The Chatbot Era Is Over - What Samsung AI Forum 2026 Said About the Agentic Shift
The Agentic Shift: From Intelligence to Impact 30 September 2026 · Samsung Electronics Seocho Office · Keynotes + Track 1 (AI Technology) / Track 2 (AX Innovation)
On 30 September 2026, Samsung AI Forum 2026 took place at Samsung Electronics' Seocho office in Seoul. This year's theme, "The Agentic Shift: From Intelligence to Impact", was a declaration that the centre of gravity has moved - from how clever a model is, to what that intelligence actually delivers.
Opened by Vice Chairman and CEO Young Hyun Jun, the day ran four shared morning keynotes plus a panel, then split into two afternoon tracks - AI Technology and AX Innovation - for six more sessions. External speakers from OpenAI, AWS, Anthropic, Dell and the University of Washington shared the stage with Samsung Research, the MX Business, the RX Business Office and the AX/PI Center. One message ran through all of it: agents are no longer a demo. They are an operations problem.
1. From Chatbots to Agentic Workflows - Richard Ho (VP, Head of Hardware, OpenAI)
OpenAI's Richard Ho opened with data showing that the shape of adoption has changed. Measuring growth in weekly active enterprise Codex users since February 2026, engineering grew 5x - while marketing grew 26x, recruiting 41x, sales 41x and legal 108x. Each figure is measured from that function's own starting population. A tool that began life in the IDE is now bleeding into professional work of every kind.
He did not sell only optimism. Under the heading "Productivity still needs direct measurement", he cited a randomised controlled trial of 16 experienced developers across 246 tasks. Indexed task completion time for the AI-allowed group came in at 119 against a baseline of 100 - that is 19% slower, not faster. Perceived productivity and measured productivity can diverge.
Fitting for a hardware lead, he closed on power efficiency: peak mixed token throughput per kilowatt of rated accelerator package power, where GPT-OSS 120B reached 85,648 against a comparison system's 44,960, with DeepSeek R1 and Kimi K2.5 also benchmarked. The unit cost of intelligence, in the end, converges on tokens per watt.
2. From Individual Productivity to Enterprise-Wide Impact - David Green (AI Technology Leader, APJ, AWS)
AWS's David Green took on the familiar problem that pilots are easy and scale is hard. He walked through a knowledge architecture handling more than 250,000 signals across 57 countries and 39 industries while holding query latency to roughly 200ms, arguing that discovery, translation, embedding and interaction have to be designed as one pipeline rather than bolted together.
The legacy modernisation example landed hardest: a 1.5 million line legacy .NET estate where the core was converted in 10 hours and weeks of testing became days. His third takeaway doubles as a summary of the whole forum.
Biggest outcomes come from fundamental rethinking / retooling.
Enterprise-scale impact does not come from layering AI onto an existing process. It comes from redesigning the process itself.
3. Visual Reasoning Beyond Language - Ranjay Krishna (Professor, University of Washington)
Professor Ranjay Krishna opened with a deliberately provocative title: visual reasoning will be bigger than language reasoning. The core idea is simple. People draw an auxiliary line to solve a geometry problem - so give the model a sketchpad too.
In the demo, GPT-4o answered a geometry proof directly and got it wrong. Given a plotting library and prompted to generate the intermediate step - "draw line DE parallel to AC" - it reached the correct answer. He also showed Molmo2 tracking multiple objects through occlusion, and a pipeline that generates a depth map first, then reasons over it with perception tokens. Together they make the case that small models can still deliver high-performance robotics AI.
4. Next-Generation Agentic AI Architecture - Yoonhyung Kim (Corporate VP, Samsung Research)
Starting from the attention mechanism, Yoonhyung Kim argued that the ceiling is close. His second direction: architectures that compress context into far fewer tokens, or internalise it into model parameters outright. A small KV cache absorbs long context so the model can keep learning over more data - continual learning rather than a fixed window.
He then set out four reasons for on-device deployment - cost, latency, privacy and "always on" availability. An application should not have to reach the internet every time it generates a response.
5. Panel - Where Agentic AI Meets Physical AI
The morning closed with a panel moderated by Juan Carlos Niebles, VP at Samsung Research. Two observations stood out: the language-centric research community and the robotics community, long separate, are now merging; and demos already exist in which large LLMs and agentic LLM loops drive physical action models to control robots. The most interesting problems, the panel agreed, live at that intersection.
Track 1. AI Technology (DX Division)
6. Efficient AI: Toward Human-Level Efficiency - Inchul Hwang (Corporate EVP, MX Business)
After showing performance curves where large multimodal models already exceed human baselines on major benchmarks, Inchul Hwang inverted the question: so why does the brain do the same work on so much less? He set out five principles of efficient information processing, each adapted to severe resource constraints.
Retrieve only the memory you need - see a friend's face in a cafe and the name surfaces instantly; you do not search every memory you hold.
Attend to what is new - on the same commute every day, it is the newly opened shop that catches your eye, not the familiar signage.
Activate only the parts you need - even as you read this sentence, only around 1% of your neurons are firing actively.
Scale effort to difficulty - 2+2 is instant; 17x24 makes you pause.
Keep learning while in use - memorising a new colleague's name does not make you forget an old friend's.
Each principle translates directly into engineering work - RAG and reranking, KV cache compression, prompt caching, dynamic routing. The conclusion was a hybrid AI architecture: on-device and cloud models bound together through a plugin and proxy structure, so the best available model can be called dynamically depending on the task.
7. Samsung's On-Device AI Platform - Yunsu Lee (Corporate EVP, MX Business)
Yunsu Lee unpacked the platform's three layers - the Personal Data Engine (PDE), the AI models, and the agent. PDE is the data layer that consolidates everything generated across a user's experience and extracts what is meaningful from it; on top of that, on-device models interpret context; and at the summit the agent platform turns that context into actual behaviour. The line he left on his final slide sums the session up.
From a phone that knows me, to a phone that acts for me.
8. Physical AI - World Models and Robot Foundation Models
Timothy Hospedales, VP at Samsung Research, covered both the promise and the perils of world models. Starting from neuroscience showing that even rats hold a world model and imagine future paths, he flagged five under-studied areas: accuracy, efficiency, action data, inductive bias and long horizon. The arithmetic was the memorable part - a robot controlled at 10Hz planning 15 minutes ahead implies a recursion depth of 10,000, which is both inaccurate and far too slow. His proposed fix: move from flat planning to hierarchical planning.
Kris Hauser, Head of Lab at the RX Business Office, gave a sober read on robot foundation models. After walking through a three-layer RFM behaviour stack, his diagnosis was blunt.
General robot AI will be a long road. 2026 is like autonomous driving's 2016 - promise shown, but tech still early to scale.
Data, evaluation, simulation, fleet operations, supply chain and talent all have to be built in parallel - and he was explicit that Samsung's north star is customer-facing robot AI.
9. Continual Learning - AI That Grows, and Does Not Forget
Kangwook Lee of Samsung Research presented self-evolving AI. The premise: across an ecosystem of hundreds of millions of connected devices and millions of apps, no two people use them the same way. The learning signals he proposed are strikingly mundane - a delete, a cancel or an undo is a dissatisfaction signal, and intent can be read from behaviour without the user ever explaining it. As learning accumulates, it is not one feature that improves but the device's behaviour as a whole.
Juan Carlos Niebles took the opposite problem - forgetting. He introduced TrustMem, a memory agent that manages external memory so edits stay auditable, and Null-Basis LoRA, which learns new tasks in the null space of existing weights so prior knowledge is left intact.
Track 2. AX Innovation (DS Division)
10. Creating Business Value Through External AX - Ajaz Munsiff (CDO APJ, Dell Technologies)
Dell's Ajaz Munsiff shared a transformation journey with Dell itself as Customer Zero. Three levers - simplify, standardise, automate - clear technical debt first; IT is then run on standards and transparency; and only on top of that does an AI-Ready Foundation go up. On the data side, the Dell AI Data Platform maps the route "from data chaos to AI outcomes".
The most immediately usable slide was use-case prioritisation: plot candidates on business value (more, better, faster, cheaper) against feasibility (data, AI model, process, people, platform), and let the matrix set the investment order. Then the conclusion.
Build an end-to-end agentic workflow, not a single agent. And design Human-in-the-Loop explicitly.
11. Putting AI to Work Safely - Sangkeun Choe (Tech Lead, Anthropic)
Anthropic's Sangkeun Choe started with what agents are actually delivering: a 100,000-line C compiler written by 16 agents that builds Linux 6.9; Stanford's Biomni agent, built on Claude, running genome-wide analysis in 20 minutes rather than months; and a 67% increase in merged PRs per engineer per day at Anthropic after adopting Claude Code.
He then laid out three layers of defence for running agents safely in engineering and R&D. The point of all three is to escape a world where a human approves every single action.
Environment - an OS-enforced isolated workspace (sandbox), with no sites approved by default and an outbound gate that only reaches an approved list. That design alone cut approval prompts by roughly 84% in internal use.
Model - a screening model that checks risky actions. In internal evaluation it caught 83% of 52 deliberately planted risky actions while blocking only 0.4% of ordinary ones, keeping the productivity cost low.
Rules - a prior rule layer that lets safe actions through and escalates repeated blocked patterns to a person.
12. Toward AI-Native DS - Jaejoon Han (Master, AX/PI Center)
In the closing session, Jaejoon Han of Samsung's AX/PI Center defined the structural complexity of the semiconductor industry in three parts: domain complexity (materials, devices, process, equipment, inspection and physical properties interacting non-linearly), data heterogeneity (images, time series, text and logs coexisting in incompatible forms), and distributed context (knowledge scattered across MES, PLM, EDA and ERP).
His framing of the bottleneck: the weakest link caps overall productivity. On one side, the cost of producing ideas keeps rising - US research productivity is roughly 1/41 of its 1930s level, and sustaining Moore's Law now takes more than 18 times the researchers it did in the 1970s (Bloom et al., AER 2020).
On the other side sits harsher arithmetic: speeding up some tasks leaves the whole roughly where it was. Apply AI to 20% of the work and make that portion 4x faster, and total time is 80% + 20%/4 = 85% - a saving of just 15%. The non-automated stretches, review, approval and system entry, still dominate lead time. The conclusion: cut the cost of generating ideas and close the execution gaps, together.
His answer is the DS AI Frontier Ecosystem: domain agents for R&D, process and defect work; MCP-based standard connectors so each system is wired once and reused many times; and an ABSORB - CONNECT - PRODUCE - SHARE cycle that stitches previously unconnected systems and database knowledge into one fabric.
What We Take Away
Twelve separate talks, but from a practitioner's seat they reduce to four sentences.
1. Agent adoption has left engineering behind. The 108x figure for legal says the "developer tool" frame has already broken. Worth asking which functions in your own organisation AI has not reached yet.
2. Measure it; do not feel it. AI can feel faster while measuring 19% slower. Adoption impact has to be measured directly, before and after.
3. Redesign the process; do not bolt on a tool. AWS, Dell and Samsung DS all said the same thing. The biggest outcomes come from end-to-end workflow redesign, not from a single agent. Keep Jaejoon Han's arithmetic in mind - make 20% of the work 4x faster and the whole only shrinks 15%. Without finding the slowest stretch, any investment gets diluted.
4. Safety design is a precondition for speed. As Anthropic's case shows, only once isolation and screening are properly in place can agents run without a human approving every step. Safety is not the brake - it is what lets you use the accelerator.
Samsung AI Forum 2026 was held on 30 September 2026 at Samsung Electronics' Seocho office in Seoul.
All sessions are available to watch again on the official Samsung Developer YouTube channel. Images in this article are screen captures from the official live stream. Track 1. AI Technology: https://www.youtube.com/watch?v=Velq-bBfX-o Track 2. AX Innovation: https://www.youtube.com/watch?v=6qISsYkakK4 Official site: https://saif2026.com
Comments