We’re building a high-performance video engine at @slateteams and hiring senior video and graphics engineers.
This is not a typical full-stack role. Real-time rendering and media pipelines, cross-platform systems.
If this is you:
https://t.co/6ZqQBaQ7Li
Context preservation shouldn’t mean dumping every quirk into a database. It should mean fixing quirks so they don’t need documenting in the first place. Otherwise you’re just turning tech debt into infrastructure.
We’re hiring at Slate.
Real product, awesome customers, genuinely cutting-edge work.
This is not your regular CRUD app. We’re building a best-in-class content creation platform.
Apply here:
https://t.co/dkUVakSqJB
@vanilagy made an extremely complex domain look simple, even though it absolutely is not.
One of the few folks on X I truly look up to and would gladly pay to learn from.
P.S. blog post, book, podcast whatever man just make it!
The majority of Mediabunny development time is spent on crafting algorithms. For example, I just released the third rewrite of the MPEG-TS seeking (random access) algorithm to be more optimized for networked resources. Thought I'd share how it works!
---
The goal: find a packet with a given timestamp. A packet contains a single video frame, and they're spread throughout the file. We don't know where they are yet.
Mediabunny first splits the entire file into 5 MiB chunks, then does a binary search on the beginning of these chunks to find the last chunk with a timestamp smaller than the search timestamp.
After this coarse phase is over, it reads forward sequentially, making sure to only partially parse data; the only data we need is the current timestamp for the track we're seeking for.
We stop reading sequentially once we find a packet with a larger timestamp than the search timestamp. This means we've found the region where our target packet lies.
Simply returning the previous packet wouldn't be correct, however: video tracks make use of B-frames, meaning the timestamps are not in order. They could go like [0, 1, 3, 2, 5, 4, ...]
To find the actual packet with the closest timestamp, we first extract something called the "reorder buffer size" from the bitstream of the codec (AVC or HEVC, in this case). Then, we rewind by that many packets, and then read forward sequentially again.
Now we simulate what the decoder would do: we keep an internal packet resize buffer, and push packets into it as we find them. Once the buffer fills up, we find the packet in the buffer with the lowest timestamp which we now know is finalized and in the right order.
We repeat this process until we emit a packet with a higher timestamp than the one we're looking for. Then we're almost done. We now read the packet in full, check if it's a key frame or not, and then finally output it to whoever called Mediabunny's API.
---
This algorithm is superior for networked resources because of the chunked binary search; it maximizes sequential, forward reads (which is what HTTP loves) and minimizes random access patterns. For small files < 5 MiB (which is most MPEG-TS files), no random access happens at all. This is the magic trick for writing performant code: know your data distribution and optimize your logic around that.
I could easily write many blog posts about this stuff but it would suck up a lot of time. Maybe I'll do it someday!
A month into agentic coding and this is exactly how I feel. Running five ideas in parallel, chasing the dopamine of fast increments is fine, but I need slow, boring, offline time for my brain to perform at its best more than ever
You need to be slowmaxxing. You need to be reading long, fat books. You need to be making 48-hour chocolate chip cookies. You need to spend hours watching wildlife, you need to spend 15+ min making your coffee. You need to breathe in and breathe out. You need to be slowwwwwwwwww.
Unpopular opinion:
Unless you’re just prototyping, you should aim to understand as close to 100% of production code generated by LLMs.
Yes, all of it.
Effective mental models are still important for humans to sustainably maintain and evolve a codebase via prompting alone.
Drove a 10K+ LoC architecture change with an agent. After a while, it kept cutting corners, escaped lint rules, adding todos, and “partially done” remarks. The solution: