Keep Pace with "DeepSeek Disruption"
Insights into Monday's DeepSeek announcement
As a lifelong technology practitioner, I have encountered countless technologies that “disrupted” the status quo at that time. Those experiences coupled with my background in Enterprise Architecture - helped evolve a number of “disruption principles” and patterns that apply regardless of the technology OR disruption. For this article I will focus on the following six “disruption principles” (not in any particular order)
Disruption is a constant
Proprietary technologies have a limited lifespan
Technical Debt is significantly higher for early/first adopters
Industry Standards are crucial in the long term
Trust but verify
Professional support isn’t optional
In this blog, I will focus on the recent DeepSeek R-1 announcement described in the following VentureBeat article - and provide my initial range of insights. As you will see, there really is a LOT to this announcement that needs to be considered.
Full disclosure
1. I have been waiting for “something” to disrupt the current state of the AI market - this is one of those somthing’s !!
2. No AI was used in creating this blog ! What was the Announcement?
In actuality - there are 2 core disruptions contained in the announcement
Cheaper Large Model Creation
DeepSeek - a Chinese AI startup company - just released a new “open source” model called DeepSeek-R1 - that is claimed to not only match OpenAI’s o1 capabilities but is approximately 90-95% more affordable than OpenAI’s o1 offering.
They did this by approaching the model creation in a different way - using “Reinforcement Learning” over Supervised fine-tuning (SFT) .
Less Processing Power
According to the article, the development of this model happened on significantly less AI hardware than what is being used today. The following are the estimates of the hardware typically used in these AI Model efforts
DeepSeek - The estimate is between 10,000 - 50,000 Nvidia GPUs
OpenAI / Google - Estimate of 500,000 NVidis GPUS’s
The 800lb Elephant in the Room
As the old expression goes - there is no such thing as a free lunch. So what isn’t immediately clear from this announcement include:
Can this information be “trusted”. Who was the information obtained from and exactly how accurate is the information. Getting more details on the expenses (or historical expenses) would provide much deeper insights.
How did an entire community of AI practitioners not experiment with this technique? I expect to see to see more articles in the coming weeks - that will give us answers to many of the questions asked here.
What are the sources of the underlying data used to build the model. Were any existing Large Language models leveraged in creating this model
Who will determine the overall viability of the model on an ongoing basis. For example, are there any biases built into the model (such as being able to answer about the Chinese government)
Who will maintain and support the model - and at what cost
How to “Keep Pace” going forward
Now that we have explored some of the basics of the DeepSeek announcement - I wanted to apply my principles to talk about the potential impact - and questions you can ask going forward:
Disruption is a constant - Technology breakthrough’s drive more breakthrough’s. Expect the techniques implemented by DeepSeek to be vetted in the coming weeks to verify/validate the reports.
Proprietary technologies have a limited lifespan - Most new technology disruptions come from a commercial entity that has some amount of financial backing. The challenge is that in the early days the costs of proprietary offerings are high (to recoup those investments). As we have seen, the initial AI cost models of proprietary models have been pricey. It is NOT uncommon to see Open Source solutions (such as DeepSeek) disrupt commercial models and help drive industry standards (resulting in lower costs)
Industry Standards are crucial in the long term - I talk frequently about this in my Webinars. While I realize that standards are usually few and far between in the early days of any new technology - AI standards are going to be critical to abstract solutions from the underlying AI implementations
Trust but Verify - What are the “best practices” (and/or products) to verify the safety of an AI Model / Implementation. Yes there are approaches (SalesForce Einstein for example) - but there is no defacto answer for this space (at least not yet)
Technical Debt is significantly higher for early/first adopters - In the absence of standards - early adopters may end up with more technical debt than they realize. This is where good Architects and/or loose coupling techniques are critical for any “proof of concept” activities
Professional Support - For companies implementing AI capabilities - they are going to need professional support. I haven’t seen much in the way of formalized support (Think RedHat as an example) - but I am sure something is coming
How to “Keep Pace” with me - My weekly AI Webinar
Every Sunday Morning (9AM EST) - Keep Pace Weekly Webinar on ChatGPT. A weekly 30-minute webinar focusing on that week’s key reading / insights.
The webinar streams LIVE on the following sites:
www.facebook.com/keeppace
www.twitter.com/keeppace
www.linkedin.com/in/keeppace
#ai #gpt #keeppace