I have 20 years of software development experience, and I couldn’t understand anything you said. Is there a dictionary for this new lingo, or am I just too mid?
He speaks very unclearly, instead of saying GPT-4-turbo he says 4.5 preview. 4.5 is invention of his.
Also mixtral medium - no idea of what he means by that.
Not to mention a claim that mixtral is as good as gpt-4. It’s on the quality of gpt3.5 at best, which is still amazing for an open source model, but a year behind openai
Mistral-medium is a model that mistral serves only via API since it's a prototype model. It hasn't been released yet and it's bigger than the mixtral-8x7b model
Sorry, but there's little that's unclear about what he said.
"mixtral medium" is just a typo: he means mistral-medium.
And GPT 4.5 is certainly not an "invention of his". Whether it exists or not (which is debatable, OpenAI said it was just mentioned in a GPT 4 hallutination and caught on), it' s a version name thrown around for like a month in forums, blog posts, news articles and such.
I just spoke all night to 8x7B and can say that it sucks much less than 3.5. It doesn’t screw up and apologize all the time (and screw up again) and doesn’t repeat what I just said verbatim. That is on topics I have a decent expertise in myself. Never had this experience of periodically forgetting that it’s not a human company with 3.5.
Local setup, “text generation webui”, TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF (Q4_K_M) on HF. You can run it on a decent intel cpu, takes around 32.5GB of ram including os (8gb for me). GPU with tensorcores can speed up few layers if you have one, but isn’t required. I get around 2.5-3 t/s with 8700 and 4070ti, that’s enough for chats that require some thinking.
Edit: I was using 2k window, a larger one would probably eat more ram. But even with 2k it didn’t feel like it loses context or something.
For macOS and Linux, Ollama is probably the easiest way to try Mixtral (and a large number of models) locally. LM Studio is also nice and available for Mac, Windows, and Linux.
As these models can be quite large and memory intensive, if you want to just give it a quick spin, huggingface.co/chat, chat.nbox.ai, and labs.pplx.ai all have Mixtral hosted atm.
I think with Mixtral Medium they mean MoE 2x13B which is on top on huggingface leaderboard? It is still not close to 8x175B, but size alone is not most important factor. With smarter training methods and data it is possible we will see performance similar to gpt-4 in open source mixture of experts of smaller sizes.
On reflection this thread is pretty clearly of general interest and my comment was more jargon that language, I hang out in ML zones too much.
For a broad introduction to the field Karpathy's YouTube series is about as good as it gets.
If you've got a pretty solid grasp of attention architectures and want a lively overview of stuff that's gone from secret to a huge deal recently I like this treatment as a light but pretty detailed podcast-type format: https://arize.com/blog/mistral-ai
Ironically, one of the best scenes from the series (Reynholm trying to wake his desktop by shouting "HELLO COMPUTER! HELLO!!" for hours and hours) feels very "ai"...
I'm trying to go deeper, I'd be curious to know what other ML zones you keep track of?
- I know r/LocalLlama, huggingface's Daily Papers and TheBloke. Most of what Youtube throws at me is horrific clickbait. I feel like there are probably whole communities I'm missing out on.
Now you know how your girlfriend feels when she hears you speak with other software people :) Excuse my assumptions if they are incorrect. I'm making projections from my own point of view.
Romeo here flexing about how he has an actual girlfriend. Sorry, Casanova, we’re busy refining mobile-ready inference models and leetcoding while you go on your so-called “dates”
A nerd finds a talking frog by a pond. The frog says, "Hey, I'm not really a frog. I'm a beautiful princess, and if you kiss me, I'll turn back into a princess, and I'll grant you one wish."
The nerd thinks for a moment, then puts the frog in his pocket and continues walking.
The frog says, "Aren't you going to kiss me and make a wish?"
The nerd replies, "Nah, I'd rather have a talking frog."
My apologies to both you and grandparent, I faux-pass'd on using such sloppy jargon in a thread of general interest. If I can be useful in decoding some of that and helping to keep this topic accessible to the broader community I'd be happy to answer any questions about what I meant. I've cited a few useful resources elsewhere in the thread and am always eager to talk shop on topics like this, so fire away if you have any questions about the field or what the hell I was talking about :)
A paragraph by paragraph "dumbed down" translation of your original words would be pretty neat to have for starters. Both to understand what you mean but also to understand the lingo.
I'm hardly the best person to give a point-by-point on how modern neural networks work. The original paper that kind of brought together a bunch of ideas that were floating around is called "Attention is All You Need" in 2017 (and those folks are going to win a Turing almost certainly) and built on a bunch of `seq2seq` and Neural Turing Machine stuff that was in the ether before that.
Karpathy has a a great YouTube series where he gets into the details from `numpy` on up, and George Hotz is live-coding the obliteration of PyTorch as the performance champion on the more implementation side as we speak.
Altman being kind of a dubious-seeming guy who pretty clearly doesn't regard the word "charity" the same way the dictionary does is more-or-less common knowledge, though not often mentioned by aspiring YC applicants for obvious reasons.
Mistral is a French AI company founded by former big hitters at e.g. DeepMind that brought the best of the best on 2023's public domain developments into one model in particular that shattered all expectations of both what was realistic with open-weights and what was possible without a Bond Villain posture. That model is "Mixtral", an 8-way mixture of experts model using a whole bag of tricks but key among them are:
- gated mixture of experts in attention models
- sliding window attention / context
- direct-preference optimization (probably the big one and probably the one OpenAI is struggling to keep up with, probably more institutionally than technically as probably a bunch of bigshots have a lot of skin in the InstructGPT/RLHF/PPO game)
It's common knowledge that GPT-4 and derivatives were mixture models but no one had done it blindingly well in an open way until recently.
SaaS companies doing "AI as a service" have a big wall in front of them called "60%+ of the TAM can't upload their data to random-ass cloud providers much less one run by a guy recently fired by his own board of directors", and for big chunks of finance (SOX, PCI, bunch of stuff), medical (HIPAA, others), defense (clearance, others), insurance, you get the idea: on-premise is the play for "AI stuff".
A scrappy group of hackers too numerous to enumerate but exemplified by `ggerganov` and collaborators, `TheBloke` and his backers, George Hotz and other TinyGrad contributors, and best exemplified in the "enough money to fuck with foundation models" sense by Mistral at the moment are pulling a Torvalds and making all of this free-as-in-I-can-download-and-run-it, and this gets very little airtime all things considered because roughly no one sees a low-effort path to monetizing it in the capital-E enterprise: that involves serious work and very low shady factors, which seems an awful lot like hard work to your bog-standard SaaS hustler and offers almost no mega data-mining opportunity to the somnobulent FAANG crowd. So it's kind of a fringe thing in spite of being clearly the future.
For a bot/spam thing, 16 years seems like a lot of planning and/or effort to put in? I don't know the typical age of people on HN these days but I can't imagine there are a lot of 2008 joins still floating around as a percentage of all commenters.
We're all too mid. Luckily, these days we hoomans have AIs to help us understand other hoomans. Here is Gpt-4-1106-preview and Perplexity.ai versions trying to shed some light what was being said. https://pastebin.com/JuxfdrLg
Hilariously neither knows who is sama (Sam Altman, the Drama King of OpenAI), nor do they recognize when they themselves are being discussed.
Reading the responses in full also gives you a glimpse on specific merits or weaknesses of these systems, namely how up to date is their knowledge and lingo, explaining capabilities, and ability to see through multiple layers of referencing. Also showcases whether the AIs are willing to venture guessing to piece together some possible interpretation for hoomans to think about.
I absolutely love pointing these things at each other and watching them go.
I screen-capped my take on this to prove* that I was actually wiring all this stuff up and plug my nascent passion/oss project, but it's really funny comparing them either way: https://imgur.com/WDrqxsz
Oh thank you, I was reading and none of that made any sense to me. I thought it could be a presentation of some dumb AI output. Now I see I’m not alone.
That gave me an idea, here is what I got from Copilot:
You have set up your system to run different AI models and compare their performance using a text editor. You are using Mixtral-8x7, a high-quality open-source model developed by Mistral AI, Dolphin, an emulator for Nintendo video games, 3.5-Turbo, a customized version of GPT-3.5, a powerful natural language model, and 4-Series Preview, a new version of the BMW sports coupe. You have noticed that the 4.5-Preview, an upcoming update of GPT-3.5, is slightly better than Mixtral-8x7, which used to be a close match. You are still waiting to access Mistral-Medium, a prototype model that is even better than Mixtral-8x7, but only available to a limited number of users.
You have discovered that Perplexity, an AI company that provides information discovery and sharing services, offers free access to Mistral-Medium through their partnership with Mistral AI. You think that Perplexity is making a mistake by giving away such a valuable model, and that they are underestimating the superiority of Mistral-Medium over the 4.5-Preview. You also think that Mistral AI is the new leader in the AI industry, and that their techniques, such as DPO (Data Processing Optimization), Alibi (a library for algorithmic accountability), sliding window (a method for analyzing time series data), and modern mixtures (a way of combining different models), are well-known and effective. You believe that the advantage of Mistral AI lies in the gap between their innovation and the ability of other developers to replicate it on cheaper and more accessible hardware. You also think that the enterprise market is not fond of the complex structure of GPT-3.5 and its variants, and that they prefer to use Mistral AI's models, which are more affordable and operable on their own premises.
You end your text with a quote from the movie Armageddon, which implies that you are leaving a situation that you dislike, but also admire.
I dont think it has anything to do with a BMW sports coupe.
What I am confused about though is it seems like the parent is mentioning models beyond the GPT4 instance I currently have access to. I checked their twitter and I have seen no anouncement for any 4.5 or 4 series previews. Is this just available to people using the API or did I miss something?
> I dont think it has anything to do with a BMW sports coupe
Well, the Paul Ricard circuit in France has a straight called Mistral. Plenty of BMWs have been there for sure, and a zillion other cars.
I wonder if that could have confused the AI a little in combination with other hints. Turbo?
If that's a thing maybe we should start picking our names not only to make them googlable but also not to confuse LLMs at least for the next few years. Months?
The OpenAI API is currently advertising their preview to my clients as `gpt-4-1106-preview`. I've been calling the Q4 2023 4-series stuff `4.5` or `4.5-preview` to distinguish it from the pre-big-announcement stuff.
I don't recall if I saw any press calling anything `4.5`, but it's a different model in some important ways (one suspects better/cheaper quantization at a minimum) and since they've used `.5` for point releases in the past it seemed the most consistent with their historical versioning.
It's a new theme and/or distribution that I'm working on under the working title `hyper-modern`. It clearly takes inspiration from things like Doom (and I'm using their modeline which is hard to improve on) but it's mostly ground up and AI-first.
It's heavily integrated with my custom model server and stuff and I'm slowly getting it integrated with other leading tools (vscode and nvim and stuff).
"Dolphin, an emulator for Nintendo video games", but in this context it refers to "Dolphin, an open-source and uncensored, and commercially licensed dataset and series of instruct-tuned language models based on Microsoft's Orca paper." https://erichartford.com/dolphin
respectfully, 20 yrs of software dev experience doesn't entitle you to understand the last 2 months of AI if you didn't spend the effort to keep up. jargon happens, its not your fault but also people need to communicate thoughts concisely given a base of knowledge. its ok to ask of course but the rest of us who have been keeping up can parse this well enough (even though I disagree with some of the assertions)
I'm snarkier than most on HN and have the scars to prove it, and I do miss the RTFM-by-default vibe of the early days, but on this one topic as you can see I'm going out of my way to apologize for being cryptic and try to include everyone in the conversation because this shit matters a lot even by our standards.
I think in this case it's much like the difference between understanding what Pokemon are, and actively playing each game so you know the names of most Pokemon.
