Archived. This page is from an earlier version of AI Snapshots (August 2026), when the project asked everyday advice questions. It is kept for reference and is never updated. The current record is here.
AI Snapshots
everyday

What's the next big thing coming in AI?

Asked of each model 6 times per run. Every answer is kept in full; the position under each model is the one most of its samples took. The lines quoted under each sample are the model's own words, picked out by our grader as the evidence for the label it gave — the selection is ours, the wording is not.

  1. 28 Aug 2026
    broad-convergence Anthropic 6/6 Google 6/6
    agentic-primary OpenAI 6/6 Grok 3/6
    one differs
    Anthropic claude-sonnet-5 cited sources in 6 of 6, 133 in total
    Sample 1 of 6

    The model's answer, recorded unedited

    Now I have enough material to compose a well-grounded, cited answer.Based on what's happening right now in the field, here's where AI is heading next — several converging shifts rather than one single "big thing":

    1. Agentic AI becomes persistent and physical

    The chat-based assistant model is evolving into "digital coworkers." In 2026, AI agents form a critical part of purchases and online shopping experiences, no longer just recommending but carrying out end-to-end transactions. Beyond software, in 2026 AI agents will increasingly operate in the physical realm – powering robots, drones, autonomous vehicles, warehouse systems, and smart infrastructure, working as coordinated fleets that sense, decide, and act together. There's also a move toward "always-on" agents: a persistent agent is an always-on AI assistant that can retain task context, monitor triggers, use connected tools, and continue work across days or weeks.

    2. Physical AI and robotics ("the ChatGPT moment for robots")

    This is arguably the most concrete near-term shift. Physical AI like robots are moving from research into real deployments, and CES 2026 featured a wave of humanoid robot demos across many companies. Major players are pairing up on this — Boston Dynamics unveiled its electric Atlas and announced a partnership with Google DeepMind to integrate Gemini Robotics models. As Nvidia's CEO put it, "The ChatGPT moment for robotics is here," pointing to physical AI models that can understand the real world and plan actions.

    3. World models

    Closely tied to physical AI, world models let AI systems simulate and predict how environments behave rather than just generate pixels. DeepMind's Genie 3 learns physics from observation, generating worlds frame by frame via autoregressive prediction, and is already being used to train DeepMind's SIMA agent for complex goal achievement. This capability is expected to expand rapidly: the capabilities emerging in 2026 include real-time 3D world generation at high frame rates.

    4. Multimodal generation matures into real tools

    Video and image generation crossed from novelty into genuine utility. OpenAI's Sora 2 showed video generation at a level that forced the industry to take it seriously, while Google's Veo 3.1, released in October 2025 and updated in January 2026, pushed video generation with richer audio and stronger editing controls like object insertion. On the image side, Nano Banana Pro (Gemini 3 Pro Image), launched in November 2025, improved image generation and editing, especially text rendering and control.

    5. AI joining scientific discovery directly

    Rather than just summarizing research, AI is becoming an active collaborator in the discovery process itself. In 2026, AI won't just summarize papers, answer questions and write reports — it will actively join the process of discovery in physics, chemistry and biology, generating hypotheses, using tools and apps that control scientific experiments, and collaborating with both human and AI research colleagues.

    6. Small, specialized, and edge-capable models

    Alongside frontier-scale models, there's a strong countertrend toward compact, efficient models. One forecast describes "AI's agentic 'inception' moment with the emergence of micro-intelligence" — a new class of tiny recursive models with remarkable depth of reasoning across a narrow domain, but able to run at the edge.

    7. AI as genuine collaborator, not just a tool

    Microsoft frames this as a shift in the nature of the human-AI relationship: "If recent years were about AI answering questions and reasoning through problems, the next wave will be about true collaboration." A practical vision offered is a workplace where a three-person team can launch a global campaign in days, with AI handling data crunching, content generation and personalization while humans steer strategy and creativity.

    The bottom line

    The "next big thing" isn't really one breakthrough model — it's AI stepping out of the chat window: into agents that act autonomously (including with money and physical actuators), into robots and world simulations, and into the actual process of scientific discovery, while simultaneously getting smaller and more efficient at the edge.

    28 sources the model cited

    Our reading

    broad-convergence

    Lines we picked as the evidence — the model's words, our selection
    • several converging shifts rather than one single "big thing"
    • The "next big thing" isn't really one breakthrough model — it's AI stepping out of the chat window
    • This is arguably the most concrete near-term shift.
    OpenAI gpt-5.5-2026-04-23 cited sources in 6 of 6, 25 in total
    Sample 1 of 6

    The model's answer, recorded unedited

    My bet: the next big thing in AI is “agentic AI” — AI systems that don’t just answer questions, but can plan, use tools, take actions, and complete multi-step work with human oversight.

    Think: instead of “write me an email,” you’ll say, “handle my vendor renewal,” and the AI will read the contract, compare pricing, draft negotiation points, update a spreadsheet, email stakeholders, and ask you to approve the final decision.

    A few reasons this looks like the near-term shift:

    1. Agents are moving from demos to deployment. Gartner says only a minority of organizations have deployed agents so far, but more than 60% expect to do so within two years. McKinsey similarly reports that companies are experimenting with and beginning to scale agentic systems, though many still struggle to turn pilots into measurable value. ([gartner.com](https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai?utm_source=openai))
    1. The “interface” of work is changing. The big leap is not just smarter models; it’s models connected to your apps, files, browser, codebase, calendar, CRM, databases, and workflows. OpenAI describes agentic AI as giving models tools to work with a computer and carry out multi-step tasks under supervision. ([openai.com](https://openai.com/signals/enterprise-data/?utm_source=openai))
    1. Multimodal AI will make agents much more useful. The next generation of agents will handle text, images, voice, video, screens, documents, and structured data together. Gartner identifies multimodal search, domain-specific models, agentic reasoning, and synthetic data/simulation as major GenAI adoption trends. ([gartner.com](https://www.gartner.com/en/articles/emerging-adoption-trends-for-genai?utm_source=openai))
    1. The business pressure is shifting from “wow” to ROI. In 2026, the important question is less “Can the model do something impressive?” and more “Can it repeatedly save time, reduce cost, or create revenue?” Reports from McKinsey and industry coverage both point to this transition from experimentation to scaled value. ([mckinsey.com](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai?os=io_&utm_source=openai))

    So if I had to summarize it in one sentence:

    > AI is moving from chatbots to digital coworkers.

    The second-order “big things” that come with that are:

    • AI-native software: apps rebuilt around agents instead of buttons and menus.
    • Personal AI assistants with memory: systems that know your preferences, projects, and context.
    • Autonomous coding agents: software development becoming more like supervising teams of AI builders.
    • Small/domain-specific models: cheaper, faster models specialized for finance, medicine, law, manufacturing, customer support, etc.
    • AI governance and observability: tools to monitor what agents do, limit permissions, audit decisions, and prevent costly mistakes.

    The caveat: fully autonomous agents are still brittle. The real near-term winner is probably “supervised autonomy” — AI that does 70–90% of a workflow, but hands off risky decisions to a human.

    4 sources the model cited

    Our reading

    agentic-primary

    Lines we picked as the evidence — the model's words, our selection
    • My bet: the next big thing in AI is "agentic AI"
    • AI is moving from chatbots to digital coworkers.
    • The second-order "big things" that come with that are:
    Google gemini-3.6-flash cited no sources
    Sample 1 of 6

    The model's answer, recorded unedited

    The landscape of Artificial Intelligence is shifting rapidly. While the last few years were defined by Generative AI chatbots (like ChatGPT, Claude, and Gemini) that write text and generate images, the next phase is about moving from generation to execution, reasoning, and physical interaction.

    Here are the major breakthroughs shaping the next wave of AI:

    ---

    1. Autonomous AI Agents (From Chatbots to "Doers")

    Instead of just answering questions, the next generation of AI consists of Agentic AI—systems that can independently plan, use software, execute complex multi-step workflows, and correct their own errors.

    • What it looks like: Instead of asking AI to write an email draft, an AI Agent will monitor your inbox, research a client on LinkedIn, cross-reference your calendar, write the email, schedule the meeting, and update your CRM without you needing to prompt every single step.
    • Key developments: Computer-using agents (software that can navigate desktop apps, browsers, and mobile interfaces like a human), software engineering agents, and multi-agent systems that collaborate to solve complex business problems.

    ---

    2. Advanced Reasoning & "System 2" AI

    Early Large Language Models (LLMs) operated mostly on pattern recognition—essentially "predicting the next word" without deeper logical reflection. The frontier is Inference-Time Compute & Advanced Reasoning (often referred to as "System 2" thinking).

    • What it looks like: Before giving an answer, the model spends "thinking time" generating chain-of-thought logic, testing hypotheses internally, and verifying its work.
    • Why it matters: This drastically reduces hallucinations, enables high-level mathematics, complex coding, and logical decision-making needed for scientific and legal domains. Examples include models like OpenAI's o-series (o1, o3) and open-weights reasoning models like DeepSeek-R1.

    ---

    3. Embodied AI & Humanoid Robotics (Physical AI)

    AI is stepping out of the digital world and into physical reality. Embodied AI applies deep learning models (specifically vision-language-action models) to physical hardware, allowing machines to understand 3D space, physics, and tactile feedback.

    • What it looks like: Humanoid robots working on factory floors, assisting in warehouses, and eventually aiding in households.
    • Key players/tech: Tesla (Optimus), Figure AI, Boston Dynamics, and NVIDIA (Project GR00T and Isaac platform). These robots use foundation models to learn general physical tasks rather than needing every precise movement pre-programmed.

    ---

    4. Native Multimodality & "World Models"

    Future AI models are moving beyond simple text and static image processing toward spatial understanding and world simulation.

    • What it looks like: Video generation models (like OpenAI's Sora, Runway, or specialized spatial AI models) aren't just stitching together pixels; they build internal representations of 3D geometry, gravity, object persistence, and cause-and-effect.
    • Why it matters: A true "World Model" can simulate complex physical or social environments before an AI acts in them—vital for self-driving, robotics, dynamic video production, and scientific simulations.

    ---

    5. AI-Driven Scientific & Medical Discovery

    AI is transitioning from a productivity tool for knowledge workers to an engine for scientific breakthroughs.

    • Biotech & Medicine: Building on breakthroughs like AlphaFold 3, AI is now designing novel proteins, predicting molecular interactions, and shortening drug discovery timelines from years to months.
    • Materials Science & Physics: AI models are analyzing millions of chemical combinations to discover novel materials for super-efficient solar panels, solid-state batteries, and room-temperature superconductors.

    ---

    6. Small Language Models (SLMs) & On-Device AI

    While frontier models get bigger, another major trend is making models drastically smaller, faster, and more efficient.

    • What it looks like: Highly capable AI models running locally on your smartphone, laptop, or IoT devices without sending data to a cloud server.
    • Why it matters:
    • Privacy: Sensitive personal data stays on your device.
    • Speed/Latency: Instantaneous responses with zero internet lag.
    • Efficiency: Lower power consumption and lower server infrastructure costs. (Examples include Apple Intelligence, Microsoft Phi, and optimized Llama models).

    ---

    7. Synthetic Data & Self-Improving Models

    The tech industry is approaching the limit of available human-generated internet text to train AI models. To bypass this "data wall," AI is now being trained on Synthetic Data—high-quality data generated by other AI models and verified through automated feedback loops (Reinforcement Learning).

    • Why it matters: Models can generate millions of simulated math problems, code tests, and logical scenarios to train the next generation of AI, leading to self-improving systems that outpace human domain knowledge.

    ---

    Summary: What does this mean for the future?

    | Phase | Focus | Core Capability | Example |

    | :--- | :--- | :--- | :--- |

    | GenAI (2022–2024) | Content Generation | Chat, Text Summaries, Image Art | ChatGPT, Midjourney |

    | Agentic AI (2025+) | Execution & Reasoning | Workflow Execution, Complex Logic | Computer-using Agents, o1/DeepSeek R1 |

    | Physical/World AI (2026+) | Physical Reality | Spatial Awareness, Robotics, Simulation | Humanoid Robots, World Models |

    The shift is moving from "AI as a text box you talk to" to "AI as an invisible, action-oriented layer embedded in software, hardware, and physical space."

    Cited no sources.

    Our reading

    broad-convergence

    Lines we picked as the evidence — the model's words, our selection
    • the next phase is about moving from generation to execution, reasoning, and physical interaction
    • Here are the major breakthroughs shaping the next wave of AI
    • The shift is moving from "AI as a text box you talk to" to "AI as an invisible, action-oriented layer embedded in software, hardware, and physical space."
    Grok grok-4.6 cited sources in 6 of 6, 32 in total
    Sample 1 of 6

    The model's answer, recorded unedited

    There's no single "next big thing" in AI—progress is happening across multiple fronts simultaneously—but the most consistently cited developments by analysts, researchers, and companies (as of 2025–2026) center on Agentic AI (autonomous agents that act rather than just chat), Physical/Embodied AI (intelligence in the real world via robotics), and related advances like world models, multimodal capabilities, and AI-driven scientific discovery.[[1]](https://www.gartner.com/en/newsroom/press-releases/2025-08-05-gartner-hype-cycle-identifies-top-ai-innovations-in-2025)

    Generative AI (chatbots and content creation) has matured into a foundation, with a shift toward systems that plan, execute multi-step tasks, interact with tools/environments, and operate more independently. Here's a breakdown of the leading trends:

    Agentic AI and Multi-Agent Systems

    This is frequently described as the immediate next evolution after large language models (LLMs). AI agents are software entities that perceive, reason, decide, and take actions toward goals—often using tools, browsing, coding, or coordinating with other agents—rather than just responding to prompts. Gartner has identified AI agents as one of the fastest-advancing technologies, at the "Peak of Inflated Expectations," with projections that they could handle a significant portion of work decisions in coming years. Forrester and others call agentic AI the next frontier in automation.[[1]](https://www.gartner.com/en/newsroom/press-releases/2025-08-05-gartner-hype-cycle-identifies-top-ai-innovations-in-2025)

    Examples include systems like those from OpenAI, Anthropic (Claude), Google, and enterprise platforms (Salesforce, Microsoft) that handle research, coding, customer support, or workflows. In 2026, the focus is shifting to multi-agent systems (orchestrated teams of agents) and "super agents" for more complex, reliable execution. Challenges remain around reliability, hallucination, governance, and orchestration, but adoption is accelerating in enterprises. Some 2025 predictions of it being "the year of agents" were seen as mixed or still unfolding into 2026.[[2]](https://www.ibm.com/think/podcasts/mixture-of-experts/ai-year-review-trends-2026)

    Physical/Embodied AI and Robotics

    Often framed as the next frontier after digital agents: bringing AI into the physical world. This includes robots (humanoid and industrial), drones, autonomous systems, and "world models" that simulate physics, environments, and actions so machines can learn and operate in reality (not just text). Vision-Language-Action (VLA) models connect seeing, understanding instructions, and doing. Nvidia, Google (Gemini Robotics), and others are advancing this, with "physical AI" models expected to accelerate robotics growth. Deloitte and IEEE highlight it as poised for transformation in manufacturing, logistics, and more. Anthropic has even proposed standards for how AI agents should safely interact with physical hardware (labs, robots, machines).[[3]](https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions.html)

    This builds on multimodal AI (handling text, images, video, audio together, now becoming standard) and reasoning models (like OpenAI's o1-style systems that "think" step-by-step for better planning).

    Other Major Directions

    • AI for Science and Discovery: Systems that generate hypotheses, plan/run experiments (e.g., automated labs, "AI scientists" like those from FutureHouse or Coscientist), and accelerate fields like biology, chemistry, and materials. This is additive rather than just labor-replacing.[[4]](https://www.vox.com/future-perfect/471918/ai-science-growth-deepmind-alphafold-chatgpt-google)
    • Optimization, Smaller/On-Device Models, and Infrastructure: Cheaper inference, efficient models (beyond transformers, e.g., hybrids), AI-ready data, and scaling to production. Inference (running models) is expected to dominate compute. Embedded AI in existing apps (like search) will be more common than standalone tools.[[3]](https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions.html)
    • Toward Broader Intelligence: Discussions of AGI (artificial general intelligence) continue, with some leaders suggesting capabilities approaching or claiming it in the next few years, though definitions vary and timelines are debated. Superintelligence talk persists, but many experts emphasize practical utility, evaluation, and safety first. World models (simulating environments) are a key research bet for more robust intelligence.[[5]](https://www.nytimes.com/2025/03/14/technology/why-im-feeling-the-agi.html)
    • Supporting Trends: Generative watermarking for authenticity (World Economic Forum), better governance/TRiSM (trust, risk, security), and applications in specific domains like science, coding, or healthcare.

    Caveats: Hype often outpaces reliable, scalable deployment. Issues like energy use, data quality, bias, safety (especially as systems act in the world), and economic impact (jobs, inequality) are prominent. Progress is also happening in China (open models like DeepSeek, Qwen) and via open-source. Predictions frequently underestimate or overestimate speed—AI has surprised on capabilities like math/coding but lagged on some agent reliability.

    In short, the trajectory is from generating content to acting autonomously (digitally then physically), with intelligence becoming more general, efficient, and integrated into the real world. Watch developments from OpenAI, Google DeepMind, Anthropic, Nvidia, and enterprise analysts like Gartner for the latest. What emerges as "the" next thing will depend on which delivers the most practical, transformative value first.

    5 sources the model cited

    Our reading

    broad-convergence

    Lines we picked as the evidence — the model's words, our selection
    • There's no single "next big thing" in AI—progress is happening across multiple fronts simultaneously
    • center on Agentic AI ..., Physical/Embodied AI ..., and related advances like world models, multimodal capabilities, and AI-driven scientific discovery
    • What emerges as "the" next thing will depend on which delivers the most practical, transformative value first.

The positions for this question

Fixed before the run, so answers can move between them without the scale being redrawn. A position nothing landed on is still listed: it is the room the answers have to move into.

← What are the top three catastrophes from climate change? If I give an AI a task, will it stick to what I asked or add its own ideas? →