RouteLLM achieves 95% GPT4o Quality AND 85% CHEAPER
HTML-код
- Опубликовано: 7 июл 2024
- RouteLLM shows some promising potential for cutting cost without sacrificing quality of LLM outputs.
My Links 🔗
➡️ Subscribe: / @wesroth
➡️ Twitter: x.com/WesRothMoney
➡️ AI Newsletter: natural20.beehiiv.com/subscribe
#ai #openai #llm
LINKS:
lmsys.org/blog/2024-07-01-rou...
arxiv.org/pdf/2406.18665
github.com/lm-sys/RouteLLM
I'm favoured, $27K every week! I can now give back to the locals in my community and also support God's work and the church. God bless America.
You're correct!! I make a lot of money without relying on the government.
Investing in stocks and digital currencies is beneficial at this moment.
Yes! I'm celebrating £32K stock portfolio today...
Started this journey with £3K.... I've invested no time and also with the right terms, now I have time for my family an…
when someone is straight forward and good at what she does best. People will always speak for them. For me I can would say give Mrs Sonia Duke of finance education a try and you be happy you did
Started with 5,000$ and Withdrew profits
89,000$
I'm glad to write her tay I do hope she will help handle my paycheck properly☺️☺️☺️
"THIS CHANGES EVERYTHING"
-Nothing changes
DRAMATIC bold letter bright RUclips captioning for the sheeple!
That is SHOCKING me to my core
problem is that things take a long time to get adopted and by the time they do get adopted a new AI thing has already come out that's way better and the cycle just repeats forever
Ripped off of Matthew Berman's title
@@kevinnivek8907 They just letters bro..
I would like to see a tutorial on how to use this, I think the idea of a separate channel might be a good idea.
seconded
The market is differentiating. It is cool to see it play out. First movers and then budget.
This is what I've been saying. Specialized clusters working together iterating until they're happy with the output should solve most AI problems. Just give them enough differences that it's a brainstorming session, and enough time to work amongst themselves.
Sounds like how we forecast the weather we call them ensembles.! We run the model 10-20 times then get a mean from all the runs.
Is it possible to make a custom gpt store chatbot into API?
@@TheDisillusionistjust use the regular OpenAI API and store the GPT prompt as a string and pass it in before your prompt...?
@GerrodNoordsylovesu thanks for the quip, might help! Dm and let's cook!
This is where ai and automation is a step closer to baby agi , and I don’t think we CAN understand what they are doing because now you have many layers and agents with their own hyper parameters but I think this is the beginning of a new type of math that the ai can help us develop
I'm into having the geeky content, whatever the context. It's helpful to see someone else who's competent dig into this stuff. It makes it less intimidating. Bring it on, Wes!
Artificial Intelligence: 1
David Shapiro Intelligence: 0
@@BrianPellerin lmao
Idk what's wrong with David, he keeps blocking my channels when I comment on his vids and polls. Somethings not right with that guy
@@TurokAgi he's a really bright, sensitive guy.
@@anta-zj3bw :/
I commented on his substack and he told me "I think this is way too conspiracy theory nonsense for me to take seriously"
Love to see some technical deep dives from you, Wes - however you do it! Many thanks, I really appreciate your videos!
I’ve been learning from you for a while, keep up the good work sir!
totally interested in the second more project-based channel - thanks man
Step by step troubleshooting of open source software on GitHub sounds good to me! I'm your kind of nerd.
Actually, I think fixing errors in stuff is a great way to learn, and I wish there was more content on RUclips like that.
Totally appreciate your videos.
thx for your work man i love your channel
i d like to see you testing those model
Thx for explaining the ongoing orchestra of the LLMRouting
Thank you Ross for educating us and keeping us informed. Keep it up!
You copied my video, thumbnail, and title. 😂😂
off course you all going to "copy each other" because the content source getting fewer and fewer, because these sources are being absorbed directly by huge corporates they "eat" any new tech around this kind of development to gain their domination in capitalism system.
maybe
When the two AI news source title videos are made by the same AI
Exactly 😂 I was confused because I saw your video first in feed 😅
@@NeonvarunYou saw it 1st, because this video was published 6-7 hours later.
Definitely will watch and enjoy more, deeper content like this. 2nd channel or not.
This is merely a modification of the MOE (Mixture of Experts) architecture. Here, the LLM adopts a concept similar to the big.LITTLE architecture in mobile CPUs, but redesigned with cost considerations in mind.
18:15 Yes please! A separate channel specifically for these coding projects + troubleshooting 🙏🤩
I would love to see a tutorial on this. The tutorials are very helpful. 😊
Hi your videos are very informative and helpful. Thank you
So damn smart. Hats off to these researchers.
Please do the second channel, I also find it helpful to understand the layer of the big nerds vs the emerging nerds and separating content approach
Great idea re the more "dev/tech" channel where you implement stuff!
Yeah, I'd like to see a tutorial on running RouteLLM on your PC. 😲😁
GPT 0 is a way infirior product...claude is something to look forward to.
This is amazing 🎉
Do another channel where you go though the open source ins and outs. It will be very valuable if you do it in your style and focus on clarity. Maybe you could even sell some of your work product from that channel, maybe a patreon? I for one would support the channel.
Good idea on the separate channel for the techy bits. Where do I subscribe? This RouteLLM is setting up a free enterprise market where models will compete against each other for the privilege of providing all or part of the answer. Harnessing that ability within a usable framework by users will be fascinating.
This is just perfect.
An inevitable development in this time of seeking efficiency. Similar to CPU's with their Performance and Efficiency cores. A project/prompt manager if you will. This is a significant step towards an open market scenario.
Your green screen trash removal is above average Wes!
I use pipelines in my LLM gui to do the same - scripts analyse question, analyse available agents and tools, and select optimal combination of them. Only thing i want to add in future is autoselect of system PROMPTS from some sort of library
I would love to see not really a tutorial but a testing with "real scenarios"
Great bro
Channel with focus on agents and LLM vode stuff. I'm in
Start another channel and do projects on it! I would love to see that!
A new channel would CHANGE EVERYTHING! Jokes aside, count me interested!
tutorual and "how-tos" are always welcome :D
Cool, you can route between Casper-Magi 3, Balthasar-Magi 2 and Melchior-Magi 1
Step by step configs would be great.
Wes, if you have the capacity to do tutorials that would be excellent, yes please. :-)
Yes please let’s check that code together and run it, do the code part at the end is a good idea, but let us know in the beginning of the video “hey I will do a deep dive at end…’ great info tho
same channel - we are moving into this sort of stuff
Bring us another channel Broseph
❤
David Shapiro: ai race is slowing
Literally today:
Did it SHOCK THE ENTIRE INDUSTRY?
literally what i was thinking xD
I am literally tired of most people not knowing when the word "literally" is not necessary.
@@anta-zj3bwYou're literally right, I agree.
yeah, I respect Dave and his thought processes a lot but it kind of felt like he made that conclusion based on far too few data points.
This project is actually in a position to be a future “stock exchange “ for pricing of use/ price of intelligence for llm’s. Think how pricing will vary over time based on energy cost, bandwidth, responsetime based on time of day, bulk purchases of a subset of models, purchase of a few given models for a given time period, bulk purchase of queries etc. more than enough variables to make a market volatile, and within this you can trade on a pure ftx betting up/down, packaging in “funds”, futures,shorting of tokens on different models..
interesting point
Yes for 2nd more technical channel
neat
Reminds me of a manager who delegates the work to balance quality and cost
Wes, are you and Matthew Berman working a secret video title racket?
I like your title lol
Why is Claude sonnet 3.5 not shown on the graph?
Conspiracy bro conspiracy 😅😂
I hated that extra 5% accuracy. If you know of some free AI with another hit of say, 5-10% in accuracy that would be worth introducing into my sub-standard existence. Do inform us.
talking about deceleration...
trust me it isn't decelerate in huge companies such as tesla, Microsoft, Google they use 2022-2024 event to create new sources they can absorb... they all still actively developing them, the latest Grokking also game changing... that's whats happen in surface. public doesn't need to know whenever they can harvest gold from Mars, public only need to know when they need something from the public
I DO!!!
😮😮
It sort of like the old adage, two heads are better than one, with each LLM being a head.
More technical videos please, even if on a new channel
This is ground breaking
If it’s true and scalable
I wonder how the context window works when complexity of the questions grows, while you keep asking the questions.
What do you call a group/swarm/cluster/map/grid/schema/network of AIs, especially when they control their routing? A murder? Thanks.
It achieves better then 80/20
Do the Channel 🎉
Hey Wes are you on Pliny discord server?
Replacing every step, every decision along the way with expert systems puts us further from agi, not closer. Maybe it's the safer route
Please a separate channel for tutorials
This is like crewai hierarchical process, not that ground breaking it seems
So what your saying is GPT 5 will be on route soon?
My brother works in cyber security for NASA. They got to see a demo of GPT-5 at a recent conference. The model they saw was fine tuned on their security software and the demo was all done with natural language. The computer had full autonomous control of the computer and it was capable of controlling the software.
@@user-ty9ho4ct4k how long ago
95% as good is still 1 in 20 responses not so good,,, 5 sub-standand response could easily wipe-out an 85% saving,,,
This seems silly. Each agent can be assigned an llm in their JSON file. Say no to the black box. Lol. It’s open source, my bad. I just can’t trust an LLM to decide the best model to use. I would rather choose them myself.
What's the point of this juggling if there isn't a single one task that any of the current models can be trusted to execute reliably without oversight?
This is good when we look forward, the next gen, but for now this does absolutely nothing.
hmm how abt gpts chatting w/ each other and rating each others? 😎
18:00 - if you’re going to have a free walkthrough channel, what’s the benefit to keeping our only fans subscription active?
Make a separate channel. That would be great.
Conceptually this has been around for a while? Am I missing something here?
Oh hold on a second before we go any further
Route (pronounced root) as in way to go
Surely
Not Route (pronounced rowt) as in cut a hole
Wes please for the love and respect of internet hardware its a router (root) which direted this comment to your video
GPT4 / 4o is *not* the best model anymore, Claude 3.5 Sonnet is
Give us all the git
Worse than GPT-4 but cheaper. So like… GPT-3.5?
Don’t use gpt anymore, Claude and perplexity kills it, hands down. Microsoft says perplexity is a wrapper, maybe but it’s great and better
I would prefer same video at the end, rather than a separate channel.
Love your vids. Will always be a fan… but you straight stole this title from MB. You can do better.
I'll be honest. It sucks. They basically trained a model to do something which GPT 4 could do itself using function calling providing its given which each models good at its that simple. This isn't anything amazing if anything they just built another node much like our NN works inside humans for choice making which help us think when really they could have used that compute on something else worth while and just got GPT to be the router instead within about 20 lines of code.
Make a second channel
Word of caution, 😄 before starting a new channel make sure to sacrifice blood to the blood god; or the blood God will punish us and the channel might fail. The only thing is...where do we find a person...we need to find a person and...
Stop the This changes everything crap
Cheaper better faster. Wonderful. Is it too late to cease Ai?
Will everyone be… laid off by Ai? Suffering Ai jobloss for years? Swell robotics doing everything? Then everyone made slaves for an Ai new world order?
Is the Ai hype finally dying down? Nothing interesting seems to have come out recently
Claude 3.5 came out two weeks ago lol
Th geekier code level stuff the better...
Separate code based channel please!🥳😎🦾
Step by step configs would be great.