On PHP, and on Realsee中文EN

Since PHP 8.0, my involvement in PHP open source has dwindled, and so has my blogging. Quite a few friends have asked about it, so I figured a post explaining where things stand is in order.

Overall, my investment in PHP has borne the results I hoped for, and meanwhile my day job has changed in ways that consume a great deal of energy — leaving little left for PHP open-source work.

About PHP:

0x00

Since joining PHP development in 2013, my pursuit has always been performance: PHP LLVM in 2014, then PHP 7.0 — the most important performance release PHP has had — and then the PHP JIT in PHP 8.0. Compared with the PHP 5.4 of the day I joined, today's PHP is several times faster.

Then there are my own projects — Yaf, Yar, Yaconf and friends — all peripherals built for PHP performance, which, admittedly, also directly improved efficiency at Weibo back when I worked there.

Lately, though, I feel my well of performance ideas has run somewhat dry. I genuinely can't find a good direction to drill deeper into — or perhaps my time simply no longer allows the deeper investment and research it would take.

And about the change in my work, let me go into a little more detail:

0x01

Since 2018, my main job has essentially been leading a small team at Beike to build VR home tours. We named it Realsee (如视), with the founding aspiration: "Reality, as you see it."

Along the way, from designing the capture hardware, to cloud-based 3D reconstruction algorithms, to applications across H5, apps and VR devices — tape measurement, VR guided tours, AI walkthroughs, spatial assessment, and much more — we've had a lot of fun, getting excited every day as new ideas came to life.

Today we have a one-stop solution spanning professional to lightweight devices: from the early days when professional capture equipment was mandatory, to now, where a single phone can reconstruct a space in 3D. For example, this is a reconstruction of my workspace, made just now with my own phone running the Realsee VR app:

We have now reconstructed over 20 million physical spaces, totaling more than 1.6 billion square meters, and the team has grown from the original 6 people to over 200 today — 80% of them in product and engineering.

0x02

But in truth, from the very beginning I never positioned Realsee as merely "VR home tours", because once a physical space is digitized, the data value it holds is enormous.

For instance, after 3D reconstruction we can generate a space's floor plan fully automatically — previously at Beike this required manual measurement and hand drawing, and the automation alone saves tens of yuan per floor plan. And with our professional Galois devices, error stays under 20mm (0.4%), so you can capture once and measure many times on the digital twin, a far more efficient workflow:

In the example above, our product's rough measurement reads 1683, while the TV's real dimension, once looked up, is 1685.

Or: since we hold dimension data for all kinds of objects in a home, physical simulations come easily — for example, air circulation:

Or again, back in 2019, by which time we had accumulated over 4 million digital spaces, we found that doing object recognition in 3D space (setting aside what can't be recognized — specific brands, styles, prices) and segmenting the results yields a functional layout scheme of a space:

Feeding this data in as a training set — input a digital space's floor-plan layout plus some user preferences, and out comes a spatial layout solution. Isn't that precisely part of what a junior interior designer does?

But to get there we were still missing a lot of data: design tooling (for manual correction), large furniture model libraries, rendering capability, and more. So in 2019 we embraced UE4 and started figuring out, bit by bit, how to make AI-designed results look "real" enough to present to users.

That is what you see on Beike as "Beike Future Home".

Today, given any floor plan, we can render a complete design scheme in a dozen seconds for users to experience in VR — design effects that once lived in the imagination can now truly be "seen".

There are many, many more examples of this data value being unlocked, and I believe we will produce many more.

0xff

As our technical and product capabilities grew richer, hoping our technology could serve more people and the value of this data be recognized by more, over the past two years we gradually began offering our capabilities to partners beyond Beike — beyond real estate and furniture/decor, more interesting scenarios have emerged.

Throughout this we kept building our openness, crafting our open platform so more people with ideas can design and develop solutions on top of our digital spaces. Today we have largely achieved low-cost, rapid, self-service embedding of our capabilities into all kinds of products via the open platform: Realsee Developer Platform

Looking back, it all seems to have followed its natural course: in 2022, the timing being right and with Beike's greater expectations, Realsee became an independent startup.

I remember years ago someone asked why I didn't consider starting a company. I said entrepreneurship seemed too grueling — too many miscellaneous things to worry about — and that quietly doing technology suited me better.

But life, more often than not, cannot be planned. Today, I've set out on the startup road anyway.

So going forward I will throw myself entirely into Realsee, which does mean I can no longer take part in substantial PHP open-source work. I can only barely keep iterating on my own few open-source projects — and one must admit, a person's energy really is limited.

Of course, I will still follow PHP's progress in my spare time; after all, it once carried a great many of my aspirations.

Finally — though I do feel some regret about PHP, and a little apologetic toward its many loyal users (PHP has plenty of excellent new developers, so my participation or absence hardly matters) — what I'm doing now is also where my own interest lies. After all, at different stages of life, one carries different expectations.

So — wish me well?