AI Future: DeepSeek Takeover

To help you understand the WTF of AI and future tech we've got a DeepSeek takeover in this issue (they've had quite the week). Let's go deeper.

This week:

😱 AI Goes Global: New LLM In Town

💭 AI Made Simple: How To Use It

✨ DeepSeek: Image Model

🔥 AI Support: Get An Agent

💡 Bad AI: Same Old?

AI GOES BIG: New LLM In Town

​

We talked about DeepSeek's new AI model a few issues ago (WTFIGOWAI readers 1, others 0), they've been having quite the moment since then.

​

They launched their latest AI model R1 rivalling OpenAI's ChatGPT, Anthropic's Claude and Meta's Llama models - but at a fraction of the cost and computational resources.

​

This development has prompted a reevaluation of AI development strategies globally and at one point this week dropped the US stock market by $1 trillion.

​

Here's what you need to know:

​

• Similar quality and outperforming leading models for less cost - and it's Open Source.
• Overtook ChatGPT to become the top free app on Apple's App Store.
• Shifted perception that smaller AI companies couldn't compete with tech giants, disproving OpenAI CEO Sam Altman's assertion that such competition was "hopeless".
• DeepSeek's breakthrough highlights the potential for more cost-effective and efficient AI development, potentially accelerating widespread adoption of AI technologies.
• Concerns remain about privacy, security and government influence.

​
Let's dig into what it does and what the challenges are.

DEEPSEEK: Made Simple

​

The DeepSeek web version has the familiar interface of other AI models. you can have a look at it here.

​

What's in the (web) box?

​

• Standard input: The default version is V3 you get solid outputs from the normal model without reasoning or web search.
• DeepThink (R1): This is where a lot of the hoo-haa has been, as its up there with the other frontier reasoning models - like o1 from OpenAI - it also shows you its thinking in a very 'human-esque' way. (sometimes the reasoning is better than the answers imo).
• Search: Use this to extend answers out to more real-time information - Deep Seek often says it has considered 40 or 50 sources - tbc on this.
• Upload clip: For documents or images - (no multimodal image output here, yet).

​

Outputs

DeepSeek-R1 delivers comparable results to other AI models, with its DeepThink mode performing at a level similar to o1. However, the model operates under Chinese government supervision, meaning certain topics - especially politically sensitive ones - may either be briefly displayed and then deleted or met with a "Not able to talk about that" response.

​

What It Means for Business

DeepSeek shot up the app store charts, so this is getting traction with the public. AI search engine Perplexity has included R1 in its model outputs and Meta has created a war room to combat it's influence.

​

Organisations should prepare for a changing competitive landscape, where pricing, access, and AI governance will look very different by the end of 2025. Talk to us about how to navigate it.

DEEPSEEK IMAGE: Here Comes Vision

​

As we said, DeepSeek are having quite the moment - they've also just launched an updated image model, moving them into multi-modal territory.

​

Here's what you need to know:

​

• Janus-Pro-7B, the larger of their models out competes established image generators like Stable Diffusion and DALL-E.
• It's open sourced under an MIT license, so it's viable as an alternative to tools like Midjourney (check T&Cs).
• Smaller models are limited to producing 384 x 384 images, which is a bit of a downside.
• Available from Hugging Face to download for now.

​

What It Means for Business

This adds more heft to the Open Source image model area, lowering costs (and potentially) energy use - making AI-generated images more accessible for businesses. However, Keep in mind privacy and copyright challenges - outputs of models have differing ownership and rights.

​

​Talk to us about how to incorporate AI into your organisational workflow.

AI SUPPORT: Get An Agent

​

AI agents are systems that understand, plan, and execute tasks to achieve specific goals. Here's what you need to know:

​

• They’re versatile: AI agents can handle tasks like customer service, scheduling, data analysis, and more.
• Efficiency wins: By automating routine workflows, AI agents free up human teams for higher-value work.
• Make a plan: Early adopters are using agents to reduce costs, and streamline operations, but implementation requires careful planning.
• Collaboration is key: Success with current agents requires aligning AI capabilities with your team’s strengths.

​
We help organisations integrate AI agents effectively into their workflow - while keeping teams in the loop to mitigate risk. Get in touch to find out more.

BAD AI: Same Old Same Old?

So, does DeepSeek represent a fundamental leap in AI progress?

​

While it has disrupted traditional AI assumptions, it hasn’t necessarily solved all the challenges like hallucinations, reliability, or inference costs, say some critics. Instead, we could see it as an economic and geopolitical shift rather than a technological revolution.

​

Some thoughts on these challenges:

​

• DeepSeek-R1 isn't actually smarter, it's just trained more cheaply​
  → The model hasn’t fundamentally improved reasoning or accuracy but has optimised efficiency, making training less expensive.
• It hasn’t solved hallucinations or reliability​
  → Like its Western counterparts, DeepSeek-R1 still struggles with misinformation and lacks the reliability needed for mission-critical tasks.
• It's restricted by government​
  → Certain topics (like Tiananmen Square) are out of bounds - what does that mean for transparency, ethics and bias?
• Inference (running the model) is still costly​
  → Even if training is cheaper, real-world use could still cost, particularly when extending AI’s "thinking" capabilities (i.e. longer run time).

​

Context

The real disruption may be economic, not technological. DeepSeek's advances will pressure OpenAI, Anthropic, Meta and the rest into a price war, undercutting their profitability - but leading to wider access. The whole 'bigger is better' bets being made by the US in particular ($500 billion project Stargate for one) will come under more scrutiny.

​

What it means for businesses

Whatever happens, this has thrown a big spanner in the US-centric AI landscape and 'brute-force' AI. This should result in more access to high-end models for less money. Test, learn and contact us about your AI initiatives.

WHATS NEW - DeepSeek Redux

​

> Perplexity can search with R1​

> Meta Deep Seek war room​

> The big bet on reinforcement learning ​

> Cyer attack halts registrations​

> Artificial integrity and your data​

G3NR8 is an impact consultancy for the exponential age. We help organisations figure out the highest impact areas to use AI for maximum impact.

​

Work with us to:

• Develop your AI strategy and policies.
• Create a competitive advantage using your proprietary data.
• Automate repetitive tasks.
• Gain actionable insight from data.
• Train your teams - we've trained 120+ large organisations in advanced generative principles and implementation.

​

​Drop us a line for a chat about productivity gains, cost savings and how AI can power growth in your business.

​