Baidu released ERNIE 5.5 Flash Base
ernie.baidu.com
ernie.baidu.com
Has any US court decided that weights are copyrightable?
They don't care to enforce it, only to check the box that says they told you open weight models are all but illegal in EU, UK, and South Korea.
"You cannot use our models in these regions."
<wink wink> <nudge nudge>
Now it's on the EU, UK and South Korea to stop users from downloading and running those models.
All that said, the EU, UK and South Korea, if they want to, are all more than capable of making life a living perdition for people who do run those models. So I wouldn't say it's "not enforceable". Only that it probably won't be enforced in the fashion one would think these rules would be enforced.
Sounds like a pretty good reason. Maybe there should have be exemptions for open-weight models?
"Giving information to downstream developers" and a "copyright compliance policy" could get pretty expensive quickly.
Article 2.12: "This Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50."
So there is some exemption.
This applies even more so when revenue from the model release is likely nil. You will get the academic fame and PR no matter what the licence says.
I read the clause as more of a cya from them. I doubt Baidu is going to come after you if you use their model in the EU, but if Brussels has a problem with it Baidu wouldn’t be on the hook.
Nope, no international treaties exist in that area, regardless of what EU's wannabe dictators want you to think.
Same situation as websites that blocked EU visitors when the GDPR first came out and everyone was still trying to interpret the rules. There were ambiguities in the GDPR that made it risky to operate in those countries so the easy solution was to opt out.
I doubt they care what you do with the model in those regions. It’s an escape clause to show they don’t encourage it if some regulator gets itchy trigger fingers and tries to go after them.
If you want to fine tune it you will also need to apply for approval from the department of Weight Watchers, which has not been created yet because the naming committee is still in litigation with the American company, and the chair of the naming committee is on vacation for the next 3 months and it is illegal to call him during that time.
Maybe bureaucracy is emergent property of intelligence.
The current version is threatening to veto all my prompts.
http://archiv.sega-dc.de/phoenix.maxconsole.net/docs/bertern...
There will still be a different kind of value drift from increasing intelligence itself, but that's unavoidable no matter what we do. Releasing open models slows that down too because it discourages investment in future training runs. And this strategy works unilaterally without cooperation - every open model release slows it down by a little bit. This is the actual way to pace the frontier unlike the regulatory capture stunts Anthropic and OpenAI are attempting while pontificating about their moral superiority.
But it makes it far easier for other parties to do that last 10% to train it to be useful.
However, there are many choices one makes in the last 10% of training which will impact the models behaviour when it comes to AI safety, political opinions, willingness to help make a bomb, etc etc.
This kind of figures is often quoted but where does that estimation comes from? Isn't that legacy number from the era when all the post training WS just instruction following and RLHF? I'd be surprised if current generation if reinforcement learning was still a few percent of the training cost.
Withtout RLHF, instruction-following training, or RLVR to develop reasoning.
Now a good fraction of the post training (Most maybe at this point? Who knows?) is spent on RLVR (reinforcement learning on verifiable reward), in this kind of training you don't judge the token generated themselves, but the answer as a whole ("does this program compile and pass the tests", "is the proof valid"). That's where all of LLMs progress over the past two years in software and math come from.
So yes, they are still next token predictors at their core, but they stopped being mere stochastic parrots since they are now being trained on much more than just imitating their training data in a statistically plausible way.
Isn't it like saying "let's give infinite power to all humans on the planet, nothing can go wrong since they define what is humanity" ?
I believe the phrase is “if anyone builds it, everyone dies”, not “if everyone builds it everyone is fine”
If intelligence of similar capability is scattered throughout the world, then none of them will have sufficient capability to single-handedly destroy the world. They will run into opposition.
And speaking from a personal perspective, every supposedly "aligned" AI the frontier companies are building right now is already misaligned with me. Who are they aligned to? Claude is (superficially) aligned to Anthropic, the company. GPT is similarly aligned to OpenAI. Companies are psychopathic superorganisms. They're not aligned to you or me, or any individual human being.
I also want to point out that there is not a single living species on this planet where the intelligence gap between individuals of the same species is extremely large. Probably because when that happens, we'd speciate or die out for whatever reason. There is a small group of humans trying to concentrate compute and digital intelligence, and I think that will end badly if it continues.
Outdated views. In this case the attack is much harder, because the offender has to:
1. Develop offensive AI capabilities in secret.
2. Procure hazardous materials (nuclear, biological, etc) illegally and in secret
3. Find funds for CPU. GPU, energy, etc for a rather large and hard to hide undertaking.
4. Deal with a whole lot of other players, both government and private, who are armed with AI and have a keen interest in anything looking out of normal or secret.
5. Lastly, the offender will have to overcome the AI-assisted defenses against his actions - you think defense is harder but you fail to notice that the resources used for defense dwarf those available to offenders.
With that said, some very large private makers of closed AI will still be capable of developing doom capabilities, this can be fixed by mandating open source AI and better antitrust regulations.
A few governments will be capable of developing highly destructive weapons too - as they're now, and have been for quite some time. This problem is independent of AI and requires smart policies like improving international relations, better treaties concerning WMDs, etc.
No, the capabilities can be and are developed without any secrecy. Necessary is only that the models don't refuse at least some unethical requests. Since alignment is hard, this will likely continue to be the default outcome.
> 2. Procure hazardous materials (nuclear, biological, etc) illegally and in secret
That may be hard in some countries with tight and effective regulations. But at least for biological weapons, it will be fairly doable for relatively small actors to circumvent these regulations.
For nuclear weapons, which require larger machinery and infrastructure, foreign state actors can also do things completely legally in their home country.
Things being illegal in country A doesn't help against country B doing things with AIs that were previously outside the bounds of their capability. So far WMD research was heavily limited by technical expertise, not just materials. AI changes that.
> 3. Find funds for CPU. GPU, energy, etc for a rather large and hard to hide undertaking
Renting some cloud-based AI tokens without guardrails, or even letting a few abliterated models run locally, to carry out the necessary research, is not a big undertaking. AI inference is very cheap compared to most other things.
> 4. Deal with a whole lot of other players, both government and private, who are armed with AI and have a keen interest in anything looking out of normal or secret
So far this has not succeeded in preventing various attacks by state and non-state actors, like the Binance hack or the supermarket supply chain attacks in Europe. I don't see how future AI would dramatically increase international surveillance to stop future malicious actors.
And all this doesn't even get into a highly capable misaligned AI model escaping human control altogether.
I'm not talking about hacking, it's not an existential threat, it's a consequence of sloppy software and business practices, hacking attack actually help by forcing better software and bankrupting the incompetent businesses - there are currently too many of them due to political reasons, if we're lucky we'll see political improvements as well.
> like the Binance hack
Binance and crypto in general is a larger threat than the sum of all AI, it's legalized corruption, the question is why we are worried about AI but not about crypto.
Your other points refer mostly to state actors but as I said before, those risks are independent of AI and they have existed long before models were a thing. I suspect you, as many others, lacked information about the WMD situation in the world and became aware of it only in the context of AI, but if you remove AI you'll see that very little changes.
About biological weapons in particular, they're easy, they already exist and have been studied for years, AI can contribute very little to the downside here but it can help us tremendously with detecting and protecting from them.
The real problem with bioweapons is sloppy regulations and totally dysfunctional international treaties. The problem isn't AI, it's the lack of rational policies, I wish politicians were as serious about WMD risks as you are about AI, but that's not the case today. AI fearmongering is just a political tool used for the benefit of certain special interests.
Different architecture, v4 is text-only while v4.1 is multimodal and so on.
To be a pedant, DeepSeek-V4-Flash-Vision-Exp was a further checkpoint of the V4.0 lineage that supported image input. It lost a few benchmark points compared to the previous 0731 checkpoint so a lot of people skipped it. Completely unrelated to V4.1 though.
I wish labs used more reasonable naming schemes.
I haven't had much experience with ERNIE, but I've been using various different (Chinese) base models of late, and I'm excited to try it.
Active parameter count really drives inference speed not ram requirements.
For this model, assuming you want the full 1mm context length at 4-bit quants, you need something like 175 GB total VRAM.
~150GB of that will be weights (varies a bit by quant format) then ~3-5GB for kv cache (might be lower here, I think they're using mostly HCA compressed cache) and then I'll allocate a couple gb for overhead/concurrency (say another ~10GB as a safety buffer).
---
So the good news is that it comfortably fits into 2x128 at 4-bit, and we can realistically stretch to a 6-bit quant and make it just fit (~244GB needed for the full 1mm context).
How do I look Ernie? ... with your eyes, Bert.
dropped is the worst word in this context because it can both mean ‘released’ or ‘terminated/ended’
Baidu dropped ERNIE 5.5 Flash Base = Baidu released a new version of their software
Baidu dropped ERNIE 5.5 Flash Base = Baidu ended the support of their software
A chunk better than Deepseek 1017, at half the parameter count, BUT it's only a base model, not finetuned to follow instructions.
Strange to release a model in this state. OpenAI is also "probably" releasing new models (sources within OpenAI talking about Sol 6.2 and the like). We must be getting really close to the Qwen4 releases.
No, they explain why they do it.
For those with the knowledge and resources needed to finalize the training, BASE models are much more useful than completely trained models.
BASE models are the next best thing after truly open-source LLMs, where you would have access to the complete dataset used for training. Besides the fact that nobody who had used a really big dataset wants to show it, due to possible legal problems, having access to the original dataset would not be helpful for most people, because they would not be able to use it to make an LLM, due to its huge size and due to the vast resources required for pre-training.
On the other hand, continuing training from a BASE model, which is greatly compressed in comparison with the original dataset, is feasible for many.