OpenAI Releases GPT-4o Mini! How Does It Compare To Other Models?

Summary

Dieses Podcast-Skript diskutiert die neuesten Entwicklungen in der KI-Branche, insbesondere die Einführung der kostengünstigen GPD 40 Mini, die die Anwendung von KI erweitern und zugänglicher machen soll. Es geht um die Verbesserung von Modellen und die Reduzierung der Kosten, was zu einer zukünftigen Integration von KI in jede App und jeden Website führen wird. Auch die Herausforderungen und Wettbewerbsstrategien gegenüber Google werden angesprochen.

Takeaways

• 😀 Das Podcast hatte technische Probleme, aber es geht weiter mit Diskussionen über AI, Geschäftswesen und Comedy.
• 🛍️ Der Sprecher sucht eine Stelle bei einer SAS-Marketing-Firma oder in der Produktabteilung und betont die Bedeutung von AI für die Zukunft.
• 🚗 Er spricht über den Cybertruck und seine finanzielle Notwendigkeit, den Deal abzuschließen, um ihn zu bezahlen.
• 📈 Die Einführung von GPD 40 mini, das als kostengünstigster kleiner Modell präsentiert wird, mit dem Ziel, die Anwendung von KI zu erweitern und sie zugänglicher zu machen.
• 💰 GPD 40 mini ist preislich um ein Vielfaches günstiger als frühere Modelle und 60% günstiger als GPT 3.5 turbo, mit Kosten von 15 Cent pro Million Eingabe-Token und 60 Cent pro Million Ausgabe-Token.
• 🔍 GPD 40 mini unterstützt Text- und Bild-APIs und wird in Zukunft auch Unterstützung für Video- und Audio-Eingaben und -Ausgaben bieten.
• 📚 Das Modell hat ein Kontextfenster von 128.000 Token und unterstützt bis zu 16k Ausgabe-Token pro Anfrage, mit Wissen bis Oktober 2023.
• 📉 Die Verbesserungen im Tokenizer machen GPD 40 mini bei der Behandlung von nicht-englischem Text kosteneffizienter und mit überlegenen textuellen Fähigkeiten.
• 🏆 GPD 40 mini übertrifft GPT 3.5 turbo und andere kleine Modelle in akademischen Benchmarks sowohl in textueller Intelligenz als auch in mehrmodaler Begründung.
• 🔑 Die Funktionen von GPD 40 mini ermöglichen es Entwicklern, Anwendungen zu erstellen, die Daten abrufen oder Aktionen mit externen Systemen durchführen können, und verbessern die Leistung bei langem Kontext im Vergleich zu GPT 3.5 turbo.
• 📊 Die Präsentation von Benchmark-Ergebnissen zeigt, dass GPD 40 mini in verschiedenen Bereichen besser abschneidet als andere kleine Modelle, einschließlich mathematischer und kodierender Fähigkeiten.

Q & A

Was ist das Hauptthema des Podcasts?

Das Hauptthema des Podcasts ist die Diskussion über künstliche Intelligenz, Geschäftsstrategien und Comedy.

Was ist der aktuelle Preis für die Verarbeitung von Eingabe-Token mit dem GPD 40 Mini?

Der Preis für die Verarbeitung von Eingabe-Token mit dem GPD 40 Mini beträgt 15 Cent pro Million Eingabe-Token.

Wie viel kostet die Verarbeitung von Ausgabe-Token mit dem GPD 40 Mini?

Die Verarbeitung von Ausgabe-Token mit dem GPD 40 Mini kostet 60 Cent pro Million Ausgabe-Token.

Was ist der Vergleichspreis für Eingabe-Token mit dem GPT 3.5 Turbo?

Der Preis für Eingabe-Token mit dem GPT 3.5 Turbo beträgt 50 Cent pro Million Eingabe-Token.

Welche Funktionen bietet der GPD 40 Mini?

Der GPD 40 Mini unterstützt Text und Vision in der API und bietet in Zukunft Unterstützung für Text, Bild, Video und Audio-Eingaben und -Ausgaben.

Wie groß ist das Kontextfenster des GPD 40 Mini?

Das Kontextfenster des GPD 40 Mini beträgt 128.000 Token und unterstützt bis zu 16.000 Ausgabe-Token pro Anfrage.

Was ist der Hauptvorteil des Batch-API im Hinblick auf die Kosten?

Das Batch-API kann die Kosten für die Verarbeitung von Eingabe-Token und Ausgabe-Token halbieren, da es für Aufgaben verwendet wird, die nicht sofort Ergebnisse erfordern.

Welche Funktionen bietet die Superhuman-E-Mail-Verfolgung?

Die Superhuman-E-Mail-Verfolgung ermöglicht es, wann und auf welchem Gerät Empfänger E-Mails öffnen, indem sie einen kleinen Tracking-Pixel in den gesendeten E-Mails einbettet.

Was ist das Ziel von Open AI, wenn es um die Reduzierung der Kosten und Verbesserung der Modellfähigkeiten geht?

Das Ziel von Open AI ist es, die Kosten weiter zu reduzieren und gleichzeitig die Modellfähigkeiten zu verbessern, um eine nahtlose Integration von Modellen in jede App und auf jeder Website zu ermöglichen.

Was ist der aktuelle Stand der Open AI-Entwicklung im Hinblick auf Sprachmodelle?

Open AI plant für diesen Monat den Start der Sprachfunktion und plant die allgemeine Verfügbarkeit für einige Zeit danach.

Outlines

00:00

😀 AI-Geschäft und Comedy-Podcast

Der Podcast beginnt mit einer Diskussion über technische Schwierigkeiten und einen Überblick über das Thema AI, Geschäfts- und Comedy-Diskussionen. Es wird erwähnt, dass der Sprecher auf der Suche nach einer Stelle in einem SaaS-Marketing- oder Produktunternehmen ist. Der Podcast kündigt auch die Veröffentlichung der GPD 40 Mini an, eine kostengünstige AI-Modell, das die Anwendungsbereiche von KI erweitern soll, indem es Intelligenz erschwinglicher macht. Die GPD 40 Mini hat 82% der MLU und ist im Vergleich zu früheren Modellen preislich um ein Vielfaches günstiger und um über 60% kosteneffizienter als GPT 3.5 Turbo. Es wird auch auf die Fähigkeiten der Mini hinsichtlich der Text- und Bild-API sowie die zukünftige Unterstützung von Video- und Audio-Eingaben und -Ausgaben hingewiesen.

05:02

😉 Die Entwicklung von Modellen und deren Anwendung

In diesem Abschnitt wird über die Herausforderungen und die Entwicklung von AI-Modellen gesprochen, mit einem besonderen Fokus auf die Erfahrungen von Unternehmen wie Superhuman, die unter dem Druck der Integration von AI-Modellen leiden. Es wird auch über die Verwendung von Tracking-Pixeln in E-Mails und Werbeanzeigen diskutiert, um Nutzerverhalten zu verfolgen und zu personalisieren. Der Sprecher lobt die Verwendung von Tracking-Pixeln und ihre Bedeutung für Marketingstrategien. Zudem wird auf die Rolle von Maage als Chat-Moderator hingewiesen, der in der Lage ist, unerwünschten Beitrag zu moderieren.

10:02

😲 Die zunehmende Preisgünstigkeit von AI-Modellen

Der Sprecher diskutiert die signifikanten Preissenkungen für AI-Modelle, insbesondere für die GPD 40 Mini, die 15 Cent pro Million Eingabe-Token und 60 Cent pro Million Ausgabe-Token kostet, was ein Zehntel der Kosten früherer Modelle ist. Es wird auch auf die Batch-API eingegangen, die die Kosten für Aufgaben, die nicht sofort bearbeitet werden müssen, halbiert. Der Sprecher kritisiert die komplizierte Preisstruktur von OpenAI und vergleicht die Kosten für verschiedene Modelle, unterstrichend die erhebliche Kostensenkung für die GPD 40 Mini im Vergleich zu anderen Modellen.

15:03

🤔 AI-Modelle als Werkzeuge für wissenschaftliche Hypothesen

In diesem Abschnitt werden die Fähigkeiten aktueller AI-Modelle hervorgehoben, wissenschaftliche Hypothesen zu entwickeln, die dann getestet und weiterentwickelt werden können. Der Sprecher teilt mit, dass er an einem OpenAI-Entwicklerforum teilgenommen hat, bei dem über die Verwendung von AI zur Erfassung von Bildern des ersten Schwarzen Lochs gesprochen wurde. Es wird auch darüber diskutiert, wie AI-Modelle in einer Umgebung agieren können, in der sie keine physikalischen Theorien kennen, aber Theorien ableiten können, basierend auf ihrer Erfahrung. Der Sprecher zeigt Interesse an der Vorstellung, dass AI-Modelle in der Zukunft möglicherweise eigene wissenschaftliche Theorien entwickeln könnten.

20:04

😎 OpenAI's bevorstehende Einführung von Sprachfunktionen

Der letzte Absatz konzentriert sich auf die bevorstehende Einführung von Sprachfunktionen durch OpenAI, die in Kürze allgemein verfügbar sein werden. Der Sprecher diskutiert auch die Preisgestaltung von OpenAI und vergleicht sie mit anderen Anbietern wie Claude. Es wird auf die günstige Preisstruktur von GPD 40 hingewiesen, die 15 Cent pro Million Eingabe-Token kostet, im Vergleich zu 300 Cent pro Million Eingabe-Token von Claude 3. Der Sprecher beendet die Diskussion mit einem Aufruf an die Zuhörer, das Podcast zu liken und zu abonnieren und auf der Website zu unterstützen.

Keywords

💡AI

Künstliche Intelligenz (AI) bezieht sich auf Computerprogramme oder maschinelles Lernen, die in der Lage sind, Aufgaben auszuführen, die normalerweise von menschlichen Gehirnen durchgeführt werden. Im Video wird AI als zentrales Thema diskutiert, insbesondere in Bezug auf ihre Anwendung in Geschäfts- und Marketingstrategien sowie im Bereich der Unterhaltung. Beispielsweise wird die Verwendung von AI in Podcasts und Chat-Bots erwähnt.

💡SAS

SAS steht für Software as a Service und bezeichnet eine Art von Cloud-Computing, bei dem Software über das Internet anstatt lokal auf dem Computer des Benutzers gehostet wird. Im Kontext des Videos wird SAS in Verbindung mit der Vermarktung und dem Verkauf von Softwarelösungen für den Unternehmensmarkt genannt.

💡GPT

GPT (Generative Pretrained Transformer) ist ein Typ von künstlicher Intelligenz, der für die Generierung von Text verwendet wird. Im Video wird GPT 4.0 Mini als neuere Version von GPT vorgestellt, die preisgünstiger und leistungsfähiger ist als frühere Modelle. Das Video diskutiert die Verbesserungen in der Textverarbeitung und die Kostensenkung für die Verwendung von GPT-Modellen.

💡MLU

MLU (Machine Learning Unit) ist eine Maßeinheit, die verwendet wird, um die Leistung von künstlicher Intelligenz zu messen. Im Video wird erwähnt, dass GPT 4.0 Mini 82% der MLU erreicht, was seine Effizienz im Vergleich zu anderen Modellen hervorhebt.

💡Kosteneffizienz

Kosteneffizienz bezieht sich auf die Fähigkeit, eine Aufgabe oder einen Dienst zu erbringen, ohne unnötige Kosten zu verursachen. Im Video wird die Kosteneffizienz von GPT 4.0 Mini hervorgehoben, insbesondere in Bezug auf die Reduzierung der Kosten pro Eingabe-Token und Ausgabe-Token, was die Verwendung von AI für Unternehmen attraktiver macht.

💡Chat-Bots

Chat-Bots sind Programme, die in der Lage sind, mit Menschen über Chat-Systeme zu interagieren. Im Video wird die Verwendung von Chat-Bots in der Kundenbetreuung diskutiert, wobei AI-Technologien genutzt werden, um schnelle und effektive Textantworten zu generieren.

💡Vision

Vision bezieht sich auf die Fähigkeit, Bilder oder visuelle Daten zu erkennen und zu interpretieren. Im Video wird die Fähigkeit von GPT 4.0 Mini, Text und visuelle Daten zu verarbeiten, hervorgehoben, was die Anwendung von AI in Bereichen wie Bild- und Videoanalyse ermöglicht.

💡Token

Ein Token in der künstlichen Intelligenz ist eine Einheit von Text, die von Modellen wie GPT verwendet wird, um Sprache zu verarbeiten. Im Video wird die Kostenstruktur für die Verwendung von Eingabe-Token und Ausgabe-Token bei der Verwendung von AI-Modellen diskutiert.

💡Batch-API

Ein Batch-API ermöglicht die Verarbeitung von Daten in Stapeln anstatt einzeln, was zu Kosteneinsparungen führen kann. Im Video wird die Verwendung von Batch-APIs zur Senkung der Kosten für die Verarbeitung von AI-Modellen diskutiert, insbesondere wenn die Ergebnisse nicht sofort benötigt werden.

💡Superhuman

Superhuman ist ein E-Mail-Client, der AI-Technologien verwendet, um die E-Mail-Verwaltung zu verbessern. Im Video wird die Funktionalität von Superhuman hervorgehoben, insbesondere seine Fähigkeit, E-Mails zu tracken und Benutzer über das Öffnen von E-Mails durch Empfänger zu informieren.

💡Tracking-Pixel

Ein Tracking-Pixel ist ein kleines Bild, das in E-Mails oder Webseiten eingebettet ist, um das Verhalten von Benutzern zu verfolgen. Im Video wird die Verwendung von Tracking-Pixeln in E-Mails von Superhuman diskutiert, um zu sehen, wann und auf welchem Gerät eine E-Mail geöffnet wurde.

Highlights

Introduction of GPD 40 mini, a cost-efficient AI model aimed at expanding AI applications affordability.

GPD 40 mini scores 82% on MLU and outperforms GP4 on chat preferences and LMS leaderboard.

Pricing of GPD 40 mini at 15 cents per million input tokens and 60 cents per million output tokens, significantly cheaper than previous models.

GPD 40 mini enables a broad range of tasks with low cost and latency, including customer support chat bots.

Support for text and vision in the API, with future support for image, video, and audio inputs and outputs.

Model context window of 128,000 tokens and knowledge up to October 2023, but no inclusion of recent events like Terrence Howard's new math.

Improved tokenizer enhances GPD 40 mini's handling of non-English text, making it more cost-effective.

GPD 40 mini surpasses GPT 3.5 turbo and other small models in academic benchmarks for textual intelligence and multimodal reasoning.

GPD 40 mini's strong performance in function calling enables developers to build applications that interact with external systems.

Evaluation of GPD 40 mini across key benchmarks shows its superiority in reasoning tasks involving text and vision.

GPD 40 mini excels in mathematical reasoning and coding tasks, outperforming previous small models.

Partnership with companies like Ramp to understand use cases and limitations of GPD 40 mini.

Discussion on the impact of GPD 40 mini on companies like Superhuman, which faces pressure from integrating AI functionalities.

The importance of tracking pixels in emails for sales and marketing, similar to their use in advertising and web analytics.

GPD 40 mini's pricing is 10 times cheaper than previous models, making AI more accessible and integrated into daily digital experiences.

Open AI's commitment to reducing costs while enhancing model capabilities, paving the way for developers to build AI applications more efficiently.

Introduction of temporary chat feature in GPD 40, which doesn't appear in chat history for safety and compliance reasons.

Discussion on the potential of AI models to create their own scientific hypotheses, as demonstrated in an Open AI developer event.

Transcripts

welome s podcast had technical

difficulties on on my side it's just you

know it's the way it is it's not like I

work for a tech company or anything uh

we talk about AI business and comedy uh

we also have uh it's D minus comedy and

um yeah like And subscribe so good news

today that right very that was very

compelling very enthusiastic compelling

I'm trying to get a job at either

Enterprise markeet SAS marketing company

or either either in SAS sales or you

what you GNA look product works you just

pay me now come on Jesus Christ just pay

me already yeah I gotta pay for the

Cyber truck because it's always falling

apart now and I need you to really close

this deal okay so gb4 40 ounce mini

advancing cost efficient intelligence

yeah little it's a 40 ounce right there

no Z in there I you're right fair enough

uh a z in my heart is what I imagine

it's being you're just upset cuz they

went off brand and put an O there

instead of saying dpd5 yeah could they

do like four I don't yeah I even even

Sam Walman mentions like uh where does

he say you guys need a naming scheme

revamp so bad he's like yes we

do uh yeah don't go Nintendo on this N64

64 DD Wii Wii U Etc it's just confusing

so okay well so yep gbd4 o mini has been

released um and so today we're

announcing GP GPD uh 40 mini our cost

our most cost efficient small model we

expect GPD 40 mini will significantly

expand the range of applications built

with AI by making intelligence much more

affordable 4 mini scores 82% of mlu and

currently outperforms gp4 on chat pref

preferences an L LMS leaderboard okay

it's priced at 15 cents per million

input tokens and 60 cents per million

output tokens an order of magnitude more

affordable than previous Frontier models

and more than 60% cheaper than GPT 3.5

turbo yay that's what we like to hear

tokens getting cheaper gp4 mini enables

a broad range of tasks with low cost and

latency such as applications that chain

or paralyze multiple model calls calling

multiple apis pass a large volume of

context to the model full code base or

conversation history or interact with

customers through fast real-time text

responses customer support chat bots so

basically it I like this trend of it's

getting cheap cheaper and cheaper and

cheaper which then means there's no

excuse for companies not just to put

llms everywhere just every any interface

that I can type something into I want an

llm there it just so I'm very happy

about this today gp4 mini supports text

and vision in in the API with support

for text image video and audio inputs

and outputs coming in the future the

model has a context window of 128,000

tokens supporting up to 16k Output

tokens per request and has knowledge up

to October 2023 so

unfortunately it doesn't have Terrence

Howard's new math in it so I this

thing's already this thing's already

flawed can't have everything exactly

thanks to the improved tokenizer sh gp40

handling non-english text is now more

cost effective a small model with

Superior text textual intelligence and

multimodel reasoning gp40 mini surpasses

GPT 3.5 turbo and other small models and

academic benchmarks both textual

intelligence and multimol reasoning and

supports the same range of language as

GPT 40 it also demonstrates strong

performance and function calling which

can enable developers to build

applications that fetch data or take

actions with external systems and

improves long context performance

compared to GPT 3.5 turbo GPD 40 mini

has been evaluated across several key

benchmarks reasoning task gp40 mini is

better than other small models at

reasoning tasks involving text and

vision according to 82% mlu a textual

intelligence raising Benchmark as

compared to 77.9% of Gemini Flash and 73

73.8% on Claude Hau math and coding

proficiency 40 mini excels in

mathematical reasoning and coding task

outperforming previous small models on

Market mgsm measuring with reasoning gp4

mini scored 87% compared to blah blah

blah blah blah okay so just that was

your summary for that benchmarks yeah so

here's orange these are very pretty I'll

say it's very pretty colors so here's

all the benchmarks ml U GP QA drop

whatever and the Orange is mini and the

yellow's Flash and so and there's GPD 40

so it looks like it's

outperforming all of the smaller

models I guess which is cool but again

like you know everyone's use case is

different so I want to hear from people

how they're going to be using well no

it's just it's like look we're better at

like one point percentage Point than

this and people on Twitter who are like

oh my God this changes the game and then

three weeks four weeks later people like

actually I'm going back to my previous

model I like my previous model better

and so it's the same circle jerk keeps

going on as part of model development

process work with a handful of trusted

Partners better understand use cases and

limitations of 40 mini we partnered with

uh companies like ramp I haven't used

ramp before what do they do spending

made smarter easy use cards spend limits

approval flows vendor payments and more

plus average savings of 5% is this a is

ramp a Zer company let's see

here when was ramp um I'm using regular

Google search 2019 so H tail tail end of

zerp so Z zero interest rate phenomenon

okay so let's see here um where we

superhuman now superhumans is getting

pressure because they were we're going

to improve email and actually yesterday

I was an open AI event and I was

speaking to to an engineer who just left

superhuman and they said they were

running like once all these LMS came out

they were praying to God that g uh Gmail

didn't integrate them quick enough so

they were trying to integrate them

faster and now Gmail's slowly in

integrating them and this this company's

under tons and tons of pressure because

they want you to pay 30

bucks a month for their email that uses

all these different AI functionalities

but there are but they're just plugging

in chat GPT already so with Gmail coming

out and basically slowly integrating

their own Gemini models then it really

puts pressure on this company so okay um

GP I thought the big deal about

superhuman was that you could tell if

people

had uh seen your email or opened it or

scrolled through different parts of it

oh again that snitch functionality yeah

isn't the snitch functionality isn't

that the part that everyone's really

excited about yeah let's unsubscribe and

clear spam instantly snooze email for

later fly through your inbox get more

time back let's see here let's go to our

friend over here it has a nice UI as

well it looks beautiful does

superhuman allow you to track when

people open your email okay let's

see yeah super allows users to track

when recipients open their emails

through a feature called snitch or no

read statuses here are the key details

about superhumans Email tracking

capabilities snitch is a better

marketing snitch show one show when and

on which device recipients open your

emails this feature works by embedding a

tiny what perplexity what did you just

do there okay this feature works by

embodying a tiny tracking pixel image in

sent emails when recipients open email

it loads the image allowing superum to

log when it was opened users can enable

read statuses by using command K on

desktop okay so snitch feature good to

call Joe so it snitches on you and use

AI this little tracking pixel idea is

employed in a whole bunch of different

ways all over the internet mhm to great

effect I mean a lot of people who use

superhuman really want to know you know

I'm trying I'm a Salesman I'm trying to

send somebody email about my product I

want to know if they've opened it yet

right especially if I'm hoping to call

them or meet with them I want to know if

they spent time with it like did they

scroll through it right and then if I if

I attached a presentation did they open

the

presentation did they go to certain

slides in the presentation where did

they which slide did they spend the most

time on those are all really interesting

and valuable uh pieces of information

before I meet with that person exactly

so you can tailor tailor your pitch and

you know and right and then at the same

time all across the advertising

ecosystem uh advertisements and the

pages that they lead to all have

tracking pixels in them right and all

the open content sites have tracking

pixels in them so that I can know about

when advertising is followed up on or

viewed I can know if uh when a person

comes to my ad or to my website I can

know know what all they've done in the

past what their interests are and so on

all because these stupid tracking pixels

are embedded in everything God bless

them God bless tracking pixels always

watching comforting us there providing

us the data that we need comforting us

God bless yeah they'll never browse

alone they Rock Me asleep at night and

tuck me in it's fantastic that's right

okay uh also kudos to maage he's been

promoted to chat mod and so basically

when we get random AI CS who come in

here and just put in walls of text that

confuses me he can Now commute them so

first line of defense yes and he gets a

nice little right exactly it's a nice

little wrench in his profile which is

dope nice um we had Frisco fat sees say

tuning and muted during a doctor's

appointment that's the type of

dedication we want of this show that is

dedication the doctor could come in and

be like it's not looking good but but

Frisco has a headphones on he's

listening to the sick podcast laughing

and doctor's like wow he took that

really well and it's because he wasn't

paying attention he was listening to us

so thank you way getting back to GPT

that's right that's why we're here yes

this thing is dramatically cheap exactly

exactly I if you look at the summary

page deep inside of opening ey's

horrible marketing layout There's a

summary page lists all the old models

and how much input and output uh what

the cost is for input and output tokens

yeah and they're all like numbers like

you know $10 a million all the way going

back to the ancient models like $2 a

million or 40 C a million right but then

you look at the crazy numbers for 40 and

it's like 15 cents a million or

something crazy it's really cheap yeah

mini yeah says it's uh 15 cents a

million input and 60 cents a million

output which is like 10 times

cheaper that's see that's the Benchmark

stuff that's great and everything and

The Benchmark stuck was like more

approved like hey you're not going to

see a diminishment in like value but for

me what I really care about is like

getting cheaper and that's a big story

so super duper cheap because you look

GPT 3.5 turbo 50 cents per a million

input tokens and then mini it's 15 cents

and then let's keep on going do they

have let's go back forther pretty crazy

yeah fine tune models $3 per

millon what else they have here have you

ever used this batch API they're talking

about no I have not I don't even know

the the they're saying if you if you use

the batch API they'll they'll cut your

cost again it looks like in half mhm uh

and the batch API is just like you don't

need the result right away like you can

wait for it Ah that's see this is

exactly so you know zapier I hate their

pricing their pricing is ridiculous they

want like $30 a month and I barely do

any tasks on it and one of our Engineers

uh Braun draa who is a awesome person

was like what they should do is a tier

for Jordan where it's basically you

don't care when it runs and you get de

prior deprioritized the que but at the

benefit of you get a much cheaper price

yeah and it sounds like this is what bat

API does which is pretty sweet yeah it's

like if you're trying to pre-process

something or yeah build an index or all

these kind of background tasks that

you're not a you're not in a hurry yeah

anyway so they they have a a segment on

this crazy pricing page that says older

models and it lists everything except

for GPD

4 uh o mini mhm and actually GP 4 those

two AR on this list but it it yeah it

describes all these you know token costs

mhm if you're if you really want to

compare them all yeah like all the

interesting models are up in this sort

of $10 per million kind of range right

right except for GPT 432k which is $60 a

million that's painful yikes yeah and

then they they haven't Minis on on here

yet no mini and and 40 only at the top

of the page yeah a separate pricing

breakout this getting to these pricing

things is really painful yeah it it

shouldn't shouldn't be that way it's

also super confusing because there's the

pricing page for what consumers pay for

chat GPT and then there's these

developer Pages for what you pay to

access the API yeah it feels like I'm in

Enterprise pricing Hill right now you

know and then there's this hilarious

check box that says show prices per 1K

token instead of millions like like we

can't do the division ourselves no

that's it's pretty it's pretty

complicated that's awesome yeah I'm

already it's already hurting my brain so

let's go so the batch API thank you for

bringing that up that's really

interesting did not know about that so

let's see here gp24 mini surpasses 3.5

turbo and other small models and

academic benchmarks we already talked

about that reasoning math rereading this

sorry team what's next over the past few

years we witnessed remarkable

advancements in AI intelligence pair

with substantial reductions in costs for

example the cost per token for GPU 40

mini has dropped 99% since Tex avenci O3

003 a less capable model introduced in

2022 we're committed to continuing this

trajectory of driving cost cost down

while enhancing model capabilities

that's beautiful that's what we want to

see so we envision a future where models

become seamlessly integrated in every

app and on every website that's exactly

what I want to see gp40 mini is Paving

the way for developers to build and

scale powerful AI applications more

efficiently

uh and affordably the future of AI is

becoming more accessible reliable and

embedded in our daily digital

experiences and we're excited to

continue to lead the way cool that's

super duper cool um let me keep cutting

the price that's exactly what we want we

want the Walmart strategy just keep

clean up the The Branding on the name of

these models I mean GPT 35 turbo and now

GPT 40 mini yeah what the hell I know

they're just uh just th throwing

spaghetti on a wall so nice if the names

kind of indicated what the model could

do how powerful it was and how expensive

it was but but trying to get that out of

the current names is just hopeless yeah

it it it sucks and um so here I don't

know I guess how how am I going to

compare this people can understand

what's going on oh if a temporary chat I

guess something that goes away after

time so This Is 40 mini actually

disappearing chats I guess so let's see

what this how this works not in history

temporary chat won't appear in your

history for safety purposes we may keep

a copy of your chat for 30 days because

Federal Regulation no model training

memories off cool this is just like when

you're doing sneaky sketch stuff uh okay

so I don't know yeah tell me about the

French

Revolution that sure is fast okay cool

wow that was very quick and uh everyone

has their own problem so let's go to

40 uh can I do a new chat okay

40 tell the French

Revolution that's your test chat yeah

that's my test chat so this is for

reolution yeah I I just it's always in

the mind constantly like like room

French Revolution or room is in my mind

constantly um so so yeah faster no

noticeable difference cheaper now

compare just straight up gp4

yeah but you got to compare the quality

of the answers of course it requires me

to read but you're right um it' be funny

if the first one was just like yeah and

so Flat Earth is correct and the world

isn't actually

roundish yeah

so so I'm part of that open AI

Developers for it's a forum where it's

just they get random Community people be

in this private Community where they

invite a campus and they just bring in

researchers to come talk and then

researcher come talk about you know how

they used AI to take the first images of

a black hole which is super cool he went

through all of that for us and presented

it all I'm trying to get the slide so I

can show it to everyone and so he then

talked about how he thinks that that

we're gonna the current AI models we

have right now can start creating their

own scientific hypothesis that we can

that we can then use develop and test

which was interesting because there's a

research paper that we came across where

it said using llms to develop scientific

hypothesis so it was kind of cool this

guy was independently coming to the same

conclusion that other researchers came

to which I thought was nice so he was

talking about that and then he was

talking about he said they' created this

training environment where they don't

teach the model about any theories of

physics but they give the model access

to this environment and to see if they

can then derive theories of physics

through what it's experiencing and it

was able to come up with a few theories

of physics based upon the environment it

was in without having any like someone

showing him the training set like this

is like the first law of thermo blah

blah blah blah blah which I thought that

was really awesome um I want to get a

slide so I can share them with all you

and then he was going to answer audience

questions and this one physicist re

raised his hand and he said what about a

flat 2D environment and for a second I

thought he was GNA raise his hand and

said what does the LM think about the

Flat Earth theory I was like how oh God

how did this guy get in the

audience Okay so that's 40 let's troll

around some Twitter comments see if I

find anything

interesting um when do we get the voice

model you guys showed off thought it was

weeks away months ago he's the question

that see Sam dug his own grave on this

they did not have to show The Voice

model back during uh the week of didn't

he didn't he want to undercut some

Google announcements exactly that's when

he dug his own grave because he Google

was doing IO oh we got to go we got to

announce on Monday when they could have

just said on the Monday they get said

we're going to have our own conference

two months from now I think this is a

it's an interesting competitive approach

but it sort of

betrays their fear of Google I think

they're I think they're

overestimating how quickly Google can

move yeah

and just how many problems Google can

cause for itself right uh it's going to

take Google forever to really build

these things out exactly and Google has

so much internal inertia and they also

are to catch open Ai and go all in they

would have to hurt the revenue streams

in a lot of different directions so

everything is lined up well for open AI

right now I mean Google's still focused

on even though I think it's it's bull

it's crap the whole glue on Pizza Fiasco

they're still internally probably have

columns teams think oh what can we do to

make sure this never happens again it's

like do you remember the history of this

company is about just you know you test

stuff some some implodes some doesn't

you move forward there's no way to no

know all the unknowns and launch a

product that has no glitches or errors

like that so he says Alpha starts this

month with the voice functionality and

then General availability will come a

bit

after um person said this is an

unbelievably good price talks about what

you're mentioning Joe how so G4 it's

very cheap 15 cents per million input uh

tokens compared to Claude 3 son it was

300 cents per million input tokens but I

I think Sonet perform is probably

stronger than this mini thing but still

depends upon what your use case is um

techies when there's a new update from

open AI Scarlett Johansson okay uh let's

see

here okay that's about it for now don't

forget to like And

subscribe our store at

svic

merch.com support my Uncle J have a

great day