Llama 2 Shakes Up the Entire AI Industry

Microsoft lands another blow + ChatGPT really got dumber + OpenAI is in trouble + We got a Sponsor!

Welcome to edition #21 of the “No Longer a Nincompoop with Nofil” newsletter.

Subscribe 🖌️ | Go Premium 🚀 | Work with me 🤝 | Discord 🤖

Free Consults for ALL!

Last week I launched Time and Money, an agency focusing on leveraging AI to help companies save time and money. This week I’m introducing completely free discovery calls. If you’re business is looking to incorporate AI or even get educated on what’s possible, let’s talk!

Here’s the tea ☕

  • Meta disrupts the landscape 🦙

Setting the stage

There have been many advancements in AI models besides ChatGPT recently. Google Bard has gotten big upgrades, including being able to read images and extrapolate information from them. Anthropic released Claude 2 which has gotten significant improvements and some even consider it to be a better programming assistant than GPT-4. It does help that its context window is absolutely massive - 100K tokens. Inflection’s Pi is a more conversational AI bot and although it may not be as good at coding as ChatGPT, it’s far and away the best at having human-like conversations (in my opinion). If you haven’t tried it, I’d highly recommend it; it even has a built-in voice feature so it’s like you’re talking to your own JARVIS.

Even with all of these advancements and other players upping their game, I don’t think OpenAI or anyone really was concerned about ChatGPT. Well, it turns out Zuck is waging a free-for-all against his competitors. First, Meta released Threads, Instagram’s answer to Twitter. The timing couldn’t have been better considering Musk announced free users were going to be limited in how many tweets they could see. Within a few weeks, Threads was launched. There are two hilarious parts to this story.

Firstly, Twitter, under Musk’s rule, open-sourced its entire source code. Anyone can go and take a look at how their algorithms work. Secondly, since Musk fired over half the employees at Twitter, citing most of them as useless, Meta, and I presume other big tech companies have been quietly hiring these same exact people to work on their own products.

But this isn’t the talk of the town today. This is just an appetiser. Zuck has blown a gigantic hole in the AI arms race. No one knows what will happen next, but one thing is certainly clear. A lot of companies, startups and even proprietary software, are in serious trouble.

Open-source was and always will be the future

Meta released Llama 2 yesterday. What is it? It’s an open-source AI model. So why is it a big deal? Well, besides the fact that it’s good, which I’ll come back to, it can be used commercially. That means absolutely anybody can do whatever they want with it. Actually, not everyone. The terms specifically state that anyone with over 700M users can’t. Sorry Google 🤷 

So let’s take a closer look shall we? There are four models, all varying in size - 7B, 14B, 34B and 70B. For general purposes’ sake, let’s look at the biggest model as it will be the most powerful. Llama 70B is in many cases better than the current ChatGPT when it’s running on GPT-3.5. It’s still not as good at coding, but there are two reasons why this doesn’t matter. Firstly, most people won’t even be using it for coding so this is negligible. Secondly, since Llama 2 is open-source, and we can do whatever we want with it, what’s stopping us from further training it on massive code datasets, like the one StarCoder used, and make it better at coding? This will 100% happen. Someone will do it. That’s the beauty of open-source.

So why is this a big deal? Well, you can now have a chatbot that is, for all intents and purposes, as good as ChatGPT’s base version running locally on your laptop. No internet connection, no API’s, nothing. This is an incredibly powerful tool. Want to train it on a specific use-case but are worried about privacy? Just do it all locally. Open-source models will only get better, and then there will be serious questions to answer for OpenAI. Why would someone pay to use a model when there’s a free one out there? There is only one drawback I’ve seen so far with Llama 2 - we don’t have access to its training data. We can’t see what it’s been trained on unfortunately.

One of the biggest potential disruptors to both OpenAI and even Google is Perplexity, and they’ve taken full advantage of the situation. They’ve added Llama 2 to their platform and it is blazingly fast. If you’d like to read more about Llama 2 and how you can run it locally, see the digest at the bottom.

Edit just before release: Meta might just update Llama 2’s coding capabilities themselves…

The curious case of OpenAI

OpenAI was originally a non-profit company. They turned for-profit and accepted over $14 Billion from Microsoft and practically relinquished ownership to them. Fair enough - Microsoft is a behemoth; they have a lot of money, compute power, influence and distribution. Azure alone rakes in over $70B annually. Following this, Microsoft themselves released Bing AI. A direct competitor to ChatGPT, but clearly inferior. Fair enough, it is part of their search engine after all.

But now Microsoft has worked with Meta to release Llama 2. This company has their hand in every single cookie jar. Microsoft is owning the AI landscape. Make no mistake, Google has dropped the ball terribly throughout this entire situation.

The good & the bad for OpenAI

Things couldn’t look worse for OpenAI even if they tried. Actually, it seems like they have been trying. You see, many, many people online have been complaining that OpenAI has made ChatGPT dumber; mainly GPT-4 in this case but even 3.5. The VP of Product at OpenAI even addressed this on Twitter and denied it. Some speculated that it only feels dumber because the glamour has worn off. Well, we might just have some evidence now.

This paper illustrates how ChatGPT may have gotten dumber recently. In my opinion, there’s a very specific reason why this is a terrible look for OpenAI, especially coupled with the release of Llama 2. How can anyone build on top of the OpenAI API’s if the underlying models themselves have been proven to get dumber? The quality of thousands of new startups depends on these models.

How can any enterprise company trust that what they build now won’t be worse than what it will be in the future? An open-source model can only get better. It can be trained in-house, it can be controlled and fine-tuned - it can be trusted to be consistent. I can’t imagine the folks at OpenAI are quite happy at the moment.

The woes continue

But wait, there’s more!

At least this time around, OpenAI has no one to blame but themselves. Remember when they campaigned for government oversight and regulation? Remember when Sam Altman spoke to Congress and practically begged for them to regulate AI; speaking in incredible vagueness, knowing full-well that not only the people listening to him have no clue what he’s talking about, no-one has an answer to AI regulation.

Well, karma’s a bitch ain’t she. The FTC is investigating OpenAI for breaching consumer protection laws and guess who’s not happy? Believe it or not, Sam Altman. The same one that was lobbying the EU to take it easy with the regulation. You reap what you sow, Sam.

Short Digest

There is simply too much to cover in a short newsletter, so I won’t go in details for certain things. Instead, you can enjoy these points for further reading and exploration. Enjoy

  • Run Llama 2 vs other LLMs to see which is better [Link]

  • Llama likes Zuck and dislikes Musk. Shocking [Link]

  • Train Llama 2 on your local machine on your own data [Link]

  • Download Llama 2 by getting an access token and going to the repo [Link] [Link]

  • It cost Meta ~ $2.6M to train Llama 2. It’s super cheap (for them, not me 😢 ) [Link]

  • Llama 2 is on HuggingChat [Link]

  • Have private convos with your docs locally using Llama 2 [Link]

  • Qualcomm is working with Meta to bring Llama 2 to mobile devices by 2024. AI will power every piece of technology soon [Link]

I release much longer digests, massive ones actually, in my premium newsletters. Next one goes out tomorrow. There is a lot happening in the world of AI, and I can’t possibly write in detail about it all. If you’d like to take a glimpse into the sheer magnitude of the current technological revolution, you can subscribe here.

✨ From our Sponsor ✨ 

I’ve been fortunate enough that I haven’t had to rely on sponsorships to run this newsletter. However, even if I do promote a tool, I will always use it myself beforehand. Ultimately, sponsorships will help me continue writing these newsletters and educating people on one of the most exciting times in human history. Enjoy 🍷 

Streamline workflows and 10x productivity!

Text Blaze is a tool that helps people eliminate repetitive typing with keyboard shortcuts. With Text Blaze, you can create smart templates that work wherever you work.

As always, Thanks for reading ❤️ 

Written by a human named Nofil

Join the conversation

or to participate.