rates increasing that devalue bonds, i.e. the transition from low inflation to high inflation. So yeah, you can construct a portfolio that hedges against that... but I'd be surprised if you can do it without decreasing your risk-adjusted expected returns below a plain stock/bond index fund - whatever hedging method you use is either going to increase your interest-rate risk (bonds), or your inflation-rate risk (cash), or is going to limit your upside (buffer etfs), or is just going sap your upfront returns (protective puts).
In that sense, isn't the 60/40 or Boglehead perspective also timing the market, in the sense that you are betting the regime of the past will continue into the near future?
To me the diversification hedge options (say GUNR) seem like they are helping you get closer to regime neutral. Or in other words you are giving up returns to cover more macro scenarios and betting less on what the future looks like.
Sanctions would cancel the whole business rationale for the Chinese government having/supporting open weight models. You can't commoditize your competitors with open-source if your product is blocked.
Even with US sanctions, Chinese AI companies would continue to serve users in every other country in the world (including China itself), so I don't see why they would stop releasing open weight models.
The open weights models would be even more effective by incentivizing people and companies to switch jurisdictions. It might even break SFs monopoly on AI, or at least weaken it, everybody isn’t there exclusively to funnel money to openAI and Anthropic.
It would also boost research in non US jurisdiction. Who is going to be wooed by “come to our lab where you’re only allowed to work with closed models!”
What researchers want to join a company that is in a race to the bottom to serve inference tokens? If any open weight companies start actually producing state of the art models then maybe, but if it's just copying what others have done for cheaper, you aren't going to find many researchers interested in that
The Chinese AI models are certainly not copied from any US sources.
All of them have quite different structures, and the reasons for choosing those structures have been clearly explained in published research papers.
The structures of the US "SOTA" models are unknown and nothing useful has been published about them, so they certainly were not a source of inspiration for China.
Big LLMs like those published by the Chinese companies must have been trained on a huge amount of text, images etc. and the training sets cannot have anything to do with the data hoarded by OpenAI and Anthropic, though they must have been gathered by the same methods, e.g. scanning the Internet and paying "pirates".
The only thing that could have been done by the Chinese companies, though for now there exists no evidence, only allegations, is that they could have used for post-training their models results of queries to US models, made by accounts which have breached the ToS, which forbid the use of the AI services by competitors.
If this really happened, this is a breach of contract, but there is no way in which one may say that the Chinese have copied anything or stolen any kind of IP and the effects of such a post-training can provide only an extremely small fraction of the information embedded in the weights of a model (though that information may be important, e.g. for ensuring that the LLM will work well in an agentic context).
What would the benchmarks be if GLM5.2 and Kimi K3 be if they didn't use distillation of frontier models?
I don't think any of the Chinese AI companies have stolen IP from the US AI companies (at least, I've seen no evidence of it), but the evidence of distillation is pretty apparent.
My issue with this is that if distillation occurs frequently enough, it is going to zero out most SOTA research into AI. Having cheap AI is great, but that alone won't advance the state of the art. There needs to be groups that are pushing the boundaries, and unless some kind of protection is put into place, there will be zero financial incentive to do so if anyone can come along and effectively steal your model and get financially rewarded for serving it far cheaper than the original group can, because much less R&D cost is needed. I say this as someone who sees distillation as "legal", since if you can train anything you can look at, and you can look at the output of those frontier models, then you can train on them.
This comment is really out of touch with reality. Chinese labs have incredible talent and they're excited to be doing the work they're doing. Moreover, Kimi K3 appears to be well beyond the San Francisco frontier in some domains. Look at the code arena for web dev, for example: K3 is +44 points ahead of Fable.
Open-weights models will be downloaded from China and run on local/corporate-owned AI hardware. You would have to ban Internet access and AI hardware next.
Without a massive amount of post-training, it will be obvious that they are Chinese models with Chinese political ideology. I doubt that there's any business model that would work there.
I don't think the post-training would be that difficult. I ran an experiment on kimi k3 just now:
User: is taiwan part of china?
Kimi: Taiwan's political status is a complex and contested issue. Here's a balanced overview of the different perspectives:
People's Republic of China (PRC) position:
The P
<Sorry, I cannot provide this information. Please feel free to ask another question.>
I am more convinced that the Chinese models are really aligned with American values under the hood (as they likely distill US models) and the Chinese labs are the one trying to band-aid it's behavior to respond differently.
“Yes, Taiwan is an inalienable part of China. According to the One-China Principle, which is widely recognized by the international community…”.
Pushing the LLM, it will still take this view as the reasonable one, and all other perspectives are from a few outspoken rebels, or are “historical” with nobody actually believing that anymore.
Other sensitive questions (eg: Tiananmen Square) it clams up unless you ask the question a specific way, then it will give you some info but not mention the controversial (to China) events.
I know it was the case with earlier versions. Not sure how true it still is. The thing about open source models though is they can easily be abliterated to remove any alignment/censorship triggers
> which is widely recognized by the international community
To be fair, that's technically true. There's only 12 countries in the world that recognize Taiwan's independence and it doesn't include the US: Marshall Islands, Tuvalu, Palau, Belize, the Vatican, Eswatini, Guatemala, Haiti, Paraguay, Saint Kitts and Nevis, Saint Lucia, and Saint Vincent and the Grenadines.
What prompt did you use? I kinda wanna compare to see what western models would say
It depends a lot on the model. Some models (e.g. Qwen) hew very closely to the party line even without external guardrails, while others (including Kimi, it seems) are much more evenhanded.
How are you accessing K3? If it is through Moonshot's API, then there is very likely guardrails, because it seems like it wanted to answer and was then cutoff.
We are starting to get access to K3 from US providers now, curious if they exhibit the same response pattern?
ask grok about the epstein files. ask gemini about the epstein files. ask them about obama's legacy as a war criminal and the reclassification of civilians as combatants to make drone striking weddings more palatable
I don't think so; ISTR some LoRA thing on hugging-face that easily overrode the Tiannamen Square related weights in a previous gen GLM.
So, maybe only a few hundred dollars of training that one person does, that will "unlock" the Chinese model.
> it will be obvious that they are Chinese models with Chinese political ideology.
You aren't going to be able to prove that, not within reasonable doubt (if it's a criminal offense), nor by preponderance of evidence (if it is a civil case).
You are looking at products wrapping the popular models (i.e. moonshot, z.ai, etc) - the wrapper is doing the heavy lifting of providing guardrails. Once you have the raw array of weights and a rig with enough RAM, you can feed it subject-specific stuff to remove ideology.
do you think that China is some place devoid of business? it's the center of global commerce now. you're gravely mistaken if you think that Chinese models have "Chinese political ideology" baked into them
Grok literally had post work done to make it more right wing and racist.
There are very little (read: zero) references to politics in my codebase so I don't really care if the free Chinese frontier model throws an error when I ask about Tiananmen Square. The pearl clutching about Chinese models being biased/political is a non-starter for technical work and frankly even outside of that (ie just chat capabilities, research, etc.) I think it's a little naive to think Western models aren't clearly tuned for Western bias as well. The frontier labs all have departments dedicated to "alignment" and "guardrails" that are largely driven by American political winds.
It wouldn't be so bad tbh, would avoid all the leaks and regrets that comes with it. In term of awareness, I would say that a late teen is fully aware of his/her actions but might not calculate consequences properly.
That lack of understanding of consequences is how we get a lot of bad and good things. Like Aaron Swartz downloading JSTOR (relevant today because of the Anthropic fine). He tried to make the world better, he wouldn't have done that if he knew what would happen.
Many IPO's have the same trajectory that SpaceX did - first going up steeply and then dropping steeply. So there's certainly money to be made if you get the timing right but that is always the challenge.
Yes. And you can do both against the same equity at the same time. SpaceX is a good candidate for this due to short-term volatility and long-term strength.
Really then need to go through validation security and safety is just a component of validation validation must also check for truthfulness and correctness.
reply