AI Playground: DeepSeek

[00:00:00] Tessa Burg: Hello and welcome to another episode of Leader Generation brought to you by Mod Op. I’m your host Tessa Burg and today we’re doing a special AI Playground episode. On DeepSeek, I’m sure everyone has seen the headlines and what we want to focus on, or at least my portion of the review is why is this important to businesses and what are we as Mod Op doing today? What do we think you need to be aware of?

[00:00:24] Tessa Burg: So first, in my opinion, it’s undeniable that DeepSeek has had a fundamental impact on the conversation about how AI foundational models are built and scaled. It introduced three technological innovations that show that you can do more with less.

[00:00:44] Tessa Burg: The numbers are in question. The training is in question, but as you’ll hear from Aaron Grando, we did dive in and start testing it. This is a real advancement and the reaction from the foundational model leaders show us that the advancement is something that they’re taking notice of and starting to rethink.

[00:01:05] Tessa Burg: How do they continue to accelerate their own product roadmaps, but differently. And this is really good for everyone. Increased competition increases access. It forces us to rethink about where investments are going and how much energy, how much processing time, how many parameters, how much of anything is really required to get high-quality results.

[00:01:30] Tessa Burg: So lots of important things happening, but for marketers. What do we need to be aware of?

[00:01:37] Tessa Burg: First, a lot of developers did jump on this. There are a lot of people developing against it. Something to be aware of when we took a look at this from a privacy and security standpoint, it did not pass our security and governance process or testing, you’ll hear more from Patty Parobek, who is our VP of AI Transformation about that. I also wanted to point out that we are not the only ones. So when you are now evaluating tools at your business, knowing what model it’s built on is extremely important. And if you don’t already, having your own governance and compliance standards and test is also very important to protect your data and to protect really all of the endpoints at your business. So some of the concerns that exist within DeepSeek, there has been evidence of data leakage. There is a lot of concern over cross data transmission, given that this is a Chinese company.

[00:02:45] Tessa Burg: There are a number of security vulnerabilities, jailbreak techniques have been successful in tests and that the code is highly susceptible to allowing in harmful content. And there has been evidence of malicious code generation. So, we are doing these tests ourselves at the same time, while this is in an environment, we’re not using our own data or any of our own client data.

[00:03:12] Tessa Burg: I don’t want to take away from the significance it’s had on the conversation, but I also wanted to make sure that in our client community, people are taking the right steps and asking their technology partners, the right questions. What models are being run? What are the security protocols in and around those models?

[00:03:30] Tessa Burg: If it doesn’t pass, there are lots of options. And if anything, this proves that competition is good, this tool, this methodology that we’re now all using as a part of our life in a positive way.

[00:03:43] Tessa Burg: That’s all I wanted to share and enjoy the perspective and the reviews from other thought leaders here at Mod Op who are investing their time in testing and evaluating the best way for us as a marketing industry to move forward with AI.

[00:04:00] Aaron Grando: How’s it going? I’m Aaron Grando and I’m VP of Creative Innovation at Mod Op. And I’ve been testing out DeepSeek R1, a new open source, large language model, that’s been making a lot of waves over the past week. And there’s a lot of reasons for those waves, big like market geopolitical force reasons. But what’s beneath those are some real practical applications that DeepSeek seems to enable. So let’s get into those.

[00:04:25] Aaron Grando: For one, DeepSeek is a reasoning model, which means that it generates its thoughts about how to answer before it generates its final answers. OpenAI pioneered this technique last year with their O1 model, but it hasn’t been available as an open source feature yet with any of the open source models that have been released up till now.

[00:04:43] Aaron Grando: So that thinking before answering approach generally means that it’s slower and more expensive to run. But the answers that it gives you tend to be higher quality. DeepSeek is available as an open source distillation of the main DeepSeek R1 model that you can run locally on your more normal, less specialized computer, like if you have a MacBook Pro.

[00:05:06] Aaron Grando: You can think of a distillation of an AI model as similar to like a compressed JPEG. It captures the essence of the model, the overall shape and capabilities of it. without having to take up an entire data center’s worth of computing power so you can run it on your local machine.

[00:05:24] Aaron Grando: Being able to download and run an open source distillation of an AI model means that we now have really, really high-quality artificial intelligence that can be used completely offline or behind a firewall without needing to send any of the data to a third-party that you might not completely trust. That’s a big deal for the enterprise.

[00:05:43] Aaron Grando: I’ve been testing out the 32 billion parameter version of DeepSeek R1 on my MacBook, and I’ve been really impressed. While the output is slower than you might get out of ChatGPT, it’s about 10 words per second, on top of the time that it takes to think before it answers.

[00:05:58] Aaron Grando: The outputs are in the same range, quality wise, as like GPT 4. 0 or O1 Mini, which have been our go-to models for the apps that we’ve been working on. That means that in less than a year, GPT 4 came out last March. The state of the art of LLMs has gone from data centers and subscription plans to free open source programs that you can run on your own computer. It’s pretty cool.

[00:06:22] Aaron Grando: So how do you use DeepSeek R1? There’s lots of ways, but the simplest and probably safest way, if your computer can handle it, is to just run it yourself. LM Studio is a nice app that lets you download, install, and start prompting DeepSeek R1 all within the same UI. And you can also test out other open source models in there as well.

[00:06:44] Aaron Grando: You just need a computer with enough RAM to be able to run the models themselves. If you don’t have a computer that can run it yourself, you can try DeepSeek’s own free chat interface at deepseek. com. Just keep in mind that DeepSeek’s privacy policy gives them some pretty broad permission to do whatever they want to do with your data. And it’s stored in China, so be careful with what you upload to it.

[00:07:06] Aaron Grando: DeepSeek is obviously still really fresh, but we’re already thinking about ways that we can fit it into the apps that we’re building. It’s good enough that it can stand in for expensive prompts that don’t really require quick completions.

[00:07:20] Aaron Grando: Think stuff like content that gets refreshed overnight or insights that update in the background. It’ll be really useful for that. And because we can run it ourselves in our secure environment, we can feed it data that we can’t feed other models. That potentially unlocks a ton of new data for us to include in the models that we’re building and the features that we’re building. Stuff like social media data and customer behavior data that we just weren’t really comfortable using before.

[00:07:48] Aaron Grando: We’re looking really hard at our roadmaps right now and thinking about the mix of models that we use overall. And if we should be adjusting our thinking to support open source models up and down the stack. So, stay tuned on that.

[00:08:02] Aaron Grando: On a scale of 1 to 5, with 5 being the most recommended, how strongly would I recommend DeepSeek? It’s definitely worth a few queries in the free chat. Just make sure you’re keeping your secure data secure. If you’ve got the setup for it, give running it locally a shot. If you have intelligence on your local machine, you can run it offline. It’s really something to see and really something that we haven’t seen before. So that’s really cool.

[00:08:28] Aaron Grando: I do think that the real impact of DeepSeek is not going to happen for at least another couple of weeks, probably more like a couple months as developers work to integrate features within DeepSeek into the apps and, you know, platforms that they’re building. So. Again, stay tuned for how that turns out over the course of the next couple of quarters.

[00:08:49] Aaron Grando: Overall, I give it four out of five stars. I’m just docking one star for those potential privacy issues with the free version of it. So, go try it out.

[00:08:59] Patty Parobek: Hello, this is Patty Parobek, the VP of AI Transformation at Mod Op. I’m talking about DeepSeek today. Yes, I did test it. I was looking for, just like anybody testing DeepSeek right now, if its capabilities were better than or greater than ChatGPT or Claude or Perplexity or Gemini or any of our favorite LLM chatbots and frontier models that have become a regular part of our workflows.

[00:09:35] Patty Parobek: And here’s what I found. I’m a marketer. I’m a strategist. I’m a consultant. So these are the types of ways that I love to use ChatGPT, OpenAI, which is my favorite, go-to model right now. And I tested many queries, many prompts against ChatGPT 4.0 and the prompts for things that I would typically ask for for personalized or better ideation with strategy development or research.

[00:10:14] Patty Parobek: So, for example. You know, act as a senior strategist or act as a marketing research professional and expert examining this landscape or examining this product line or this customer segment and create for me a priority X, Y, and Z. So, those are the types of prompts that I love for exploratory research and finding customer segments. And I hadn’t thought of, especially if it’s an industry that I’m not, very advanced in.

[00:10:53] Patty Parobek: And. When I tested those types of queries, those types of prompts between DeepSeek and ChatGPT 4.0, I didn’t find a remarkable material difference between the two. In fact, there were some that were very nearly identical, which makes sense to me logically if the rumors are true that DeepSeek trained a lot of its model on the outputs of, , open AI, but it was really sound.

[00:11:29] Patty Parobek: And I know that there’s a lot of buzz around price efficiency and speed and things like that when it comes to DeepSeek and speed of development. But I got to tell you, the cost is not just financial. When you’re thinking about using a model like DeepSeek, you have to think about the risks and I just I have to read you some of the risks so that you’re fully aware.

[00:12:00] Patty Parobek: If you plan on testing DeepSeek, these are the things that you’re signing up for. Okay. So listen to this list. Number one, privacy issues, data privacy issues. DeepSeek collects extensive user data. So things like chat histories, search queries, your device information, keystroke patterns, I.P. addresses, uh, Internet activity, maybe from other applications.

[00:12:29] Patty Parobek: So there’s some serious potential for misuse here. And then when you think about data storage, data storage happens in China, which sparks some concerns around what, is the government going to access and are there surveillance risks then against the data that it’s correcting? It’s a particularly worrisome given where China’s data protection laws and government controls over companies that DeepSeek operates.

There’s unauthorized data transmission concerns, cybersecurity vulnerabilities, AI driven information. I think it was maybe, right, a week after the big headlines of DeepSeek were emerging that there was already breach concerns. So there’s a lot to process there. And unless if you’re bringing, if you’re thinking about testing this personally, you should know that anyway, don’t put anything in there that you don’t want that organization to have or don’t use it if you don’t want it studying your behavior for whatever purposes will come out of that.

[00:13:43] Patty Parobek: But professionally, oh, man, you know, really think about what you’d be giving up. By using that and understand if you’re willing to take those risks on behalf of your business or your organization, your staff and your clients, and then maybe think again, Do your research certainly look at the terms privacy policy, uh, run them against ChatGPT or your favorite frontier model choice, and have it explain what the risks are to you. Perplexity is my favorite go-to source when it comes to, hey, explain me the risks of this new tool or this new application, apart from just taking the straight-up terms and reading through them. And Perplexity is the one who gave me this list, so there you go.

[00:14:38] Patty Parobek: But when it comes to DeepSeek, yeah, it’s great. It’s a great model. Just like ChatGPT 4.0 is. I think that OpenAI’s next models will be even better. I think DeepSeek will launch something after that that will be even better. I think Claude and Gemini are in the same boat. I think it’s just up to us to continue to test for our roles for the workflows that are going to be best for us and keep a very close eye on security and risk appetite and make decisions responsibly. And for now, Open AI. In my workflows, especially enterprise instances, or where you can connect directly to the API and have it live in your environment. That’s the way to go for me.

[00:15:26] Patty Parobek: So just, you know, AI responsibly my friends.

[00:15:33] Patty Parobek: To wrap up, we always ask at the end of these reviews on a scale from 1 to 10, would you recommend it for broad use? For DeepSeek? No. If there was a 0, I would probably select that. I can’t assume the risk of broad use.

[00:15:50] Patty Parobek: For something like that, I just wouldn’t be able to take the chance that, I would be able to put enough controls and protections in place so that 400, 500 people would be able to use it responsibly. And I probably wouldn’t refer to it a friend for their same reason. I don’t want them uploading pictures of their children for whatever reason.

[00:16:11] Patty Parobek: And, um, and that’s that. So thank you for another spicy episode of AI Playground.

[00:16:22] Sasha Dookhoo: Hello, my name is Sasha Dookhoo and I’m a VP of PR at Mod Op. I recently had the opportunity to look at DeepSeek and thought it was a disruptive platform in the AI market. I would rate it a four out of five. DeepSeek delivers high performance AI capabilities at a fraction of the cost compared to competitors. The platform provides open source models allowing developers to access, modify, and adapt the code for various projects, providing extensive customization.

[00:16:54] Sasha Dookhoo: In terms of efficiency, DeepSeek’s models such as R1 offer impressive performance in tasks like coding and reasoning, often surpassing established LLMs. On the other hand, there are some data privacy concerns. User data is stored on servers located in China, raising potential security issues. DeepSeek’s chatbot is also programmed to avoid discussions on topics considered sensitive by Chinese authorities, which may limit its utility.

[00:17:24] Sasha Dookhoo: As DeepSeek is still developing its ecosystem, users may encounter integration challenges as well. Overall, DeepSeek presents a compelling AI platform that is worth exploring with information that is not sensitive in nature.

[00:17:39] Tessa Burg: That concludes this week’s episode. We hope you enjoy the latest from our AI playground. If you have any ideas for apps that we should review, email us at podcast at Mod Op. com that’s P O D C A S T at Mod Op M O D O P. com. Until next time, have a great week.