Just dropping the image..

[-]

maifee@reddit

It's 2025, and the heatmap remains almost the same And deepseek joined, that's one great news though

Reply

[-]

uhuge@reddit

HA web service is their new Open..

Reply

[-]

Ne_Nel@reddit

OpenAI being full closed. The irony.

Reply

[-]

At this point, OpenAI is being sustained by hype from the public who are 1-2 years behind the curve. Claude 3.5 is far superior to GPT-4o for serious work, and with their one-release-per-year strategy, OpenAI is bound to fall further behind. They're treating any details about GPT-4o (even broad ones like the hidden dimension) as if they were alien technology, too advanced to share with anyone, which is utterly ridiculous considering Llama 3.1 405B is just as good and you can just download and examine it. OpenAI were the first in this space, and they are living off the benefits of that from brand recognition and public image. But this can only last so long. Soon Meta will be pushing Llama to the masses, and at that point people will recognize that there is just nothing special to OpenAI.

Reply

[-]

andreasntr@reddit

As long as OpenAI has money to burn, and as lomg as the difference between them and competitors will not justify the increase in costs, they will be widely used for the ridicuolously low costs of their models imho

Reply

[-]

Minute_Attempt3063@reddit

When their investors realize that there are better self host able options, like 405B (yes you need something like AWS, would still be cheaper likely) they will stop pouring money into their dumb propaganda crap "The next big thing we are making will change the world!" Was gpt4 not supposed to do that? Agi is their wet dream as well

Reply

[-]

andreasntr@reddit

Yeah I don't like them either, unfortunately startups are kept alive by investors who believe almost everything they are told. Honestly, people are already moving away from Azure OpenAI since the service is way behind the OpenAI api and performance are very bad, and that's another missed source of revenues. I hope MSFT starts to be more demanding soon

Reply

[-]

JustSomeDudeStanding@reddit

What do you mean about the performance being very bad? I’m building some neat applications with the Azure OpenAI api and gpt4o has been working just as well as the OpenAi api. Seriously open to any insight, I have the api being called within excel, automating tasks. Tried locally running Phi3 but computers were simply too slow. Do you think using something like llama 304b being powered through some sort of compute service would better?

Reply

[-]

Sad_Rub2074@reddit

I contract with a large company that has agreements with Microsoft. Honestly, Azure openai with the same models has more hallucinations and doesn't perform as well as direct to openai. We won't leave azure since we have a large contract with them and infra, but we might end up contracting with openai directly for their apis. I am currently reviewing the other models though to see if it's worth creating an agreement with openai directly. We also have contracts with AWS and GCP. It's easy enough to switch, so after a bit of benchmarking and testing we will see. Might end up using azure openai for the easier tasks and switching to another model for the heavy lifting (405b). If that doesn't work out, then will go directly to openai for the more complex tasks. Azure ran out of the model we are looking for in ALL regions. Crazy.....

Reply

[-]

JustSomeDudeStanding@reddit

Very interesting, thanks for the response. Biggest driving force for me choosing Azure is the data security that comes with it. I’m kind of using it like agents, multiple calls to the api which act as context for other calls. Been working fine for that. I might look into using AWS so I can deploy a fine tuned model

Reply

[-]

Sad_Rub2074@reddit

Are you using Node.js?

Reply

[-]

andreasntr@reddit

Exactly what we are experiencing, thank for the thorough explaination

Reply

[-]

andreasntr@reddit

Azure is months behind in terms of functionality. Ust to cite some missing features: gpt-4o responses cannot be streamed when using image input, stream_options is not available (which is vital for controlling your queries cost token by token)

Reply

[-]

Minute_Attempt3063@reddit

Only reason why i use ChatGOT right now, is for spelling corrections for when i need to answer tickets of clients, and for format the words in a bet better way. Works good for that, at least.

Reply

[-]

Sad_Rub2074@reddit

405B on AWS is slightly more expensive than 4o. While I do use 4o for a few projects it's mostly garbage for more complex tasks. 405B is actually pretty good and for more complex tasks I normally use 1106. I'm benchmarking amd testing to see if it's worth moving some of my heavier projects over to 405B. There is talk that openai isn't doing too hot and definitely dipped with metas latest release. Microsoft is drooling right now.

Reply

[-]

Minute_Attempt3063@reddit

AWS might be a bit more expensive, sure, but you can self host Metas model, and you are not relying on some odd company. No one has to pay Zuck to use the model. You just pay for the hosting and that's it. And I think that is just better for everyone. Sure you might pay a bit more to hosting, at least you don't. Red to pay CloseeAi

Reply

[-]

Sad_Rub2074@reddit

Yes. I was just saying that it is not less expensive for most people. I agree with the main point of the post and most of the replies. OpenAI definitely fell out of favor for me as well. Azure OpenAI also doesn't perform as well with the same models -- more likely to not follow directions. 4o is terrible for more complex tasks. I still prefer 1106. At the enterprise I work for, though, it's worth paying for the models we need/use. Of course cost is still a factor. Definitely use the big 3 + openai. Had access to Anthropic directly, but didn't make sense. We already have large contracts with AWS, GCP, and Azure -- so receive steep discounts. Definitely a fan of open-source and use/support when I can. Just released a new NPM module for pricing. Only 11kb and easy to add other models.

Reply

[-]

Lissanro@reddit

Honestly I do not even care if "OpenAI" achieves AGI - if they do, it will be closed and cannot relied upon. In the past, when ChatGPT was just released, I was its active user at first. As time goes by, I noticed that things that used to work started failing, or working too differently, breaking existing workflows, and even basic features like editing AI responses were not available, making it even harder to get high quality output. So I just migrated to open models, and never looked back. Even though OpenAI tries to pretend closed models are "safer", they proven that the opposite is true, it is literally unsafe for me to rely on a closed model if it can break at any moment, or my access can get blocked for any reason.

Reply

[-]

-p-e-w-@reddit

All it takes is for interest rates to go up a little more, and investors will be demanding ROI from OpenAI, because otherwise they'll be better off just carrying their money to the bank. Collecting tens of billions of dollars on the vague promise that *someday*, investors might get *something* back is an artifact of the economy of the past few years, and absolutely not sustainable.

Reply

[-]

JoyousGamer@reddit

At which point OpenAI will be snapped up by someone. Its the backbone to a variety of AI tools out there in the enterprise space currently.

Reply

[-]

Physical_Manu@reddit

Can it easily be done so because of the unusual legal structure? Whoever is doing the merger or acquisition would have to be top of the field.

Reply

[-]

deadweightboss@reddit

sorry but as someone who does this kind of thing for a living, startups and rates are totally orthogonal. good startups have closest to zero beta out there

Reply

[-]

Camel_Sensitive@reddit

>sorry but as someone who does this kind of thing for a living Are you sure? >startups and rates are totally orthogonal. Yes, as long as you completely ignore late state valuations, investor sentiment, and borrowing costs. >good startups have closest to zero beta out there Literally zero startups have a beta of zero. many of them have negative beta, which is why otherwise good investors throw money at bad ideas. Any asset class that actually achieves zero beta is instantly restrained by capacity, which has never been the case in the start up world.

Reply

[-]

deadweightboss@reddit

also the “negative beta“ you’re talking about is much more akin to theta. how many years in are you?

Reply

[-]

Camel_Sensitive@reddit

>also the “negative beta“ you’re talking about is much more akin to theta. No, it's not. [A negative beta describes an investment that tends to increase in price when the general market price falls and vice versa.](https://sharegain.com/securities-lending-explained/beta/#:~:text=A%20negative%20beta%20describes%20an,lending%20rates%20will%20generally%20rise) In fact, negative beta and theta are not related in any sense at all. They actually apply to completely different financial instruments. Using theta to describe an ongoing concern isn't just silly, it's literally impossible. [Theta, the Greek letter θ, is used to name an options risk factor concerning how fast there is a decline in the value of an option over time.](https://www.investopedia.com/terms/t/theta.asp)

Reply

[-]

deadweightboss@reddit

ok you don’t work in the industry lmao.

Reply

[-]

deadweightboss@reddit

oh ok. sorry i must be ignoring the hundreds of billions of dollars in committed capital to privates which is restrained by capacity. there’s a reason why dry powder is dry powder. also, you’re not valuing startups with daily or monthly marks. Marks are quarterly at most. Nothing i’m saying is controversial. try explain why 08 vintage funds did so well.

Reply

[-]

psychicprogrammer@reddit

Given the current inflationary environment, expectations are for rates to decrease.

Reply

[-]

andreasntr@reddit

I'm not saying it's sustainable, just saying also users have very strict spending needs (i'm talking about companies) and can't ignore the price/performance tradoff

Reply

[-]

3-4pm@reddit

WSJ article late yesterday about low ROI for M$ AI.

Reply

[-]

Caffdy@reddit

Is Llama 405B really as good as ChatGPT 4o?

Reply

[-]

Physical_Manu@reddit

Not in terms of languages other than English, formatting, or trivial knowledge but other than that I would say they are fairly on par.

Reply

[-]

True-Surprise1222@reddit

4o is quite literally worse than 4 was on its day of launch.

Reply

[-]

nh_local@reddit

llama 3 is not fully multimodal. gpt4o yes. Currently there is no company that has presented a model with such capabilities, open or closed

Reply

[-]

Significant-Turnip41@reddit

I think we haven't really seen what the multimodal training will yield. You are right the competition has definitely caught up but I would bet money before the year is over we may see that gap widen again

Reply

[-]

kurtcop101@reddit

I am honestly shocked that they have not rushed something out to challenge 3.5. I am suspecting they're riding the wave and wanting to see Opus 3.5 first so they know how to market the next model. I suspect the last thing they want is to release something that upstages sonnet 3.5 only for Opus to sweep them out. If Opus releases first, they can target it better - if Opus is still better then they will come in and run it much cheaper or fluff about the tools you can use.

Reply

[-]

JoyousGamer@reddit

Well except for multiple large enterprise providers use OpenAI as a the default for their tools. As an example Co-Pilot is built on OpenAI and that is one of a wide variety that are using it. So no OpenAI is not being sustained by hype from the public.

Reply

[-]

West-Code4642@reddit

at this point, Anthropic is OpenAI 2.0, except that their CEO is a researcher and not a showboat.

Reply

[-]

AmericanNewt8@reddit

Anthropic is honest about what they're doing, at least. I don't have any problems with there being commercial software in the business per se, OpenAI just... god, they're so *annoying*.

Reply

[-]

West-Code4642@reddit

you're right. I mean OpenAI 2.0 from the sense of being an improved version of OpenAI. they've also kind of led the charge in interpretability research, which caused others (google, oai) to follow

Reply

[-]

Admirable-Star7088@reddit

>They're treating any details about GPT-4o (even broad ones like the hidden dimension) as if they were alien technology, too advanced to share with anyone, which is utterly ridiculous considering Llama 3.1 405B is just as good and you can just download and examine it. At the end of the day, it's all about gaining an edge and making bank for OpenAI. But saying that outright might not go down too well, so they opt for arguments like the ones you've heard. They gotta make ends meet somehow, especially since ChatGPT is their only cash cow (as far as I know), unlike tech giants like Microsoft, Google, or Meta. The one thing that grinds my gears is their choice of company name. It's very misleading.

Reply

[-]

unplannedmaintenance@reddit

Does Llama have JSON mode and function calling?

Reply

[-]

Thomas-Lore@reddit

Definitely has function calling, not sure about json.

Reply

[-]

fivecanal@reddit

How? I only use prompts to control it, but the jsons I get are always invalid one way or another. I don't think most other models have a generation parameter that can guarantee the output is valid JSON.

Reply

[-]

mr_birkenblatt@reddit

Besides constraint generation like others have said you can also just use prompts to generate json. You have to provide a few examples of how the output should look like though and you should specify that in the system prompt

Reply

[-]

Nabushika@reddit

Its not a product of the model, it's literally just the sampler, enforcing that the model can only output tokens that fit to the "grammar" of json. Any model can be forced to output tokens like this.

Reply

[-]

DooDiDaDiDoo@reddit

Constrained generation means anyone with a self hosted model could make JSON mode or any other format with a bit of coding effort for a while now. Llama.cpp has grammar support and compilers for JSON schemas, which is a far superior feature to plain JSON mode.

Reply

[-]

unwitty@reddit

I don't know but it doesn't matter when you can just use [guidance](https://github.com/guidance-ai/guidance), [LMQL](https://lmql.ai/), or manual token filtering to achieve the same thing without any of the constraints from black box API endpoints.

Reply

[-]

nagarz@reddit

Pretty much the tesla's of LLMs, they became big, got big stacks of cash, and have kinda become a laughingstock.

Reply

[-]

CeFurkan@reddit

100% Claude is way way better. Only problem is , it is more censored. Like don't answer medical question like gpt4

Reply

[-]

BearRootCrusher@reddit

But what about whisper?

Reply

[-]

UnionCounty22@reddit

Open a eye

Reply

[-]

firest3rm6@reddit

well, as daddy elon once tweeted

Reply

[-]

Danmoreng@reddit

Best quote from Zuckerberg Bloomberg interview. https://youtu.be/YuIc4mq7zMU?t=14m58s

Reply

[-]

Drited@reddit

Wait if OpenAI is not open....then maybe it's not AI either!!! Maybe it's just Storybots behind the scenes. They need a new name: ClosedBots.

Reply

[-]

Hambeggar@reddit

InaccessibleAI RestrictedAI LimitedAI ExclusiveAI UnavailableAI ProhibitedAI BarredAI BlockedAI SealedAI LockedAI GuardedAI ControlledAI SelectiveAI PrivatizedAI SequesteredAI

Reply

[-]

Lissanro@reddit

You forgot ClosedAI.

Reply

[-]

According-Channel540@reddit

Hit the mark

Reply

[-]

Sad_Rub2074@reddit

I own one of these xD

Reply

[-]

the_mighty_skeetadon@reddit

I feel like you used Gemma 2 to create this list

Reply

[-]

Downtown-Case-1755@reddit

Feels more like a Mistral response

Reply

[-]

Spirited-Pin-7378@reddit

ClosedAI

Reply

[-]

eljokun@reddit

ironic innit

Reply

[-]

OrganicMesh@reddit

Just want to add: - Whisper V3 was released in **November 2023**, on the OpenAI Dev Day.

Reply

[-]

ijxy@reddit

Oh cool. It is open sourced? Where can I get the source code to train it?

Reply

[-]

Aureliony@reddit

You can't. Only the weights are open sourced, not the training code.

Reply

[-]

Amgadoz@reddit

You actually can. HF has code to train whisper. Check it out

Reply

[-]

Aureliony@reddit

That's code written by third parties, but no official code from OpenAI

Reply

[-]

Amgadoz@reddit

You don't need official code. It is a pytorch model that can be fine-tuned using pure pytorch or HF Transformers. LLM providers don't release training code for each model. It isn't needed.

Reply

[-]

Aureliony@reddit

That's a really bad excuse to not release the code.

Reply

[-]

Amgadoz@reddit

I guess? But really this is the least irritating thing they have done so far.

Reply

[-]

ijxy@reddit

Ah. So, it is closed source. Got it.

Reply

[-]

Aureliony@reddit

It wouldn't be too difficult to write your own training code as the model architecture is open: https://github.com/openai/whisper/blob/main/whisper/model.py. The difficult part is getting the training data.

Reply

[-]

lime_52@reddit

Fortunately, the model is open weights, which means that we can generate synthetic training data

Reply

[-]

ijxy@reddit

Ah, so like reverse engineering Microsoft Word.

Reply

[-]

pantalooniedoon@reddit

Whats different to Llama here? Theyre all open weights, no training source code nor training data.

Reply

[-]

ijxy@reddit

No difference.

Reply

[-]

a_beautiful_rhind@reddit

A lot of models are open weights only, so that's not the gotcha you think it is.

Reply

[-]

ijxy@reddit

Open weights != open source.

Reply

[-]

Hubi522@reddit

Whisper is really the only open model by OpenAI that's good

Reply

[-]

CeFurkan@reddit

True After that open ai is not open anymore They don't even support Triton on windows

Reply

[-]

7734128@reddit

[https://openai.com/index/clip/](https://openai.com/index/clip/) is great.

Reply

[-]

Sushrit_Lawliet@reddit

(C)open(n)AI

Reply

[-]

Hambeggar@reddit

CopennAI

Reply

[-]

PwanaZana@reddit

That's a city in Denmark

Reply

[-]

Sad_Rub2074@reddit

CopenhagenAI

Reply

[-]

dampflokfreund@reddit

Pretty cool seeing Google being so active. Gemma 2 really surprised me, its better than L3 in many ways, which I didn't think was possible considering Google's history of releases. I look forward to Gemma 3, possibly having native multimodality and much longer context.

Reply

[-]

Cool-Hornet4434@reddit

I've been hooked on Gemma 2 27B. I always start out a fresh chat with a model introducing myself and asking "what's your name?" to see if they baked in any kind of personality, and Gemma is brimming with personality. Gemma is relatively good at translation, follows instructions pretty well, and is even good on Silly Tavern Roleplay. The only disappointing thing is that it's only 8K context and the sliding context window is actually about 4K, so when I try to refer something back to the earliest part of a chat at the 8K limit, gemma tells me her memory is fuzzy or maybe she hallucinates it. Other than that though Gemma is my new favorite. I'd love to see a 70B (but with only one 24GB VRAM card I'd need a 2.25BPW version of a 70B)

Reply

[-]

DogeHasNoName@reddit

Sorry for a lame question: does Gemma 27B fit into 24GB of VRAM?

Reply

[-]

rerri@reddit

Yes, you can fit a high quality quant into 24GB VRAM card. For GGUF, Q5\_K\_M or Q5\_K\_L are safe bets if you have OS (Windows) taking up some VRAM. Q6 probably fits if nothing else takes up VRAM. [https://huggingface.co/bartowski/gemma-2-27b-it-GGUF](https://huggingface.co/bartowski/gemma-2-27b-it-GGUF) For exllama2, these are some are specifically sized for 24GB. I use the 5.8bpw to leave some VRAM for OS and other stuff. [https://huggingface.co/mo137/gemma-2-27b-it-exl2](https://huggingface.co/mo137/gemma-2-27b-it-exl2)

Reply

[-]

perk11@reddit

I have a dedicated 24GB GPU with nothing else running, and Q6 does not in fact fit, at least not with llama.cpp

Reply

[-]

Brahvim@reddit

*Sorry, if this feels like the wrong place to ask, but:* How do you even run these newer models though? :/ I use textgen-web-ui now. LM Studio before that. Both couldn't load up Gemma 2 even after updates. I cloned llama.cpp and tried it too - it didn't work either (as I expected, TBH). Ollama can use GGUF models but seems to not use RAM - it always attempts to load models entirely into VRAM. *This is likely because I didn't spot options to decrease the number of layers loaded into VRAM / VRAM used, in Ollama's documentation*. I have failed to run CodeGeEx, Nemo, Gemma 2, and Moondream 2, so far. How do I run the newer models? Some specific program I missed? Some other branch of llama.cpp? Build settings? What do I do?

Reply

[-]

perk11@reddit

I haven't tried much software, I just use llama.cpp since it was one of the first ones I tried, and it works. It can run Gemma fine now, but I had to wait a couple weeks until they they added support and got rid of all the glitches. If you tried llama.cpp right after Gemma came out, try again with the latest code now. You can decrease number of layers in VRAM in llama.cpp by using -ngl parameter, but the speed drops quickly with that one.

Reply

[-]

Gab1159@reddit

Yeah, I have it running on a 2080 ti at 12GB and the rest offloaded to RAM. Does about 2-3 tps which isn't lightning speed but usable. I think I have the the q5 version of it iirc, can't say for sure as I'm away on vacation and don't have my desktop on hand but it's super usable and my go-to model (even with the quantization)

Reply

[-]

martinerous@reddit

I'm running bartowski\_\_gemma-2-27b-it-GGUF\_\_gemma-2-27b-it-Q5\_K\_M with 16GB VRAM and 64GB RAM. It's slow but bearable, about 2 t/s. The only thing I don't like about it thus far is that it can be a bit stubborn when it comes to formatting the output - I had to enforce a custom grammar rule to stop it from adding double newlines between paragraphs. When using it for roleplay, I liked how Gemma 27B could come up with reasonable ideas, not as crazy plot twists as Llama3, and not as dry as Mistral models at \~20GB-ish size. For example, when following my instruction to invite me to the character's home, Gemma2 invented some reasonable filler events in between, such as greeting the character's assistant, leading me to the car, and turning the mirror so the char can see me better. While driving, it began a lively conversation about different scenario-related topics. At one point I became worried that Gemma2 had forgotten where we were, but no - it suddenly announced we had reached its home and helped me out of the car. Quite a few other 20GB-ish LLM quants I have tested would get carried away and forget that we were driving to their home.

Reply

[-]

Cool-Hornet4434@reddit

I can use a 6BPW to get it to fit. 8BPW is too big, and I could go lower but 6BPW fits with 4Bit Cache applied and even rope scaled up to 24K context... BUT since Gemma's sliding context window (for attention I guess?) Is only 4K, there's not a whole lot of extra benefit.

Reply

[-]

Nabushika@reddit

Should be fine with a ~4-5 bit quant - look at the model download sizes, that's gives you a good idea of how much space they use (plus a little extra for kv and context)

Reply

[-]

Admirable-Star7088@reddit

Me too, Gemma 2 27b is the best general local model I've ever used so far in the 7b-30b range (I can't compare 70b models since they are too large for my hardware). It's easily my favorite model of all time right now. A happy surprise from Google, since Gemma 1 was [total shit](https://www.youtube.com/watch?v=1Mn0U6HGLeg).

Reply

[-]

SidneyFong@reddit

I second this. I have a Mac Studio with 96GB (v)RAM, I could run quantized Llama3-70B and even Mistral Large if I wanted (slooow\~), but I've settled with Gemma2 27B since it vibed well with me. (and it's faster and I don't need to worry about OOM) It seems to refuse requests much less frequently also. Highly recommended if you haven't tried it before.

Reply

[-]

Wooden-Potential2226@reddit

Same here - Gemma-2-27b-it-q6 is the best model you can put on 2xp100 currently.

Reply

[-]

pc_g33k@reddit

> Pretty cool seeing Google being so active. Shh... Google is known for aggressively shutting down projects.

Reply

[-]

crusainte@reddit

They get you hooked in hopes that you would use the GCP ecosystem.

Reply

[-]

Open_Channel_8626@reddit

Gemma 2 beating llama 3 is something I really did not see coming

Reply

[-]

EstarriolOfTheEast@reddit

Google has always been active in openly releasing a steady fraction of their Transformer based language modeling work. From the start, they released BERT and unlike OpenAI with GPT, never stopped there. Before llama, before the debacle that was Gemma < 2, their T5s, FlanT5s and UL2 were best or top of class for open weight LLMs.

Reply

[-]

forwardthriller@reddit

I stopped using them , gpt4o is utterly unusable for me , it rewrites the entire script every time. I don't like its formatting. I always need gpt4 to correct it

Reply

[-]

Inevitable-Crow-1675@reddit

Open ai is cooking something

Reply

[-]

Hearcharted@reddit

Llama 3.1 405B is The Boogeymodel that kills The Boogeymodel 😳

Reply

[-]

nh_local@reddit

I don't know if they asked - but what about Microsoft?

Reply

[-]

8braham-linksys@reddit

I despise Facebook and Instagram but goddamn between cool and affordable VR/XR with the Quest line and open source AI with the llama line, I've become a pretty big fan of Meta. Never would have thought I'd say a single nice thing about them a few years ago

Reply

[-]

Downtown-Case-1755@reddit

He hero we need, but don't deserve. All their stuff is funded by Facebook though, so......

Reply

[-]

XhoniShollaj@reddit

Meanwhile Mistral is playing tetris with their releases

Reply

[-]

empirical-sadboy@reddit

I mean, they are considerably smaller orgs. Some of what's depicted here is just due to Google and Meta being so much larger than Mistral

Reply

[-]

Combinatorilliance@reddit

<3 tetris

Reply

[-]

SavaLione@reddit

Does Meta have open source models? Llama 3.1 doesn't look like an open source model.

Reply

[-]

the_mighty_skeetadon@reddit

They say open source, but it's more correctly an "open model" or "open weights model" -- because the training set and pretraining recipes are not open sourced at all.

Reply

[-]

SavaLione@reddit

They say so but it doesn't mean that the model is open source The issues with the Llama 3.1 I see right now: 1. There are a lot of complaints on huggingface that access wasn't provided 2. You can't use the model for commercial purposes

Reply

[-]

the_mighty_skeetadon@reddit

This is not correct -- you can use Llama 3.1 for commercial purposes. It's not as permissive as Gemma, but it is free for commercial use.

Reply

[-]

SavaLione@reddit

Ok, now I get it, thanks It's free for commercial use if you don't exceed 700kk monthly active users

Reply

[-]

the_mighty_skeetadon@reddit

It's even more complicated -- it's tied to a specific date: > 2\. Additional Commercial Terms. If, **on the Llama 2 version release date**, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion, and you are not authorized to exercise any of the rights under this Agreement unless or until Meta otherwise expressly grants you such rights. So specifically targeted at existing large companies. Tricky tricky

Reply

[-]

LinkSea8324@reddit

To be fair, OpenAI gave us Whisper.

Reply

[-]

levraimonamibob@reddit

just the most open AI company ever, they're open-absolutists i tell ya

Reply

[-]

Leading_Bandicoot358@reddit

This is great, but calling llama 'open source' is misleading "Open weights" is more fitting

Reply

[-]

Raywuo@reddit

But code is also available to run these weights! The only part that is not available are terabytes of texts used for training, (which can and have been replicated by several others), obviously to avoid copyright issues.

Reply

[-]

danielcar@reddit

In the spirit of open source, one needs to be able to build the target. Open weights is great.

Reply

[-]

Leading_Bandicoot358@reddit

The code that creates the weights is not available

Reply

[-]

Raywuo@reddit

From what I know, yes it is! Not just one version but several of them. It is "easy" (for a python programmer) to replicate LLama. There is no secret, at most, there are little performance tricks?

Reply

[-]

Leading_Bandicoot358@reddit

You are mistaken on this matter

Reply

[-]

No_Comparison1589@reddit

We got this all wrong. Open AI is open for making money with AI.

Reply

[-]

ab2377@reddit

edit the image and change the 4th one to ClosedAI ty.

Reply

[-]

shroddy@reddit

I wonder what Cohere is cooking these days...

Reply

[-]

choronz333@reddit

Rebrand to ClosedAI? Nothing "Open" about OpenAI at all...

Reply

[-]

trakusmk@reddit

Oh the philosophical burden of contradictions in this world

Reply

[-]

Floating_Freely@reddit

Who could've guessed a few years ago that we'll be rooting for Meta and Google ?

Reply

[-]

PrinceOfLeon@reddit

If this image showed models released under an actual Open Source license, only Mistral AI would have any dots, and they'd have fewer. If this image showed models which actually included their Source, they'd all look like OpenAI.

Reply

[-]

BoJackHorseMan53@reddit

No one has released their training data. They're all closed in that regard

Reply

[-]

PrinceOfLeon@reddit

That's acceptable. Few folks would have the compute to "recompile the kernel" or submit meaningful contributions the way that can happen with Open Source software. But a LLM model without Source (especially when released under an non-Open, encumbered license) shouldn't be called *Open Source* because that means something different, and the distinction matters. Call them Open Weights, call them Local, call them whatever makes sense. But call them out when they're trying to call themselves what they definitely are not.

Reply

[-]

BoJackHorseMan53@reddit

Well, llama 3.1 has their source code on GitHub. What else do you want? They just don't allow big companies with more than 700M users to use their llms

Reply

[-]

the_mighty_skeetadon@reddit

They don't have training datasets or full method explanation. You could not create Llama 3.1 from scratch on your own hardware. It is not Open Source; it is an Open model.

Reply

[-]

Blackclaws@reddit

Should change August 2025 when the AI Act of the EU forces you to either do that or pull your LLM from the EU.

Reply

[-]

BoJackHorseMan53@reddit

Pulling open source llm from EU doesn't mean anything. People can always torrent models.

Reply

[-]

Blackclaws@reddit

Any LLM that wants to operate in the EU will have to do this. Unless Meta/Google/OpenAI/etc. want to all pull out of the EU and not do services there anymore they will have to comply.

Reply

[-]

NeedsMoreMinerals@reddit

We should start putting the Open of OpenAI in quotes. "Open"AI

Reply

[-]

Crazyscientist1024@reddit

Here's what I don't get about OpenAI, just open source some old stuff to get your reputation back. If I was Sam and I wanted people to stop joking about "ClosedAI" just open source: DALLE-2, GPT-3.5 (Replaced by 4o Mini), GPT-3, maybe even the earliest GPT-4 checkpoint as LLaMa 405B just beats it. They're probably not even making money from all these models anymore. So just open-source it, get your rep back and probably more people would start liking this lab.

Reply

[-]

omercelebi00@reddit

The higher you are, the more spectacular your fall.

Reply

[-]

3-4pm@reddit

I blame the pandemic.

Reply

[-]

protector111@reddit

They Should make them Change the Title to closeAi

Reply

[-]

nscavalier@reddit

ClosedAI

Reply

[-]

artsybashev@reddit

Open is the new Close. Resembles all those "Democratic People's Republic of ..." countries.

Reply

[-]

mrdevlar@reddit

Such places are also run by a cabal of people who suffer from self-intoxication.

Reply

[-]

Icy_Foundation3534@reddit

Open

Reply

[-]

Everlier@reddit

What if we normalise the charts accounting for team size and available resources? To me, what Mistral is pulling off is nothing short of a miracle - being on par with such advanced and mature teams from Google and Meta

Reply

[-]

AnomalyNexus@reddit

>What if we normalise the charts accounting for team size and available resources? I'd much rather normalize for nature of edits. Like if you need to fix your stop tokens multiple times and change the font on the model card that doesn't really count the same as dropping a new model.

Reply

[-]

kmouratidis@reddit

I'd also like to normalize for different model versions. E.g. taking a look at some of Google's latest releases: - gemma-scope-2b-pt-res - gemma-scope-2b-pt-mlp - gemma-scope-2b-pt-att - gemma-2-2b-gguf - gemma-2-2b-it-gguf - gemma-2-2b - gemma-2-2b-it - gemma-2-2b-pytorch - gemma-2-2b-it-pytorch - gemma-2b-awq These shouldn't count as 10 different models...

Reply

[-]