In case you have been in an undiscovered place far removed from the AI principalities and powers that have kept you oblivious of what took place on Monday January 20, when a meteor-like LLM based AI reasoning model descended with cataclysmic force for immediate global consumption. It instantly became the most downloaded and active model to date. That release shook the very core of investors riding high on the AI clouds along with consequences to its deserving crown prince Nvidia.
That “meteor” was DeepSeek-R1, developed by a once unknown 2023 Chinese startup by the same name, led by a recently unknown 40-year-old Liang Wenfeng and funded by a Chinese hedge fund called High-Flyer! While by November DeepSeek had teased out their V3, the kicker is that at the landing of the R1, it still felt completely stealth. No drama, no fuss, no restrictions, just boom! There it was and it is open to all!
So, what is DeepSeek-R1 and what gives it such profundity:
– It is an advanced LLM based AI model that can perform complex reasoning tasks in coding, mathematics and science etc. Reasoning models employ cognitive processes using logic, inference and problem-solving techniques to analyse information, make decisions and draw conclusions.
Until recently OpenAI’s o1 was regarded as the most famous reasoning model, adept at math, coding and scientific problem-solving. Though, Claude’s Sonnet 3.5 along with three Chinese companies (including DeepSeek, Alibaba and Kimi) have been known to point out o1 performance as their target.
It’s always about the cost, eventually!
It is kind of refreshing to be able to finally bring cost into the equation of technology development, something we have always had to consider, at least until this era of AI to AGI pursuit where hundreds of millions of dollars were now the starting point and billions and potentially trillions of dollars were entertained as potential development cost.
What makes this moment so profound is not that another company claims to rival o1 on key benchmarks like logical reasoning, scientific reasoning, coding skills and problem-solving skills; but that DeepSeek-R1 claims to do so with a much lower compute budget (pre train, post train and inference) and hence at a fraction of the cost. To put this in numbers, late 2024, Open AI closed a whopping $6.6B round, the largest ever single raise by a technology firm and were already (and still) in talks about raise of another $35B or so.
In contrast while we have no full data on how much DeepSeek has raised so far, they claim to have spent about $6 million to train their AI models that are comparable in performance to the Open AI’s. We all remember the publications claiming a whopping $100m price tag for the training of ChatGPT4 and about $200m for Gemini 1. The question before us is has DeepSeek succeeded in building a very good agentic AI in R1 at a much lower cost?
Let us think about that for a moment! Granted, coming behind an impressive o1 helps all emerging reasoning agents. Reviewing o1’s 6-point chain-of thought can be summarized at a high level as: PROBLEM ANALYSIS (Understand, interpret and reformulate), then TASK DECOMPOSITION (break into smaller subtasks), followed by SYSTEMATIC EXECUTION (using step-by-step solution) then ALTERNATIVE SOLUTIONS (explore good options) then SELF-EVALUATION (via results verification) and finally SELF CORRECTION (address and fix errors). While this word-based flow or sequence of activities might seem straightforward, the implementation is not trivial by any measure. To achieve the semblance of robustness and the breadth of quality reasoning an agent like o1 is capable of, requires much technical resource.
But despite that, the modality of achieving performance can be improved upon and be re-engineered to use less resources, more strategic use of key techniques, better alignment of reasoning chains to queries (i.e. instead of a one-size fits all approach; simple queries may not need elaborate reasoning chains, while structured-thinking methods target complex problems) and so much more. Activities that eventually culminate in a high-performance reasoning agent at a more conservative cost.
This appears to be the R1 story, a story that grows on us with each passing day, as more organizations large and small, free lancers, independents and corporates take hold of the o1 competitor as a comparable solution beautified by the cost savings it brings with it.
Cost and capital are at the top of the barrier to entry, access, or opportunity in nearly every framework known to humans.
Open = Open-Source AI
When OpenAI came on the scene, for many it was thought to be an open-source AI platform for its ChatGPT and GitHub Copilot platforms. But the reality quickly dawned on most that none of OpenAI’s technology is open source or available to the public. All the code behind the tools is closed and only viewable, modifiable and reuseable by its developers. In addition, while some of OpenAI’s tools are free to use, most of its products and services are fee based.
On the contrary, DeepSeek-R1 is a fully open-source software platform which allows developers outside DeepSeek, reuse, modify, deploy, optimize and offer improvements to the model for their use and in benefit to others. This was a bold move that will significantly contribute to the advancement of model building going forward.
While cost and capital are at the top of the barrier to entry, access, or opportunity in nearly every framework known to humans. Technology access or the lack of it can define the likelihood for large scale prosperity for large swathes of people or the lack thereof.
What do we look for when we Benchmarks Models – The answer is plenty, but here are some useful parameters:
Figure 1 – Source DeepSeek R1 research paper
Figure 2 – www.analyticsvidhya.com
From the graphs in Figure 1, it is notable that on the Tasks representing Logical Reasoning, Scientific Reasoning, Coding Skills and Problem-Solving Skills, the performance differential is not significant between R1 and o1. In other testing, using financial data of SPY investments (SPY is an Exchange Traded Fund / ETF that tracks the S&P 500), the SQL (Structured Query Language; used to manage and access databases) query generation for data analysis showed both R1 and o1 demonstrating high accuracy with excellent performance in generating algorithmic trading strategies. It is noteworthy that R1 showed some edge as it outperformed the S&P 500 by including superior Sharpe and Sortino ratios (these ratios are used to determine risk adjusted returns of an investment) compared to the market for better cost-efficiency. It is important to note that R1 also produced invalid SQL queries and timeouts though mitigated by its self-correcting logic.
The most significant most glaring differential though is in cost. According to the charts in Figure 2, DeepSeek-R1 input cost per million tokens is $0.14 compared to o1’s $7.50 and the output cost has R1 at $2.9 per million tokens and o1 at $60. Based on published numbers, R1 users on average will spend about 3.6% or less of what users spend on o1. The scale of cost disparity is not even funny!
The logical question that follows is …and so what?
This story is particularly interesting because it underscores the realization that some bright people can take a big problem, leverage existing information, and augment key techniques and complex models to be more efficient to achieve desired goals, and innovate on significant milestones within the confines of the limitations and challenges they face. Or simply stated, produce a lower-cost, high-performance, more-accessible solutions that can target applications formally relegated to existing very high-cost, high-touch technologies.
DeepSeek-R1 has introduced a robust AI reasoning agent with lower compute demands, lower infrastructure resource costs with freedom of access and simultaneously diminished the first-mover advantage of OpenAI with o1 and others.
It has unleashed accelerated adoption by so many developers and shifted the “center of cost gravity” from the “highfalutin” infrastructure to the “humbler” application layers defined by quality.
The out-of-the-box thinking, that leverage efficient lower cost architectures result in lower resource demand in compute power, hardware costs in GPU, CPU, RAM etc. and lower data center management and other infrastructure costs.
This is the classic disruption theory framework at play, where the cheaper, more efficient technology emerges as a real threat to a much higher ranked, much more expensive one already in implementation, simply by meeting market requirements, while being readily available to most.
Cost and efficiency remain the engines that drive the competitive advantages of accessibility, which in turn drives long term sustainability, and scalability:
Lower costs make advanced AI technology available to more companies, researchers, inventors, and users especially those with modest economics. Lower cost per user means innovation applications can be accessible to diverse user groups including the underserved at a modest cost per user. This is critical to long term sustainability and efficient scalability. The more communities that can gain access to a technology the better, the richer, the more versatile and the more representative that technology becomes. This is what DeepSeek-R1 is fulfilling with its arrival.
R1 broke the expectation that to do this all-important AI work, a raise of high digit hundreds of millions of dollars and even billions if you are o1 was without a doubt the only path to success. The need for an ever-increasing access to the deepest pockets, to hire engineers at any cost some at high seven-digit salaries, incur astronomical expenses in hardware, infrastructure and training data, was always justifiable. Regardless of the knowledge that such behavior culminates in an assurance that the resulting very high barrier to access, will ensure that populations that need the innovations the most will only see it happen to them, not for them.
R1 potentially makes it possible that instead, the opportunity to participate and create solutions most impactful to diverse populations now exists.
The Bottom Line is People!
The truth is it will take some time to fully vet R1, but they have delivered to us a lot of promise and as an open source software (released under MIT open-source license) many will eventually contribute to make it better, fix the many kinks, and even close the many security gaps that will be discovered along the way.
R1 has lowered the barrier to entry to many, building a bridge rather than the chasms of participation wrapped in deep-pocked investors, well-heeled startups that speak in dollar numbers that ought to be reserved for long-standing blue-chip companies, expensive hardware, and training data.
That said, the panic sell-off on Nvidia was an over-reaction that underscores the fragility of the dependence on the prediction of analysts who often make no room for anything except what is obvious. The truth though is that Nvidia and its breath of technology will play many major roles in the development and advancement of the AI industry.
Amid all this, the biggest phenomenon is the mobility of users. Within a week of R1’s announcement without hesitation uncountable users jumped on board, eager to use, learn and work with R1. They were less interested in where it originated, who funded it or who created it. This demonstrates that the true “holy grail” for technology is “build it with inclusivity” and they will come.
It is ironic that it comes at a time that the US is choosing to deepen inequity by doing away with Diversity, Equity, and Inclusion popularly known as DEI (a program that has especially helped white women gain deep inroads into large corporations and industry). It is even more ironic that China is the pioneer of this inclusive technology strategy!
It bears repeating that cost and capital are at the top of the barrier to entry, access, or opportunity in nearly every framework known to humans. Technology access or the lack of it can define the likelihood for large scale prosperity for large swathes of people or the lack thereof.
Finally, let me close out with a shout out to all the other teams of bright people out there stealthily building the most ingenious and amazing technology solutions yet to be released, know this – you will have the global reception to give the world a significant and important leap forward!
About Ngozi Obikwere
Ngozi Obikwere has both private and public sector experience from leadership to practitioner, and developer to end-user. She has participated in settings ranging from multinationals to startups. She resides with her family and is continuously encouraged by the inexhaustible possibilities made possible by God Almighty!