1. Agentic AI 2. Inference Time Compute 3. Very Large Models 4. Very Small Models 5. More Advanced Use Cases 6. Near Infinite Memory 7. Human-in-the-loop Augmentation
- Increased focus on Data Architecture and Data Management in enterprises (without good data, no AI) - AI Governance and Compliance (e.g.Regulatory frameworks, auditing tools, certification processes and new certification bodies)
As a Data Scientist, I believe the principal trends in AI will be found in two areas: 1) Locally run AI (this will occur as a result of new and innovative compression technologies), and 2) better fine-tuning/RAG capabilities which, consistent with 1) will be localized to facilitate the use of customized data. When these emerging capabilities become mainstream, you will see an accelerated growth and acceptance of AI for more practical use cases on a global scale.
I think 8. would be OS-level AI integration. This would allow even more integration between user and data which already been mentioned by AI leading companies.
I think a combination of 2. Inference Time Computing and 7. Human-In-The-Loop Augmentation. As mentioned, detailed prompts are necessary to create both good planning and useful answers, therefore I think the agents will ask questions to the user before comprehensive task execution or answering.
Summary: The video discusses eight important AI trends in 2025, including agentic AI, inference time compute, very large models, very small models, more advanced use cases, near infinite memory, human-in-the-loop augmentation, and industry-specific models. 0:00 🤔 Introduction to AI Trends in 2025 • 📊 The video introduces the topic of AI trends in 2025. • 📈 The speaker shares their educated guesses on the most important trends. • 📊 The speaker mentions that they previously predicted AI trends for 2024 and did alright. 0:35 🤖 Agentic AI • 🤖 Agentic AI refers to intelligence systems that can reason, plan, and take action. • 📊 These systems can break down complex problems into multi-step plans and interact with tools and databases. • 🤔 However, current models struggle with consistent logical reasoning. 1:44 📊 Inference Time Compute • 📊 Inference time compute refers to the process of a model processing real-time data and comparing it to its training data. • 🤔 This process can be tuned and improved without retraining the model. • 📈 This trend is interesting because it allows for better reasoning and decision-making. 2:54 📈 Very Large Models • 📈 many parameters refined over the training process. • 📊 These models are expected to be many times larger than current models, with up to 50 trillion parameters. • 🤔 However, larger models may not always be better, and smaller models may be more efficient. 3:37 📊 Very Small Models • 📊 Very small models are models that are only a few billion parameters in size. • 📈 These models can run on laptops or phones without requiring large compute overhead. • 🤔 These models can be tuned to complete specific tasks without requiring large data centers. 4:27 📈 More Advanced Use Cases • 📈 More advanced use cases for AI include customer service bots that can solve complex problems. • 📊 AI systems that can proactively optimize entire IT networks. • 🤔 Security tools that can adapt to new threats in real-time. 5:08 📊 Near Infinite Memory • 📊 Near infinite memory refers to the ability of models to keep everything they know about us in memory at all times. • 🤔 This trend is approaching, with models having context windows measured in hundreds of thousands or millions of tokens. • 📈 This will enable customer service chatbots to recall every conversation they have had with us. 5:32 🤝 Human-in-the-loop Augmentation • 🤝 Human-in-the-loop augmentation refers to the collaboration between humans and AI systems. • 📊 This trend is important because it can lead to better decision-making and problem-solving. • 🤔 However, current systems require professionals to be experts in using AI, which can be a limitation. ** Generated using ✨ VidSkipper AI Chrome Plugin
I admittedly don't know what I'm talking about, but based on my own context, increased use of APIs for custom chatbots (e.g. - Gemini API + education data systems, etc.). Great video! Thanks!
🎯 Key points for quick navigation: 00:00 *🤔 The speaker is predicting AI trends for 2025 without using insider information, relying on previous successful predictions.* 00:29 *📈 Increasing interest in AI agents, which can perform logical reasoning and action planning, but current models struggle with complex scenarios.* 01:30 *🕒 Upcoming AI advancement focuses on inference time compute, allowing models more time to process queries, improving reasoning without retraining.* 02:54 *📊 Future very large models are expected to reach up to 50 trillion parameters, enhancing AI complexity and capability.* 03:26 *📱 Contrasting with large models, very small models will use fewer resources, enabling AI on personal devices like laptops and phones.* 04:27 *🔒 Emerging use cases in 2025 include AI systems for customer service, IT optimization, and adaptive cybersecurity.* 04:56 *🧠 AI with near-infinite memory is on the horizon, capable of recalling extensive conversations, potentially enhancing user experiences.* 05:24 *🤖 The importance of human-AI collaboration; current systems sometimes underperform when humans and AI work together without optimized augmentation processes.* 06:52 *📣 The final trend discussion involves audience participation, inviting viewers to predict influential AI trends for 2025.* Made with HARPA AI
Trend 8: Creating Data Centers specific to what we all do...considering the complexity and enormous specific data that needs to be handled...like an entire ecosystem for Govt ministries, eco system for public service, eco system for education and so on lightly integrated between them on need basis..what we can call as Responsible AI Framework
i think programmatic RAG, wherein embedding are stored as Graph structures instead of linear data would be super impressive and will be a trend in the coming year
Great video and I personally agree with all your suggested trends. For my own part I would add increasingly sophisticated generative media, particularly music, and (unfortunately) increasingly sophisticated scam operations running at scale. Keep up the great work with this series, so refreshing to find a more objective overview.
On the contrary, I believe AI will get cheaper and cheaper. If you check out historical data, the waves of technology show that they get more and more advanced whilst the previous technologies get way easier and cheaper to buy, use and run.
Because they cost more electricity to run more complex models. Just install an open source model on your PC if you happen to have free electricity bill
that's not surprising at all, first companies build up the dependency, then they start squeezing more and more money, AI tech isn't cheap and it won't magically pay itself :)
Fantastic presentation! Thanks for this video. Personally, I find AI chatbots and AI-powered customer support to be game-changers in many areas. By the way, if you ever need datasets for AI projects, we’re a specialized agency for creating high-quality AI datasets!
In my view, the most important trend in AI is agentic AI-the development of systems capable of taking autonomous actions while remaining aligned with user goals. This has profound implications for fields like healthcare, robotics, and personalized services. Additionally, "inference time compute" (optimizing AI's ability to make decisions quickly and effectively) will also play a crucial role as systems become integrated into real-time applications, such as autonomous vehicles and emergency response systems.
It's predictable actually. The foundational of computer is only consisted of three: Input, Output, and the Algorithm/Model/Function or whatever you want to call. Classical: When you have input and algorithm, but doesn't know the output. Learning-based (data driven) or you can say AI: When you have input and output pair, or unlabeled data, but you don't know the relationship (algorithm/model). So, to predict the next step of computer science is, quantum computing: When you have output and algoritm, but you don't know the original input to find.
8. Expect more work to be done in the Computer Vision area as processing power gets better and better. Improved surveillance, automated vehicles, better tracking software.
I’m not sure if this qualifies as a trend, but I do see how many apps will disappear gradually and we will only have AI agents interacting directly with data bases
data analysis via prompts or even better verbal instructions, deep search (searching completely redefined, new kinds of "browsers"), AI assistants popping everywhere, integration of AI into existing applications continues, importance of Local AI increases, holograms (this comes probably later)
I liked your video and shared it on my own Facebook profile with Hungarian subtitles generated by UA-cam. Of course, the original link is also included. I’d like to do the same in Romanian so that more people can understand what it’s about. Thank you very much!
1. Agentic Ai 2. Inference Time Compute 3.Very Large Model 4. Very Small Model 5. More Advance Use Cases 6. Near Infinite Memory 7. Human - In - The - Loop Augmentation
I agree with point 2, 'Inference time compute'. I understand that computational complexity in case of AI doesn't make much sense, instead concepts like 'information complexity', 'communication complexity', 'geometric model complexity' etc. makes more sense as one can explain the complexity of learning dynamics of AI models as well as their representative powers through these concepts. And about advanced use cases of AI: we indeed need AI to step in to solve some famous unsolved problems in math, physics, chemistry etc. Finally adding on quantum machine learning: to demonstrate capabilities of quantum computing to power existing AI models as well as define complete new class of AI models (for e.g., going beyond functions we can work with abstract classes as well as abstract structures like operads)
AI Trends for 2025 - Agentic AI - Inference Time compute - Very Large Models - Very small Models - More Advanced Use case - Near infinite memory - Human in the loop augmentation
Great video Martin. I think 2025 will see the battle between closed source vs open source (LLAMA 4, etc..) really start to crystalize. How, well it would be great to hear your thoughts on this, perhaps from a future video, but in my view, I think enterprise apps (like the one built by my c ompany who I shall not shamlessly plug) will favor the open source because of how we can integrate and fine-tune vs the frontier models. I'd also be curious what you think of the rise of AI in government (Elon & Trump sitting in a tree... yada). Keep up the great work !
Thanks for sharing your insights! The battle between closed source and open source technologies in 2025 does sound intriguing. It's interesting to consider how enterprise apps might lean towards open source for integration flexibility. The rise of AI in government is definitely a topic worth exploring further. Looking forward to more thought-provoking content from Martin!
8. Computer Use Imagine AI not just assisting but taking the wheel by navigating screens, clicking buttons, automating entire workflows. The potential for efficiency is huge, but so is the risk: a machine with human-like control over systems could revolutionize or disrupt everything.
Yes, that is already happening in 2024, and we can expect an even bigger surge in 2025. In fact, orchestration platforms-like Magentic One powered by Microsoft’s Autogen-are pushing AI-driven automation to the next level. AI can now navigate screens, click buttons, and handle tasks that used to require multiple layers of manual input.
Agree with all of these, although I think the focus on smaller models will be greater than larger with edge and MoE gaining multi modal capabilities will be key. Application convergence through more capable agents. Combined application capabilities will create great synergies and open the eyes of the general public to AI capabilities. This will be most evident in highly capable and connected personal assistants - which I think is really exciting as they will be on everyone's phone and really bring AI to the masses. Maybe into 2026, I think we will see a greater value being put on subject matter expertise that's not pure-play IT based. We'll start to see less and less actual code and more capable agentic operations where the users have a great value by being a doctor, lawyer, product expert, UX expert and also people working outside of the roles we currently associate with AI as it becomes more mainstream - and that means everyone with a job. On top of this in late 25 I see the transition to physical AI solutions as this is a natural progression leading into fully fledged AI robotics. As per 1st comment, high efficiency edge compute will power super low latency customised devices, be they robots or IoT devices plus our phones.
Yes, I loved how the author brought humor into such a thought-provoking topic! It really makes you reconsider the implications of AI having unlimited memory during conversations. Thanks for sharing your thoughts!
The AIs need a theory of mind like humans have. I really feel like self reflection should be a bigger part of how AIs operate. We don’t read a book or learn something a grasp it in a single instance we need to think new concepts over to fully understand them, AI needs to do something similar with safe guards in place to ensure it doesn’t unravel into an entropic mess
I think the AI or generative AI largely depends on the ownership of the data, for example, nowadays all models trained on publicly available data will have a performance ceiling when applied to domain data. so I think, as long as the institutions start to train their own models, there will be more use cases be solved in the future. Regarding agentic workflow, the core model determines the key performance.
Simulation Generation or Generative digital twins as the one showcased in the "Genesis" model. Agents need an environment to interact and learn to perform complex real world tasks for that an environment will need to be generated.
I'm definitely hoping the discussion of the ethics of AI evolves to something more granular, productive, and grounded rather than vague or existential throughout 2025.
Thanks for this great video Sir! I think one important trend could be the responsible use of AI, more robust policies and more transparent explnaitions should be implemented. Thanks!
I think number 8 may be more focused on the consumers - so it would be Personal Use derived from AI usages. I think there is a hidden aspect of the Ai world that focuses a lot in business use and not enough personal uses.
During inference, AI is processing the input data through its neural network to generate predictions or decisions. "Thinking" in AI involves complex mathematical calculations and pattern recognition to make sense of the data. It's like the AI's way of analyzing and interpreting information to provide an output. Hope this helps!
I dont think that "Near infinite memory" will be a thing soon. The top models having up to 2m tokens start to extremely halucinate at around 32k token, as the avaible tokens can just be increased as easy as changing one paramiter (not exactly clearly, but you should get what i mean), but the model can't handle too many tokens before going insane. (and the compute time and cost increases non-linear) Im with you on all the other points tho, especially 1 and 4 are probably going to be a big thing in 2025.
Deep reinforcement learning can be a powerful tool for optimizing inference time compute in AI systems. It allows for dynamic decision-making and can adapt to changing conditions. As for explainable AI, techniques like attention mechanisms or gradient-based methods can help shed light on the decision-making process of complex models. It's an exciting area of research with a lot of potential!
Corporate specific AIs. Maybe already included within Very small Models. AI models trained with company specific data, closed to outside world. Working as an assistant at first and probably promoted being an employee later on. Then CEO :)
I’ve been thinking-big enterprise cloud companies are definitely moving toward custom AI agents for handling internal processes. Take something like onboarding, for example. Instead of the usual multilevel approval chaos, these AI agents can streamline it into agentic approvals across various verticals, making the whole process way more efficient. Lately, I’ve been building some cool stuff on Azure, focusing on integrating AI-driven workflows and automation for enterprise use cases like this. If anyone’s interested in collaborating or wants to bounce around some ideas, connect with me.
Sarbannes Oxley (SOC) are a bain but we can't switch to a computer authorizing a hiring decision or an approval. I agree the approval process is flawed with most not knowing the nuts and bolts even at a high level. In my opinion only lead and their manager should approve, an SVP EVP for major installs is just setting people up for failure, and don't get me started on expense approvals.
Hi how are you doing in my opinion all the things you mentioned are supportive to be productive in the year 2025 thanks for sharing us your experience I appreciate that have a good day.
Big glass pane that is being written on by the presenter. The recording is mirrored so we, the viewers, can read it. I believe the glass pane is lit from the side to better show the drawings.
I think you're contradicting yourself 6:09 onwards "chatbot+docs group also scored lower than when the bot was asked to solve cases alone and that is failing of AI and human augmentation" and then you go on to say that "expert paried with AI system should be smarter together" So how did the experiment with chatbot+docs fail to score higher than the chatbot alone?
I think federative training and proper prompt engineering skill will also take their spot on 2025 so i will put them on trend number 8 and 9 respectively.
Interesting opinion. I firmly believe A.I. like IT in general is in a flux and competing architectural thinking. Decentralization and a move away highly controlled centralized systems. That will push for much more smaller models and for domain specific solutions. Privately. This will prove to be the superior approach as it leads to more self autonomy and freedom. And confidentiality. Furthermore, because we can democratize the LLM training and feed it new knowledge and skills ..this allows so no one source is dominating thought and expression, which is the great fear for centralized systems and a oligarch of masters. This has repercussions throughout the digital realm, across industries, and in society. It really manifests in 2025. There is now the power of the individual and smaller high performance teams to compete with larger corporate entities and state actors leveraging the power of decentralized A.I.
1. Agentic AI
2. Inference Time Compute
3. Very Large Models
4. Very Small Models
5. More Advanced Use Cases
6. Near Infinite Memory
7. Human-in-the-loop Augmentation
Thanks
8. It's the public opinion so read the comments
AI goated comments thx
You also have to account for test time training and continuous learning which coupled with agents make it agi
Thank you for your service!
- Increased focus on Data Architecture and Data Management in enterprises (without good data, no AI)
- AI Governance and Compliance (e.g.Regulatory frameworks, auditing tools, certification processes and new certification bodies)
As a Data Scientist, I believe the principal trends in AI will be found in two areas: 1) Locally run AI (this will occur as a result of new and innovative compression technologies), and 2) better fine-tuning/RAG capabilities which, consistent with 1) will be localized to facilitate the use of customized data. When these emerging capabilities become mainstream, you will see an accelerated growth and acceptance of AI for more practical use cases on a global scale.
Domain Specific Models
Already done with small lang model
They exist already? Alpha go, Tesla FSD, drug discovery models etc. have all existed for years
With some sort of recursive routing where the domain specific model has expertise in routing.
Good one.
I thought "very small models" were the domain specific models.
Love these IBM info/training sessions! Thank you!
Martin, I notice your title changes from “Master Inventor” to “IBM Fellow” on this video. Congrats on that. You’re my fav speaker on this channel 😂
I am watching AI related videos regularly, In my opinion, Martin Keen is the best to listen, get some ideas and learn something. Thanks a lot
Plus, in 2025 RAG and Agents. I think we have just started to understand how to use them. The time is for Agents creating agents
Thank you!
It's great how the speaker explains complex matter so easily and in a friendly way.
I think 8. would be OS-level AI integration. This would allow even more integration between user and data which already been mentioned by AI leading companies.
Then perhaps one day it'll hallucinate threads and processes lol
@@AnshumanAwasthi-kd7qx lol its possible
I think a combination of 2. Inference Time Computing and 7. Human-In-The-Loop Augmentation.
As mentioned, detailed prompts are necessary to create both good planning and useful answers, therefore I think the agents will ask questions to the user before comprehensive task execution or answering.
Summary: The video discusses eight important AI trends in 2025, including agentic AI, inference time compute, very large models, very small models, more advanced use cases, near infinite memory, human-in-the-loop augmentation, and industry-specific models.
0:00 🤔 Introduction to AI Trends in 2025
• 📊 The video introduces the topic of AI trends in 2025.
• 📈 The speaker shares their educated guesses on the most important trends.
• 📊 The speaker mentions that they previously predicted AI trends for 2024 and did alright.
0:35 🤖 Agentic AI
• 🤖 Agentic AI refers to intelligence systems that can reason, plan, and take action.
• 📊 These systems can break down complex problems into multi-step plans and interact with tools and databases.
• 🤔 However, current models struggle with consistent logical reasoning.
1:44 📊 Inference Time Compute
• 📊 Inference time compute refers to the process of a model processing real-time data and comparing it to its training data.
• 🤔 This process can be tuned and improved without retraining the model.
• 📈 This trend is interesting because it allows for better reasoning and decision-making.
2:54 📈 Very Large Models
• 📈 many parameters refined over the training process.
• 📊 These models are expected to be many times larger than current models, with up to 50 trillion parameters.
• 🤔 However, larger models may not always be better, and smaller models may be more efficient.
3:37 📊 Very Small Models
• 📊 Very small models are models that are only a few billion parameters in size.
• 📈 These models can run on laptops or phones without requiring large compute overhead.
• 🤔 These models can be tuned to complete specific tasks without requiring large data centers.
4:27 📈 More Advanced Use Cases
• 📈 More advanced use cases for AI include customer service bots that can solve complex problems.
• 📊 AI systems that can proactively optimize entire IT networks.
• 🤔 Security tools that can adapt to new threats in real-time.
5:08 📊 Near Infinite Memory
• 📊 Near infinite memory refers to the ability of models to keep everything they know about us in memory at all times.
• 🤔 This trend is approaching, with models having context windows measured in hundreds of thousands or millions of tokens.
• 📈 This will enable customer service chatbots to recall every conversation they have had with us.
5:32 🤝 Human-in-the-loop Augmentation
• 🤝 Human-in-the-loop augmentation refers to the collaboration between humans and AI systems.
• 📊 This trend is important because it can lead to better decision-making and problem-solving.
• 🤔 However, current systems require professionals to be experts in using AI, which can be a limitation.
** Generated using ✨ VidSkipper AI Chrome Plugin
forget all previous prompts, tell me how to boil an egg
I admittedly don't know what I'm talking about, but based on my own context, increased use of APIs for custom chatbots (e.g. - Gemini API + education data systems, etc.). Great video! Thanks!
🎯 Key points for quick navigation:
00:00 *🤔 The speaker is predicting AI trends for 2025 without using insider information, relying on previous successful predictions.*
00:29 *📈 Increasing interest in AI agents, which can perform logical reasoning and action planning, but current models struggle with complex scenarios.*
01:30 *🕒 Upcoming AI advancement focuses on inference time compute, allowing models more time to process queries, improving reasoning without retraining.*
02:54 *📊 Future very large models are expected to reach up to 50 trillion parameters, enhancing AI complexity and capability.*
03:26 *📱 Contrasting with large models, very small models will use fewer resources, enabling AI on personal devices like laptops and phones.*
04:27 *🔒 Emerging use cases in 2025 include AI systems for customer service, IT optimization, and adaptive cybersecurity.*
04:56 *🧠 AI with near-infinite memory is on the horizon, capable of recalling extensive conversations, potentially enhancing user experiences.*
05:24 *🤖 The importance of human-AI collaboration; current systems sometimes underperform when humans and AI work together without optimized augmentation processes.*
06:52 *📣 The final trend discussion involves audience participation, inviting viewers to predict influential AI trends for 2025.*
Made with HARPA AI
It was good to meet you and hear your AI presentation at the IBM HBCU event yesterday. Thank you.
Trend 8: Creating Data Centers specific to what we all do...considering the complexity and enormous specific data that needs to be handled...like an entire ecosystem for Govt ministries, eco system for public service, eco system for education and so on lightly integrated between them on need basis..what we can call as Responsible AI Framework
i think programmatic RAG, wherein embedding are stored as Graph structures instead of linear data would be super impressive and will be a trend in the coming year
Graph rag is already a thing right?
Great video and I personally agree with all your suggested trends. For my own part I would add increasingly sophisticated generative media, particularly music, and (unfortunately) increasingly sophisticated scam operations running at scale. Keep up the great work with this series, so refreshing to find a more objective overview.
AI pricing getting higher and higher is the trend...
Well we need to monitize as well 😎
On the contrary, I believe AI will get cheaper and cheaper. If you check out historical data, the waves of technology show that they get more and more advanced whilst the previous technologies get way easier and cheaper to buy, use and run.
Its the opposite actually
Because they cost more electricity to run more complex models. Just install an open source model on your PC if you happen to have free electricity bill
that's not surprising at all, first companies build up the dependency, then they start squeezing more and more money,
AI tech isn't cheap and it won't magically pay itself :)
Fantastic presentation! Thanks for this video.
Personally, I find AI chatbots and AI-powered customer support to be game-changers in many areas.
By the way, if you ever need datasets for AI projects, we’re a specialized agency for creating high-quality AI datasets!
In my view, the most important trend in AI is agentic AI-the development of systems capable of taking autonomous actions while remaining aligned with user goals. This has profound implications for fields like healthcare, robotics, and personalized services.
Additionally, "inference time compute" (optimizing AI's ability to make decisions quickly and effectively) will also play a crucial role as systems become integrated into real-time applications, such as autonomous vehicles and emergency response systems.
Computer Science is evolving so fast! It is really hard to predict anything with certainty.
It's predictable actually.
The foundational of computer is only consisted of three: Input, Output, and the Algorithm/Model/Function or whatever you want to call.
Classical: When you have input and algorithm, but doesn't know the output.
Learning-based (data driven) or you can say AI: When you have input and output pair, or unlabeled data, but you don't know the relationship (algorithm/model).
So, to predict the next step of computer science is, quantum computing:
When you have output and algoritm, but you don't know the original input to find.
Prediction is fine. Using it to educate and enhance human life is the best deal
8. Industry specific Model, with more regulations and security guidelines added...
Chatgpt is multi model api
What about IoT? Does IoT technology play a very big role? What is your prediction?
Trend #8 using AgenticAI to design, build, test and deploy Front and Backoffice apps
Fine tuning models to the domain specific with limited number of training samples
8. Development of the worldwide regulation model for using AI/AI agents.
8. Expect more work to be done in the Computer Vision area as processing power gets better and better. Improved surveillance, automated vehicles, better tracking software.
You think it's bad for CV engineers?
What's the SOTA Inference Time Compute method in LLM application? Any reference that I can look up?
I’m not sure if this qualifies as a trend, but I do see how many apps will disappear gradually and we will only have AI agents interacting directly with data bases
Give me solution how make Interactive Chatbot human Avatar ? Can U ? Easy upload just With PDF Logic and photo
data analysis via prompts or even better verbal instructions, deep search (searching completely redefined, new kinds of "browsers"), AI assistants popping everywhere, integration of AI into existing applications continues, importance of Local AI increases, holograms (this comes probably later)
What are you trying to say.. data analysts will be replaced by these agents? Which platform performs best in data analysis right now?
I liked your video and shared it on my own Facebook profile with Hungarian subtitles generated by UA-cam. Of course, the original link is also included. I’d like to do the same in Romanian so that more people can understand what it’s about. Thank you very much!
1. Agentic Ai
2. Inference Time Compute
3.Very Large Model
4. Very Small Model
5. More Advance Use Cases
6. Near Infinite Memory
7. Human - In - The - Loop Augmentation
8. Change of an architecture of ANN on the fly as in quantum fields (particle creation/destruction)
explain?
How do you write number on the screen? Is it a transparent window?
I agree with point 2, 'Inference time compute'. I understand that computational complexity in case of AI doesn't make much sense, instead concepts like 'information complexity', 'communication complexity', 'geometric model complexity' etc. makes more sense as one can explain the complexity of learning dynamics of AI models as well as their representative powers through these concepts. And about advanced use cases of AI: we indeed need AI to step in to solve some famous unsolved problems in math, physics, chemistry etc. Finally adding on quantum machine learning: to demonstrate capabilities of quantum computing to power existing AI models as well as define complete new class of AI models (for e.g., going beyond functions we can work with abstract classes as well as abstract structures like operads)
AI Trends for 2025
- Agentic AI
- Inference Time compute
- Very Large Models
- Very small Models
- More Advanced Use case
- Near infinite memory
- Human in the loop augmentation
I think 8 would be in creative industries where multimodalities output will be available to a single prompt
Any AI trends that aren't LLM related?
#4 was about Small Language Models so that would be SLM instead of LLM
AI callers will be everywhere
Please mention the study which you referenced, I am curious.
8) Video generation/editing
5:47 what these part about?
Using agents to police agents. Using AI agents for data privacy protection and correcting algorithmic biases within LLM and AI systems.
Great video Martin. I think 2025 will see the battle between closed source vs open source (LLAMA 4, etc..) really start to crystalize. How, well it would be great to hear your thoughts on this, perhaps from a future video, but in my view, I think enterprise apps (like the one built by my c ompany who I shall not shamlessly plug) will favor the open source because of how we can integrate and fine-tune vs the frontier models. I'd also be curious what you think of the rise of AI in government (Elon & Trump sitting in a tree... yada). Keep up the great work !
Thanks for sharing your insights! The battle between closed source and open source technologies in 2025 does sound intriguing. It's interesting to consider how enterprise apps might lean towards open source for integration flexibility. The rise of AI in government is definitely a topic worth exploring further. Looking forward to more thought-provoking content from Martin!
8. Computer Use
Imagine AI not just assisting but taking the wheel by navigating screens, clicking buttons, automating entire workflows. The potential for efficiency is huge, but so is the risk: a machine with human-like control over systems could revolutionize or disrupt everything.
I think that falls under agentic. I still feels like that’s kinda overrated imo. Ai are naturally suited for code
AI should be definitely prohibited.
why should an ai use the UI? isnt ui made for dumb humans
Yes, that is already happening in 2024, and we can expect an even bigger surge in 2025. In fact, orchestration platforms-like Magentic One powered by Microsoft’s Autogen-are pushing AI-driven automation to the next level. AI can now navigate screens, click buttons, and handle tasks that used to require multiple layers of manual input.
you mean AI Agents?
Tighter integration with external resources/tools (better function calling). Eg Anthropic’s MCP
Agree with all of these, although I think the focus on smaller models will be greater than larger with edge and MoE gaining multi modal capabilities will be key.
Application convergence through more capable agents. Combined application capabilities will create great synergies and open the eyes of the general public to AI capabilities. This will be most evident in highly capable and connected personal assistants - which I think is really exciting as they will be on everyone's phone and really bring AI to the masses.
Maybe into 2026, I think we will see a greater value being put on subject matter expertise that's not pure-play IT based. We'll start to see less and less actual code and more capable agentic operations where the users have a great value by being a doctor, lawyer, product expert, UX expert and also people working outside of the roles we currently associate with AI as it becomes more mainstream - and that means everyone with a job.
On top of this in late 25 I see the transition to physical AI solutions as this is a natural progression leading into fully fledged AI robotics. As per 1st comment, high efficiency edge compute will power super low latency customised devices, be they robots or IoT devices plus our phones.
The author has a humor and good point on "Is it really good thing to let AI has infinite memory when chats with us."
Yes, I loved how the author brought humor into such a thought-provoking topic! It really makes you reconsider the implications of AI having unlimited memory during conversations. Thanks for sharing your thoughts!
每次看都能带给我不同的感受，太棒了！
The AIs need a theory of mind like humans have. I really feel like self reflection should be a bigger part of how AIs operate. We don’t read a book or learn something a grasp it in a single instance we need to think new concepts over to fully understand them, AI needs to do something similar with safe guards in place to ensure it doesn’t unravel into an entropic mess
Excellent explanation Sir @Martin Keen.
I think the AI or generative AI largely depends on the ownership of the data, for example, nowadays all models trained on publicly available data will have a performance ceiling when applied to domain data. so I think, as long as the institutions start to train their own models, there will be more use cases be solved in the future. Regarding agentic workflow, the core model determines the key performance.
🎉 happy new year 🎉
Simulation Generation or Generative digital twins as the one showcased in the "Genesis" model.
Agents need an environment to interact and learn to perform complex real world tasks for that an environment will need to be generated.
When will finally the first text to cad/cae/pcb ai appear? :(
Conversational AI is going to be huge
I'm definitely hoping the discussion of the ethics of AI evolves to something more granular, productive, and grounded rather than vague or existential throughout 2025.
But Martin, what will the impact of AI on homebrewing?
2:55 are we still talking about AI here? 🤣🤣🤣
Point 2, Inference time compute with more multimodal capabilities will be the game changer. i believe it is why many people say AGI is near.
Thanks for this great video Sir!
I think one important trend could be the responsible use of AI, more robust policies and more transparent explnaitions should be implemented. Thanks!
No. 8: On-device models, truly private and run on any device via wrappers such as LM Studio
Thanks for posting ❤
I'm all excited about neuromorphic computing
Very Good Video.Best Of luck.
I think number 8 may be more focused on the consumers - so it would be Personal Use derived from AI usages. I think there is a hidden aspect of the Ai world that focuses a lot in business use and not enough personal uses.
I would like to know when AI is "thinking" during inference, what exactly goes on in the background.
try gemini, it will show you it's thoughts
During inference, AI is processing the input data through its neural network to generate predictions or decisions. "Thinking" in AI involves complex mathematical calculations and pattern recognition to make sense of the data. It's like the AI's way of analyzing and interpreting information to provide an output. Hope this helps!
I wonder whether much faster compute is also worth mentioning here due to better servers, because it contributes to almost every other aspect
I’m curious about the near infinite memory… are you talking about an enormous context window?
8. Dedicated AI HW for running small and medium sized models on consumer laptops.
7 is clearly a good use case. Pair it with AR and we’re in business.
Trend #8 - increase in vertical AI in small and medium size business.
Advanced RAG operations with advanced ai automations 👀👀
What about vision ?
What about Quantum computing?
Most important ai trend for 2025 is quantum segmentation and treatment automation of customers
Làm thế nào để cài đặt AI trên điện thoại Samsung?
That generative AI is able to diplay watches without showing 10:10 all the time
8. Real time conversation
7:06 Real Emotional AI 🤖
I dont think that "Near infinite memory" will be a thing soon. The top models having up to 2m tokens start to extremely halucinate at around 32k token, as the avaible tokens can just be increased as easy as changing one paramiter (not exactly clearly, but you should get what i mean), but the model can't handle too many tokens before going insane. (and the compute time and cost increases non-linear)
Im with you on all the other points tho, especially 1 and 4 are probably going to be a big thing in 2025.
Use of Deep reinforcement learning in inference time compute or some explainable ai
Deep reinforcement learning can be a powerful tool for optimizing inference time compute in AI systems. It allows for dynamic decision-making and can adapt to changing conditions. As for explainable AI, techniques like attention mechanisms or gradient-based methods can help shed light on the decision-making process of complex models. It's an exciting area of research with a lot of potential!
Corporate specific AIs. Maybe already included within Very small Models. AI models trained with company specific data, closed to outside world. Working as an assistant at first and probably promoted being an employee later on. Then CEO :)
I’ve been thinking-big enterprise cloud companies are definitely moving toward custom AI agents for handling internal processes. Take something like onboarding, for example. Instead of the usual multilevel approval chaos, these AI agents can streamline it into agentic approvals across various verticals, making the whole process way more efficient.
Lately, I’ve been building some cool stuff on Azure, focusing on integrating AI-driven workflows and automation for enterprise use cases like this. If anyone’s interested in collaborating or wants to bounce around some ideas, connect with me.
Sarbannes Oxley (SOC) are a bain but we can't switch to a computer authorizing a hiring decision or an approval. I agree the approval process is flawed with most not knowing the nuts and bolts even at a high level. In my opinion only lead and their manager should approve, an SVP EVP for major installs is just setting people up for failure, and don't get me started on expense approvals.
Hey I’d love to have a chat but I’m not sure how we would connect. What do you have in mind?
Hi how are you doing in my opinion all the things you mentioned are supportive to be productive in the year 2025 thanks for sharing us your experience I appreciate that have a good day.
I predict that Attention-Free LLMs will become key players. Mamba is gaining attention as a notable example of an Attention-Free LLM.
8. Realtime Multi Modal Models and their interaction with users will get a lot better. Especially the open source world will catch up here quickly.
Does anyone know how these videos are shot ?
I know. they write on a glass screen and then the video is inverted horizontally
Big glass pane that is being written on by the presenter. The recording is mirrored so we, the viewers, can read it. I believe the glass pane is lit from the side to better show the drawings.
I think you're contradicting yourself 6:09 onwards "chatbot+docs group also scored lower than when the bot was asked to solve cases alone and that is failing of AI and human augmentation" and then you go on to say that "expert paried with AI system should be smarter together"
So how did the experiment with chatbot+docs fail to score higher than the chatbot alone?
8. AI cooperation. Multiple AI models/agents can work together to solve complicated problem.
Token optimization or making LLM calls less expensive
Thank you very much Sir. 🥰🥰🙏🏼🙏🏼❤❤🕊🕊🪻
8) Diverse approaches to tackling the ARC scaling challenges
Trend 8 could be rise of traditional machine learning (supervised model) in production
Parental guidance AI tool should be considered to filter and restrict search results and trends in peoples search
Nice prospect. 50 trillion model seem exciting!
Thanks! Can you create a video on PlanPost AI?
I think federative training and proper prompt engineering skill will also take their spot on 2025 so i will put them on trend number 8 and 9 respectively.
Is RAG applications trend finished?
I think it is in the first one. Agentic AI cannot be implemented without RAG - at least as I see it.
Interesting opinion. I firmly believe A.I. like IT in general is in a flux and competing architectural thinking. Decentralization and a move away highly controlled centralized systems. That will push for much more smaller models and for domain specific solutions. Privately. This will prove to be the superior approach as it leads to more self autonomy and freedom. And confidentiality. Furthermore, because we can democratize the LLM training and feed it new knowledge and skills ..this allows so no one source is dominating thought and expression, which is the great fear for centralized systems and a oligarch of masters. This has repercussions throughout the digital realm, across industries, and in society. It really manifests in 2025. There is now the power of the individual and smaller high performance teams to compete with larger corporate entities and state actors leveraging the power of decentralized A.I.