If you had told a developer in early 2020 that by the end of 2026 most of their peers would use AI every working day, that software agents would open pull requests, and that a model would be better than a PhD-level human baseline on science questions while still fumbling an analog clock, they would have asked what you were smoking.
Well. It's mid-November 2026.
So let's do something we rarely do in this industry: stop, zoom out, and measure how far AI has actually travelled in six and a half years. Not hype. Numbers, milestones, and one uncomfortable pattern in the data that I think every developer should think about.
And at the end, there's a very concrete story about what all this means for a tool millions of people use, and a deadline that is 15 days away.
First, a correction: AI didn't "arrive" in 2020
Dev.to will eat me alive if I skip this. AI research is decades old. Deep learning broke through in 2012, and the Transformer architecture dropped in 2017.
But 2020 is the year the curve bent for everyone else. GPT-3 showed that if you scale a Transformer far enough, you get something that follows instructions from a handful of examples, with no task-specific training. That one idea (you can "program" a model with text) is the root of almost everything that followed. So I'm using 2020 as year zero of the era we're all living in.
The six-year timeline
| Year | The moment | Why developers cared |
|---|---|---|
| 2020 | GPT-3 arrives (mid-year). AlphaFold 2 dominates the CASP14 protein-folding challenge (late-year). | Few-shot prompting is born. And AI cracks a 50-year-old science problem. |
| 2021 | DALL·E (Jan). GitHub Copilot technical preview (Jun). | For the first time, AI lives inside the editor. |
| 2022 | DALL·E 2, open-weights Stable Diffusion, then ChatGPT on Nov 30. | Image generation goes open source. ChatGPT turns "AI" into a text box anyone can use; analysts estimated ~100M users within about two months. |
| 2023 | GPT-4, Claude, Llama 2 (open weights). | Multimodal and long-context models; RAG becomes a default architecture; "prompt engineer" becomes a job title (briefly). |
| 2024 | Sora preview. Nobel Prizes in Physics and Chemistry both honour AI-related work. | Legitimacy. AI is now science, not just product. |
| 2025 | Reasoning models. DeepSeek-R1 open weights (Jan). Terminal and IDE coding agents go mainstream. | Models "think" before answering. Costs fall. Developers start delegating, not just autocompleting. |
| 2026 | Agents inside everyday workflows. Stanford's AI Index says capability is accelerating, not plateauing. | The question changes from "can it do it?" to "can I trust it, and who's accountable?" |
What the numbers say
Timelines are fun; surveys are honest. Stack Overflow has been asking developers about AI since 2023, and the trajectory is steep.
Developers using AI tools (Stack Overflow):
| 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|
| 44% | 62% | 79% | only ~17% say they don't use AI tools at work |
AI agents are the faster story. In the 2025 survey, 31% of respondents used agents. A smaller pulse survey in April 2026 found agent use had almost doubled to 59%, with Claude Code usage among agent users rising from 41% to 55%.
And outside our bubble, Lightcast's read of the Stanford data shows 2.5% of all US job postings now mention AI skills: up 55% in a year and roughly 300% over a decade.
So: massive adoption. Done, right? Not quite.
The uncomfortable pattern: usage up, trust down
Here's the part that doesn't fit the hype narrative.
- In 2025, 46% of developers said they distrust AI output, versus 33% who trust it. Only 3% said they "highly trust" it.
- In the 2026 survey, nearly half trust AI output when they can easily verify it. Just 6.6% would trust it with an important decision.
- The most common frustration in 2025: solutions that are "almost right, but not quite."
- Governance is lagging: only 24% of respondents said their employer had published approved AI tools or guidance.
I'd argue this is the defining dynamic of 2026: AI has become infrastructure faster than it has become trustworthy. We use it constantly and check it constantly.
Stanford's 2026 AI Index captures the same weirdness from the capability side: models now beat human baselines on PhD-level science questions, yet still fail at tasks like reading an analog clock. Researchers call this jagged intelligence: brilliant peaks, baffling potholes.
My takeaway (opinion, not data): the most valuable developer skill of the next few years isn't prompting. It's verification: writing tests, reading diffs, designing systems where a wrong-but-plausible answer gets caught.
The quiet shift nobody is tweeting about: AI left the data center
In 2020, running a serious neural network meant renting GPUs and calling an API. That model (your data goes up, the answer comes down, and you pay per call) defined the first half of this era.
The second half is different. Thanks to WebAssembly, WebGPU, and runtimes like ONNX Runtime Web, a browser tab can now run real models locally:
import * as ort from "onnxruntime-web"; // Simplified: load a model once, cache it, run inference on-device const session = await ort.InferenceSession.create("/models/segmenter.onnx", { executionProviders: ["wasm"], // or "webgpu" where supported }); const result = await session.run({ input: imageTensor }); // Nothing was uploaded. No API key. No per-call bill. Why this matters:
- Privacy: sensitive data never leaves the device.
- Cost: inference runs on the user's hardware, so there are no per-image credits.
- Latency and availability: after the first download there is no round trip, and some tools work offline.
- Longevity: your tool can't be switched off because someone else's API changed pricing.
That last point just became very real for millions of people.
Case study: the end of an era
For years, remove.bg was the answer to "how do I remove a background?" It popularised one-click AI background removal and was a textbook example of the cloud-API era.
Now its standalone website is closing on December 1, 2026, with the functionality moving into Canva (which has owned it since 2021). Reporting based on remove.bg's own FAQ says the closure is at 9:00 CET, unused pay-as-you-go credits expire, and the self-serve API moves to Leonardo.Ai (also Canva-owned). If you rely on it, export your usage reports and invoices before the cutoff.
There's a lesson here that has nothing to do with any single company: if your workflow depends on someone else's server, it has an expiry date you don't control.
If you're already in the Canva ecosystem, Canva is the official successor and will be fine for you. But if you want something that doesn't ask for an account or a credit card, or send your photos anywhere, read on.
Meet APIC: the on-device approach, in practice
Full disclosure: I built APIC, so weigh my opinion accordingly. But everything below is verifiable in about 60 seconds, which is exactly why I'm comfortable showing it to a room full of skeptical developers.
APIC (Advanced Image Processing Center) is a free, browser-based photo editor where the AI runs on your device. No upload. No credits. No account required to edit or export (there's an optional Google sign-in, nothing more).
Background removal, the remove.bg use case
- On-device AI cut-out with three model sizes: Light (~5 MB, built for phones), Fast (~44 MB), and Maximum (~176 MB, multi-pass, best on a PC). Models download once and are cached.
- A Restore / Erase brush to fix hair, fur and awkward edges, plus one-click replacement with white, black or a blurred copy of your photo.
- Full-resolution output as transparent PNG or WebP.
- Batch mode: remove backgrounds, resize, compress and convert many images, then download one ZIP.
And because I got tired of juggling ten tabs, it's eleven tools in one
- Remove Object: paint over a person, sign or blemish; inpainting runs locally (MI-GAN ~28 MB, or LaMa ~208 MB for big, complex fills).
- AI upscale (2x / 4x) with Real-ESRGAN.
- Compress to an exact file size ("under 200 KB" for that upload form), resize, and convert between PNG, JPEG, WebP, AVIF, BMP, ICO, PDF and PSD, with HEIC photos accepted as input.
- Pro colour tools: histogram, RGB curves, HSL mixer, 3-way grading, and 3D LUT (.cube) export for DaVinci Resolve, Premiere and Photoshop.
- Social presets for 15+ networks with safe-zone guides; Layers with blend modes and PSD import/export; Hide area to blur or pixelate faces, plates and card numbers before you share.
- A Windows app that runs fully offline, no account needed.
The honest trade-offs (because you'd find them anyway):
- The first run of any AI tool downloads its model, so it's slower once and fast afterwards.
- The heaviest models are best on a desktop-class machine. On phones, use Light.
- It's a browser tool, not a replacement for Photoshop's full ecosystem.
- I'm one developer shipping this, not a 500-person company. Feedback genuinely shapes the roadmap.
Try it with your hardest photo
Don't test it on a clean studio shot. Take the image remove.bg struggled with (wispy hair, glass, a busy background) and run it through APIC's Remove BG tab on Maximum. Use the Restore brush if needed. If it doesn't hold up, tell me where it failed. If it does, you've just saved yourself a subscription.
Make it your default (it takes 20 seconds)
With 15 days left on the old default, now is the moment to build a new habit:
- Open apic-studio.vercel.app and drop in a photo.
- Bookmark it and pin the tab, so "remove background" now means one click on your own bookmarks bar.
- Send it to the one person in your team, family or client list who still says "just use remove.bg."
- Tell me what's missing: a comment below, or find me on GitHub / dev.to. I read everything.
(I also wrote a shorter, practical guide on the remove.bg shutdown specifically: remove.bg Is Shutting Down — Here's a Free, No-Upload Alternative.)
So, how much AI is there in six years?
Started as a research demo behind a waitlist in 2020. By 2026 it's in the editor, the terminal, the job postings, the Nobel Prizes, the survey data, and increasingly, in your browser tab, running on your own hardware.
The next six years won't be about who has the biggest model. They'll be about three questions:
- Can I verify it? (trust)
- Who controls it? (open weights, on-device, no kill switch)
- Does it respect the user? (privacy by architecture, not by policy)
The tools that win on all three won't need a marketing budget. They'll just be the ones people quietly never switch away from.
I'm betting on that. Come try APIC, and tell me where I'm wrong.
Your turn: what was your "oh, this changes everything" AI moment between 2020 and 2026: GPT-3, Copilot, ChatGPT, or something else? Drop it in the comments. 👇
Sources: Stack Overflow Developer Survey 2025 · Stack Overflow Developer Survey 2026 coverage · Stanford HAI AI Index 2026 · Lightcast · remove.bg shutdown reporting


