The DeepSeek Edition - The Dispatch #2
Reasoning models and why it's different, AI calorie tracker, and personal biohacking
Welcome all 5870 subscribers to another edition of The Dispatch. Today's edition is about the most hotly spoken about app that is charting the US App store and has caused a mini financial meltdown in the world - DeepSeek.
Here is The Dispatch #2
DeepSeek
DeepSeek is a new open source reasoning model that's come from China and was apparently trained and built for a budget of less than $6M. That is several orders of magnitude less than what the other SOTA models from OpenAI, Anthropic, and Gemini are trained for. And it's matching and even exceeding benchmarks of these models.
There are a few things interesting about this model as a consumer and from a macro perspective. From a consumer point of view, it's a reasoning model that is available for free. o1 and o1 mini are only on the paid ChatGPT plus accounts. This exposes more people to the reasoning aspect of LLMs and that's a gamechanger.
Being able to see the unfiltered thought process of an LLM to consider all cases and then answer a question is eye opening. It personalizes the LLM and you can see how it has got something correct or how there are gaps in thinking. This builds trust in an LLM. Seeing the train of thought also allows you to tweak the subsequent responses to address the gaps or add more context to get an even more nuanced response.
See an example of DeepSeek:
Whereas o1 and o1 mini offer a very summarized version of this thinking.
Here is another example of the level of thinking that DeepSeek does.
But one thing is for sure, it's hard to go back to using LLMs the old way after seeing the train of thought.
I am a Claude Sonnet 3.5 super user. And am just waiting for it to launch it's reasoning model. Till then I have used custom instructions to simulate this thinking process in the output. Here is how you can set it up.
I am yet to see applications apart from math and coding, where a reasoning model like o1 or deepseek-r1 can outperform GPT4o or Claude Sonnet 3.5. For example for some writing tasks I've stuck to gpt4o for simplicity.
Deepseek r1 will be a game changer in tools like Cursor, Replit, etc. They will make coding more effective, because you can see what the LLM is considering while writing the code.
Anyway, I encourage you to try out Deepseek for yourself.
From a Macro lens, it's a fundamental shift in the market where it is now clear that new companies without much of a budget can reproduce SOTA models thanks to DeepSeek's breakthroughs at a small budget and open source tech.
There are a lot of takes on Twitter/X. but I found Yishan's take that it's not a Sputnik moment where the USSR (Soviet Union) made breakthroughs in space tech before the US and the western world could, but more like Google in 2004 when it launched their cost-effective and optimized search engine, took market share, and went public.
Read the rest of the thread here
🚀 Spotlight: AI Calorie Tracker
I’ve always struggled to count my calories and hence have a consistent diet. To solve my own problem, I've launched an AI calorie tracker app where you take a picture of your food and it automatically tracks calories. It's reasonably accurate — I tried to make it more accurate through prompt engineering. You can also add height, weight, goals, and accordingly calorie and protein targets will be calculated.
This is something that I built for myself, please try it out and give me feedback!
🍱 Check out the AI Calorie tracker here
Lastly, a quick biohacking update: I'm going to the sauna frequently, and I think it's really helping me mentally and physically. My sleep has improved tremendously - I'm actually sleeping more than I normally do, and it's giving me mental clarity and peace in the morning.
That's it for today, thanks for reading! Have you tried DeepSeek? What's your take on it? Reply or comment below.
In case you or someone you know would like to have an paid consulting call with me about product or AI, get in touch with me by replying to this email. You can check out my value add here. Alright then, talk soon!