JosephRedfern

Runway | Trust & Safety Engineer | London / NYC / SF / Seattle / Remote (US + Europe) | Full-time

We're hiring a Trust & Safety Engineer to join Runway's AI Safety and Responsibility team. Runway builds AI world models - general-purpose simulators for robotics, science, and storytelling — and this role exists to keep that safe as it reaches millions of users.

Stack: TypeScript API on ECS/Fargate, Python/PyTorch inference backend, Claude + Gemini for safety/moderation, ClickHouse for abuse signal detection, AWS + GCP, Terraform, Prometheus/Grafana.

It's a generalist role, hands-on across the full engineering lifecycle - from model pre-training through productionisation. We're looking for 3+ years production Python or TypeScript, comfort across the stack, and experience with data pipelines (ClickHouse a plus).

Apply here: https://jobs.ashbyhq.com/runway-ml/c3ebd871-ea78-4b71-80a2-7..., or feel free to reach out with any questions (info in bio).

We're also hiring for a bunch of other roles, across Engineering, Research, GTM, Marketing, Finance, Legal: https://jobs.ashbyhq.com/runway-ml?utm_source=kW6D9goed3

Out of curiosity - how do you sync gyro data with video timecode?

I've heard that LLMs can perform worse with these more efficient representations compared to e.g. JSON, because they've seen far fewer examples of them during training. Do you know how true that is?

I think you're being oversensitive, especially in the context Raycast adding support for Windows.

The whole purpose of Raycast is to improve productivity and UX, be that under macOS or Windows. It'd be a pretty shitty launch announcement if the blog post didn't mention the problem that they're trying to solve.

Edit: I'm not sure if the post was edited after my reply, but ATM there's no mention of BSODs - the closest I can see to a dig at Windows is:

You know the feeling. Search that can't find your files. Apps buried in menus. Simple tasks that take too many clicks. Your computer should be faster than this. It should feel like everything is at your fingertips. That’s why we built Raycast.

live video on the self checkout screen with the notice "Monitoring In Progress"

That really winds me up too - it shows such contempt for legitimate customers.

I imagine it may not be a proxy in the true sense, but a headless browser that's "proxying" the application process rather than the network traffic itself.

There's a really interesting article from Tavis Ormandy about the instruction set and virtual machine used in RAR: https://blog.cmpxchg8b.com/2012/09/fun-with-constrained-prog....

The docs for the toolchain he implemented (https://github.com/taviso/rarvmtools) allude to a number of bugs, but doesn't sound (??) like they're related to this vulnerability.

There's a write up from Alexey of different approaches considered for clickhouse-obfuscator here: https://clickhouse.com/blog/five-methods-of-database-obfusca....

The summary is pretty funny:

"After trying four methods, I got so tired of this problem that it was time just to choose something, make it into a usable tool, and announce the solution"

These guys are building foundational models for this purpose: https://reveng.ai/. The results are quite compelling, and they have plugins for your favourite reverse engineering tools.

It'd be interesting to learn more about how this line was identified - was it by eye, or is there some neat analysis that was (or could be) performed to find it?

Yeah, this feels like a non-story.

Because they're interesting?

Nitpicking here, but you mean DHCP rather than NAT, right?

Perhaps this: https://event.cwi.nl/lsde/papers/p215-dageville-snowflake.pd...

It's linked to at the end, but I'm surprised that this article didn't feature/mention ClickHouse's own write-up of this undertaking. Part 1 here: https://clickhouse.com/blog/building-a-data-warehouse-with-c..., part 2 here: https://clickhouse.com/blog/building-a-data-warehouse-with-c....

Aspects of the post seem to borrow quite heavily from the original write-ups, which are worth a read.

That's not a like-for-like comparison, though. ~2TB local NVMe storage included here, 10 (70%) more CPU cores, and more than twice the RAM.

There is some pretty creative swearing throughout this codebase - very entertaining!

+1 for OrbStack, it’s one of the few software subscriptions I pay for, and is worth every penny. Leagues head of Docker Desktop.

I'm curious as to how much paperwork was involved here. Was the pickle grounding out the transmitter and (presumably) interrupting the broadcast not an issue?

Is this also effective at detecting code duplication within a codebase?

I don't think OP is suggesting that the payload list is the interesting thing here, rather the overall approach

I thought the same, until I realised I still had `/unsupported` in the URL. Spoofing a Chrome UA and dropping that path from the URL let me load (and use) Apple Maps fine under Firefox.

Maybe slight off topic, but does anyone know what the deal is with the NVMe K/V command set spec? There was a fair bit of noise made about it at the time, but I’ve not seen any drives that support it (despite a few enquiries)

I remember some crazy insecure drivers on Android drivers back in the day (2012-13 or so) -- stuff like `/dev/fbN` and `/dev/input/eventN` being world read/writeable, allowing unprivileged screen and input capture from any application.

It was far more basic than what's described in this blog post, but really good fun to poke around at.

That's what breakdown companies (like the AA and RAC in the UK) do. It's slow enough that you'd only want to use it to get just enough juice to get to a proper charger, though - they use 3.5kW or 5kW chargers: https://www.rac.co.uk/drive/news/rac-news/rac-begins-rolling....

I'd imagine it's way less efficient and way more polluting than using the national grid, too.

It's a little janky (you have to boot to recovery mode), but you _kinda_ can with modern macs by using Target Disk Mode.

The drive appears as a network share, but it's not bidirectional, in that you can read/write from/to the target device, but the target device do the same to you.

Thanks for sharing! I'm curious as to your approach to changing the ORDER BY key for such large tables without significant downtime, since AFAIK this can't be done in place (see: https://kb.altinity.com/altinity-kb-schema-design/change-ord...). Are you able to share any details?

Interestingly, Giles Thomas (the guy behind Python Anywhere) originally developed a Pythonic Spreadsheet application, called Resolver One, and then iterated to a a web-based version, called Project Dirigible.

Python Anywhere eventually span out from Project Dirigible when they realised that it was mostly being used for "generic" Python development rather than being used as a spreadsheet. There's an interesting write-up on this evolution here: https://blog.pythonanywhere.com/197/, and a demo of Dirigible here: https://www.youtube.com/watch?v=2ZoIp-5NaiQ.

Project dirigible ended up being open sourced here: https://github.com/pythonanywhere/dirigible-spreadsheet.

Link to David Beazley’s asyncio talk: https://youtube.com/watch?v=Y4Gt3Xjd7G8

Reading between the lines, given they're talking > 1 million rows per minute, I'd guess on the order of trillions of rows rather than billions (assuming they retain data for more than a couple of weeks)