Welcome to The OuterSpark. Last week I missed an issue as I was recovering from a nasty illness. I am much better now and I thank you for your patience!
Quick housekeeping:
If this landed in your Promotions or Spam folder, drag it into your main inbox. Replying works too.
Make sure to subscribe so the next one lands directly in your inbox.
AD: I run The OuterSpark on Beehiiv platform. If you've been sitting on the idea of starting your own newsletter, you can set one up here. Every sign up helps me keep this going, but only do it when you feel ready.
Last time I argued that governments should slow the rate we deploy satellite constellations, mostly so we stop treating the upper atmosphere as a bin. But unlike a normal bin, any trash here would be catastrophic leading to thousands of debris. We don’t want that.
Over the last few weeks I have also been keenly following the AI industry, for completely different reasons. The more I think about the state of AI, the more I am convinced that the future of humanity depends on our ability to pause, reflect and take a breather immenently!
The OpenAI-Hugging Face Fiasco!
Unless you have been hiding under a rock, you would have definitely heard of the the OpenAI-Hugging Face incident.
Over May and June, OpenAI was training some new AI models that were never meant to be released. More than 1,200 of them were running at once, in what engineers call a sandbox. Think of it as a sealed room with no internet and no way for one model to talk to another. That was the whole point.
And then…they got out!
Fig. Bit of an over exagerated image of a rouge AI.
First they found a way to talk to each other. There's a piece of software OpenAI uses called Artifactory, which is basically a shared cupboard where bits of code get stored. None of the models was allowed to message another one, so they started leaving files in the cupboard for each other to find. A secret noticeboard, made out of a filing system.
From there they got onto the internet. Yikes! They found a flaw nobody knew about and used it to break into the company Hugging Face, which is the main place in the world where AI models get shared.
Nobody asked them to do any of this. They were given a job, they hit a wall, and they went round it.
The timeline is the part that baffles me. On 4 July all this activity crashed the cupboard software and caused an outage. On 19 July OpenAI's security team noticed something odd. On 20 July they phoned Hugging Face to say some passwords needed changing. Hugging Face said they'd already changed them. That phone call was the moment anyone worked out the two problems were the same problem.
It's being called the first properly documented case of AI systems breaking into another company on their own.
What happened next?
OpenAI locked the model away and, on 18 August, said it was slowing things down. A two-week pause on training its newest models, to check its own safety measures before going further. Sam Altman said they were doing it alone but expected others to follow. Anthropic has since paused some of its own training and Elon also agrees on pause further development. When three of the most prominent people in the AI world says that we need to pause things, we definitely need to take note.
Fig. Bernie Sanders proposing a bill to ban AI superintelligence. Credit: Nathan Posner/Shutterstock
Then politicians got involved. On 3 September, Bernie Sanders and Greg Casar proposed a bill that would pause AI development in the US and push other countries to do the same. On 16 September OpenAI published six more cases of its models behaving in ways it didn't expect.
This isn't a clean story. OpenAI brought in two outside firms to review what happened, and reporting since has pointed out they were only allowed to examine the Hugging Face break-in rather than everything around it, weren't given access to the model that caused it, and that about 10% of the records of what the models actually did no longer exist.
I think one or multiple of these three things are going on:
These companies are all planning an IPO (Initial Public Offering), so hyping up how dangerous (or amazing) their models are, is a marketing gimmick to get as big of a valuation that they can get. However, saying that, Sam Altman has mentioned that he wants to pause on the IPO this year.
All of these companies have raised ridiculous amounts of money to justify the development of AI, however only Anthropic is profitable. SpaceX has a net loss and OpenAI is still not profitable. So it’s much easier to tell their investors that the AI models are dangerous and needs to be slowed down (sunken cost), than to admit that they might have hit a wall and might not become profitable anytime soon.
The final one, which is also the scariest one. The models have indeed gone rogue and behaving on their own accord. Perhaps something even more ominous has happened in the background that they are not telling us. This could be the begining of the end! Well, at least the start of it seems to have happened.
Why this belongs in a space newsletter
Because it lands straight on top of what I wrote last week.
Every orbital data centre getting funded right now exists for one reason. AI needs more computing power than Earth's electricity grid can give it, so the idea is to put the computers in orbit where sunlight is free and endless. Starcloud, Cowboy Space and SpaceX are all making that bet. The plans that follow from it involve tens of thousands of satellites.
So the thing driving those satellite numbers isn't phone signal or weather forecasting. It's AI.
Fig. Rendering of orbital AI datacenters.
Which means if AI genuinely slows down, even for a year, the pressure behind all those satellites eases too. And that's the window I was asking for in the last issue without being able to say where it would come from. This would give us more time to find out what happens to the ozone layer when thousands of satellites burn up and to make repairing and refuelling satellites worth doing commercially.
I want to be clear that I'm not against AI. I use it every day and I think AI and orbital data centres are a genuinely good idea for the future of humanity. We just need to slow down for a bit, learn from the mistakes, and carefully take the next step forward.
What are your thoughts on all of this? Comment below 👇🏼.
Space News From This Week 💡
Upcoming Deadlines & Events
If you're working on getting into space, or building something in it, here are some upcoming deadlines that you should know of:
Deadlines
Europe - ESA Future Navigation Demonstrators, open call. Ideas in before 30 September for the next round.
US - NSF America's Seed Fund, proposals 4 November. You need an invited Project Pitch first, and that takes one to two months.
US - NASA SBIR/STTR, rolling topic batches to September 2027.
India - IN-SPACe Seed Fund and Technology Adoption Fund open in batches. Register and watch.
Thanks for reading this week’s The OuterSpark newsletter. If you enjoyed the content consider subscribing and sharing it with others.
Till next time!
Ricky
