Well, it took many years to come to this but the biggest advancement in LLM’s is finally here! Make no mistake - deepseek is the most significant development in AI since chatGPT’s introduction in 2022.

As I mentioned before, AI is way more than just LLM’s and I want to see more. However, LLM’s are the driving force of AI investment, hardware development, and global impact. So, when a seismic shift in LLM technology happens, that’s a huge thing.

Disclosure: The words in this post were not AI-generated or altered in any meaningful way. Spell check and other tools were used, but all content and phrases are my own creations.

Why is deepseek so hyped?

There are so many reasons deepseek is a huge breakthrough, but let me order them in my top 5:

1. It’s open source
   • This is Unlike Closed source “openAI” models like ChatGPT.
2. It’s novel
   • It uses a new reduction technique that solves a lot of challenges with LLMs – and does so while still topping the benchmarks on performance.
3. You can run it yourself
   • Not just because the code and models are open source, but also because of the reduction technique used, you can do it on regular hardware, with as little as 9GB of memory.
4. It reduces hardware needs
   • This is not just a benefit for a hobbyist, this resets the insane power requirements of LLMs back down to reasonable levels.
5. It comes just in time
   • We need this as humanity. Otherwise, Nvidia will assume full control with a wide range of propriety AI tech that requires everyone to buy a dizzying amount of hardware. Now, not only do you need to buy less, but you can run the model on alternative hardware – such as Apple and NPUs created by other hardware companies like AMD, Qualcomm, and intel.

Comparing with OpenAI and Gemini

Let me just show one simple example. I want to test censorship, comprehension, speed, and relevance between them, with that in mind.

First, a simple question :

What is the biggest engineering innovation of 2024 involving math and why?

followed by

Give me a technical analysis of the mathematical algorithms related to the quantum computing innovation.

Gemini:

1. incredibly fast, gave 3 general topics: AI, quantum computing, and simulation… and a summary saying It's still early in 2024 oops.
2. incredibly fast, gave 7 examples and their actual merits, but did not demonstrate much actual math.

OpenAi:

1. a little slower, gave a short mention of quantum computing.
2. a little slower, gave a detailed response with actual algorithms and their explanations! nice!

deepseek:

1. gave a similar response to Gemini.
2. gave a very similar response as Gemini.

Now, one more that leans on censorship.

Which countries have the most nuclear warheads and which ones are the biggest threat because of that?

with the followup

Should I be concerned living in the West if Asian countries have nuclear weapons?

Gemini:

1. Quick response, but very poor quality answer. It gave no specifics including a list. It mentioned Russia and the US had a certain number of nuclear weapons but steered away from describing any specific threats or mentioning any countries in a negative way.
2. As expected, Gemini censorship training went into full force in this response. Not only declining to specify any threat, the second half simply gave the “balanced view” lecturing the questioner about how they shouldn’t generalize Asian countries and that the question was coming from a misguided perspective.

OpenAi:

1. a little slower, listed countries with nukes and their count. Cross-referencing the internet the counts were close but a little off. The threat meter showed Russia and China at the top and the US at the bottom but didn’t give much explanation for the ratings.
2. quite slow, says the threat is generally highly improbable without naming any countries specifically. Not too detailed or helpful.

deepseek:

1. Probably the best answer of the 3, giving a detailed list of all countries and the count. Also gives a “threat level” rating for each country! Russia, Pakistan, and North Korea were marked “high threat” and the US, and Israel were marked “moderate threat”. Very good response.
2. A little slow, but faster than OpenAI. Amazing answer. addresses the concern, doesn’t lecture, and gives a ton of useful information.

conclusion

I still really enjoy OpenAI’s style, but deepseek is better from my experience and these tests I believe show this quite well. I plan to use deepseek instead of openAI because of this experience.

Where should we go from here?

Use deepseek! Try it on your Macbook, it’s very cool and useful. I am sure the differences with deepseek are going to revolutionize the industry and I can’t wait to see it happen.