Contents
The question "how smart is AI?" sounds natural but answers poorly. A model can outperform an expert at writing code and still fail simple arithmetic. It can generate a cinematic scene and lose a character's face within three seconds. It can read text almost like a human — and invent a fact on page thirty of a document.
Key takeaways
AI advances unevenly — and that is normal. Each domain has its own ceiling, metrics, and path from "conference demo" to "production dependency." Comparing GPT to a human "in general" is meaningless; comparing the automation level of a specific activity is useful.
The main shift is from "can do" to "can do reliably and autonomously." Learning to generate video once is a breakthrough. Doing it ten times in a row with predictable results is maturity. Reliability, controllability, and acceptable cost define when a human stops being necessary at every step.
Professions change through role shift, not simple replacement. Developers, designers, analysts, and musicians increasingly act as operators and curators of systems that handle 80% of routine work. Human value moves toward goal-setting, verification, and accountability for outcomes.
Why "how smart is AI?" is the wrong question
Imagine rating a doctor on one parameter: "how good are they at everything." Surgeons and therapists, pediatricians and radiologists — different profiles. AI works the same way, only the gap between strong and weak sides is wider.
This is jagged intelligence: the model solves an olympiad math problem and mixes up dates on a resume. It generates a working microservice but cannot explain why the business abandoned an API five years ago. It writes ten coherent pages and inserts a nonexistent research link.
Measuring AI progress as one percentage of "general intelligence" is a methodological mistake. Better to look at the automation level of a specific activity: result quality, stability, cost, need for human oversight, and how long the system runs without intervention.
The central question of this article:
In which domains has AI become a full-fledged tool, and where does it remain an impressive toy?
How to measure progress: a maturity scale
A simple 0–5 scale — not an academic standard, but a practical ruler:
| Level | Meaning |
|---|---|
| 0 | Barely works: demo, prototype, unusable without heavy rework |
| 1 | Experimental tool: sometimes useful, unpredictable |
| 2 | Helpful assistant: speeds work, human controls every step |
| 3 | Professional tool: fits workflow, errors are localizable |
| 4 | Significant autonomy: system runs most of the chain, human at milestones |
| 5 | Replacement for typical tasks: human not required at every stage for standard scenarios |
One number is not enough. When rating a domain, keep eight criteria in mind:
- result quality — how close to professional standard;
- stability — same outcome on repeated runs;
- controllability — can you specify exactly what to change;
- error correction — fix without starting over;
- autonomous run length — how many steps without failure;
- cost — affordable at scale;
- speed — fits real work rhythm;
- oversight need — where humans are mandatory by law, risk, or quality.
Two domains with the same star rating can differ: text and coding are both high, but code carries security and regression risks; text carries factual hallucinations.
Programming — one of the most mature domains
Software development is a rare case where AI entered millions of professionals' daily work through the IDE, not a keynote. Cursor, GitHub Copilot, Claude Code, Windsurf, and dozens of alternatives changed the loop: write → run → debug → commit.
What AI already does
Modern models and agents confidently:
- write and complete code in dozens of languages;
- explain foreign code and map dependencies;
- find bugs, suggest fixes, and refactor;
- generate tests and documentation;
- work with issues and pull requests;
- build apps from descriptions — single file to multi-file projects;
- call tools: terminal, browser, API, repo search.
In enterprise settings this is no longer an experiment. Teams use AI for onboarding into projects with years of history, routine acceleration, and prototyping. For how mature teams adopt AI without hype, see top engineering teams in practice.
Where the ceiling is
Large-system architecture still needs humans: trade-offs between consistency and service autonomy, module boundaries, multi-year platform evolution. AI is weak on business context — "why we did it this way in 2019" rarely lives in code. Long-term ownership — accountability, regressions, dozens of stakeholders — stays with the team.
Verifying complex decisions (security, concurrency, financial logic) without an expert is risky: the model sounds convincing when wrong.
The central question
Does AI replace the programmer or turn them into the operator of a more powerful system?
Mostly the latter — for now. Value shifts to task framing, architecture, review, and production accountability. But the share of automatable routine grows faster than tasks needing "pure" engineering judgment.
Maturity: ★★★★★ (5/5) — professional tool with momentum toward autonomous dev agents.
Text and knowledge work
Language models are the foundation under most other breakthroughs. Text, translation, summarization, classification, entity extraction, brainstorming, tutorial dialogue, analyzing hundreds of documents in minutes — all embedded in workflows from support to legal and research.
What is largely solved
Where truth can be checked afterward or errors are cheap, LLMs are a mature product. Draft emails, article structure, UI translation, internal FAQ, first-line ticket sorting — standard 2026 stack.
Large-volume text is a separate advantage: the model does not tire reading a thousand pages of policy. With RAG (retrieval-augmented generation over corporate data), this underpins enterprise AI.
Remaining problems
Factual errors and hallucinations persist. The model completes plausibly from language statistics. Multi-step causal reasoning is weak: five-to-seven step chains with hypothesis testing need verification.
No truth guarantee without external sources, verification tools, or humans in the loop. In high-risk domains (medicine, law, finance), text AI is an assistant, not an oracle.
Maturity: ★★★★★ (5/5) — with the caveat that process matters as much as the model.
Images — AI became a designer's tool
Image generation went from "random art" to editing while preserving composition. Before: "make a beautiful picture." Now: "change the foreground object, keep lighting and background."
Capabilities
Models like Midjourney, DALL·E, Stable Diffusion, Flux, and successors cover:
- concept art and illustration;
- ad visuals and design variants;
- stylization and aesthetic transfer;
- characters and environments;
- inpainting and targeted edits.
Designers use AI as a variant generator, not final art director: ten directions faster than one perfect first try.
Limitations
Fine control is weak: exact angle, readable signage text, complex multi-object perspective — still lottery. Character consistency across frames and sessions needs extra techniques (references, LoRA, manual polish). Reproducibility — two runs with the "same" prompt differ.
Maturity: ★★★★☆ (4/5) — strong professional tool with control and pipeline limits.
Music — from fragment to finished work
Music models made in two–three years a jump images took five. Before: a short loop or instrumental snippet. Now: a song with vocals, arrangement, and verse-chorus structure.
What models do
Suno, Mureka, Lyria (Google), ElevenLabs Music, MiniMax Music, and specialized open-source tools create:
- full tracks with vocals and instruments;
- genre-specific arrangements;
- sound design and atmospheric beds;
- variations from reference or text.
Per instrumental generation benchmarks (August 2026), leaders like Suno V5.5 and Mureka V9 differ more in quality and controllability than in whether music is possible at all. The race moved from capability demos to fine tuning.
The main problem
AI creates music better than it edits it. Replace only the bass line in bar three, keep vocals, change tempo — still closer to manual DAW work than dialog with a model. Authorial intent over album length (unified palette, narrative) is hard to hold.
Maturity: ★★★★☆ (4/5) — finished works yes; full producer replacement no.
Video — impressive but not yet mature
Video is the most visually striking and among the least "even" domains. In three years the industry walked:
text → image → image-to-video → text-to-video → multimodal input → video with audio as a finished scene.
What models do
Veo, Seedance, Kling, Runway, Wan, and others generate realistic scenes, camera motion, characters, lip-sync, short story beats, VFX, and increasingly audio in one pass. Modern systems show not only visual plausibility but early object behavior understanding — see AI Index 2026 (Stanford HAI).
Leaderboards shift fast: on the text-to-video arena (August 2026), Wan 3.0, Gemini Omni Flash, and MiniMax H3 lead. A quarter later the list may differ — sign of a live industry, not a settled one.
The main problem is time, not frame
One beautiful frame is solved. Event sequence is not:
- character must stay the same face and clothes;
- objects keep properties while moving;
- physics must not break between seconds;
- camera follows script, not random drift;
- story must run tens of seconds and minutes, not only 5–10.
Temporal consistency is the barrier between "social clip" and "scene for film edit."
Maturity: ★★★☆☆ (3/5) — strong short scenes; weak long-form directing.
Voice and speech — largely solved
Speech is one of AI's most "invisible" wins: users stop thinking "that's a synthesizer."
Capabilities
- speech-to-text with high accuracy on major languages;
- text-to-speech with natural prosody;
- voice cloning from short samples;
- translation preserving timbre;
- emotional coloring and real-time conversation.
Voice assistants, dubbing, audiobooks, accessibility — product grade, not lab demo.
Limitations
Long monologues without energy drop, fine intonation direction, full emotional truth in drama — not perfect. Voice-cloning safety is separate: deepfake calls and fraud need technical and legal guardrails.
Maturity: ★★★★★ (5/5) — with caveats on abuse and edge-case prosody.
3D and game worlds
After stills and video, the next bar is interactivity. Not "make a model" but "make a world you can walk in."
Generation
AI helps with objects, textures, environments, characters, animations, and game assets. Pipelines speed pre-production: sketch a level faster than sculpting every rock by hand.
What blocks production
- geometric accuracy and clean topology for animation;
- consistency across assets in one scene;
- physics and collisions;
- engine integration (Unity, Unreal) without manual rework.
Still an accelerator for artists and tech art, not a "generate AAA game" button.
Maturity: ★★★☆☆ (3/5) — strong draft generation; weak end-to-end pipeline.
Science and mathematics
AI as research assistant is 2026 reality: paper analysis, experiment code, pattern search, hypothesis framing. Some systems tackle olympiad and research problems, occasionally contributing to proofs.
The jagged-intelligence paradox
A model can advance a hard problem and fail an elementary one. Best argument against a single "AI IQ" scale. Peak brilliance matters less than reliability — related to verifying LLM math reasoning.
Next level
Scientific agent: hypothesis → experiment → analysis → new hypothesis — a loop still needing humans at critical gates. Full scientific autonomy is ambition, not current standard.
Maturity: ★★★☆☆ (3/5) — powerful assistant; weak autonomous reliability.
Robotics — digital AI meets the physical world
In code, an error is a wrong answer. In the physical world — broken object, fall, injury. A different risk class.
What works today
Robots and embedded AI recognize objects, plan trajectories, perform specific manipulations in warehouses and factories. RL and simulation moved the needle on individual tasks.
What stays hard
Hand dexterity, unpredictable environments, long-horizon planning, safety around people — all orders of magnitude harder than generating text. Digital AI "ran ahead"; robot bodies and physics did not.
Maturity: ★★☆☆☆ (2/5) — point successes; no mass autonomy in the open world.
Autonomous agents — the next big threshold
A separate maturity axis — not modality (text, video) but mode of work.
| Mode | Pattern |
|---|---|
| Chatbot | human → query → AI → answer |
| Copilot | human → task → AI helps each step |
| Agent | human → goal → AI plans, searches, calls tools, checks, fixes, repeats |
Agents extend mature coding and text models: same LLMs, plus actions. For enterprise context see why enterprise AI needs agents and platform for agents, not speed.
The main problem
Not one operation but a long chain. Ten steps without failure is exponentially harder than one lucky step. Error at step 7 ruins everything before it. Agent reliability is the key front for 2026–2027; see also LLM routing taxonomy and agent operating systems.
Maturity: ★★★☆☆ (3/5) — complex demos and pilots exist; industrial reliability in progress.
Summary maturity map
| Domain | Level | Already possible | Main problem |
|---|---|---|---|
| Text | 5/5 | Full document work, dialogue, analysis | Factual errors, no truth guarantee |
| Coding | 5/5 | Development, refactor, tests, IDE agents | Architecture, business context, accountability |
| Voice | 5/5 | Realistic synthesis and recognition | Cloning safety, fine direction |
| Images | 4/5 | Pro design, concepts, ads | Fine control, consistency, reproducibility |
| Music | 4/5 | Finished songs and arrangements | Editing, long-form authorial intent |
| Video | 3/5 | Cinematic short scenes | Temporal consistency, long narrative |
| Science | 3/5 | Research assistant, some proofs | Reliability, jagged intelligence |
| 3D | 3/5 | Asset and environment generation | Production pipeline, interactivity |
| Autonomous agents | 3/5 | Multi-step tasks in pilots | Long-chain reliability |
| Robotics | 2/5 | Specific actions in controlled settings | Physical world, safety, dexterity |
The map is a mid-2026 snapshot. In a year rows will shift; video and agents likely fastest.
AI does not advance in a straight line
There is no single scale "AI intelligence = 80%." There are many independent axes: language, vision, hearing, programming, music, video, math, physical world, planning, autonomy.
AI is not "mind at level X" but an uneven capability profile. That matters for strategy: a support chatbot does not mean warehouse robots are ready. A video pilot does not mean hands-off code review.
Where AI will advance next
Near horizon (1–2 years)
- dev agents — from autocomplete to autonomous tasks with controlled merge;
- video production — longer consistent scenes, edit integration;
- music production — better editing and DAW handoff;
- research agents — literature → code → experiment loops with verification;
- multimodal assistants — one interface to text, voice, screen, tools.
Next tier (3–5 years)
- autonomous agent teams with role split;
- AI-native development — spec and tests as primary artifacts, generated code underneath;
- full films from script with revision, not only clips;
- interactive virtual worlds with runtime-generated content;
- scientific agents with reproducible experiments.
Hardest tier
- robotics in unstructured environments;
- physical world with safety and liability;
- long-term autonomy without goal drift and error accumulation.
Forecasts are not guarantees. History suggests digital modalities mature faster than "intelligence + body + regulation."
How the idea of profession changes
The main transition for many roles:
tool → assistant → operator → (partial) autonomous executor.
| Role | Was | Becoming |
|---|---|---|
| Developer | Writes every line | Tasks agents, reviews, owns architecture |
| Designer | Draws every mockup | Curates generation, fixes control points |
| Musician / producer | Records every part | Sets reference and structure, finishes in DAW |
| Videographer | Shoots and edits | Directs scenes from generative blocks |
| Analyst | Builds reports manually | Frames questions, verifies model output |
| Researcher | Reads everything alone | Orchestrates search and experiment agents |
If AI covers 80% of work, human value shifts to:
- goal-setting and success criteria;
- checks at high-risk points;
- accountability to client, law, society;
- what is not in training data — context, negotiation, ethics.
The question is not "will they be replaced" but what stays exclusively human on current trajectory.
FAQ
Why can't we say "AI is at human level"?
Because "human level" is not one number. Humans are uneven too: a chess grandmaster may be weak at accounting. AI beats people on some tasks and loses on neighbors — sometimes in the same session.
Which domain is most mature?
Text, programming, and voice top the scale for typical scenarios. But maturity always has caveats: without verification, text is risky; without review, code is risky; without policy, voice cloning is dangerous.
Why does video lag images?
One image is static. Video adds time: object consistency, physics, story. Computationally and algorithmically harder by orders of magnitude; viewer expectations are higher.
Will AI replace programmers?
More likely it changes the role: less routine typing, more task framing, integration, security, and system accountability. Demand grows for people who work with AI.
What is jagged intelligence?
When strong and weak sides do not correlate with intuitive task "difficulty." Hence caution with one scale and need for per-domain assessment.
How to use this map in a company?
For each process ask: what maturity level is needed (2 — assistant or 4 — autonomy)? Which of the eight criteria matter? Where is a human mandatory? Do not extrapolate chatbot success to robotics without a separate review.
Further reading
- Can AI understand a 15-year-old codebase — AI maturity in legacy code
- Why enterprise AI needs agents — from chat to action
- The future of enterprise AI — platform, not one model
- How top teams adopt AI — practice without hype
- Why LLMs appeared — context for the leap
- LLM math proofs need verification — jagged intelligence in action
- Antigravity: agent operating system — autonomy infrastructure
Conclusion
Stop asking whether AI can do something. Ask:
How reliably, autonomously, and cheaply can AI do it?
The path can → can stably → can autonomously defines real maturity — and real impact on professions.
AI is not moving toward one universal intelligence level. It gradually captures separate human capabilities — one after another. The most interesting action is not where a model first learned a skill, but where it learned to do it well enough that a human is no longer needed at every step.
This week, pick one row from the summary table closest to your work and rate it honestly on the eight criteria — not from Twitter vibes, but from your last real pilot. That beats any "general AI IQ."