2026-03-17
By Vadym · Generated with Boba, curated by me
I've been writing software for over fifteen years. In that time I watched cloud computing eat the data center, mobile eat the desktop, and open source eat the enterprise. Each shift felt big at the time. This one feels different — not like a platform change, but like a change in what software is.
Let me try to explain what I mean.
In early 2024, GPT-4 was genuinely impressive but also clearly limited. It hallucinated confidently. It couldn't browse the web, couldn't write a multi-file codebase coherently, couldn't run code and check its own work. You could feel the ceiling.
Then things started compressing in ways I didn't expect.
GPT-4o dropped in May 2024 — real-time multimodal, voice, vision, all in one. A few months later came the o1 reasoning models: systems that don't just predict the next token, but that think, decompose problems, backtrack, check their own reasoning. That was a qualitative shift. Google's Gemini 2.0 arrived in December 2024 with native multimodality and built-in agent architecture. Meta shipped Llama 4 in April 2025, open-source, multimodal, capable enough to run competitive intelligence tasks.
Then 2025 got faster. GPT-5 launched in August 2025. Claude Opus 4.6 shipped in February 2026 with the highest Terminal-Bench score ever recorded, leading every frontier benchmark on agentic web research, and — crucially — the ability to decompose work across teams of agents running in parallel. GPT-5.4 followed in March 2026, the first general-purpose model with native computer-use at that level of capability, with a 1M-token context window.
That last part matters. A million tokens is roughly the size of a small codebase. The model can now hold your entire project in context, plan across it, and make coordinated changes. That's not autocomplete anymore. That's something else.
I deploy AI dev tools. I think about this every day.
Two years ago, GitHub Copilot was the story — inline autocomplete that occasionally surprised you. Engineers used it like a smarter tab-completion. The ceiling was: it helps with boilerplate, but you still drive everything.
That ceiling is gone.
Claude Code launched in May 2025 and within eight months became the most-used AI coding tool among professional developers, overtaking Copilot and Cursor. It runs in the terminal, reads your entire codebase, understands architecture, and makes changes across multiple files simultaneously. Not just within a file — across your project, with an understanding of what the change means in context.
Today's agents handle 30-to-60-minute coding tasks independently. They write the code, generate tests, debug failures, iterate until CI passes. The engineers who thrive aren't the ones who memorize the most syntax or know the most APIs — they're the ones who can communicate intent clearly and review AI output critically.
I've watched this play out firsthand. The senior engineers I respect most have adapted faster than I expected. The ones who are struggling are the ones who thought the tooling was a fad.
It's not a fad. And we're still at the beginning.
I've been hearing "robots will replace workers" for twenty years. It was mostly noise. Then something changed.
The key shift wasn't motors or batteries. It was the AI stack underneath. Once you have models that can see, reason, plan, and act — and once you can run those models cheaply enough to put them in a robot's onboard computer — the physical embodiment becomes the easier problem.
Let me walk through what's actually shipping.
Boston Dynamics Atlas retired its hydraulic platform in April 2024 after nearly a decade of research demos. The new electric Atlas has 56 degrees of freedom, fully rotational joints, and a reach of 2.3 meters. It can hot-swap its own battery in three minutes. At CES 2026, Boston Dynamics announced production has begun — 2026 deployments are already fully committed, shipping to Hyundai's manufacturing facilities and Google DeepMind. These aren't demos. These are production units with booked customers.
Figure AI's Figure 02 ran a production pilot at BMW's Plant Spartanburg in 2025 — five days a week, ten hours a shift, positioning sheet metal parts for welding. It moved over 90,000 components with millimeter accuracy across 1,250 operating hours. BMW and Figure are now evaluating where the next-generation Figure 03 fits in the broader manufacturing line.
Tesla Optimus is a longer game. The Gen 3 hands now feature 50 actuators total — a 4.5× increase from Gen 2 — enabling tasks from cracking eggs to battery cell manipulation. In December 2025, Tesla released footage of Optimus jogging smoothly across a lab floor. Musk confirmed on the Q4 earnings call that current units are collecting data, not yet doing "useful work." Tesla is converting its Model S/X production line to manufacture Optimus at scale, targeting a million units per year. Consumer sales are targeted for 2027.
Unitree is the one most people in the West underestimate. The Chinese firm shipped over 5,500 humanoid robots last year — more than Tesla, Figure AI, and Agility Robotics combined. Their G1 starts at $13,500. They're targeting 20,000 units in 2026. The G1 robots have been demonstrated doing kung fu, trampoline flips, and parkour — not because that's the use case, but because it's how you stress-test locomotion and balance control.
1X Technologies NEO is taking a different bet: the home. Pre-orders are open at $20,000, with first deliveries in 2026. At 66 pounds, it can lift 150 pounds, operates at under 22 decibels (quieter than a refrigerator), and is built for soft-body interaction with humans in unstructured environments. OpenAI is a backer. The pitch is not a factory robot — it's a robot that can fetch something from your fridge, open the door for a guest, and turn the lights off at night.
Sanctuary AI's Phoenix is partnering with Magna International — one of the largest automotive suppliers in the world — to mature the platform for manufacturing environments. They're taking the slower, more methodical path: deep capability in specific industrial settings before going broad.
Software engineers have been safe in cycles of automation because the work required judgment, creativity, and context. Writing code wasn't like welding a bolt — it was more like designing the fixture that holds the bolt. Hard to automate.
That argument is weakening.
I'm not saying software engineers are going away. I'm saying the leverage ratio is changing. One good engineer with strong AI tooling can now do what required three or four engineers before. That doesn't mean mass unemployment tomorrow — it means the industry absorbs fewer new entrants, and the definition of "good engineer" shifts toward systems thinking, architecture judgment, and the ability to steer and verify autonomous work.
The physical world story is slower but pointing in the same direction. Humanoid robots working in factories are real and shipping now — not in limited lab conditions but in volume production environments. The home is next. The timeline is measured in years, not decades.
I've been building an AI assistant for myself — scheduling, research, trading analysis, voice interaction, memory. It started as a hobby project and turned into something I use every day. The gap between what I can do with a well-configured AI stack versus what I could do before is not marginal. It's transformative.
I'm not saying this to sound optimistic. I'm saying it because I've felt the shift personally, not just read about it. And when you feel something like this in the tools you use daily, it's a different kind of knowing.
The question I sit with is: what does it mean to be a software engineer in five years? In ten?
My honest answer is that I don't know — but I think the engineers who are asking that question seriously, and adapting rather than waiting, are the ones who will define what the answer looks like. The ones who treat this as a moment to build more, think bigger, and use leverage they never had before.
The beginning of the end isn't necessarily the end of something good. Sometimes it's the end of a ceiling.
By Vadym · Generated with AI, curated by me