48 | ChatGPT will have news updates, Apple releases an open source model, Midjourney V6, and many more AI news for the week ending on December 24th

0:00Hello, and welcome to a short news weekend edition of Leveraging AI, the podcast that shares practical ethical ways to leverage AI, to improve efficiency, grow your business, grow your business, and advance your career. This is Isar Maitis, your host. And while we're all winding down for the holidays at the end of the year, the AI world is not. There's still an incredible amount of really important and interesting news that happened this past week. So let's get started.

0:31We will start this week's news with the biggest dog around, which is OpenAI. And there's a lot of news coming from them this week. The first piece of news I want to touch is the partnership between OpenAI and Axel Springer, which is one of the larger journalism organizations in the world. They own Politico and Business Assider, as well as Bild and Welt from Germany. And they signed a partnership with OpenAI. The partnership is supposed to provide quotes and direct access to news coming from Axel Springer's sources direct into ChatGPT, including direct quotes and links to articles. The goal of this is obviously to provide ChatGPT access to real-time accurate news, which it does not have today.

1:15So it does not have news because its latest data tracks now back to April of 2023. And also it hallucinates. And this new aspect will allow ChatGPT users and probably shortly after Microsoft users to get access to immediate news, including this specific information that comes from them with links to go and check that information. It will also expand Axel Springer's usage of OpenAI tools within their processes. This is not a surprising move. I think it's very clear to everyone right now that the days of go grab everything you can off the internet and you can train your models on that are over, which means training of new models will involve licensing deals.

1:56And in this particular case, it's a very interesting licensing deal because it provides OpenAI access to accurate real-time news, including paid behind the firewall news that only people who have subscriptions to these journals actually get access to. So very interesting deals, but I expect to see a lot more of those moving forward. Another fun and weird piece of news about ChatGPT comes from the fact that more and more users through the month of December complained that ChatGPT became lazy and that it's providing less of an outcome compared to what it did before and sometimes even does not provide any answers.

2:32And people started wondering whether it's taking, quote unquote, winter break, meaning does the fact that people are winding down through the holiday season also impacts ChatGPT because it's trained on human data. and a user on X called Rob Lynch actually tested that and it asked ChatGPT to write code telling it it's May versus telling it it's December. And it wrote significantly more code when it thought it was the month of May versus the month of December, which is really weird. And yet that is the reality. So ChatGPT is quote unquote getting lazy in December. But the good news is that if the reason is really the fact it thinks it's winter break, it should go back to full swing as soon as we all get back to work through the second week of January.

3:26Two interesting pieces of news about OpenAI which relates to safety that I'm not sure if it's related to the whole snafu they have with the board or not, but either way, they are buffing up their safety measures and looking to be safer in the future. This comes from two different aspects. One is they're growing and strengthening their internal safety process and their internal governance for risky AI systems. There's a new safety advisory group that will make recommendations to the leadership on risky models. The board was granted a veto power on the decision whether to release or not release potentially harmful AI models.

4:06The models would be rated for things like cybersecurity and disinformation and other potential for negative impact, and any model that will be defined as high risk will be banned from being deployed to the public. So this is one aspect. The other aspect is that OpenAI is now offering what they call super alignment research grants. Any research nonprofit and or university can apply for these grants for a total of 10 million. The grants are there to support research in super alignment. So those who don't know what alignment of models mean, if you think about when a model is developed or open source models, they're developed and then they're released to the public, which means they have no guardrails and no quote unquote alignment to whatever rules and regulations that humans want them to have.

4:53So aligning models are very important to keep them safe and within what we want them to perform versus steering away from that. Super alignment has to do with aligning super intelligent models, which means models that are more capable and more powerful than humans. So OpenAI is providing grants between$100 ,000 to$2 million to researchers to help them in figuring out super alignment, focusing on managing models that are potentially surpassing human level intelligence. Now, if I combined the last two things that I said together with everything that happened in OpenAI's leadership that we still don't know exactly what happened, it leads me to believe that they have models that are way more advanced than we know of that are potentially already more capable than humans.

5:50So they may have achieved AGI or aspects of AGI already, and hence before they're releasing these models to the public or before they're enhancing them even further, they want to invest as they should in additional safety. The problem with that is that we are allowing the cat to guard the cream. And what I would really like to see is some kind of a third party government or a quasi-government body that is overlooking at all of this for every company, whether open source or closed, and providing feedback on the level of risk that these models provide before any company or organization are allowed to release them.

6:31That might be wishful thinking, but I really hope we're going to get there very quickly because I feel that we're getting closer and closer to the point of no return. OpenAI also announced a new and very helpful feature, which is archive for chat. So if you're like me and you're a heavy user, you now have thousands of these rows on the left of your chat of all your chat history. Well, many of them you will never ever want to go back to and yet you don't want to delete them just in case you will ever want to reference them in something. So there's now an archive function which allows you to remove them from the list of history of chats on the left without actually deleting them.

7:10This functionality is now available for the web and iOS versions of ChatGPT and support for Android is coming soon. As I mentioned, very helpful if you're a heavy user that wants to manage, that wants to manage your history. I really hope that sometime soon they will also introduce folders that will allow us to better organize chat history. And in support of all of us, ChatGPT released their best practices on how to improve your prompts. And these include six tips on how to do better. The first one is write clear instructions. That's pretty self-explanatory. You need to be very specific in detailing what is it that you're trying to achieve, what's the desired persona, what's the length of the output that you want, et cetera, et cetera, but be very clear in your instructions.

7:55The second thing is to provide reference texts. It helps the model be more focused and reduces hallucinations. So the more context you're going to give the model with other texts that you provided, the more accurate and better the results are going to be. The third recommendation is to split complex tasks into a sequence of smaller prompts and subtasks in order to get better results. I've been doing this for a very long time, and I can tell you it makes a very big difference. Now, it's okay to still write longer prompts as far as background, but as far as the specific instructions you want to achieve, if there's multiple instructions, break it down into different steps.

8:35The other benefit of that, that they're not referring to, but that I use a lot, is if one of the steps is not doing exactly what you wanted it to do, you can stop it there basically and redo just that step without having to redo the whole thing, which is very helpful to progress in the right direction. The fourth one is interesting and they're saying you need to give the model time to think, meaning allowing the model to figure out a process rather than just giving you an answer gives you better answers. And one of the ways to do this is to ask the model to use chain of thought as part of your request for the answer.

9:13The fifth suggestion is to use external tools and plugins in order to get better results. So they're referring to things like evaluating images and using data analysis capabilities from what used to be external plugins and now is just built into ChatGPT, but enhancing the capabilities of ChatGPT with specific capabilities for very specific tasks and topics. And the last tip that they provided is to systematically test changes and evaluate the prompts across multiple use cases. What they're basically saying is the fact that a prompt is working very well on one use case doesn't necessarily mean it works best for other use cases.

9:53And also the model keeps on changing. So they're suggesting to keep on evaluating and testing variants in your prompts, especially for different kinds of scenarios and use cases. So these are the six tips coming straight from the horse's mouth, so straight from OpenAI on how to get better results with your prompting. I will most likely record a separate episode with a lot of examples on each and every one of those components in order to give you even more depth. But for this news episode, what we shared should be enough. And from OpenAI to their biggest partner, Microsoft, Microsoft has teamed up with a startup called Suno to add song creation to Microsoft Copilot.

10:35So users can now turn text and ideas through description into full musical compositions with lyrics, instruments, et cetera. This is obviously very cool or disturbing, depending on how you view all of these generative AI capabilities. But from my view, it's allowing creative people with great ideas who do not have the skills, whether in graphic design, writing, or in this case, music creation, to create more stuff that's overcoming their skill barriers for creation. So it's democratizing the creation of music, which I found really cool. Obviously, this can create issues of replicating songs of existing people, etc., which is a whole different issue that needs to be addressed through regulation and legislation.

11:23But from a creative expression perspective, I think it's fantastic. Bigger news from Microsoft is that they've just launched what they call GPTRAG, which is a machine learning library, which enables enterprises to create their own large language models using RAG methodology much easier than before as part of the Azure OpenAI platform. The goal is to allow large enterprises huge productivity gains when it comes to creating their own large language models while ensuring governance and data safety. The bottom line is it will allow large organizations to take their data that's already stored on Azure and use it as training data for large language models that they can use internally or externally while keeping their data safe, which I think is a goal of more or less every large organization in the world today.

12:12So a great move there by Microsoft. And from Microsoft to another giant that has been playing an interesting role in the AI world, but not in the forefront like the other big players, IBM. So IBM just teamed up with a company called Jackson AI. Jackson AI has been playing for a few years in developing AI solutions for military applications, which means the accuracy of information is critical. It cannot hallucinate. And together, IBM and Jackson are building a new infrastructure that they call DSAIL, which stands for Domain Specific AI Language. But the critical aspect of it is this new architecture allows to more or less eliminate hallucinations from AI models, which will obviously differentiate those models from most AI models that we know today.

13:03This is an interesting development, and I think we're going to see a lot more of that coming into 2024, where all the big players are probably going to move forward to try to reduce the amount of hallucinations of their models as much as possible. That being said, there are many experts out there that claim that actually hallucinations are not a problem. They're actually the benefit of large language models compared with search engines that just provide you accurate information. We want large language models to be creative, to think out of the box, to be able to brainstorm with us different ideas versus just provide us existing information in the new package.

13:39So in the show notes, I will link to an article called Hallucinations are a Feature, Not a Bug by Daniel Miner. It's a great article that will give you a lot of food for thought on that topic. And from IBM to another huge giant that we don't hear a lot of, but does a lot in the AI world, which is Apple. So we got used to talking about OpenAI and Microsoft and Google and Meta, and we don't hear much from Apple, but Apple has been obviously playing with AI for a very long time. And they're now starting to release more and more things. And the first piece of news comes that Apple has teamed up with EPFL and they've released a multi-modal, very powerful open source framework that right now it's for research use only, but it's a framework that allows building multi-modal AI models, including taking into account scalability of data and architecture and training across various aspects, including powerful vision capabilities.

14:40And this architecture allows them to excel on multiple tasks that the model has not seen before. So very powerful, very capable model released by Apple. The company that is known for keeping its cards close to its chest and never releasing anything has released an open source model, which is A, exciting, and B, maybe showing the side that they're going to take in this future battle between closed and open source models. Another interesting piece of news from Apple is that the researchers found new techniques to run large language models straight on iPhones with very limited memory. They're basically switching from running on the device's RAM memory, which was a traditional way to do this, to running on flash memory.

15:28And I'm not going to get into all the technical details, but the biggest thing about it is that it allows to run much more advanced and more capable AI models straight on phone hardware, similar to maybe what Google is planning to do with the small version of Gemini. And from Apple's perspective, it builds on the rumors of their anticipated Ajax project, which is supposed to be ChatGPT competitor. So while Apple has stayed quiet in the years since ChatGPT came out, you can be sure that they've been working very, very hard to develop competitive capabilities. And I think we'll start seeing them come out and become available to us in the very near future, but definitely through 2024.

16:12Another company that we've talked about several times in the past few episodes is Mistral. Mistral is a French company that has been consistently releasing very powerful and capable open source AI models. They released their latest one a few months ago, and it's as good and in some cases surpassed GPT 3.5 and Llama 2.70B from Meta. So very capable models as open source, but they're claiming that in 2024 and potentially early in 2024, they're going to release an open source model that is going to be as powerful as GPT-4. That's obviously a very serious statement since everybody that claimed that so far did not achieve that goal yet of developing a model or releasing a model that is as good as GPT-4.

16:57That comes in as part of rumors that OpenAI are preparing to release GPT-4.5 in early next year. So the race is definitely on with more and more players, both on open source and closed platforms. Now, Mistral recently raised$415 million at a$2 billion valuation that allows them to continue fueling their development and the powerful models that they have been releasing. And if we're talking about raising money, there's news on both OpenAI and Anthropic raising or about to raise crazy amounts of money to continue their development. Anthropic has announced they're planning to raise$750 million at an 18.4 billion valuation.

17:42That's four and a half X on their previous valuation of$4 billion that was just earlier this year, which shows you how much anticipation and how much real expectations there are from these companies to produce in the long-term future. And there are rumors that OpenAI are planning another raise in the near future that will value them at$100 billion, showing how much trust there is in the financial side of businesses in this new technology called AI for the masses. One more company that we talked about a lot that I really like is Midjourney. Midjourney just released version six to the world. We talked about last week that they're allowing people to access it through a website.

18:30So that might be the most obvious difference that to access it, you don't need to use Discord anymore. But in addition, there's some major upgrades on how it interprets prompts and its ability to actually understand more parameters in the prompts than before. the initial testing that I've done and that I've seen a lot of people doing are absolutely stunning. So while there's a few parameters that are not working as good as they did on Discord, the results that people are getting and myself included using this new platform are out of this world. And I'm pretty sure they will bring back the other features again.

19:07So huge improvement on the image generation capabilities of Me Journey in their competition with other platforms such as DALI-3 and Stable Diffusion and many others. Something that is definitely worth mentioning is the capability of Me Journey 6 to now generate text, which it was not able to do before. So now you can add stylized text within your graphics generated by Me Journey, which is obviously a huge upgrade. And I want to end on a positive note with something that happened with AI this week. Researchers using deep learning models were able to discover a new antibiotics type that is working against drug-resistant bacteria.

19:47This is the first new antibiotics discovered in over 60 years. So the way they did it is they used AI to evaluate over 12 million different compounds. From that, AI was able to identify five compound classes that will most likely be effective against existing drug-resistant bacteria. The process that was just done with AI relatively quickly used to take over a decade with traditional non-AI tools. And it definitely brings forward the potential positive benefits of using AI across multiple aspects of our lives as humans. In this particular case, creating new drugs that will enable us to fight different bacteria that we cannot fight right now.

20:32This episode goes live just before Christmas. So Merry Christmas to everyone. Happy New Year to anybody who's listening and your loved ones. I wish all of you to have a great time with your families and your loved ones and to have a kick-ass 2024. If you want to start 2024 with the right foot forward, you can join our AI Business Transformation course, and I will put the link in the show notes, and you can get 24 % off to start the year of 2024 with the right foot forward with the promo code KICKOFF24. I'm most likely not going to record and use episode next week. I will release an episode this coming Tuesday with an interview, as always, deep diving into a specific topic.

21:15And I'll be back recording the weekend news episodes at the beginning of 2024. So again, happy new year for everyone. Happy holidays. Keep exploring AI and have an amazing holiday week.