Can AI's Hallucination Problem Be Solved?
Confident, yet inaccurate answers are seemingly part of the charm of LLMs. But how serious is the AI hallucination problem and will it ever be completely solved?
This is some text inside of a div block with Cleo CTA
CTASigning up takes 2 minutes. Scan this QR code to send the app to your phone.
A quick, non-technical look at open AI models and why the matter so much
In the last few weeks, you’ve probably seen a few headlines talking about open source AI models
Or maybe you haven’t because you don’t work at an AI company and aren't addicted to Twitter. (Sorry - still Twitter)
Either way, what’s happening around the world is that lawmakers are all thinking about AI and how it should be regulated.
From questions about training models on copyrighted material to the threat of misinformation at scale– it’s going to be tricky to find a balance that’s pro-public welfare and pro-innovation.
Amidst this uncertainty and the fact that AI is probably going to change every aspect of our lives, there's been an equal number of questions raised around who should own these algorithms and how can we make sure large groups of people aren't left behind.
So. Here's a couple of high-level hot takes about open source AI models and why they matter.
Let’s start with the basics.
The meaning of open source, as it relates to software, is that the original source code is accessible to the public at no cost and can be modified or redistributed however you like.
If there was such a thing as a free lunch, it would be open source.
Some better known open source projects would be Firefox, the coding language Python, and Linux (which we’ll talk more about at the end). Oh and the freaking internet.
The opposite of open source is what’s called proprietary software. This is where the code is kept confidential and you have to have a license or buy the software to use it.
*cough* *cough* Adobe.
But there’s a really important nuance we have to recognize between open source software as we traditionally think about it and open source LLMs. (If you’re not sure what an LLM is – we’ve got your AI 101 education covered. Spoiler: It means large language model and is pretty much another name for AI)
With open source software, it’s like looking into an ant farm. You can see the source code, the logic and most importantly – you’re able to predict what the system will do. This is called deterministic.
You can know the output by looking at the input.
Ok now things get a bit nuanced. When we talk about open source LLMs there’s not one specific meaning.
It could mean that the model is free to use. It could mean the source code or training data is transparent as well.
Usually it’s some combination of the above, but more often than not the code and data are not visible. Or if the data is available, it’s usually massive.
Where this gets even more tricky is that these aren’t deterministic systems. So even when you’ve got access to the model or the code, it’s still hard to know exactly what the output will be.
(Side note: this is also a big part of our work on Cleo is understanding when your chats with her are actually helpful and if it’s giving you the information you want.)
So it’s not apples to apples when comparing open source software to open source AI, especially in terms of transparency and predictability, but the cost is often the same – free.
Although, even that gets a bit murky.
Pretend you had the resources and the questionable sanity to start a business.
Would you work 100 hours a week and take on silly amounts of debt in an area that lawmakers are still a bit like “hmm, this seems dangerous?”
No - you wouldn’t.
Now put on your AI lawmaker goggles for a moment. Which type of software would terrify you less? The one where there is some degree of transparency or the black box?
In fact, France’s strategy to be a major center for artificial intelligence within the heavily regulated EU is to embrace the development open source LLMs.
There’s also the teensy problem that when you build on someone else's software they can, ya know, just change it. Or start charging you for it. Censor your products. Or literally anything else they want to do.
This is a huge point which we could probably do 10 more blog posts on.
Do you know what’s smarter than a room full of super geniuses making 800k a year?
The world.
From leaked internal memos at Google one senior developer said
“...the uncomfortable truth is, we aren’t positioned to win this arms race and neither is OpenAI. While we’ve been squabbling, a third faction has been quietly eating our lunch. I’m talking, of course, about open source. Plainly put, they are lapping us. Things we consider ‘major open problems’ are solved and in people’s hands today.”
He then goes on to list several features that the global open source AI community has been able to solve what the Bardians have not. Just some small things like LLMs on a phone
or a scalable personal AI you can literally fine tune your own AI at home in a few hours.
Power to the people indeed.
It’s really a brilliant read and I’m not doing it justice here. So if you need a bit of hope for the future give it a look.
Speaking of hope for the future…
What could possibly go wrong?
We won’t pound the table here or pull out some Cambridge Analytica, but the point is, if an LLM exists in the public domain in a transparent manner it becomes harder to abuse. It’s inherently more supportive of privacy and less likely to security issues.
That said, a recent grilling of Mark Zuckerberg revealed that maybe putting these powerful tools in the public’s hands could potentially be more dangerous than beneficial
Earlier this year in September, Mark Zuckerberg was having a closed door chat with 7 members of congress on the dangers of AI, particularly Meta’s (kind of) open-sourced Llama 2 model.
We know what you’re thinking. But it’s a llama. They’re so great.
Well first of all, you’re thinking of Alpacas:
Secondly, during the meeting it was revealed that Llama 2 gave step by step directions on how to make Anthrax– a nasty nerve gas.
The senator’s point was that is it really in the public’s best interest to be able to access this kind of information?
Fair point. But the policy team at Cleo (there is none) would counter that point and say that access to information is very different from access to materials.
You can technically find out how to make a nuclear weapon without using an AI. Good luck getting that uranium though 👍
Ok, we don’t want to make light of what is a serious topic. To buy potentially dangerous things, like a hand gun, there’s a background check and licensing system in place. How does that get applied in this situation? Can it?
Yet we do come back to the point that if someone is truly set on causing harm, they’ll find a way to do so, regardless of the restrictions in place.
So how much of this is a genuine concern and how much of this is the bigger players with private models leaning on their pals in congress to protect their moats?
Time will tell.
It’s hard to imagine but back before most of us were born, the 90s, the internet was just getting started.
You had some big companies, specifically Microsoft and Sun Valley Systems (don’t ask) both trying to create the “web” using their own, proprietary operating systems..
This was also around the time that a guy called Linus Torvald released his own open source operating system called Linux.
For the sake of speed, I’m going to skip to the end of the story.
Because Linux was free to use, transparent, and widely accessible it became the backbone of the internet we know today.
Literally greater 70% of websites and 90% of cloud servers use Linux.
Can you imagine if anyone who wanted to do anything related to the internet had to get a license from Microsoft? No offense, but no thanks.
Since then, thousands, if not millions of companies have come along and build on top of the open standards and protocols the internet represents. And it was made possible because for the reasons just mentioned.
We know there’s some important nuances when discussion open vs. closed LLMs, but from first principles – open source is the way.
Looking to work for an AI company?
Confident, yet inaccurate answers are seemingly part of the charm of LLMs. But how serious is the AI hallucination problem and will it ever be completely solved?
Can an AI be conscious? Alex Brody as he unpacks one of the many ethical questions AI researchers and modern philosophers are trying to answer
Explore how AI money apps are shaping the future of finance 💸