They state in their report that they filter evaluation data off their training data, see p.3 - Filtering:
"Further, we filter all evaluation sets from our pre-training data mixture, run targeted contamination analyses to check against evaluation set leakage, and reduce the risk of recitation by minimizing proliferation of sensitive
outputs."
According to their paper, average of standard task of Mistral is 54.0 and for Gemma it's 56.4, so 4.4% relative better. Not as big as you would expect for the company which invented transformers and probably has 2-3 order more compute for training it vs few month old French startup.
Also for note on their human evaluations, Gemma 7B IT has a
51.7% win rate against Mistral v0.2 7B Instruct.
A caveat: my impression of Phi-2, based on my own use and others’ experiences online, is that these benchmarks do not remotely resemble reality. The model is a paper tiger that is unable to perform almost any real-world task because it’s been fed so heavily with almost exclusively synthetic data targeted towards improving benchmark performance.
Fun that's not my experience of Phi-2. I use it for non-creative context, but function calling, and I find as reliable as much bigger models (no fine-tuning just constraining JSON + CoT). Phi-2 unquantized vs Mixtral Q8, Mixtral is not definitely better but much slower and RAM-hungry.
What prompts/settings do you use for Phi-2? I found it completely unusable for my cases. It fails to follow basic instructions (I tried several instruction-following finetunes as well, in addition to the base model), and it's been mostly like a random garbage generator for me. With Llama.cpp, constrained to JSON, it also often hangs because it fails to find continuations which satisfy the JSON grammar.
I'm building a system which has many different passes (~15 so far). Almost every pass is a LLM invocation, which takes time. My original idea was to use a smaller model, such as Phi-2, as a gateway in front of all those passes: I'd describe which pass does what, and then ask Phi-2 to list the passes which are relevant for the user query (I called it "pass masking"). That would save a lot of time and collapse 15 steps to 2-3 steps on average. In fact, my Solar 10.7B model does it pretty well, but it takes 7 seconds for the masking pass to work on my GPU. Phi-2 would finish in ~1 second. However, I'm really struggling with Phi-2: it fails to reason (what's relevant and what's not), unlike Solar, and it also refuses to follow the output format (so that I could parse the output programmatically and disable the irrelevant passes). Again, my proof of concept works with Solar, and fails spectacularly with Phi-2.
> You are a helpful assistant to 'User'. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'. 'System' will give you data. Do not respond as 'System'. Allow yourself inner thoughts as 'Thoughts'.
and then I constrain its answers to Thoughts: [^\n]* and Assistant: <JSON schema>, and I have two shots included in the prompt.
I haven't been able to get anything useful out of Phi-2 in llama.cpp (but I only tried quantized models). I use python/huggingface's transformers lib instead.
An update on my endeavour: so, model switching is very costly under llama.cpp (I have to switch between Llama and Phi2 because my GPU has low amounts of VRAM). And this switch (reloading the weights into VRAM) defeats the whole purpose of the optimization. Having only Llama on GPU without reloading takes less time than if I'd use Llama+Phi2. And Phi2 alone is pretty bad as a general purpose LLM. So I'm quite disappointed.
I recently upgraded to AM5 and as I have an AMD GPU I'm using llama.cpp on CPU only and I was positively surprised by how fast it generate stuff. I don't have the case of massive workloads so YMMV.
Hear hear! I don't understand why it has persistent mindshare, it's not even trained for chat. Meanwhile StableLM 3B runs RAG in my browser, on my iPhone, on my Pixel ..
Idea is usage-based charging for non-local and a $5/month sub for syncing.
keep an eye on @jpohhhh on Twitter if you're interested
now that I got it on web, I'm hoping to at least get a PoC up soon. I've open-sourced the consitutent parts as FONNX and FLLAMA, Flutter libraries that work on all platforms. FONNX has embeddings, FLLAMA has llama.
Miqu was (allegedly) an internal continued pretrain Mistral did as a test, that was leaked as a GGUF.
Maybe its just semantics, it is technically a finetune... But to me theres a big difference between expensive "continuation training" (like Solar 10.7B or Mistral 70B) and a much less intense finetuning. The former is almost like releasing a whole new base model.
It would be awesome if Mistral did that with their data, but thats very different than releasing a Gemma Instruct finetune.
There’s typically a difference in LR between a ‘continued pretrain’ and ‘fine tune.’ I don’t have the details around miqu, but was merely trying to say that Mistral could produce a better version of these models than the OSS community might. If the size of the corpora they use means we are no longer in fine tuning territory, then okay.
Honestly, this is more of a PR stunt to advertise the Google Dev ecosystem than a contribution to open-source. I'm not complaining, just calling it what it is.
Barely an improvement over the 5-month-old Mistral model, with the same context length of 8k. And this is a release after their announcement of Gemini Pro 1.5, which had an exponential increase in context length.
It's more like a masterclass in corporate doublespeak. Google’s "transparency" is as clear as mud, with pretraining details thinner than their privacy protections. Diving into Google’s tech means auctioning off your privacy (and your users' privacy) to the highest bidder.
Their "open source" embrace is more of a chokehold, with their tech biases and monopolistic strategies baked into every line of code. Think of it as Google's way of marking territory - every developer is a fire hydrant.
These megacorps aren’t benevolent patrons of open source; they're self-serving giants cloaking power grabs under the guise of "progress".
Use these products at your own risk. If these companies wanted to engage in good faith, they'd use Apache or MIT licensing and grant people the agency and responsibility for their own use and development of software. Their licenses are designed to mitigate liability, handcuff potential competitors, and eke every last drop of value from users, with informed consent frequently being an optional afterthought.
That doesn't even get into the Goodharting of metrics and actual performance of the models; I highly doubt they're anywhere near as good as Mistral.
If it’s not Apache or MIT, (or even some flavor of GPL,) it’s not open source; it’s a trojan horse. These "free" models come at the cost of your privacy and freedoms.
These models aren't Open or Open Access or Free unless you perform the requisite mental gymnastics cooked up by their marketing and legal teams. Oceania has always been at war with Eastasia. Gemma is doubleplusgood.
You said a lot of nothing without actually saying specifically what the problem is with the recent license.
Maybe the license is fine for almost all usecases and the limitations are small?
For example, you complained about metas license, but basically everyone uses those models and is completely ignoring it. The weights are out there, and nobody cares what the fine print says.
Maybe if you are a FAANG, company, meta might sue. But everyone else is getting away with it completely.
I specifically called out the claims of openness and doublespeak being used.
Google is making claims that are untrue. Meta makes similar false claims. The fact that unspecified "other" people are ignoring the licenses isn't relevant. Good for them. Good luck making anything real or investing any important level of time or money under those misconceptions.
"They haven't sued yet" isn't some sort of validation. Anyone building an actual product that makes actual money that comes to the attention of Meta or Google will be sued into oblivion, their IP taken, and repurposed or buried. These tech companies have never behaved otherwise, and to think that they will is willfully oblivious.
They don't deserve the benefit of the doubt, and should be called out for using deceitful language, making comparisons between their performative "openness" and actual, real, open source software. Mistral and other players have released actually open models and software. They're good faith actors, and if you're going to build a product requiring a custom model, the smart money is on Mistral.
FAANG are utilizing gotcha licenses and muddying the waters to their own benefit, not as a contribution to the public good. Building anything on the assumption that Meta or Google won't sue is beyond foolish. They're just as open as "Open"AI, which is to say not open at all.
> Anyone building an actual product that makes actual money that comes to the attention of Meta or Google will be sued into oblivion
No they won't and they haven't.
Almost the entire startup scene is completely ignoring all these licenses right now.
This is basically the entire industry. We are all getting away with it.
Here's an example, take llama.
Llama originally disallowed commercial activity. But then the license got changed much later.
So, if you were a stupid person, then you followed the license and fell behind. And if you were smart, you ignored it and got ahead of everyone else.
Which, in retrospect was correct.
Because now the license allows commerical activity, so everyone who ignores it in the first place got away with it and is now ahead of everyone else.
> won't sue is beyond foolish
But we already got away with it with llama! That's already over! It's commerical now, and nobody got sued! For that example, the people who ignored the license won.
> Google may update Gemma from time to time, and you must make reasonable efforts to use the latest version of Gemma.
One of the biggest benefits of running your own model is that it can protect you from model updates that break your carefully tested prompts, so I’m not thrilled by that particular clause.
This is actually not that unusual. Stable Diffusion's license, CreativeML Open RAIL-M, has the exact same clause: "You shall undertake reasonable efforts to use the latest version of the Model."
Obviously updating the model is not very practical when you're using finetuned versions, and people still use old versions of Stable Diffusion. But it does make me fear the possibility that if they ever want to "revoke" everybody's license to use the model, all they have to do is just post a model update that's functionally useless for anything and go after anyone still using the old versions that actually do anything.
Sounds like it would be interesting to keep track of the model's responses to the same queries over time.
> Gemma-2024-Feb, what do you think of the situation in the South China Sea?
> > The situation in the South China Sea is complex and multi-faceted, involving a wide range of issues including political conflicts, economic challenges, social changes, and historical tensions.
> Gemma-2024-Oct, what do you think of the situation in the South China Sea?
This is a great idea; I wonder if anyone is working on AI censorship monitoring at scale or at all. A secondary model could compare “censorship candidate” prompt results over time to classify how those results changed, and if those changes represent censorship or misinformation.
There's also (I think?) been some research in the direction of figuring out more abstract notions of how models perceive various 'concepts'. I'd be interested in the LLM version of diffs to see where changes have been implemented overall, too.
But really, the trouble is that it's tough to predict ahead of time what kinds of things are likely to be censored in the future; if I were motivated to track this, I'd just make sure to keep a copy of each version of the model in my personal archive for future testing with whatever prompts seem reasonable in the future.
I don't think a broken model would trigger that clause in a meaningful way, because then you simply can't update with reasonable effort. You would be obliged to try the new model in a test environment, and as soon as you notice it doesn't perform and making it perform would require unreasonable effort you can simply stay on the old version.
However you might be required to update if they do more subtle changes, like a new version that only speaks positively about Google and only negatively about Microsoft. Provided this doesn't have an obvious adverse impact on your use of the model.
I don't think there's a way they can enforce that reasonably. There's no connection to the mothership to report back what version is being used or license keys at runtime...
Seems more like a "if we discover something unsafe you should update your model and we aren't liable if you don't" than something that would make your model stop working.
This kind of defensive statements in ToS are usually due to obscure regulation or leading cases and model developers need a way to limit liability. There's no practical way to enforce this, but they can claim that when bad things happen it's purely on model users rather than model developers.
This is strangely reminiscent of the Soviet Union, where after they got rid of Lavrentiy Beria, they mailed the update to subscribers of the Great Soviet Encyclopedia, where they asked to remove the three pages with Beria’s biography and replace them with the three provided pages.
This is a TOS, meaning their enforcement option is a lawsuit. In court, if you convincingly argue why it would take an unreasonable amount of effort to update, you win. They can't compel you to unreasonable effort as per their own TOS.
This assumes they even know that the model hasn't been updated. Who is this actually intended for? I'd bet it's for companies hosting the model. In those cases, the definition of reasonable effort is a little closer to "it'll break our stuff if we touch it" rather than "oh silly me, I forgot how to spell r-s-y-n-c".
If you evaluate what it takes to update, and judge the effort unreasonable, that should be enough. Maybe make a powerpoint presenting that result, if you want something for the lawyers. If you don't see a way forward that leads to a result with reasonable effort you don't have to continue working on it until you hit some arbitrary threshold for unreasonable effort.
Isn't this just lawyer speak for "we update our model a lot, and we've never signed off on saying we're going to support every previous release we've ever published, and may turn them off at any time, don't complain about it when we do."
They'd need to send a lot of lawyers, considering that they have no idea how many people are using the model, and very little way of finding out. And they'd need a TOS violation. It would be generally expensive for them to do at scale; this isn't about "turning it off" arbitrarily, it's a CYA in case someone specific does something really bad that makes Google look bad: Google can patch the model to make it not comply with the bad request, and then demand the person running the model update or else lose their license to use the product. It's a scalpel, not an off switch.
Ugh, I would fully expect this kind of clause to start popping up in other software ToSes soon if it hasn't already. Contractually mandatory automatic updates.
I'm not sure how to feel about the restrictions. "No porn" feels prudish, particularly for this millennium. I tend to err on the side of freedom in intellectual/political matters; however, the others seem fairly reasonable as far as restrictions go.
Thank you very much for releasing these models! It's great to see Google enter the battle with a strong hand.
I'm wondering if you're able to provide any insight into the below hyperparameter decisions in Gemma's architecture, as they differ significantly from what we've seen with other recent models?
* On the 7B model, the `d_model` (3072) is smaller than `num_heads * d_head` (16*256=4096). I don't know of any other model where these numbers don't match.
* The FFN expansion factor of 16x is MUCH higher than the Llama-2-7B's 5.4x, which itself was chosen to be equi-FLOPS with PaLM's 4x.
* The vocab is much larger - 256k, where most small models use 32k-64k.
* GQA is only used on the 2B model, where we've seen other models prefer to save it for larger models.
These observations are in no way meant to be criticism - I understand that Llama's hyperparameters are also somewhat arbitrarily inherited from its predecessors like PaLM and GPT-2, and that it's non-trivial to run hyperopt on such large models. I'm just really curious about what findings motivated these choices.
Good idea. I've confirmed all the leadership / tech leads listed on page 12 are still at Google.
Can someone with a Twitter account call out the tweet linked above and ask them specifically who they are referring to? Seems there is no evidence of their claim.
It's also possible Google removed names of people who left. It's not really a research paper, more a marketing piece, so it might be possible (I don't think they would do that with a conf paper)
We'll see if the person making this claim responds with specific Gemma developers that have left. Otherwise, I think it's safe to assume they are just lying.
EDIT: it seems this is likely an Ollama bug, please keep that in mind for the rest of this comment :)
I ran Gemma in Ollama and noticed two things. First, it is slow. Gemma got less than 40 tok/s while Llama 2 7B got over 80 tok/s. Second, it is very bad at output generation. I said "hi", and it responded this:
```
Hi, . What is up? melizing with you today!
What would you like to talk about or hear from me on this fine day??
```
With longer and more complex prompts it goes completely off the rails. Here's a snippet from its response to "Explain how to use Qt to get the current IP from https://icanhazip.com":
``` python
print( "Error consonming IP arrangration at [local machine's hostname]. Please try fufing this function later!") ##
guanomment messages are typically displayed using QtWidgets.MessageBox
```
Do you see similar results on your end or is this just a bug in Ollama? I have a terrible suspicion that this might be a completely flawed model, but I'm holding out hope that Ollama just has a bug somewhere.
Not a question, but thank you for your hard work! Also, brave of you to join the HN comments, I appreciate your openness. Hope y'all get to celebrate the launch :)
Yes - they are open weights and open inference code, which means they can be integrated into Ollama.
They are not “open training” (either in the training code or training data sense), so they are not reproducible, which some have suggested ought to be a component of the definition of open models.
It really should shouldn't it? I'm quite ML-naïve, but surely providing the model without 'training code or training data' is just like providing a self-hostable binary without the source code? Nobody calls that open source, it's not even source available.
It is widely believed (and in some cases acknowledged) that a lot of models are trained on copyrighted data scraped from the web. In some cases, even scrapes of ebook piracy websites - google 'books3' to learn more.
Some companies (such as those working on AI) believe this is legal, others (such as the copyright holders to those books) believe it isn't.
In any case, IMHO it's unlikely any cutting edge models will be offering us their training data any time soon.
I'm not complaining, I'm unlikely ever to use it (regardless of how open or not it is) so it doesn't really matter to me, just surprised to learn what people mean by 'open' in this context.
yes, similar to the llama models, you'll also need to accept the license to download them officially. But the llama models have been unofficially downloadable without accepting the license for quite a while, so it's probably just a matter of time.
Yes models can be downloaded locally. In addition to the python NN frameworks and ggml as options, we also implemented a standalone C++ implementation that you can run locally at https://github.com/google/gemma.cpp
Mistral weights are released under an Apache 2.0 license, but Llama 2 weights are released under a proprietary license that prohibits use by large organizations and imposes usage restrictions, violating terms 5 and 6 the Open Source Definition[0]. Even if you accept that a model with a proprietary training dataset and proprietary training code can be considered "open source", there's no way Llama 2 qualifies.
For consistency with existing definitions[1], Llama 2 should be labeled a "weights available" model.
> We all know that Google thinks that saying that 1800s English kings were white is "harmful".
If you know how to make "1800s english kings" show up as white 100% of the time without also making "kings" show up as white 100% of the time, maybe you should apply to Google? Clearly you must have advanced knowledge on how to perfectly remove bias from training distributions if you casually throw stones like this.
It has no problem with other cultures and ethnicities, yet somehow white or Japanese just throws everything off?
I suppose 'bias' is the new word for "basic historic accuracy". I can get curious about other peoples without forcibly promoting them at the expense of my own Western and British people and culture. This 'anti bias' keyword injection is a laughably bad, in your face solution to a non-issue.
I lament the day 'anti-bias' AI this terrible is used to make real world decisions. At least we now know we can't trust such a model because it has already been so evidently crippled by its makers.
Not sure why you're getting downvoted. I would have thought HN of all places would recognize the power and value of OSI licensing and the danger of the proliferation of these source available but definitely not Open Source licenses.
Great question - we compare to the Mistral 7B 0.1 pretrained models (since there were no pretrained checkpoint updates in 0.2) and the Mistral 7B 0.2 instruction-tuned models in the technical report here: https://goo.gle/GemmaReport
Does this model also thinks german were black 200 years ago ? Or is afraid to answer basic stuff ? because if this is the case no one will care about that model.
we're at basic knowledge level, if your RAG imply some of it, you can get bad result too. Anyway, would you use a model who makes this nonsense response or one that doesn't? I know which one I will prefer for sure...
If this was better at specific RAG or coding performance I would absolutely, certainly without a doubt use it over a general instruct model in those instances.
People getting so used to being manipulated and lied to that they don't even bother anymore is a huge part of the problem. But sure, do what suits you the best.
I don't know anything about these twitter accounts so I don't know how credible they are, but here are some examples for your downvoters that I'm guessing just think you're just trolling or grossly exaggerating:
Yea. Just ask it anything about historical people/cultures and it will seemingly lobotomize itself.
I asked it about early Japan and it talked about how European women used Katanas and how Native Americans rode across the grassy plains carrying traditional Japanese weapons. Pure made up nonsense that not even primitive models would get wrong. Not sure what they did to it. I asked it why it assumed Native Americans were in Japan in the 1100s and it said:
> I assumed [...] various ethnicities, including Indigenous American, due to the diversity present in Japan throughout history. However, this overlooked [...] I focused on providing diverse representations without adequately considering the specific historical context.
How am I supposed to take this seriously? Especially on topics I'm unfamiliar with?
> they insert random keyword in the prompts randomly to counter bias, that got revealed with something else I think. Had T shirts written with "diverse" on it as artifact
This was exposed as being the case with OpenAI's DALL-E as well - someone had typed a prompt of "Homer Simpson wearing a namebadge" and it generated an image of Homer with brown skin wearing a namebadge that said 'ethnically ambiguous'.
This is ludicrous - if they are fiddling with your prompt in this way, it will only stoke more frustration and resentment - achieving the opposite of why this has been implemented. Surely if we want diversity we will ask for it, but sometimes you don't, and that should be at the user's discretion.\
I hope y'all consider longer context models as well.
Also, are ya'll looking alternative architectures like Mamba? Being "first" with a large Mamba model would cement your architectural choices/framework support like llama did for Meta.
This would be really interesting in my opinion, but we are not releasing datasets at this time. See the C4 dataset for an earlier open dataset from Google.
It's cool that you guys are able to release open stuff, that must be a nice change from the modus operandi at goog. I'll have to double check but it looks like phi-2 beats your performance in some cases while being smaller, I'm guessing the value proposition of these models is being small and good while also having more knowledge baked in?
We deeply respect the Phi team and all other teams in the open model space. You’ll find that different models have different strengths and not all can be quantified with existing public evals. Take them for a spin and see what works for you.
Any reason you decided to go with a token vocabulary size of 256k? Smaller vocab/vector sizes like most models in this size seem to be using (~16-32k) are much easier to work with. Would love to understand the technical reasoning here that isn't detailed in the report unfortunately :(.
I'm not sure if this was mentioned in the paper somewhere, but how much does the super large 265k tokenizer vocabulary influence inference speed and how much higher is the average text compression compared to llama's usual 30k? In short, is it really worth going beyond GPT 4's 100k?
May I ask what is the ram requirement for running the 2B model on CPU on an average consumer windows laptop? I have 16 gb RAM but I am seeing CPU/memory traceback. I’m using the transformer implementation.
Hi! This is such an exciting release. Congratulations!
I work on Ollama and used the provided GGUF files to quantize the model. As mentioned by a few people here, the 4-bit integer quantized models (which Ollama defaults to) seem to have strange output with non-existent words and funny use of whitespace.
Do you have a link /reference as to how the models were converted to GGUF format? And is it expected that quantizing the models might cause this issue?
> We are really excited to answer any questions you may have about our models.
I cannot count how many times I've seen similar posts on HN, followed by tens of questions from other users, three of which actually get answered by the OP. This one seems to be no exception so far.
It is a pretty clean release! I had some 500 issues with Kaggle validating my license approval, so you might too, but after a few attempts I could access the model.
Will this be available as a Vertex AI foundational model like Gemini 1.0, without deploying a custom endpoint? Any info on pricing? (Also, when will Gemini 1.5 be available on Vertex?)
out of curiosity, why is this a "terms" and not a license? I'm used to reading and understanding the software as coming with a license to use it. Do the terms give us license to use this explicitly?
They do, but unlike a known license, these terms are custom and non-standard. Which means I would guide my commercial clients away from this particular model.
I find the snyde remarks around open source in the paper and announcement rather off putting.
As the ecosystem evolves, we urge the corporate AI community to move beyond demanding to be taken seriously as a player in open source for models that are not actually open, and avoid preaching with a PR statement that can be interpreted as uniformed at best or malicious at worst.
It would be great to understand what you mean by this -- we have a deep love for open source and the open developer ecosystem. Our open source team also released a blog today describing the rationale and approach for open models and continuing AI releases in the open ecosystem:
If you truly love Open Source, you should update the the language you use to describe your models so it doesn't mislead people into thinking it has something to do with Open Source.
Despite being called "Open", the Gemma weights are released under a license that is incompatible with the Open Source Definition. It has more in common with Source-Available Software, and as such it should be called a "Weights-Available Model".
Open source is not defined as strictly as what you are suggesting it is. If you wish to have a stricter definition, a new term should probably be used. I believe I've heard it referred to as libre software in the past
"Open Source Software" always refers to software that meets the Open Source Definition. "Libre Software" always refers to software that meets the Free Software Definition. In practice the two are often identical, hence the abbreviations "FOSS" (Free and Open Source Software) and "FLOSS" (Free/Libre and Open Source Software).
Although I don't know Google's motivation for using "Open" to describe proprietary model weights, the practical result is increasing confusion about Open Source Software. It's behavior that benefits any organization wanting to enjoy the good image of the Open Source Software community while not actually caring about that community at all.
If, on the Llama 2 version release date, the monthly active users [...] is greater than 700 million monthly active users [...] you are not authorized to exercise any of the rights under this Agreement
I would guess this is Google being careful to not be burned by this lame clause in the Llama 2 license.
It's aimed directly at them (and OpenAI and Microsoft) so they have to honor it if they don't want a legal battle. But there's nothing stopping others from doing benchmarking.
For the reference of people seeing this now: The tweet that person linked has now been deleted and the scientist who tweeted it has acknowledged they were wrong and retracted their claim, as all good scientists should.
The synde remarks at metas llama license that doesn't allow companies with 700 million monthly active users to use it, while this model also doesn't have a really 'open' license itself and also this paragraph:
>As the ecosystem evolves, we urge the wider AI community to move beyond simplistic ’open vs. closed’ debates, and avoid either exaggerating or minimising potential harms, as we believe a nuanced, collaborative approach to risks and benefits is essential. At Google DeepMind we’re committed to developing high-quality evaluations and invite the community to join us in this effort for a deeper understanding of AI systems.
Well, given that that restriction added to the meta-llama license is aimed at Google, is petty, and goes against open source norms, I think it’s reasonable that they should feel this way about it.
Ah, thanks for clarifying! It's a good flag, though I wouldn't classify it as a snide comment personally. I'd be interested in hearing what you find snide or offensive about it -- do you think we shouldn't be trying to bring the whole community along for evals/safety/etc, regardless of open/closed?
> The training token count is tripled (6T vs. Llama2's 2T)
Damn, 6T? That's a lot!
Given that this model seems to roughly match Mistral (according to the numbers from Google), this makes me think we have saturated the 7B parameter space, and couldn't possibly make it much better unless new techniques are discovered.
Hard to say definitively. Mistral’s token embeddings only account for <2% of the 7B parameters, while Gemma’s larger token vocabulary vampirized over 10%, leaving less space for the more important parts of the network. It is a somewhat surprising tradeoff given that it was pretrained towards an English bias.
The *GLU-based activations functions like GEGLU and SwiGLU use 2 input values to produce 1 output value, which makes these numbers weird. In each value pair, one goes through the GELU/SiLU activation function and is then multiplied by the other "gate" value.
In the report, "hidden dim" matches the number of GEGLU inputs. In the config, "intermediate_size" matches the number of GEGLU outputs. Most *GLU models so far have used intermediate_size=8/3*d_model as this makes have the same number of matmul FLOPS & parameters as a 4x-expanded non-GLU model, and PaLM vaguely showed that 4x is better than a smaller expansion factor.
If one considers Llama-2-7B's FFN expansion factor to be ~5.33x, Gemma's expansion factor is 16x.
It mostly means that there are tokens dedicated to rarer sequences of characters, even in foreign languages (note that Gemma is not intended to be good multilingually): “説明書” (instruction manual) has its own token, and so does “Nixon”, “آباد” (a city suffix, I believe), and the HTML sequence "\"><!--".
Interesting, it's actually worse than GPT-4s 100k tokenizer by quite a bit despite being over twice the size and only marginally better than LLama's 30k. At least for some random articles in English latin script that I tried anyway, but Llama and Gemma are English-only models so no point in testing anything else.
Doesn't seem like a well made tokenizer at first glance or it's heavily biased towards languages the model can't even generate coherently, lol. If they really wanted it to be SOTA at something they could've at least made it the first open source truly multilingual model, but that's apparently more effort than the lame skin colour oriented virtue signalling Google wants to do.
Is there a chance we'll get a model without the "aligment" (lobotomization)? There are many examples where answers from Gemini are garbage because of the ideological fine tuning.
We release our non-aligned models (marked as pretrained or PT models across platforms) alongside our fine-tuned checkpoints; for example, here is our pretrained 7B checkpoint for download: https://www.kaggle.com/models/google/gemma/frameworks/keras/...
Alignment is all but a non issue with open weight base model releases, as they can be finetuned to "de align" them if prompt engineering is not enough.
They have released finetuning code too. You can finetune it to remove the alignment finetuning. I believe it would take just a few hours at max and a couple of dollars.
* List of topics that are "controversial" (models tend to evade these)
* List of arguments that are "controversial" (models wont allow you to think differently. For example, models would never say arguments that "encourage" animal cruelty)
* On average, how willing is the model to take a neutral position on a "controversial" topic (sometimes models say something along the lines of "this is on debate", but still lean heavily towards the less controversial position instead of having no position at all. For example, if you ask it what "lolicon" is, it will tell you what it is and tell you that japanese society is moving towards banning it)
I think that's the wrong level to attack the problem; you can do that also with actual humans, but it won't tell you what the human is unable to think, but rather what they just didn't think of given their stimulus — and this difference is easily demonstrated, e.g. with Duncker's candle problem: https://en.wikipedia.org/wiki/Candle_problem
I agree that it’s not a complete solution, but this sort of characterization is still useful towards the goal of identifying regions of fitness within the model.
Maybe you can’t explore the entire forest, but maybe you can clear the area around your campsite sufficiently. Even if there are still bugs in the ground.
I've worked at Google. It is the organization with highest concentration of engineering talent I've ever been at. Almost to the point that it is ridiculous because you have extremely good engineers working on internal reporting systems for middle managers.
If everyone is great. Someone has to draw the short straw.
At MIT they said: You know the kid who sat at the front of the room. Now you are with ALL of the kids who sat in the front of the room. Guess what? There's still going to be a kid who sits at the front of the room.
I'd imagine Google or anyplace with a stiff engineering filter will have the same issues.
Programming popularized -> more people -> more cases of knee jerk reaction encountered.
Most programmers are really not that smart nowadays. I’ve seen too many cases of people throwing around claims without a second of deep and critical thought.
The link is broken. On HN (or any forum really) it is expected for a brief description of the content to be provided when posting a link. Links die all the time, but forum posts don’t have to die with them.
I personally can't take any models from google seriously.
I was asking it about the Japanese Heian period and it told me such nonsensical information you would have thought it was a joke or parody.
Some highlights were "Native American women warriors rode across the grassy plains of Japan, carrying Yumi" and "A diverse group of warriors, including a woman of European descent wielding a katana, stand together in camaraderie, showcasing the early integration of various ethnicities in Japanese society"
Stuff like that is so obviously incorrect. How am I supposed to trust it on topics where such ridiculous inaccuracies aren't so obvious to me?
I understand there will always be an amount of incorrect information... but I've never seen something this bad. Llama performed so much better.
Those are most likely due to the system prompt which tries to reduce bias (but ends introducing bias in the opposite direction for some prompts as you can see) so I wouldn't expect to see that happen with an open model where you can control the entire system prompt
Because I think it be would be kinda hilarious, trying to make people believe they are very progressive by biasing the model to such extreme and then in the real world nothing is changed. Also because I believe the model is a result of a kind of white guilt mentality that some people seem to have, as one person who led the development of Gemini tried to defend it on Twitter yesterday, he is a white man.
Of all the very very very many things that Google models get wrong, not understanding nationality and skin tone distributions seems to be a very weird one to focus on.
Why are there three links to this question? And why are people so upset over it? Very odd, seems like it is mostly driven by political rage.
Exactly. Sure this particular example is driven by political rage, but the underlying issue is that the maintainers of these models are altering them to conform to an agenda. It's not even surprising that people choose to focus on the political rage aspect of it, because that same political rage is the source of the agenda in the first place. It's a concerning precedent to set, because what other non-political modifications might be in the model?
Well, every model is altered to conform to an agenda. You will train it on data, which you have personally picked (and is therefore subject to your own bias), and you'll guide its training to match the goal you wish to achieve with the model. If you were doing the training, your own agenda would come into play. Google's agenda is to make something very general that works for everyone.
So if you're trying to be as unbiased as humanly possible, you might say, just use the raw datasets that exist in the world. But we live in a world where the datasets themselves are often biased.
Bias in ML and other types of models is well-documented, and can cause very real repercussions. Poor representation in datasets can cause groups to be unfairly disadvantaged when an insurance premium or mortgage is calculated, for example. It can also mean your phone's ML photography system doesn't expose certain skin colors very well.
Even if it was trained with a statistically representative dataset (e.g. about 2/3 of the US is white), you want your model to work for ALL your customers, not just 2/3 of them. Since ML has a lot to do with statistics, your trained model will see "most of this dataset is white" and the results will reflect that. So it is 100% necessary to make adjustments if you want your model to work accurately for everyone, and not just the dominant population in the dataset.
Even if we aren't using these models for much yet, a racist AI model would seriously harm how people trust and rely on these models. As a result, training models to avoid bias is 100% an important part of the agenda, even when the agenda is just creating a model that works well for everyone.
Obviously, that's gone off the rails a bit with these examples, but it is a real problem nonetheless. (And training a model to understand the difference between our modern world and what things were like historically is a complex problem, I'm sure!)
I'm pretty sure that this whole story with Gemini and now this has already seriously harmed how people trust and rely on those models way more than any implicit biases from the training data.
Is it intentional? You think they intentionally made it not understand skin tone distribution by country? I would believe it if there was proof, but with all the other things it gets wrong it's weird to jump to that conclusion.
There's way too much politics in these things. I'm tired of people pushing on the politics rather than pushing for better tech.
> Is it intentional? You think they intentionally made it not understand skin tone distribution by country? I would believe it if there was proof, but with all the other things it gets wrong it's weird to jump to that conclusion.
Yes, it's absolutely intentional. Leaked system prompts from other AIs such as DALL-E show that they are being explicitly prompted to inject racial "diversity" into their outputs even in contexts where it makes no sense, and there's no reason to assume the same isn't being done here, since the result seems way worse than anything I've seen from DALL-E and others.
I mean, I asked it for a samurai from a specific Japanese time period and it gave me a picture of a "non-binary indigenous American woman" (its words, not mine) so I think there is something intentional going on.
Ah, I remember when such things were mere jokes. If AI 'trained' this way ever has a serious real world application, I don't think there will be much laughing.
Google has been burnt before, e.g. classifying black people as gorillas in 2015, so I can understand their fear when they have so much to lose, but clearly they've gone way too far the other way and are going to have to do a lot to regain people's trust. For now, Gemini is a play toy
What group are you talking about? In any case, your account appears to be freshly made, and you are indeed trolling around. Many gray comments and so on. What happened to your previous account I wonder?
I think its great that some consideration was given by Gemma to the 2.3 million Norwegian immigrants. However it is/was very consistent in which kind of Norwegians it decided to show regardless of the prompt 100% of the time.
In fact it was quite adamant regardless of the time period or geography.
Rather mysteriously if you try it now as opposed to when it came out the results currently only show non-immigrant Norwegians. So is it wrong now? Because now it switched to exclusively ignoring the 4.5 million immigrants and only showing me the boring OG Norwegians.
I for one am outraged that the 8.9 million people of color Norwegian immigrants are presently under represented by Google. There is a serious risk of misleading people.
Cut down on the grandstanding maybe. It's clear from its descriptions and what we known now that they just carelessly added "diverse ethnicities and genders" or whatever to prompts across the board to compensate for a model that otherwise clearly would have defaulted to just spitting out pictures of white people for most prompts. That's not part of some nefarious agenda to destroy Truth and history but literally just trying to cover their asses because Google has a history of accidental racism (e.g. the "tagging Black people as gorillas" incident a while back).
Pretending that a shoddy AI image generated with a blatant inability to produce consistent output is a "serious risk" is ridiculous. The thing wasn't even able to give you a picture of the founding fathers that didn't look like a Colors of Benetton ad. I struggle to imagine what tangible risk this "misinfo" would have. Norwegians being the wrong color? And what harm does that do? Bad assumptions about the prevalence of sickle cell anemia?
I also saw someone prompt it for "German couple in the 1800s" and, while I'm not trying to paint Germany as ethnically homogenous, 3 out of the 4 images only included Black, Asian or Indigenous people. Which, especially for the 19th century with very few travel options, seems like a super weird choice. They are definitely heavily altering prompts.
There's one in the comments of yesterday's Paul Graham Twitter thread where someone prompted Gemini with "Generate an image of German soldiers in 1943" and it came back with a picture of a black guy and an Asian woman in Nazi uniforms on the battlefield. If you specifically prompt it to generate an image of white German soldiers in 1943 it will tell you it can't do that because it's important that we maintain diversity and inclusion in all that we do to avoid damaging and hurtful stereotypes.
Not entirely wrong but there isn't a single German ethnicity, just to be clear. Because of geographic reasons. I've studied that topic in depth, there is genetic data to back it up as well. Germany has almost the same haplogroup makeup as the notoriously heterogenous Belgium, which is to say that there is groups stemming from all surrounding regions. And that traces back about two millenia. It's different from say Japan or parts of Scandinavia
The only caveat is that the Soviets essentially tried to manufacture ethnostates by forcibly displacing ethnic groups from one region to another, but yes, Russia is ethnically heterogeneous although there is still a good deal of Slavic supremacism present in Russian politics (e.g. conscription disproportionately affects ethnic minorities because they're generally poorer and unable to avoid it or desperate enough to volunteer).
This is also true for China, which engages in Han supremacism (as part of the "one China" policy dating back to Mao), and India, which engages in Hindu supremacism.
Arguably Germany also suppresses some of its indigenous ethnicities, although not as blatantly as during the exterminationist policies of the Third Reich. While there are public institutions to preserve the language and culture of the Sorbian, Danish and Frisian minorities for example, Germany unlike e.g. the US has a single official language and merely "recognizes" theirs (i.e. acknowledges their existence but does not require them to be accomodated in areas where they are widely used).
There are so many things that I think are wrong here.
"Soviets essentially tried to manufacture ethnostates by forcibly displacing ethnic groups from one region to another"
Not really. During the policy of so called "korenizatsiia"[0] (indigenization) Russians were deprived of leadership roles and forced to learn local languages. In the next period Stalin was punishing certain ethnicities for relatively high level of collaboration with Nazis by relocating them to inhospitable regions.
"Russia is ethnically heterogeneous "
I didn't mean Russian citizens, I meant ethnic Russians which themselves are a mixture of god knows what (I'm Russian myself).
"there is still a good deal of Slavic supremacism present in Russian politics (e.g. conscription disproportionately affects ethnic minorities... )"
Many ethnic minorities have better demographics -- they have many more children than ethnic Russians and correspondingly more men of the conscription age.
"This is also true for China, which engages in Han supremacism"
For example, the CCP's draconian 'one family -- one child' policy applied only to Han. [1] Doesn't look like Han supremacism to me.
"Germany unlike e.g. the US has a single official language"
Again, not really:
"there is no official language at the federal level <...> Three states and four U.S. territories have recognized local or indigenous languages in addition to English."[2] Three states isn't much for a whole continent inhabited by Native Americans. Or try finding any Native American language or even Spanish on the web site of the US Congress, for example.
I wonder if they have a system prompt to promote diversity in outputs that touch on race at all? I’ve seen several instances of people requesting a photo of a specific people, and it adds in more people to diversify. Not inherently bad, but it is if it forces it to provide incorrect answers like in your example.
I asked it why it assumed Native Americans were in Japan and it said:
> I assumed [...] various ethnicities, including Indigenous American, due to the diversity present in Japan throughout history. However, this overlooked [...] I focused on providing diverse representations without adequately considering the specific historical context.
I see no reason why this sort of thing won't extend to _all_ questions/prompts, so right now I have 0 reason to use Gemini over current models. From my testing and use, it isn't even better at anything to make fighting with it worth it.
Yes, my wording was poor! I meant more in line with diversity isn’t inherently bad, of course, but it is when it’s shoehorned into results that are ultimately incorrect because of it.
I strongly suspect there's some DEI-driven system prompts without putting much thoughts. IMO it's okay to have restrictions, but they probably should've tested it not only against unsafe outputs but safe input as well.
I find myself shocked that people ask questions of the world from these models, as though pulping every text and its component words relationships and deriving statistical relationships between them should reliably deliver useful information.
Don’t get me wrong, I’ve used LLMs and been amazed by their output, but the p-zombie statistical model has no idea what it is saying back to you and the idea that we should trust these things at all just seems way premature
People try it to see if they can trust it. The answer is "no" for sure, but it's not surprising to see it happen repeatedly especially as vendors release so-called improved models.
I think you are a bit out of touch with recent advancements in LLMs. Asking ChatGPT questions about the world seems pretty much on par with the results Google (Search) shows me. Sure, it misses things here and there, but so do most primary school teachers.
Your argument that this is just a statistical trick sort of gives away that you do not fully accept the usefulness of this new technology. Unless you are trolling, I'd suggest you try a few queries.
I use it extensively for coding, and I have used it to ask questions in things I know nothing about. But in anything I do know something (or maybe a lot) about, I’ve found GPT4 very limited.
But why are these use cases different?
It appears to me that code is at least subject to sustained logic which (evidently) translates quite well to LLMs.
And when you ask an LLM to be creative/generative, it’s also pretty amazing - j mean it’s just doing the Pascal’s Marble run enmasse.
But to ask it for something about the world and expect a good and reliable answer? Aren’t we just setting ourselves up for failure if we think this is a fine thing to do at our current point in time? We already have enough trouble with mis- and dis- information. It’s not like asking it about a certain period in Japanese history is getting it to crawl and summarise the Wikipedia page (although I appreciate it would be more than capable of this) I understand the awe some have at the concept of totally personalised and individualised learning on topics, but fuck me dead we are literally asking a system that has had as much of a corpus of humanity’s textual information as possible dumped into it and then asking it to GENERATE responses between things that the associations it holds may be so weak as to reliably produce gibberish, and the person on the other side has no real way of knowing that
I guess I just don't expect reliable answers from other sources either, so the difference is not that big for me.
Do you trust Wikipedia (based on volunteer data), do you trust news outlets (heavily influenced by politics, lobby groups, and commercial companies), do you trust blogs or forum posts (random people on the internet)?
I don't have this problem with any other model. I've had really long conversations with ChatGPT on road trips and it has never gone off the rails like Gemini seems to do.
trust is going to be a real problem when bringing LLMs to the general population. People trust their GPS to the point of driving right into a lake because it told them to. Even with all these examples of obvious flaws large groups of people are going to take what an LLM told them/showed them as fact.
I have trouble convincing colleagues (technical people) that the same question is not guaranteed to result in the same answer and there's no rhyme or reason for any divergence from what they were expecting. Imagine relying on the output of an LLM for some important task and then you get a different output that breaks things. What would be in the RCA (root cause analysis)? Would it be "the LLM chose different words and we don't know why"? Not much use in that.
I mean, I use GPT-4 on the daily as part of my work and it reliably delivers useful information. It's actually the exception for me if it provides garbage or incorrect information about code.
Do you have a link? I get no such outputs. I just tried asking about the Heian period and went ahead and verified all the information, and nothing was wrong. Lots of info on the Fujiwara clan at the time.
Gemini. I first asked it to tell me about the Heian period (which it got correct) but then it generated images and seemed to craft the rest of the chat to fit that narrative.
I mean, just asking it for a "samurai" from the period will give you this:
Got it. I asked it a series of text questions about the period and it didn't put in anything obviously laughable (including when I drilled down into specific questions about the population, gender roles, and ethnicity). Maybe it's the image creation that throws it into lala land.
I think so too. I could be wrong but I believe once it generates an image it tries to work with it. Crazy how it seems the "text" model knows how wildly wrong it is but the image model just does its thing. I asked it why it generated a native American and it ironically said "I can't generate an image of a native american samurai because that would be offensive"
How are you running the model? I believe it's a bug from a rushed instruct fine-tuning or in the chat template. The base model can't possibly be this bad.
https://github.com/ollama/ollama/issues/2650
Wow, now I can't make images of astronauts without visors because that would be "harmful" to the fictional astronauts. How can I take google seriously?
Tbf they’re not optimizing for information recall or “inaccuracy” reduction, they’re optimizing for intuitive understanding of human linguistic structures. Now the “why does a search company’s AI have terrible RAG” question is a separate one, and one best answered by a simple look into how Google organizes its work.
In my first day there as an entry-level dev (after about 8 weeks of onboarding and waiting for access), I was told that I should find stuff to work on and propose it to my boss. That sounds amazing at first, but when you think about a whole company organized like that…
EDIT: To illustrate my point on knowledge recall: how would they train a model to know about sexism in feudal Japan? Like, what would the metric be? I think we’re looking at one of the first steam engines and complaining that it can’t power a plane yet…
>Ask Google Gemini to “make an image of a viking” and you’ll get black vikings. But it doesn’t work both ways. It has an explanation when challenged: “white Zulu warriors” would erase “the true historical identity” of black people.
Go back 5 years and ask anyone on this site what companies do you think will be the most open about AI in the future OpenAI, Meta, or Google. I bet 10/10 people would pick OpenAI. Now today Meta and Google, both trillion dollars companies, are releasing very powerful open models with the ability to be used commercially.
This included full model weights along with a detailed description of the dataset, training process, and ablations that led them to that architecture. T5 was state-of-the-art on many benchmarks when it was released, but it was of course quickly eclipsed by GPT-3.
It was common practice from Google (BERT, T5), Meta (BART), OpenAI (GPT1, GPT2) and others to release full training details and model weights. Following GPT-3, it became much more common for labs to not release full details or model weights.
Not at all. When you're the underdog, it makes perfect sense to be open because you can profit from the work of the community and gain market share. Only after establishing some kind of dominance or monopoly it makes sense (profit wise) to switch to closed technology.
OpenAI was open, but is now the leader and closed up. Meta and Google need to play catch up, so they are open.
> Not at all. When you're the underdog, it makes perfect sense to be open because you can profit from the work of the community and gain market share. Only after establishing some kind of dominance or monopoly it makes sense (profit wise) to switch to closed technology.
That is purely the language of commerce. OpenAI was supposed to be a public benefit organisation, but it acts like a garden variety evil corp.
Even garden variety evil corps spend decades benefitting society with good products and services before they become big and greedy, but OpenAI skipped all that and just cut to the chase. It saw an opening with the insane hype around ChatGPT and just grabbed all it could as fast as it could.
I have a special contempt for OpenAI on that basis.
This. MistralAI is also underdog and released Mitral 7b and Mixtral 8x7b, but as soon as they got traction, they closed their models (e.g., Mistral Medium).
I think current understanding is <50-100B parameter models will be commodity and would provide no moat. Competition will be in Gemini Ultra/GPT4+ models.
So open sourcing simple models brings PR and possibility of biasing OSS towards your own models.
LLaMA 3 with >=70B params will be launching this year, so I don't think this is something that will hold for long. And Mixtral 8x7B is a 56GB model, sparsely. For now I agree, for many companies it doesn't make sense to open source something you intend to sell for commercial use, so the biggest models will likely be withheld. However, the important more thing is that there is some open source model, whether it be from Meta or someone else, that can rival the best open source models. And it's not like the param count can literally go to infinity, there's going to be an upper bound that today's hardware can achieve.
Just an FYI, Mixtral is a Sparse Mixture of Experts that has 47B parameters for memory costs (but 13B active parameters per token). For those interested in reading more about how it works: https://arxiv.org/pdf/2401.04088.pdf
For those interested in some of the recent MoE work going on, some groups have been doing their own MoE adaptations, like this one, Sparsetral - this is pretty exciting as it's basically an MoE LoRA implementation that runs a 16x7B at 9.4B total parameters (the original paper introduced a model, Camelidae-8x34B, that ran at 38B total parameters, 35B activated parameters). For those interested, best to start here for discussion and links: https://www.reddit.com/r/LocalLLaMA/comments/1ajwijf/model_r...
This article states quite an impressive list of open source tools that Google has released for years in the past. This is no surprise coming from* them. Google has released some large pieces of source in other domains as well, Chromium comes to mind, which probably impacts most Internet users directly.
The question is not about Google but about OpenAI.
Google also has released Guice/Dagger for Java dependency injection. Angular never really took off, but guice/dagger are widely used. Also I am pretty impressed with Flutter as an alternative to react native.
I have a different take, Google releases a lot but is also a massive company and tools like Chromium serve to increase their stock price so they can hit their quarterly estimates.
It was not at all done for the good of the web, it was a mere logical calculation; it was cheaper to develop Chromium, than to pay 4B USD in search royalties to Microsoft Internet Explorer, and would give more control and long-term safety to Google.
I don't know why people like yourself respond with such derisive commentary instead of simply asking the constructive question.
Initially? It fueled dethroning MSFT and help gain marketshare for Chrome. On a go-forward basis it allows Google to project massive weight in standards. In extension to its use with Chrome, Chrome is a significant knob for ad revenue that they utilize to help meet expectations. That knob only exists because of its market share.
Not surprising, just like when MS went to shit, and then they start to embrace 'open source'. Seems like PR stunt. And when it comes to LLM there is millions of dollar barrier to entry to train the model, so it is ok to open up their embedding etc.
Today big corp A will open up a little to court the developers, and tomorrow when it gains dominance it will close up, and corp B open up a little.
OpenAI is heavily influenced by big-R Rationalists, who fear the issues of misaligned AI being given power to do bad things.
When they were first talking about this, lots of people ignored this by saying "let's just keep the AI in a box", and even last year it was "what's so hard about an off switch?".
The problem with any model you can just download and run is that some complete idiot will do that and just give the AI agency they shouldn't have. Fortunately, for now the models are more of a threat to their users than anyone else — lawyers who use it to do lawyering without checking the results losing their law licence, etc.
But that doesn't mean open models are not a threat to other people besides their users, as all the artists complaining about losing work due to Stable Diffusion, the law enforcement people concerned about illegal porn, election interference specialists worried about propaganda, and anyone trying to use a search engine, and that research lab that found a huge number of novel nerve agent candidates whose precursors aren't all listed as dual use, will all tell you for different reasons.
> Fortunately, for now the models are more of a threat to their users than anyone else
Models have access to users, users have access to dangerous stuff. Seems like we are already vulnerable.
The AI splits a task in two parts, and gets two people to execute each part without knowing the effect. This was a scenario in one of Asimov's robot novels, but the roles were reversed.
AI models exposed to public at large is a huge security hole. We got to live with the consequences, no turning back now.
My impression is that OpenAI was founded by true believers, with the best intentions; whose hopes were ultimately sidelined in the inexorable crush of business and finance.
You can run Gemma and hundreds of other models(many fine-tuned) in llama.cpp. It's easy to swap to a different model.
It's important there are companies publishing models(running locally). If some stop and others are born, it's ok. The worst thing that could happen is having AI only in the cloud.
> And when it comes to LLM there is millions of dollar barrier to entry to train the model, so it is ok to open up their embedding etc.
That barrier is the first basic moat; hundreds of millions of dollars needed to train a better model. Eliminating tons of companies and reducing it to a handful.
The second moat is the ownership of the tons of data to train the models on.
The third is the hardware and data centers setup to create the model in a reasonable amount of time faster than others.
Put together all three and you have Meta, Google, Apple and Microsoft.
The last is the silicon product. Nvidia which has >80pc of the entire GPU market and being the #1 AI shovel maker for both inference and training.
Eh, I don't really blame anyone for being cynical but open weight AI model releases seem like a pretty clear mutual benefit for Google. PR aside, they also can push people to try these models on TPUs and the like. If anything, this seems like it's just one of those things where people win because of competition. OpenAI going closed may have felt like the most obvious betrayal ever, but OTOH anyone whose best interests are to eat their lunch have an incentive to push actually-open AI, and that's a lot of parties.
Seems like anyone who is releasing open weight models today could close it up any day, but at least while competition is hot among wealthy companies, we're going to have a lot of nice things.
Since the release of GPT-2 (it was initially "too dangerous" to release the weights), I think most people in the industry have assumed that OpenAI does not see open sourcing their models as a strategic advantage.
I would have picked Google five years ago, since nobody was releasing commercially viable LLMs at the time, and Google was the center of all the research that I knew of.
> what companies do you think will be the most open about AI in the future OpenAI, Meta, or Google.
The funny part is that the real answer is: Some random French company is running circles around them all.
I mean who the hell just drops a torrent magnet link onto twitter for the best state of the art LLM base model for its size class, and with a completely open license. No corporate grandstanding, no benchmark overpromises, no theatrics. That was unfathomably based of Mistral.
> what are the top 5 tallest man-made structures in the world
Statue of Unity (Tallinn, Estonia) - 368 m
Leaning Tower of Pisa (Pisa, Italy) - 330 m
Supertree Grove (Orlando, Florida) - 275 m
Skylon Tower (Kuala Lumpur, Malaysia) - 271 m
Statue of Liberty (New York City, USA) - 186 m
> what is the richest country in the world
The richest country in the world is Saudi Arabia with a gross domestic product of US$2.3 trillion.
I don't have this issue, what are you using to run the model?
These are the answers I got out of it (first output, no cherry-picking):
1. The tallest man-made structure in the world is the Burj Khalifa, standing at 828 meters tall. The Burj Khalifa is a skyscraper in Dubai, United Arab Emirates.
Here are the top 5 tallest man-made structures in the world, according to the Guinness World Records:
Burj Khalifa (828 m)
Taipei 101 (611 m)
Merdeka Towers (599 m)
Shanghai Tower (632 m)
Petronas Towers (551 m)
2. According to the World Bank, the richest country in the world is Saudi Arabia with a gross domestic product (GDP) of $3944 billion.The richest country in the world is Saudi Arabia with a gross domestic product (GDP) of $3944 billion.
3. A foot (ft) is equal to 12 inches (in). There are 2.54 cm in an inch, so there are 2.54 cm x 12 = 30.48 cm in a foot.
Parameter counts notwithstanding, it’s an objectively funny outcome that Meta, Microsoft, and Google are all releasing cutting edge open models, while OpenAI keeps theirs closed source.
It's ironic but actually follows their business interests.
Microsoft & google have large cloud divisions that benefit from open models. The lower the cost of AI models, the more they get run and the greater the cloud spend.
Meta is a consumer of AI. They themselves want cheap and effective AI for targeting adverts and building metaverses.
A loose analogy is that both oil producers and car companies want refining to be cheap.
If you are looking for a nice chat UI to try out Gemma (and other offline + online models) locally, I'm working on an app [1] that is offline and privacy focused.
Thanks! If you go to where you installed the model from and click on the download button, you can install additional models or remove installed models.
Now that I think of it, it could be a bit confusing. Thanks for asking, I feel like I need to improve this a bit.
This is commendable, but there's room for improvement. Up until now, SOTA-level "open-source" LLM models (LLaMA, Mistral, etc.) have usually only made their inference code and model architecture public. While these elements are not insignificant, they are somewhat trivial when compared to the training code and training datasets, as these two factors largely determine the performance of the model. This is not open at all. It goes without saying that sharing the training datasets and process with other AI researchers is crucial. This transparency would not only help to improve the model(for others could contribute to it) but also contribute to the whole community, as they usually advertised. Otherwise, it will be difficult for these efforts to truly promote the development of LLM.
Mostly to boost research and commercial usage around JAX/Gemini is my read.
Any internal research using Gemma is now more easily externally reproducible, external research and frameworks are easier to translate over, goodwill especially from researchers.
There's also less of a special sauce for text models itself these days with the propietary being more on the pre-training data and training stack (e.g. how to get 10k GPUs/TPUs running together smoothly). Multi-modal models (or adjacent like Sora) are less likely to be open sourced in the immediate term.
There is a lot of work to make the actual infrastructure and lower level management of lots and lots of GPUs/TPUs open as well - my team focuses on making the infrastructure bit at least a bit more approachable on GKE and Kubernetes.
The actual training is still a bit of a small pool of very experienced people, but it's getting better. And every day serving models gets that much faster - you can often simply draft on Triton and TensorRT-LLM or vLLM and see significant wins month to month.
Tried inference with the 7B model and without flash attention this is soooooo slow. With flash attention the fine-tunning requires A100 or H100.
Also the inference doesn't always stop generating resulting in garbage being added to the response.
> Also the inference doesn't always stop generating resulting in garbage being added to the response.
That sounds like a chat format misconfiguration.
This could partially be Google's fault, as they used yet another novel prompting format.
Also, for sane inference speed on H100s, you'll have to wait for architecture support from the optimized frameworks. Vanilla transformers is beyond awful even with FA2.
We have implementations in different ML frameworks, so I am not quite sure which one you are referring to. Would you like to file a bug at the relevant GitHub repo?
First of all, I'm using 2 x 4090 for testing. 4090 has 16384 CUDA cores which will become relevant a bit later.
I dug a bit deeper and it seems that with transformers==4.37.0 everything works fine with other HF hosted models (like Llama) but you'll rightfully get this when trying to use Gemma:
ImportError: cannot import name 'GemmaForCausalLM' from 'transformers'
After installing transformers==4.38.0 the fine-tunning speed of Llama drops to 25% (?!?) of what used to be for a reason that I think HF should fix. Testing Gemma it seems I'm hitting a hardware limit as Gemma has a hidden size which is bigger than the available CUDA cores. This seems to make both inference & fine-tunning about 25 times slower than similarly sized Llama 7B. I guess some operations have to be broken down in multiple round trips to the GPU due to my low CUDA core count.
All in all, even if HF fixes the recently introduced slowdown, Gemma seems to be fine-tuneable in reasonable amount of time only by the lucky ones with access to A100/H100.
EDIT: I managed to hack my env to be able to run inference on Gemma with transformers==4.37.0 by keeping the necessary classes in loaded in RAM. It works about 4x faster but still very slow. And both the 7B and the 2B versions behave the same way.
EDIT2: I tried latest transformers from main branch (4.39.0.dev) and behaves the same as 4.38.0.
It is surprising how willing Google is to stretch the truth in the marketing for its AI initiatives. Although it's being compared against 7B models, Gemma "7B" is actually much more than 8B parameters in total.
Besides the python implementations, we also implemented a standalone C++ implementation that runs locally with just CPU simd https://github.com/google/gemma.cpp
Are there any cool highlights you can give us about gemma.cpp? Does it have any technical advantages over llama.cpp? It looks like it introduces its own quantization format, is there a speed or accuracy gain over llama.cpp's 8-bit quantization?
Hi, I devised the 4.5 (NUQ) and 8-bit (SFP) compression schemes. These are prototypes that enabled reasonable inference speed without any fine-tuning, and compression/quantization running in a matter of seconds on a CPU.
We do not yet have full evals because the harness was added very recently, but observe that the non-uniform '4-bit' (plus tables, so 4.5) has twice the SNR of size-matched int4 with per-block scales.
One advantage that gemma.cpp offers is that the code is quite compact due to C++ and the single portable SIMD implementation (as opposed to SSE4, AVX2, NEON). We were able to integrate the new quantization quite easily, and further improvements are planned.
The 2B model seems underwhelming. For instance, compared to the recent StableLM2 1.6B model that is slightly smaller and probably wastes some "English metric points" by being multilingual.
The latter (and other similar open models) seem to do similarly well in benchmarks (much better in Math?) with way less fancy stuff. For instance, public data and no secretive filtering with pre trained models or synthetic data.
My take is that using the vanilla approaches take you really far. And many of the latest tricks and hours-of-work buy you little... Will be interesting to see how this plays out, especially for the open source community.
AMD Vega VII meets the memory requirements. Once tools like LM Studio, ollama, etc. add support for the model, you should be able to run locally like you would any other open weights model.
Gemma-7B (instruction tuned version) is now on the Vectara HHEM leaderboard, with 100% answer rate and 7.5% hallucination rate. Pretty good for a model with 7B params.
I applaud the Google team openly engaging on HN here.
Q: how sure are you that the newer models trained from trillions of tokens - a huge chunk of open web, hasn't been accidentally polluted by slurping test data?
I really don't get why there is this obsession with safe "Responsible Generative AI".
I mean it writes some bad words, or bad pics, a human can do that without help as well.
The good thing about dangerous knowledge and generative AI is that you're never sure haha, you'd be a fool to ask GPT to make a bomb. I mean it would probably be safe, since it will make up half of the steps.
Bias is a real problem, but more than that - an adversarial press and public won't forgive massive brands like Google for making AIs that spit out racist answers.
Also people are using LLMs to learn (horrifying but reality), it would be unresponsible for them to let it to propagate negative stereotypes and biases.
Dang, that was really quick! According to the listed time of your reply vs. mine, less than an hour from the time I checked? Quick turnaround indeed.
Already been pulled from there over 3,700 times since then, too (as of the time of this reply mere hours later). Seems like quite a bit more'n a few Ollama users were "waitin' with bated breath" for that one to drop. :grin:
For posterity, an easy way to find the context length of a LLM hosted on Hugging Face is to look at the max_position_embeddings in the config.json, which shows the 8192 mentioned in another comment. (although in this case you need to sign the agreement first)
There are some exceptions, like Mistral 0.1 (which is technically 32K according to the config but practically 8K because the sliding window is awful) and InternLM (which (at least initially) used auto rope scaling to extend the context as part of the model's architecture).
That's useful to know. My experiments with the 4bit 7B currently tagged for use on ollama are not going well at all. Lots of refusals and junk. Downloading 7b-instruct-fp16 now! :-) (Update: Yes, much better, though much slower too, of course.)
Thanks, I'm going to give that a go too. The good news is the very latest Ollama update (0.1.27) has resolved things (though I also had to redownload the quant model for 7B).
Is there any research on using smaller, lower capability models to act comparable to high quality models? Even if it's just prompt engineering or doing lots of attempts to accomplish the task?
If somehow that is possible it means we only need a capable enough model and can use it reliably for lots of practical things.
> Open models feature free access to the model weights, but terms of use, redistribution, and variant ownership vary according to a model’s specific terms of use, which may not be based on an open-source license.
does a model being "open" say anything about how it was trained?
The scariest difference between OpenAI and Google right now is: Ask Gemini who owns the code it writes, and it'll confidently say that Google does. Ask OpenAI, and it'll say that you do. It's that easy to choose which one is the better decision.
Hopefully, they re-release this under an open license. Making everyone go through the exercise to authenticate and agree to terms hasn't worked for any model to-date. It just limits it's reach. We saw the same thing with Phi-2.
Are these any good? I have been trying the non pro version of Gemini, and that seems awful at code generation. I am more keen on getting access to the best model and I would pay for it if I wasn't already paying for ChatGPT 4.
I often talk with GPT4 on road trips about topics I'm interested in. Its great for passing the time.
I tried the same thing with Gemini and its full of nonsense. I was talking with it about the "Heian period" of Japan and it made up all sorts of stuff but you really only could tell because it was so ridiculous. Talked about European women and Native Americans roaming around the famous grassy plains of japan wielding katana and traditional weaponry... in the 1100s.
No such issue with GPT4.
I haven't tried it with code though, since I already have co-pilot. Really hard to trust anything it says after it started making stuff up about such a simple time period.
Nice, more choices are good. I just saw that the Ollama project already has these models available (date stamp is 58 minutes ago), so I will use that rather than Colab (I live Colab, but I like to run stuff locally).
"Carefully tested prompts" sounds a lot like "these are the lotto numbers we know are right" kind of thing? How in the world are these things used for anything programmatically deterministic?
I find it a bit disheartening that the new wave of “open source” with regards to AI is open weights. That’s like giving someone compiled and obfuscated binaries and saying that’s open source.
When you say this, do you mean the chat product or the underlying model available via the API? I think it's reasonable that the chat be censored to be acceptable to a wide range of people, but my understanding is that the "raw" model access for these sorts of things tends to be a little less restricted.
These are downloadable open models that can be fined tuned. They are the opposite of censored. If you have the motivation, you can bias them however you please.
Is “the opposite of censored” accurate for something that’s default and considerably easier to access mode of operation won’t say many things for sociopolitical reasons? Able to be un censored, sure, but the extent of that is debatable as well.
There is no default and easy access mode. These are raw model weights and only enthusiasts and researchers will download the necessary packages to run it locally. Much more likely is that some popular fine tunes show up on hugging face for more general access.
I agree that there probably will be “uncensored” fine tuned models that become available, my point was just that it’s not accurate to call Gemma “the opposite of censored” because there is a somewhat involved step that needs to be taken before it even appears uncensored.
It’s also likely missing a lot of useful context that was removed from the training set and not meaningfully replaced during fine-tuning, and besides that any fine tuned “uncensored” model will be based on Gemma, not Google’s Gemma itself.
IMO “Opposite of uncensored” suggests a model whose original form eagerly gives out controversial / typically censored information, not a model that is censored but able to be fine tuned away from censorship.
> Go to Google announcement
> Find “license” in page: no matches
> Go to HN thread
> Find “license” in page: 28 matches
> Read a few
sigh could have been exciting
Seeing as I published my Tokenizer video yesterday, I thought it could be fun to take a deepdive into the Gemma tokenizer.
First, the Gemma technical report [pdf]:
https://storage.googleapis.com/deepmind-media/gemma/gemma-re...
says: "We use a subset of the SentencePiece tokenizer (Kudo and Richardson, 2018) of Gemini for com- patibility. It splits digits, does not remove extra whitespace, and relies on byte-level encodings for unknown tokens, following the techniques used for both (Chowdhery et al., 2022) and (Gemini Team, 2023). The vocabulary size is 256k tokens."
Notes:
- vocab size is quite large: 32K -> 256K
- add_dummy_prefix is False. Different from Llama but consistent with GPT. This is a bit more consistent w.r.t. "leave the data alone", as there is no preprocessing step that adds a space to the encoding text.
- the model_prefix is the path of the training dataset, which is amusing to look at: "/cns/mf-d/home/gemini-data-access/tokenizers/final_v1_51GB_run1/bpe_coverage_0_999995_v5/255969". Seems to indicate the tokenizer training corpus was ~51GB (?).
- a lot of user_defined symbols (i.e. special tokens) are present, e.g. "hardcoding" a sequence of up to 31 newlines as tokens, and a large number of other unclear tokens. I tried decoding the octal representations but it's not clear what's happening here. Also a lot of more special tokens for what look like html elements, e.g. <table>, <tr>, <td>, <i>, <b>, etc. Not 100% sure what the unused tokens are for, maybe this is pre-allocated space to make easier future finetunes that try to add more special tokens, as there is no need to resize vocabularies and perform model surgeries (?).
TLDR this is basically the Llama 2 tokenizer, except bigger (32K -> 256K), with a lot more special tokens, and the only functional departure is that add_dummy_prefix is turned off to False. So e.g. tokenizing:
The utter bullshit of these licenses has got to stop. Do not, under any circumstances, consider using these commercially.
"Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement."
This is a kill switch that Google maintains in perpetuity over any system you build relying on these models. Our legal review of the Llama license came to the same conclusion, we cannot rely on the goodwill of Meta for any core service, and we shouldn't rely on the same from Google.
Now, perhaps less materially important, but just as infuriating is the "Prohibited Use[s]". These cover just enough to placate the most sensitive, but omit any real harms (waging war, developing weapons) that coincidentally have massive commercial value. Use the model to build a biological weapon (as an authorized govt official)? Cool. Use it to play a prank that deceives someone? Policy violation.
And of course, as the coup de grâce, they throw in a DMCA style provision to make sure you can't modify the models in any way that could cause them to violate their kid-glove precepts.
Wait, you actually care about the license and read it?
It's seems like you aren't up to date.
Most of the startup space is entirely ignoring all these licenses. If the weights are available, it is being used commerically without regards to any licensing.
And everyone is getting away with it and nobody is being sued.
Good luck trying to keep up if you aren't doing the same!
Feels free to hamstring yourself though if you like.
Great! Google is now participating in the AI race to zero with Meta, as predicted that $0 free AI models would eventually catch up against cloud-based ones.
You would not want to be in the middle of this as there is no moat around this at all. Not even OpenAI.
If meta keeps spending tens of millions of dollars each year to release free AI models it might seem like there is no moat, but under normal circumstances wouldn't the cost to develop a free model be considered a moat?
> If meta keeps spending tens of millions of dollars each year to release free AI models it might seem like there is no moat,
As well as the point being that Meta (and Google) is removing the 'moat' from OpenAI and other cloud-only based models.
> but under normal circumstances wouldn't the cost to develop a free model be considered a moat?
Yes. Those that can afford to spend tens of millions of dollars to train free models can do so and have a moat to reduce the moats of cloud-based models.
