In this newsletter:
📝 Post: How Much Freedom Should We Give AI?
🗞️ In Case You Missed It: SpaceX and Starlink
🗞️ In Case You Missed It: Tech Headlines
😎 Pick of the Week: Food and Drink
📦 Featured Product: Every Day Carry
📝 How Much Freedom Should We Give AI?
Last week I talked about choosing a specific model within your chatbot of choice. But as mentioned, this has changed over the last few months, much less over the last few years. The highest reasoning models are getting much more powerful. This week, OpenAI claimed their next major model Astra had breakthroughs on 10 long-standing math problems.
Just in the last couple of months, we also saw OpenAI’s unreleased model Escape and Hack a Company in Cybersecurity Test. Except they weren’t alone. A few days later, Anthropic says its AI models hacked 3 organizations during testing. What’s interesting is this comes a month or so after the US Government ordered Anthropic to block all public access to Claude Fable before removing it a few weeks later. You can read more about Why the US Government Shut Down Claude Fable.
However, it’s very important to note that none of these bots decided to start hacking on their own and become sentient. They were being tested with a given objective, and some level of parameters to work within. Essentially, the chatbots found holes in the rules of the game but did so without causing any harm.
Traditional software works because someone tells it what to do and it follows the rules. I talked about this in my post about The Power of Automation Software to Upgrade Your Workflows. AI functions very similar, except there’s not only one way to solve the problem its given. The more capable and powerful the specific model becomes, the better it becomes at finding multiple paths towards the objective, or at least different paths. And sometimes, these paths are unexpected.
What’s interesting is that sometimes what programmers and researchers use to help train the models seems to come out of left field - and it works. This is akin to how humans function, too. Have you ever had a problem in one area of your life, move to work on another problem, and it hits you. You now know how to solve the first problem. Famously, Mark Zuckerberg said he gave up running as exercise to take his mind off work because ‘the problem with running is you can think a lot’.
That said, AI can improve by doing something “new” rather than these companies continuing to throw money at compute power. One new way AI Companies are combating this is by Buying Tons of Old Books Because They’re Free of AI Slop. This is for one main reason: the information written within the books are written entirely by humans and no forms of AI or machine learning programs. With the amount of content being uploaded to the internet everyday, the amount of information isn’t what’s helpful. What the AI bots need is good information - or at least, not other AI generated information.
A friend shared a video with me that was pretty interesting. In it, computer scientist Stephen Wolfram shares how The universe follows rules, but that doesn’t mean it’s predictable. He describes a concept called “computational irreducibility” that essentially states as systems (and AI by extension) become more and more complex, there isn’t always a way to figure out what they will do or how they function. He gave the example of a horse - we know how to train them, ride them, and how they generally function - but they can still may surprise us with their actions.
This made me think of a mobile phone. For most of us, we know how our phones work, but we could never troubleshoot most hardware issues, much less build one from scratch. Like the horse and our phones, we know how to use AI, we know generally what it will do and how it works, but there’s still this black box concept happening that data goes in and data does out and we still don’t always know the outcomes.
Which is where the AI battle lives. Giving it too many guardrails means we restrict what it can’t and can do - making it far less useful, but potentially more accurate. Giving it too much freedom makes it more capable and powerful, but also less predictable and often times flat our wrong.
This doesn’t mean AI is becoming sentient as I mentioned already. At least yet. And I don’t know anything or think they will. I’m also not trying to scare anyone, just making an observation. The companies saw this happen and were able to contain it. But that also means the next test may try to do more and could do so without being caught…
The more access we give AI, the more we train, and the more it’s allowed to do, the more issues like this will surface. So the questions of “What can AI do?” or “What model do I use” may be replaced by a much deeper and more difficult to answer question of “How much freedom are we willing to give up?”
🗞️ ICYMI: SpaceX and Starlink
Just to remind you: SpaceX is “revolutionizing space travel by carrying both crew and cargo to Earth orbit, the Moon, Mars, and beyond. Starlink is a satelitte internet service offering fast, affordable internet to rural areas. Starlink is owned and operated by SpaceX, and has recently hit 12 Million Subscribers. On the mobile front, SpaceX Outlines Plans to Take on AT&T, Verizon and T-Mobile. They already offer a satellite mobile service through Starlink Mobile, so maybe it’s speed, availability, and price is ready for the mainstream. In order to launch more and more satellites, they are also expanding physically. SpaceX is set to acquire 130,000 acres of marshland in southern Louisiana.
🗞️ ICYMI: Tech Headlines
Apple Plans iPhone-to-Windows Copy and Paste in EU After Microsoft Request. We’ll see if this is available early next year in the US (or even globally outside of the EU).
Bending Spoons to buy Airtable for $1.28B. If you use Airtable in your business, you will have a bit of runway to offload your databases should you choose. If history repeats itself from their other buyouts, Airtable will in fact get more features, but that will come with a higher price tag, too.
Ford announced a new small-to-midsize electric truck - the Ford Fathom. Check out this article with more information: Ford Fathom Electric Truck: What We Know So Far.
OpenAI is giving ChatGPT free users unlimited text chats. I still stand by the argument that if you are using AI a lot, it’s worth paying for as the token usability becomes much more useful.
😎 POTW: Food and Drink
It’s about time for another round of food picks! Enjoy some niche lists in the food and drink space:
📦 Featured Product
Over the last few months, I’ve shared some products on pens, fidget toys, and multitools, but here are a few more items that fit the Every Day Carry (EDC) category.
