What is truth to a machine?

Welcome to edition #12 of the “No Longer a Nincompoop with Nofil” newsletter. I missed a week because of a terrible flu but I’m back now, and with a very special announcement.

ps. For $5/month I’ll send you a second newsletter and provide exclusive access to tutorials I’m working on that leverage AI. If you’re interested you can sub below

Here’s the tea ☕

• ChatGPT vs TruthGPT 🧐 

• Open source is getting serious 📈

• Google sure is dancing 💃 

Truth is subjective

There’s a reason why so many people are worried about privatised LLMs becoming the norm. Private companies having a monopoly on AI models means they get to subject everyone to their own biases. You have to understand something - the bias inherently within ChatGPT comes from the fact that a few hundred engineers and designers in OpenAI decide what type of responses it should produce. This is why competition and open source are so important moving forward.

There is, however, a slight problem. Creating and training an LLM like ChatGPT is expensive. Very, very expensive.

Which is why someone like Elon Musk has no issue creating his own ChatGPT alternative - TruthGPT. To put into perspective just how expensive it is to try and create your own AI model, Musk spent a minimum of $100M to buy 10,000 GPU’s. Why? Musk wants to create an AI that “understands the nature of the universe” and is “truth telling”. A very noble cause I must say, but there’s just one problem with this. Who’s truth will “TruthGPT” convey? Truth and reality aren’t concrete concepts that exist in the digital age. To think an LLM is capable of truth telling is nonsensical when you understand the way in which it is trained.

Regardless, it’s still a good thing Musk is doing this. Competition is good, especially in AI. I just can’t help but find it funny that this comes just a few weeks after Musk signed an open letter calling for a pause on giant AI experiments. Some say it was just to slow down OpenAI so he could catchup. Regardless, it’s become clear now why Musk even purchased Twitter in the first place. What better place to get the gigantic amount of text data required to train an LLM than a platform with over 237 million daily users. Twitter as a company doesn’t even exist anymore - it’s all under a new company Musk filed very recently called X.AI.

Regular folk like you and me can’t ever be a part of this battle between tech giants and billionares on who gets to create and control the most important piece of tech ever made. This is why the future must be open source.

Open Source is winning

I didn’t think I would be saying this, but open source contributions have been amazing recently. Meta released SAM (Segment Anything Model) that can cut out any single object from an image. They also released a model to animate childrens drawings. I’m genuinely impressed with Meta’s work in releasing open source models.

Do I think they have something to gain in doing so? Absolutely. You see, Meta has been investing heavily in VR tech over the last few years and the sudden boom in AI advancements means better experience for VR users. So much so that I’m convinced people will be addicted to simulated VR worlds in the future. I’ll be talking more about this in Thursday’s newsletter.

Meta also released LLaMA, a large language model similar to the models powering ChatGPT. They released four models, all with varying sizes. It’s great work by Meta but there is a catch. These models can’t be used commercially. I can’t train my own AI model that uses LLaMA and then try to sell this model to businesses to use. So in reality, although very useful for the general public, it’s not entirely “open” source.

I previously spoke about LAION, a non-profit company in Europe working to democratise and open source powerful AI models for everyone to use. Last week they released by far the best open source ChatGPT alternative I’ve used called Open Assistant. It’s really good. It amazes me that a company that relies on research grants and donations is able to release an AI model that is almost on par with ChatGPT, and it’s all open source. You can see the code for the data, the model, the weights - everything. The problem however is that any tool built on top of Meta’s LLaMA model can’t be used commercially. That means I can’t train it for my very specific use case and sell it to businesses for them to use.

So will we ever have a proper, fully open source ChatGPT alternative? Well, I did miss a week of newsletters, and in the AI world, that’s like 6 months of progress. A monumental release occured that flew under the radar. Most people aren’t aware of it or don’t understand its significance. Why? There’s no demo, no fancy videos or release events. So what was released? Data. A massive amount of data.

Red Pajama is an open source project that’s going to replicate Meta’s LLaMA model and make it open source. The company behind this project, called Together, is led by professors from Stanford, ETH Zürich and former employees of Apple, Google and other big tech companies. They’re slowly releasing the data they’re going to use to train and build AI models that can be used by anyone, for anything.

To put it simply - lets say I deal with senstive data that I can’t give to ChatGPT. With Red Pajama, I could train my own personal AI model for my specific use cases and not have to worry about sending my data to a private company. This is the future of AI - a place where each and every person and business will have the ability to train an AI to do whatever they want. An AI model trained by the people for the people.

There was actually a whole lot more that happened in the open source world but I’ll cover those in future editions.

Google is.. moving?

A few months ago, Microsoft CEO Satya Nadella said that he wants people to remember that they made Google dance. Well after years of mediocrity and sluggishness, Google is finally doing things, releasing products and features.

When Google first announced Bard, their competitor for ChatGPT, not only did they lose in the naming department (Bard!?) but also in the capability department, specifically coding. Anyone who codes and has used AI to help them code understands that the future of programming is a combination of human and AI. So when Bard couldn’t code for the life of it, people were understandably very disappointed.

That’s not the case anymore with Bard getting an upgrade, and it’s a good one. Bard can now write code, edit code and even answer questions on code and repositories. It’s almost on par with GPT-4’s coding capabilities. This really shows how far Google is being pushed by OpenAI. The startup is dragging the big tech giant out of its slumber.

But that’s not all. I know announcements aren’t very useful and Google is the king of simply making announcements without actually shipping anything, but things are different now. Google must ship. So what will they be releasing? It’s called Project Magi - an AI powered search engine with the purpose of creating a more personalised user experience.

The good thing for Google is that distribution will never be a problem for them. They simply need to release a product and people will use it because the barrier to using it can’t be lowered. Obviously the product still needs to be good (rip Google Hangouts). This is a clear response to not only ChatGPT, but new AI search platforms like Perplexity and Neeva that are similar to Google, but without the ads and better UX.

I must admit, when ChatGPT first got announced and became big, I also got sucked into the hype and thought perhaps Google will become the next Kodak or Blockbuster. I forgot the power of distribution. Do I think Google will take losses from all the new competition? Absolutely. Will Google become obsolete? I think it would take colossal failures and inaction for this to happen, and I don’t see it happening.

Larry Page, co-founder of Google, has openly stated numerous times that the reason Google exists is to reach AGI (Artifical General Intelligence). Google pioneered the technology that ChatGPT uses. They have been at the forefront of AI research for over a decade. I think it would be silly to count them out just yet.

If you’re new, you can subscribe here 🙂 

I create weekly roundup posts on reddit. You can check out the last two here and here.

As always, Thanks for reading ❤️

Written by a human named Nofil