Deepseek's New AI Has A Secret

Author: Two Minute Papers · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Deepseek’s New AI Has A Secret · Channel: Two Minute Papers · 05:08 · Watch on YouTube

Topics: Model inference and optimization, AI infrastructure and hardware, AI models, Local and open-source models

Summary

The video discusses a new DeepSeek AI release that reportedly performs much better at building software and using a computer terminal, despite a small version-number change. Its price per output token is low, though the host cautions that its greater token use can make the actual cost per task less favorable. The host suggests that learning from self-feedback may help explain the capability gains, while noting this is an interpretation rather than a confirmed explanation. The video also highlights the potential for open weights to be distilled into smaller systems, and ends with a promotion for Lambda’s AI computing services.

Key points

The new DeepSeek version reportedly makes a substantial improvement in software creation and terminal use.
The system is available through an API for a little over a dollar per million output tokens, according to the video.
Comparing cost per million tokens can be misleading because different systems may use different numbers of tokens for the same task.
In the host’s independent test, DeepSeek used two to three times more tokens than Claude Opus to accomplish the same thing.
The host says DeepSeek’s cost per task is still good, though less favorable than the headline price suggests.
The model weights can be downloaded for free, but running them requires expensive hardware.
The host proposes learning from self-feedback as a possible explanation for the capability improvement, while stressing there are no guarantees.
Self-feedback could let an AI critique and improve its own work through repeated attempts, such as finding and fixing problems in a game it creates.
The video argues that competition between frontier models and open-weight alternatives could make AI more capable and affordable.

Chapters

Start	Chapter	Summary
00:00	A surprisingly strong DeepSeek update	The host introduces the new release and describes reported gains in software building and terminal use despite its modest version label.
01:30	Pricing, token use, and open weights	The host explains why cost per task matters more than cost per million tokens and discusses the expensive hardware needed to run downloaded weights.
02:30	A possible role for self-feedback	The host proposes that learning from self-feedback may explain the capability jump and illustrates how an AI could iteratively critique and improve a game it builds.
04:05	The promise of AI competition	The host reflects on the difficulty and potential of self-evaluation, celebrates rapid progress and competition, then promotes Lambda’s AI computing services.

Quotes

It builds much better working software — 00:30

we should always think in cost per task, not per million tokens. — 01:00

This means that the AI is able to critique its own work and learn from it. — 03:02

we can create situations where the AI can learn from its own experiences, learning like a human. — 04:05

Concepts and entities

DeepSeek, DeepSeek V4.1 Flash 1017, Claude Opus 5.5, self-feedback, distillation, API, Lambda, NVIDIA, Károly Zsolnai-Fehér

Full transcript (original language)

00:00 Good news everyone, a new DeepSeek AI system is available. Hooray! So, is this a small or a big jump in AI capabilities? Well, doesn’t sound big because the previous version was 4.1 flash and this one is called 4.1 flash 1017. I know. Names. Only about 5 weeks have passed since the last version and they didn’t even want to increment the version number. So, I guess it’s not a big jump,

00:30 right? >> [laughter] >> Boy, was I wrong when I thought that. Holy matter of papers, look at that. It builds much better working software and [clears throat] is way better at using the terminal to operate the computer. And it’s really good at other things, too. You can run it from the API for just a bit more than a dollar per million output tokens. Wow, but can that really be? This is 15 times cheaper than

01:00 Claude Opus 5.5. Finally, intelligence too cheap to meter. Well, that sounds a bit better than it really is. You see, we always measure dollars per a million tokens, but different AI systems use a different amount of tokens for the same task and DeepSeek is not among the best in this. In my independent test, I saw that it can easily take two to three times more tokens to accomplish the same thing as

01:30 Claude Opus. So, we should always think in cost per task, not per million tokens. Now, the cost per task for DeepSeek is still quite good, just a bit less good as what is advertised. Now, here is the best part. You can download and own the weights for free and own it forever. You will need, hmm, let me check my notes. Ah, yes. Exactly one bazillion dollars to get the

02:00 hardware to run it. So, I can’t run that myself at home. However, that’s not the point. The point is that after a new AI system appears, it starts a race to create something close to as good but at a fraction of its size and cost. Deep Seek does that to the frontier models while other systems distill Deep Seek down into smaller and cheaper systems. So, yes. First, one bazillion dollars

02:30 will be needed and maybe less than a year later, you find yourself being able to run it on a beefy laptop. But, wait. The architecture did not change at all in this new version. So, how on earth did they get such a huge jump in capabilities? Dear fellow scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. So, how? Now, I looked at the language they used in their announcement and cross-checked it against other real research to see what

03:02 they may have meant. No guarantees though, but I think the key term here is learning from self-feedback. What’s that? That’s the opposite of what everybody is talking about which is needing to obtain more training data. This is the opposite. This means that the AI is able to critique its own work and learn from it. That is absolutely amazing. Now, hold on to your papers, fellow scholars, and imagine that it could code up a

03:32 racing game. Then, it starts playing it and notices, “Wait a minute, the collision detection does not seem to work. Things go through each other. Well, that’s not great. Let’s fix that.” Compile the new game and keep playing and improving it. Now, there’s no acceleration. Fix that. Then, visual detail is quite poor. Fix that, too. And so it goes. So, if you remember one thing about this episode, let it be this. We keep talking about giving more

04:05 and more training data to AI systems to learn from. But, previous research works have shown that we can create situations where the AI can learn from its own experiences, learning like a human. And maybe this is it coming online in a real product we can own and use. Ooh, that is incredible and very hard to pull off because the AI essentially judges and rewards itself. And it’s not always easy to tell when it

04:36 is doing well or not. The frontier AI models are improving at an incredible pace and the open weights alternatives are right on their heels. And all this improvement in just 5 weeks. I am stunned. Amazing. It is lovely to see that there is intense competition in this space. That is amazing news for all of us fellow scholars. More intelligence for all of us, cheaper. What a time to be alive. I use Lambda to

05:08 reproduce AI research papers often in minutes. It’s also great to train your own models or fine-tune an existing one. Run inference or text to image or video, easy peasy. Running a deep seek chatbot or agent, super fast, super reliable. Lambda gives you powerful Nvidia GPUs to run your own experiments. I test ideas from the papers I cover and moments later, results. Love it. Seriously, try it out now at lambda.ai/papers.