We are in Q1 2026, nearly twenty-five quarters since I last made a blog entry. The tech world has completely changed post-COVID – with the advent of the new world of deep learning/LLMs, AI Agents and Robotics. I have watched this evolution as an active embedded observer, after having been “doing AI” in some shape or form for the past three decades. Finally, there is some overall coherence (would not call it real maturity yet) as to where things are headed in this crazy world of AI. I thought I would kick off a new thread of blogs, firstly, with an assessment of where is the new AI/Robotics headed and what does GOFAI (Good old-fashioned AI) bring to this journey. First, a few observations to set the stage for further discussion, organized along the following themes – a) Why the AI tech explosion now? b) What is the technical maturity of the latest AI tech? and c) How does GenAI adoption look thus far from multiple viewpoints?
Why the AI explosion now? Primarily a few key technology, business and consumer behavior trends have converged to drive this phenomena. Key triggers are:
- End users (consumers, engineers, technocrats) accustomed to and comfortable with “uncertainity” – uncertainity in systems behavior/uncertainity in “answers” – such as search, uncertainity because there is so much information – it is only fair etc. Also with the notion that uncertainity can be overcome with iteration towards some goal – try again (and again) looking for what you want (it is only compute!). Finally with the notion that failing fast and failing is ok in these systems (mission critical or otherwise), the bar on system stability, reliability and more are much lower than the past in both consumer and business markets. Overall the notion of “reliability” is being upended.
- Current LLM-driven AI is completely dependent on “costly” compute and storage – GPUs, CPUs and memory chips – so it is a boon for semi-conductor companies. Investments in AI related HW companies, energy generation companies – startups and mature players is beyond normal measures.
- The software development stack has matured quite well over the past decade. Building and managing distributed systems at scale, compute at the centre and at the edge is well-understood, tooling to build and manage these systems is well developed. Data center management, VM management, cyber security and cloud adoption is reasonably mature. Open-source systems (for storage, compute, search, pipeline management) provide reasonable starting points to build big data systems at reasonable scale and experiment based on workloads. We have reached an inflection point looking for the next frontier of SW development to reduce the costs overall – Can we scale/automate coding and maintenance?
- Finally, data scale, maturity of data science and deep learning lend well to the R&D required for building the next generation of AI systems (both HW(robotics) and SW.
- Businesses are looking for the next framework/tool to reduce their overall operating costs – people and processes. Any reasonable and reliable level of automation and intelligence would help in this effort and are eager to try the latest AI technologies in the real world. Recent lay-offs for the past 18-24 months posit the advance of AI as a reason. However, this seems to be a good time to reset the people and skilll resources looking ahead for an AI world – whatever that may be.
What is the overall maturity of current AI technology? Would a “real” aka old-school engineer building reliable systems use this “AI technology” as a subsystem?
Since 2022, the LLM-based technologies continue to evolve on many different dimensions. Industrial large-scale adoption is also being promoted at scale, and businesses exploring various “use-cases” of this technology. Enormous investments have been made in scaling infrastructure – GPUs. memory, compute and data centers to allow the large-scale experimentation by technology companies, consumers and businesses alike. The whole tech world is functioning like a big global lab without a clear signal yet that the technology is as reliable as your every day IC engine in your automobile or the motor in your washing machine. In summary,
- AI technology as of today is still a work-in-progress. Many open threads of work still remain to develop a reliable, robust technology sub-system or component. Issues with AI have been addressed elsewhere – safety, alignment, ethics, hallucinations and more. Many philosophical engineering and systems questions still remain which I will outline briefly further below.
- Building a system with AI introduces various complexities, things that traditional engineering had taken for granted based on the “embedded” determinism in the system-being built. Developing processes to “extract” determinism from a non-deterministic/stochastic substrate is an extremely difficult problem and computationally costly.
- LLM-based systems and Agentic AI address the same problem of building distributed systems with intelligence of the past (circa early 2000s) – the tooling has vastly improved enabling everyone to participate in this global experiment. One-off “claimed” successes do not make a reliable system. We do not know all the “glue” and human tooling required to get LLM-based systems into production.
- Automation implicitly depends on determinism and repetition. Variation in the underlying system – is either designed out or the system handles the inherent variation by bucketing behaviors. LLM-based systems go against this basic principle – leading to major system complexity in maintainenance and debugging. What happens and what to do when something fails? Playbooks to maintain these systems are yet to be developed. Current solutions involve a human in the loop or trying to create the most appropriate dataset to “fix” the unwanted behavior.
- Though the overall idea is to minimize human involvement in maintaining these systems, when such intervention is required, the expertise called for is orders of magnitude more complec – you need an expert to fix the issue – rather than a regular engineer. If the expert was not invoved in building the system or the available knowledge is not openly available, it is highly risky.
How does GenAI adoption look ? What is the real-storyline? Many claims of success are being promoted on social media, early boosters of this technology in big tech, the startup world and businesses and elsewhere are committed to this technologies success, ongoing large-scale investment for AI seems to suggest one may be missing something etc. Here are some early indicators –
- Content management – especially english content – textual is reasonably mature for summarization. Search – where you go through every hit – can now combine multiple hits into a reasonable summary. However, content validation still requires a human in the loop. One can see this in action in all the tools offered by Google in its application stack. Image, audio and video management is still under progress but reasonably good in a large number of use-cases. Semantic validation (aka alignment) is an open issue. Given that the “chat” interface has become norm de jour for interacting with AI systems and with easy access to “human knowledge”, the whole role of the education system is being questioned.
- Content generation – Getting a basic template fleshed out for all standard docs is possible to get started, Multiple iterations are required to clean the strawman up to make the content contextually relevant. Video and audio generation, translation, transcription, speech-to-text and text-to-speech are continually progressing – finding the right match of a LLM model/version for your use-case is the difficult part. The next rev of the model may not match! – so overall improvements are not guaranteed to be monotonic. You may need to redo the whole thing again.
- SW Coding – building off of open-source and private repos – LLM-based coding aims to speed up the coding lifecycle. Standardized patterns of UI, process motifs (mobile app, web app etc), data retrieval motifs and underlying libraries are utilized via prompting to converge iteratively to a working codebase. Tools such as ClaudeCode, Codex, Cursor speed up this process and it feels quite empowering for an engineer/business analysis with no programming experience to generate code. Though useful at some level for prototyping etc, how much of this code would be deployed in a production system, would be maintainable etc is up for debate. Furthermore, this assumes all necessary “coding motifs” already exist in the repos and no new ones are required – which is a very strong assumption. Overall folks claim productivity improvements and adoption but a fully built piece of commercial software based on “vibe” coding is yet to launched.
- Data Analysis- one of the big use cases for modern AI technology is data analysis across different formats, structured/unstructured, qualitative/quantitative, text/numerics/images/videos in different domains. Though there are many claims of success, very few fully documented case studies exist where the domain expert is not facilitating or interpreting the intermediary and final results to coordinate the overall process. This includes the efforts on building automated scientists in a domain and more.
- Process Coordination – Workflow execution and orchestration – enterprise AI and domain-specific tasks critically rely on this. There are numerous variations in such systems – the type of org, their data models and existing systems, decision points and human-interface points, how the overall process is triggered and run etc. A variety of “agentic platforms” aim to provide the toolkit to do this. Much of it is work-in-progress with ad hoc implementations across orgs. A clear story on adoption and successful, scalable deployment is still yet to be seen.
- Robotics and related applications – The whole world is seen as a collection of sequence of symbols – which could be text or actions or signals etc. which correspond to “things” in the real world. Capturing this set of “regularities” – in a statistical sense and stitching together actions that guide robots is an open area of research attracting a lot of funding. Most existing robotic systems still rely on traditional reasoning techniques.
Overall, every aspect of this AI technology trend seems to promise a lot but to realize those promises is going to take a lot of work on multiple dimensions. Many questions need to be answered from multiple perspectives before society at large is comfortable with AI-embedded systems and products. All the above discussion needs to be viewed in the context of seemingly simple questions but address deeper philosophical viewpoints from multiple perspectives. Most technological innovations usually have few stakeholders whereas the current AI juggernaut has the whole society as its stake-holder – making this more than a technology issue. Many questions are yet to be answered with some level of clarity such as the following –
- Is the past a “good” enough representation of the future? All of data-driven learning based systems implicitly assume this – Is this true? To what extent is this true? What happens in domains when/where it is not? How/when do we develop new solutions for new problems – instead of “recycling” old solutions? Even adapting old solutions piece-meal into a new one requires knowledge that is outside of the system. We assume that solutions from the past are good enough for future solutions – when and where is this true?
- What does “innovation” mean? The notion of discovering or inventing something needs to be re-thought.
- What is the role of the experts in a domain ? As automated systems come into play and experts fade away – what would be the new world? Where would expertise come from? What role would humans play?
- The economics of AI across the spectrum is still evolving – we yet do not know who/what/how will be affected. Various entities are taking actions – short term and long term – much is yet to be known. The world is a big petri-dish – many concurrent experiments underway.
- Given the far reach implications of this technology – philosophical concepts such as “safety”, risk, bias, alignment etc . have been propounded and attempts at operationalizing them (including quantifying them) – in the context of building AI systems is still underway.
- Most of the LLM systems and foundational models rely on publicly available however copyrighted data. The whole framework around copyright, credit assignment and more is up for discussion. It is also unclear what are the boundaries of intellectual property.
- AI systems also impact the current notion of labour and expertise – SW engineering aka coding being only one of them. Across domains there are record keepers, documentors, analyzers, scribes, knowledge gatekeepers and more who are going to be affected. Both knowledge work and physical labor are going to see a major change when these AI-based systems work reliably else some combination of system plus people is needed. Many of these things are yet to evolve at scale though we have some examples – such as drones being controlled remotely, automated cabs and more.
- The above leads to a big one – what should the role of education be? How should it look? What should be taught? Who or what makes an expert etc. Our whole notion of knowledge/expertise and more are up for discussion.
Overall, depending on your mental model and outlook, the whole world looks like an opportunity – changes at every level of everyday existence. On the other hand, it is also extremely daunting, as this whole techno-driven worldview- affects every aspect of daily living. Also given the present day global geo-politics, resource and enviromental constraints, much of what the world will look like in the next few decades will be determined by how the promises and perils of modern day AI are realized.
Keep checking in periodically as I discuss related themes on this AI journey across the spectrum of AI technology, product development, adoption and more.