2025 Renaissance of AI Reasoning?
_______________________________________________________________________________________________________________________________________________________________
The sudden rise of AI reasoning models and systems?
On December 5th 2024 OpenAI o1 caused a stir by kicking off with its ability to spend time 'thinking' before giving you a response. It claims to go beyond capabilities of LLMs to a new era of thinking and reasoning systems to solve more complex problems. I was intrigued when it told me it took 7 seconds to think and then took more time to analyse and provide answers to my pretty detailed prompt about knowledge philosophy.
On December 11th 2024, Google announced its new 'Deep Research' feature, as a part of the new Gemini 2.0 platform.
Gemini Deep Research "will create an entire research plan before executing it and dramatically improve the quality of outputs for more complicated requests ". Google suggested that this new breed of reasoning systems will dispel growing fears of a 'black box' and growing concerns about hallucinations that are based on probabilistic AI systems. Reasoning systems replace this fear and uncertainty with both logical and creative approaches. This will provide step by step transparency that will certainly build trust in organisations using AI .
In the first week of January 2025, I was made aware of DeepSeekv3, an open source AI reasoning model from DeepSeek a Chinese startup.
At the moment it is entirely free to test, with 671 Billion parameters with an added 'DeepThink' capability that might make it a serious open source competitor to OpenAI o1 and Google Gemini 2.0. It claims to be "designed to handle very complex queries.
I asked DeepSeekv3 to apply DeepThink to my prompts and explain the method for how it was doing it and the results, step by step. DeepSeek replied that it involved breaking down the problem presented, explored various perspectives, and synthesized insights to provide a comprehensive response. DeepSeek said it approached the problem with the following reasoning steps:
Understanding the Core Concern ( understanding the context)
Gathering Information (and asking for more information needed)
Analysing the Nature of the Core Issues
Analysing Key Themes
Formulating and Exploring Different Perspectives
Refining the Perspectives
Proposing Practical Steps
Synthesising Insights
Providing a Comprehensive Response
Encouraging Further Exploration
Final Thoughts
DeepThink told me that it used:
Iterative Reflection: to revisit initial thoughts multiple times, refining them as it considered new angles of information.
Holistic Approach: to balance technical insights with ethical and societal considerations to ensure a well-rounded perspective.
Critical Analysis: to question assumptions, explore alternative viewpoints, and sought to understand the broader implications.
The results, described step by step, were simply extraordinary!
I cannot stop thinking about the implications, as a result of this new breed of reasoning and generative AI tools, from OpenAI, Google and DeepSeek in December 2024. They will significantly impact organisations with even more effective knowledge management and elevate human knowledge work to new dimensions! Already, OpenAI are talking about the release of OpenAI o3 during 2025 (apparently o2 was dropped to avoid copyright issues).
Finally, I then asked OpenAI o1, Google Gemini 2.0, DeepSeekv3, and then added Anthropic Claude 3.5 Sonnet, to each tell me about their thinking approach and methods for performing inductive, deductive and abductive reasoning. All models confirmed their ability to do all three and I will detail this claim and their results in a following blog post shortly.
My aim, in this 2025 series of blog posts and video's, is to demystify AI by discussing significant developments, opportunities and risks.
As at January 6th 2025, I would summarise the AI developments to date as:
The years 1956 to 2010 were the years of discovering and developing schools of both classical symbolic expert systems and neural networking AI models that acted in isolation from one another and with relatively slow computing power and data.
2010 to 2022 were the years of exponentially increasing computing power and big data for image and pattern recognition (ImageNet & AlexNet) and major step change achievements like Deep Mind's AlphaGo and AlphaFold, (domain specific AI developed to solve specific domain problems that started to combine logic and creative probability approaches)
2023/24 were clearly the formative years of generative AI through Large Learning Models.
Will 2025 be the year of 'Agentic AI' which claims to have the ability to autonomously make decisions, act, iterate and adapt through extraordinarily rapid learning, to pursue goals based on both their logical programming and /or learned objectives?
Will 2025/26 be the years of establishing organisational AI Reasoning Systems that integrate logic and reasoning with generative AI?
Please contact me below if you are interested in knowing more about receiving this new series of human knowledge and AI blog posts, and / or let me know if you have other key questions and issues to discuss.
Ron Young
7 January 2025