GPT-4o mini Prompt Chain: Legit TRICK for DIRT CHEAP AI with SOTA Accuracy

Video ID: 0Z2BQPuUY50

YouTube URL: https://www.youtube.com/watch?v=0Z2BQPuUY50

Added At: 13-06-25 21:18:49

Processed: No

Sentiment: Positive

Categories: Tech, Education

Tags: AI, Machine Learning, Natural Language Processing, Prompt Chains, Fusion Chains, GBD4 o Mini

Summary

• The price of intelligence is going to zero with the release of gbd4 o mini, a cost-effective high-performing model.
• This model is 30 times cheaper than GPT-40 and 20 times cheaper than Cloud 3.5 Sonic.
• The author demonstrates how to use prompt chains and fusion chains to take advantage of cheap high-performing models.

Transcript

as predicted the price of intelligence
is going to zero gbd 40 mini just came
out and is the most coste effective
model while maintaining a majority of
the performance of a state-of-the-art
model I think the numbers here are
actually really really staggering the
gb40 mini is insanely close to the
performance of GPT 40 in many of these
cases and many of these benchmarks with
a trick we've been looking at on the
channel you can maintain the exact
performance gains that you can get from
gpg 40 and even Claude 3.5 Sonic that's
what we're going to look at in this
video but we really do have to express
how crazy this is gb40 mini really is
insanely affordable intelligence 30
times cheaper than gbt 40 this is
roughly averaging their input and output
token costs and then it's roughly 20
times cheaper than 3.5 on it and again
the crazy thing here is the performance
gains don't drop off a cliff Gemini
flash is the only model that's come as
close but with every one of these other
models once you go down in size once you
make your model cheap you lose all of
the state-of-the-art performances this
is why gb40 mini is so incredible you'll
notice here you know 82 to 88 that's a
5% accuracy difference 13% we have a 3%
difference here 6% difference 3%
difference this is really really crazy
considering the costs right we're
talking about percentages often smaller
than 10% right the accuracy is at the
worst a 133% drop off and again with
this trick I want to show you in this
video you can maintain a lot of the
benefits of using one of these models
we're going to talk about that in just a
second here but it's really really wild
that by only a drop of 10% you have a
gain of 30,000. 30X and a 20x so you
know if you're using Cloud 3.5 Sonic
which is already priced at $3 per
million tokens it's only going to be a
20x but still uh only is a silly word to
use there because a 30 times Improvement
in price and a 20 times Improvement in
price are both absolutely incredible so
you know gbd4 o mini confirms what we've
been kind of betting on on the channel
the price of intelligence is going to
zero so what does that mean for us how
can we utilize this how can we play with
both the absolute state-of-the-art
models like claw 3.5 Sonet and also how
can we take advantage of the cheaper
second tier but still high performing
model like gbd4 mini everything starts
out with a prompt but you can take the
prompt and chain together the results
and by chaining your prompts together
you can accelerate the results by having
your models think step by step blow up
their context window solve problems one
step at a time you've seen this we've
discussed this on the channel we pushed
this a step further in our previous
video you can do this with multiple
chains running across different models
so you can have the state-ofthe-art
results running you take your inputs and
instead of just running a single prompt
chain to get that accelerated result you
run 2 3 4 five or more prompt chains and
then what you do at the end is you merge
the results with an evaluator function
this evaluator function is very
important because it forces you to say
this is what it means to get a good
result out of my llms out of my prompts
and out of my prompt chains so in the
end it looks like this right prompt one
feeds into prompt two two to three and
this can go all the way to n as many
prompts as you need to chain together
you can but then you take your results
and you merge them together with the
evaluator this is called the fusion
chain link is going to be in the
description we talked about this a lot
in the previous video this allows you to
get the state-of-the-art results out of
the best performing models of course at
the price of running all of these models
on your same prompts and on your same
prompt chains in the end this is going
to be more expensive the incredible
thing here is thanks to gp40 mini now
that we have a high performing model
that is extremely cheap what we can
actually do here is take our Fusion
chain and just run a gb24 mini fusion
chain the price of this chain
effectively has dropped to zero while we
maintain most of our results because
we're using prompt chains and we're
using the fusion chain here right so
we're going to get three takes of a step
by step task by task prompt we're going
to then evaluate the results using a
clean evaluator function and then we're
going to finally get that that last
output so you know running nine prompts
of these you know higher-end models can
now be done at a fraction of the cost
every single one of these prompts is
going to cost us 20 to 30 times less so
that means we can really really exploit
this idea of using cheap high-
performing models with prompt chains and
fusion chains so this is a big idea that
I've been playing with a lot and I've
been getting some really really high
class results so what does this look
like in action let's go ahead and look
at an application we've been building on
the Channel this is called zero noise
it's an application that allows me to
fetch information from blogs from change
logs from tools from webites only when
there are relevant changes when there's
new content this allows me to aggregate
information faster and only consume
information when it's time to consume
information I think one of the most
important things you can do in the age
of AI is make sure that your information
diet is as clean as possible there's
going to be so much content generated
there already is so much content you
feel the on a daily basis you know what
this is like I think it's important to
build and use tools that help you filter
out the noise keep yourself in a low
noise focused environment in previous
videos we built up the fetch workflow
where we go scrape all these blogs for
Relevant content for new relevant
content we built up the learn workflow
and in this video I want to show off the
Run recommendations the recommendation
workflow so we have a recommendations
prompt chain with multiple versions of
gbt 40 Mini running in every single
prompt chain and all it's going to do
here is based on the content we have
already set up in our fetch so you know
you can see in here we have the cursor
change log we have the AER blog we have
Simon W's blog shout out Simon W and
then we have a couple additional ones
here what this will do is take related
keywords and go ahead and perform a SEO
type search I'm using ex AI you can use
really whatever you want scrape some
related results based on some related
keywords so thanks to gp24 mini this
agent workflow is nearly free when
before it was you know 20 30 times more
expensive so you can see here I'm
getting some great and some not so great
results but that's kind of the whole
point of this right I have a nice
variety of recommendations that I can
now look through that my agentic
workflow has surfaced for me
automatically based on these related
tags now I can hit thumbs up and thumbs
down and it's going to save a brand new
recommendations feedback data type which
will then change the flow of the
application in the future based on the
thumbs up or the thumbs down that I give
so this is really cool this is what it
looks like let me go ahead and just show
you what the code looks like just a
little bit I want to focus on the prompt
chain and show you how cool this is that
we can take gpg 40 mini a cost effective
high accuracy model and generate some
still very state-of-the-art prompts and
prompt chain so we have this workflow I
like to use this agentic pattern here
where we retrieve we're on our agentic
we act learn and notify and all the
magic is happening right here in the
recommend workflow so let's go ahead and
walk through this a little bit so in our
agentic we're taking all the scraped
content from every one of our you know
currently
existing uh information sources that
we're interested in we're creating
content in markdown format we're then
Gathering our previous positive feedback
items and our negative feedback items
that's created by hitting the thumbs up
and the thumbs down here from previous
runs and then we're just you know
creating some context for these prompts
then we're actually running our promp
and you can see here we have an
evaluator and here is the important part
so we're using this Fusion chain which
allows us to run a series of prompts
over a series of models there's
something kind of interesting happening
I am just running a prompt chain with
four gbto mini models so I'm running the
fusion chain over these two prompts let
me just show you these two prompts real
quick so based on all of the scraped
content we're going to extract keywords
and then based on our positive and
negative feedback back we're going to
filter out those keywords that are not
relevant and make sure that we maintain
and explore the keywords that are
relevant to the positive feedback right
so the thumbs up and that means that you
know we have two prompts here instead of
three prompts we have two prompts and we
have one more chain and they're running
into the evaluator method and the
evaluator all that's doing here is we
can just go ahead and minimize a lot of
this the evaluator is basically just
taking all of the last results just
taking the results from the last last
layer of prompts right so this last
layer here and it's merging them all
into the evaluator function and all
we're doing here is we're going to get
all the keywords and all the items that
we do actually want to search for and
then it's going to return that response
and then a score for each one of our
final outputs the evaluator is really
really powerful it forces you to Define
what it means to get a great result out
of your prompt chain and out of your
llms basically what we do is given the
final layer of every single prompt chain
you merge them together using whatever
logic you like but the interesting part
here is that this Fusion chain is
powered by four gbg 40 Minis and it's
giving us this like great set of results
based on the script content from the
existing URLs right so what does this
look like in detail let's go ahead and
look at some of the log files we have
the fusion results here the top response
here is that fuse response from our
evaluators so you can see we have a list
of URLs here which we pulled we have an
entire explanation for every single
keyword and then here you can see all
the keywords we have right so these
keywords then create additional SEO
searches and then the SEO searches are
surfaced here and then I can just click
in and see what I'm interested in right
so we have some uh we have a fireworks
uh cursor blog here and looks like
there's a coll lab happening here and
yeah this just happened recently so very
interesting I actually did not know that
cursor was working with fireworks and
they achieved this really really crazy
1K tokens per second um so you know this
is really cool so if I was interested in
this type of content this type of
information and I wanted to see more on
cursor prediction I would just come in
here thumbs up and now what's happening
is based on my thumbs up here this is
going to save in my configuration. Json
file and on subsequent recommendations
it's going to see that I thumbs this up
and specifically look for more content
like this right so we can see exactly
what that looks like in the
configuration if we close our providers
you can see here we have some
recommendation feedback and I'll just
collapse all this so we can just see
that most recent one that I just added
but you can see that that's here right
so we have the URL coming in we have the
keywords and then we have positive
feedback negative feedback so this is a
positive one right since I gave this a
thumbs up and now this cursor prediction
SEO keyword is going to be prioritized
in subsequent search is so if I was now
interested in fireworks content their
blogs I could come in here take a good
look at their uh you know content and
then I could you know run the previous
workflow which we covered in another
video a different agentic workflow will
run and start to put together the
information that we need to know if
there are updates on this blog that I
haven't seen before with every agentic
workflow they start to feed back into
each other right so here's a quick
diagram that kind of shows that off so
in the beginning we have our config.js
on and you know that's this file f it
contains our recommendations and it has
our you know providers here so you can
see we have the cursor change log there
and and we have the HTML elements that
are used to you know find new content we
load that we scrape those websites we
run our gp40 mini fusion chain that
gives us our keywords we then run our
keyword search generate the
recommendations as you saw and then we
displayed them so you know that was you
know this UI here and then want to hit
thumbs up or thumbs down this is where
we get into this really intering
feedback loop where this actually
updated the config.js right with some
new recommendations which then fuels and
populates the subsequent workflows right
and you can see that in one of the
prompts here we're actually pulling in
you know all of our positive feedback
and our negative feedback from our
config and then this is passed in as
context into the prompt chain right so
this gets passed in down here into the
fusion chain and we're running the uh
parallel run here which will run all of
these prompt chains at the exact same
time and we can dig into the prompts a
little bit so the first prompt is
keyword extraction in the markdown
format giving it some rules to follow
placing the scrape content and asking
for the output format ending with a you
know leading sequence to lead the model
into the right results we're using the
five key elements of the prompt here
that's our keyword extraction prompt and
then our prioritization and filtering
prompt is doing the work of of you know
listing the positive items listing the
negative items and then we're just
asking you know based on the previous
prompt which will be the you know Json
result of the keywords that were
extracted from the scraped content we're
saying you know filter out we're saying
keep and for anything new just leave it
in there there are unexplored keywords
we just want all those to surface right
this is a simple prompt this could be a
lot more complex you might be thinking
you can just you know match on keywords
and just you know pull them out yes and
no really there are many keywords that
your prompts could extract that will be
similar and the whole idea is to give
more of these capabilities to your llms
to your prompts to your AI agents let
them do the work you just ask exactly
what you want to happen and it's going
to be a point where you're going to want
to feel this workflow by a prompt
instead of with code right the whole
idea is to be practicing building out
these identic workflows so that we are
coding less and prompting more so that's
that workflow we pass into positive
negative then we ask for that same
output over time this will get more
sophisticated and our prioritization and
filtering logic will improve but then
this is really cool you know instead of
passing in you know for the big top tier
models I'm just passing in four GPD 40
mini models and getting you know
basically the exact same results as
running the big three models right so
instead of running 40 Gemini 1.5 we're
just running om mini and saving a ton of
ton of money we're then saving this we
can go ahead and just finish looking at
the fusion result here so you can see
all the keywords from the top result
this is the part of the prompt chain
because llms these models they're not
deterministic right so every single gbt
40 mini call is going to be different
every one of these arrays here
represents a
gbg4 mini prompt chain right so we have
two prompts here where it returned the
keywords that we're looking for right in
the first run of gbd4 mini we got this
result here right for that first prompt
and you know so extracted some good
keywords we have ai API comparison you
know code lening gener of AI we have it
explain every SEO keyword and how it
relates to the content from the URL so
that the llm is you know thinking a
little bit more about the decisions it's
making and then it runs the second
prompt right so this is the filtering
prompt it looks like nothing got changed
between these two runs so you know that
meant that our feedback loop didn't
include or exclude any one of these
keywords right so that's the first
prompt chain and then we have another
prom chain here right and we got
completely different keywords right
cursor editing AER llm models open Ai
and if we just go and collapse this you
can see that we did filter out open AI
because I had the open AI in the
negative keyword if we go ahead hop over
to Json you can see I had open AI here
in the negative feedback on a previous
run I don't really need to know anything
about the open AI API so I gave this a
thumbs down and you know this prompt
chain of gp40 mini actually just
filtered that out right and this is
really nice because if I was writing
code I wouldn't have been able to filter
on this exact keyword right but since
I'm using llms and they have good
reasoning ability it's all open Ai and
it just filtered it out for me right
that's really awesome so you can see you
know also this prompt chain working you
know here's the first set of keywords in
the first prompt and then after the
filtering with my recommendations you
know got rid of some items which is
exactly what I wanted so this is just a
way that you can utilize both prompt
chains and gp40 mini to get some really
really incredible results I've run this
workflow and I've run you know the big
three plus the mini and I can tell you
that during four mini models into this
Fusion chain is effectively doing the
exact same job as the state-of-the-art
models and you know this is all a big
shout out to you know coming full circle
here all the work that open AI is doing
really leading the pack here with the
release of gp4 many you know keep your
eyes on prompt chains I know that you're
likely working with AI agents and llm
libraries but prompt chains and chaining
in general is a really really powerful
way to kind of bridge the gap between
second tier and top tier Model results I
do think in the future we're going to
see this trend where anthropic Google
and open AI are going to keep pushing
out like very very topend models and
really charging for it it's going to get
cheaper of course but they're going to
charge for the best model and then
they're always going to have some type
of secondary model right we saw that
with Gemini Flash and Gemini Pro and
we're seeing it now with gbd 40 mini
open ey was really the first to kind of
kick that off with uh gbt 3.5 and then
gbt 4 and then everyone followed we have
you know Claude Haiku Sonet and Opus so
I think this is going to be a trend that
we continue to see and I just want to
give you a technique to kind of bypass
all the noise right you can using the
second tier models like gbg4 oh mini and
a great Fusion chain or even just a
prompt chain you can really really get
the top state-of-the-art Model results
when you're using the right prompt
chaining techniques let me know if you
like this idea let me know if that makes
sense let me know if you're
experimenting with prompt chains in the
comments below we are on a journey to
building intelligence that works on our
behalf we're building software that is
living that works while we sleep if that
interests you hit the like hit the sub
and I will see you in the next one