[00:00]
A tiny Chinese startup just humiliated
[00:02]
OpenAI. They built a model that matches
[00:04]
GPT4's performance. But here is the
[00:06]
crazy part. They did it for onetenth of
[00:09]
the training cost and they are giving it
[00:10]
away for free. Is the era of paid AI
[00:13]
over? Today we put Deepseek V3 to the
[00:17]
test to see if free finally means
[00:18]
better. Before we dive into the tests,
[00:20]
[music] it's crucial to understand what
[00:22]
makes Deepseek so different. In a world
[00:25]
flooded with AI models that are
[00:27]
essentially variations on a theme,
[00:29]
DeepSeek [music] broke the mold. This
[00:31]
isn't just another copycat trying to
[00:33]
chase GPT4's [music]
[00:34]
tail. They fundamentally rethought the
[00:37]
architecture from the ground up, leading
[00:39]
to a breakthrough in [music] efficiency
[00:40]
and power. The secret source, it's a
[00:44]
concept that's been around for a while,
[00:45]
but has only recently been perfected.
[00:47]
They're using a sophisticated
[00:49]
architecture called a mixture [music] of
[00:50]
experts ore.
[00:53]
To understand why this is a gamecher,
[00:55]
imagine a traditional model like GPT4 as
[00:58]
one giant monolithic brain. It's
[01:01]
incredibly powerful, but it's also
[01:02]
incredibly [music] expensive. Every time
[01:04]
you ask it a question, no matter how
[01:06]
simple, [music] the entire massive brain
[01:09]
has to power up and process it.
[01:11]
Deepseek, on the other hand, is like a
[01:13]
team of 64 highly specialized experts.
[01:16]
When you ask a coding question, [music]
[01:18]
a smart routting network instantly
[01:20]
identifies the two best experts for the
[01:22]
job, say the Python Pro and the [music]
[01:24]
Algorithm Ace, and only wakes them up.
[01:27]
The other 62 experts remain dormant,
[01:29]
[music] saving an immense amount of
[01:31]
computational power. This makes it not
[01:33]
just faster, but radically [music]
[01:35]
cheaper to train and run. We're talking
[01:37]
about achieving top tier performance
[01:39]
while using only a fraction of the
[01:41]
computational resources.
[01:42]
>> [music]
[01:43]
>> This efficiency is what should be
[01:44]
terrifying for big tech. Think about it.
[01:47]
If [music] Deep Seek can offer this
[01:49]
level of intelligence for mere pennies
[01:51]
on the dollar through an API, [music]
[01:52]
why would anyone continue to pay a
[01:54]
premium? Open AAI's entire $20
[01:58]
subscription model, which subsidizes
[01:59]
[music] the immense cost of their giant
[02:01]
brain, is suddenly on very shaky ground.
[02:05]
This isn't just a new competitor. It's a
[02:07]
potential extinction level event for the
[02:09]
old way of doing AI. Talk is [music]
[02:11]
cheap. Let's put these models to the
[02:13]
test with a realworld coding challenge.
[02:16]
On one [music] side, we have the
[02:17]
reigning champion GPT40. On the other,
[02:20]
the Challenger [music] Deep Seek Coder
[02:22]
V2. This isn't just any model. It's an
[02:25]
open-source [music] mixture of experts
[02:27]
model trained on a colossal two trillion
[02:29]
tokens of code and natural language. It
[02:32]
boasts [music] top scores on benchmarks
[02:34]
like human eval
[02:37]
claiming to rival proprietary models at
[02:39]
[music] a fraction of the cost. But
[02:41]
benchmarks are one thing. Practical
[02:43]
application is another. So [music] I
[02:45]
asked both to write a Python script for
[02:47]
a classic snake game using the Pi game
[02:49]
library. To make it interesting, I added
[02:52]
a twist. The snake must speed up every
[02:54]
[music] time it eats an apple. This
[02:56]
tests not just basic code generation,
[02:58]
but [music] also state management and
[03:00]
logical implementation.
[03:02]
Right away, Deepseek's performance was
[03:04]
[music] impressive. It generated the
[03:06]
complete functional code about 20%
[03:08]
faster than [music] GPT40.
[03:11]
For developers, that speed translates
[03:13]
directly [music] to productivity,
[03:15]
enabling faster iteration and problem
[03:17]
solving. But speed is meaningless if the
[03:19]
code is broken. So the real question is,
[03:22]
[music] does it actually work? And the
[03:24]
answer is a resounding yes. Deepseek's
[03:27]
[music] code ran perfectly on the very
[03:29]
first try. Zero errors, zero debugging.
[03:33]
GPT40, however, stumbled. It missed a
[03:36]
crucial [music] variable definition,
[03:37]
throwing a syntax error that broke the
[03:39]
program. I had to go back and prompt it
[03:42]
a [music] second time to get a working
[03:43]
fix. This initial test highlights a key
[03:46]
difference reliability.
[03:48]
While both models eventually produce the
[03:50]
correct [music] code, Deepseek delivered
[03:52]
a flawless solution faster and on the
[03:55]
first attempt. [music] For any developer
[03:57]
on a deadline, that's a gamecher in this
[04:00]
round. [music]
[04:01]
Deepseek isn't just a contender. It's
[04:03]
looking like the new heavyweight
[04:04]
champion of coding. Now for the second
[04:06]
test, reasoning and logic. This is where
[04:09]
many models, especially earlier
[04:11]
open-source ones, fall flat. They can
[04:13]
perform complex calculations, but often
[04:16]
miss the simple realworld context
[04:18]
[music] that humans grasp instantly.
[04:20]
This is a crucial hurdle for AI to
[04:22]
overcome if it's going to be genuinely
[04:24]
useful. So, I set a classic logic trap
[04:27]
to see if Deep [music] Seek could think,
[04:29]
not just calculate. I asked, "If I dry
[04:33]
five shirts in the sun and it takes 4
[04:35]
[music] hours, how long does it take to
[04:36]
dry 10 shirts?" The trap is obvious. A
[04:40]
purely mathematical brain might double
[04:42]
the time to 8 hours. It's a simple
[04:44]
question, but it's a fantastic test for
[04:46]
[music] contextual understanding.
[04:48]
Deep Seek answers 4 hours. It
[04:51]
immediately understands that drying is
[04:53]
[music] a simultaneous event. The shirts
[04:55]
all dry together, so adding more shirts
[04:57]
doesn't extend the time, assuming you
[04:59]
have enough space. [music] This ability
[05:01]
to handle nuance and implicit
[05:03]
assumptions is incredibly impressive.
[05:05]
It's a sign of sophisticated training on
[05:07]
diverse, highquality data. This isn't
[05:10]
just a one-off trick. [music] This
[05:12]
reasoning power extends across the
[05:14]
board. In my tests, it excelled at
[05:16]
debugging code, planning multi-step
[05:18]
projects, and even breaking down complex
[05:20]
scientific concepts. [music]
[05:22]
It feels far less robotic than other
[05:24]
open-source models. It's not just
[05:26]
regurgitating [music]
[05:27]
data. It's connecting dots and
[05:29]
demonstrating genuine problem-solving
[05:31]
[music] skills. In fact, in five out of
[05:34]
the seven logic and reasoning tests I
[05:35]
ran, it either matched or outright beat
[05:38]
the current [music] industry leader, GPT
[05:40]
40. That is a monumental achievement for
[05:43]
a model that's completely open- source.
[05:46]
And remember [music] the best part, you
[05:47]
aren't paying a single cent for this
[05:49]
level of intelligence. Why does this
[05:51]
matter if you aren't a [music]
[05:52]
developer? Because of competition. For
[05:55]
the last 2 years, we accepted that smart
[05:57]
AI costs $20 [music] a month. Deep Seek
[06:01]
just proved that intelligence is
[06:02]
becoming a commodity like electricity.
[06:04]
It's getting cheaper every day. This
[06:07]
forces Open AAI and Google to either
[06:09]
lower their prices or release something
[06:11]
significantly better. Either way, we
[06:14]
win. One caveat. This is a Chinese
[06:17]
model. If you are working on top secret
[06:19]
[music] government data, maybe stick to
[06:21]
local models. But for learning, coding,
[06:24]
and general tasks, [music] it's a
[06:25]
no-brainer. If you want to run AI
[06:28]
completely privately on your own
[06:29]
computer, check out this tutorial next.
[06:32]
The revolution is open source.