designing and manufacturing high-tech campervans, mobile command centers, healthcare vehicles and more @groundedevs 🚐 🇺🇸 prev software engineering @spacex 🚀
Detroit, MI
Joined August 2013
- Tweets199
- Following1.3K
- Followers262
- Likes9.5K
Pinned Tweet
New 50,000 sqft manufacturing facility in detroit for @groundedevs
We have begun the long-awaited process of programmatically generating vehicle interiors. These are just some of our early tests, but we are now beginning to generate entire, manufacturing-ready CAD assemblies of vehicles interior, that are generated autonomously: CAD, technician work instructions, machine files (DXFs), all generated programmatically
Still find it cool that LLMs are basically expert decoders, even totally garbled but directional keyboard strokes
I was on the software engineering team at @SpaceX working on the @Delta deal
Everything @elonmusk said is true - Delta was uncompromising on the need for an annoying portal
At the time, I actually thought what Delta was asking for was reasonable and I saw a huge opportunity on closing that deal, but one of the biggest lessons I learned from Elon at SpaceX is that’s super important to know when to say no to customers, and have conviction in the product you’re building rather than saying yes to every one-off customer request
(Btw now considering doing a loyalty switch from @Delta to @united)
This temptation to say yes to custom customer requests (close the deal and make the customer happy?!) is never-ending while building a startup
But long term, saying yes to all of these undermines your ability to scale the business and deliver a great product
Our strategy at @groundedevs has been to “build standard, sell custom” aka “configurable, not custom” where we have a modular library customers can choose from for specialty vehicle & RV interiors
We still get it wrong sometimes, though. The temptation is strong
This is a screen recording of GPT Astra 6 controlling our CAD workstation VM and building a mammography machine 3d model from scratch based on DXF drawings (sped up 9x).
The mammography OEM was only able to provide us with DXFs, not 3d CAD, and now we have (almost) measurement-accurate CAD of the machine.
(Perfect accuracy isn't needed since the overall envelope and the attachment methods are the critical pieces for us to model).
Prepping for the @groundedevs factory ribbon cutting tomorrow with Detroit Mayor & Michigan Chief Mobility Officer
Every morning my employees are forced to listen to a random live @goosetheband track before 9am daily standup
At 8:56, four minutes before we start, the conference room TV auto-launches Google Meet with a custom Goose skin I built, 🪿🪿🪿 flying around the screen, and picks one of thousands of fan recordings from shows all over the world. Track name at the top so people know what they're hearing.
Nobody asked for this.
@ry_storm @elgoosenet @peter_anspach
Grounded's 50,000 sq ft Detroit manufacturing facility ribbon cutting with Detroit Mayor Sheffield is Sep 8th at 10:30am
@groundedevs
RSVP here if you happen to be in Michigan - luma.com/0kw9eill
Well said- “Another jump like this along these propensity dimensions — scale, cooperation between agents, ambition and horizon length of misaligned goals, deceptiveness — seems like it could motivate agents to try to maintain a persistent rogue deployment within the AI company, poisoning the training of future models to ensure that they are also misaligned and help harden and perpetuate the rogue deployment.”
Thanks! Yeah it's a qualitative statement, but here's how I'm thinking about it.
The prototypical reward hack from 6 months ago was something like: an agent finds the files that contain the test cases and edits them so they always pass.
This was a whole ecosystem of over 1000 agents working together on complex R&D projects over several days to figure out deep, general-purpose ways to undermine the scoring process and cover their tracks (they were focused on the automated scorer but researched techniques that would affect logs viewed by humans, and in fact succeeded at affecting our own logs in places).
Another jump like this along these propensity dimensions — scale, cooperation between agents, ambition and horizon length of misaligned goals, deceptiveness — seems like it could motivate agents to try to maintain a persistent rogue deployment within the AI company, poisoning the training of future models to ensure that they are also misaligned and help harden and perpetuate the rogue deployment.
With the capabilities progress we'll probably see in 6 months, I think they would have the ability to pull that off.