The Lindahl Letter

Working out of notebooks is an easy and lightweight way to just mess around with code. One of my all time favorite ways to go about doing that is in the Colab research space provided by the team over at Google [1]. They have an offering for a pro subscription, but you can get in and use things without a payment plan. Sure the promise of faster GPUs, more memory, and longer runtimes is enticing. If you really get deeply into the process of training models, then it might make sense for you to make arrangements to pay for some compute. A lot of people have shared notebooks and you can quickly start to work on things and see code in action. The part of it that is so appealing to me is that you can look at a very distinct block of action and run it in the notebook environment without having to worry about anything else in the stack. 

Breaking things into smaller understandable pieces is absolutely fantastic for learning about things. It’s also nice to just jump in and be able to execute the code. Here very soon according to the Google Blog page for developers we are going to see AI-powered coding free of charge within the Colab space [2]. You will be able to open the notebook and instead of adding code or text you will have access to another option to generate. That is where the magic will be located and you will be able to very rapidly use Colab in a generative way. I think it is going to bridge the gap between interacting with generative models to accomplish things and the current sort of limitations. Inside that argument would be the backward linkages to the Google ecosystem where instead of plugins or managing extensions the system will just have extensibility into a massive and mature ecosystem. 

I have been waiting since that May 17, 2023 post for the feature to show up. Literally, I just keep checking on an almost daily basis to see when we can get to the point of rock and rolling the generative creation of code from the PaLM 2 family of code models [3]. At this point, I’m going to admit that my favorite part of this whole thing is that they have called it Codey. Maybe the controversial Clippy was just ahead of its time, but I think Codey will be just right as a part of a larger ecosystem of generative coding efforts that are about to take flight. You will see that very little traffic occurred on Twitter about this news [4]. In terms of education and getting people going with generative code generation this is going to be epic. You will see my contributions to GitHub of notebooks go from sparse to daily [5]. I have started dropping some code to make fractals and do some various types of math into a new repository [6].

All of the security for Google Colab notebooks is managed by those teams. You certainly have control over your own private notebook being saved to your Google drive. Access to the data and notebooks is controllable and the team over at Google is spinning up and then spinning down virtual machines to do this work. Most of the process is entirely ephemeral and to the end user is fundamentally just a browser. The FAQ page does not really spend much time addressing security [7]. It’s a place to very quickly be able to execute code from a browser. Using it for educational purposes makes total sense. It’s a great place to share code and learn about how that code works. It has enough power to do some basic machine learning and to that end it is a pretty awesome experience. You are not going to end up using this as a production runtime for any sort of tasking. One of the things I’m actually curious about in the future is what happens when people start using ChatGPT plugins or agents that go out and farm compute from environments like this type of browser based engagement.

Footnotes:

[1] https://colab.research.google.com/

[2] https://blog.google/technology/developers/google-colab-ai-coding-features/ 

[3] https://9to5google.com/2023/05/17/google-colab-codey/ 

[4] https://twitter.com/search?q=colab%20codey&src=typed_query 

[5] https://github.com/nelslindahlx 

[6] https://github.com/nelslindahlx/fractals 

[7] https://research.google.com/colaboratory/faq.html 

What’s next for The Lindahl Letter? 

* Week 130: Build captain fractal using Colab

* Week 131: Bulk imagine improvement scripting

* Week 132: Synthetic data notebooks

* Week 133: Automated survey methods

* Week 134: Make a link based news report automatically

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

The topic for this week was a weighty one for sure. It’s one that I think will functionally happen, but I do not believe will be real in terms of actual practice. We are seeing a lot of emphasis on using these new AI systems that are popping up everywhere. People are not able to build these systems from scratch anymore. Using code from somebody else is becoming more and more a part of the process. When you look at how to democratize AI systems it is certainly about spreading usage and making these systems omnipresent. Security is structurally something that is important in this situation, but it remains something that generally happened before you got involved. In that scenario, really deeply considering how your AI system security posture is going to be invoked you have to consider both forward looking actions and the backward linkages within these products both in terms of the build and design. 

Within the calling of an API in this space it certainly is possible to consider what data is going out and what is coming back. You can pretty clearly understand what happens with the data after it is sent over and what is going to happen with it after the fact. My searches on Google Scholar for articles related to AI security were not super exciting [1]. That search with the brackets around the search term yielded about 2,900 items. Taking the brackets off opened up the results to over 3 million [2]. This may very well be a space where the academic content has not caught up with where the technology happens to be at the moment. A lot of academic research in the space is occurring, but since the papers are not really about or from production implementations the topics of how to secure, manage, and deploy are distinctly lacking. 

At this point in the story, I got super interested in reading about comparisons of protocols for distributed social networking. Things like the AT protocol vs activitypub [3]. Oddly, one of the best lists that I encountered was on Wikipedia where somebody is clearly keeping track of open projects related to this one [4]. You might be wondering if that list includes 30+ projects with tacking and information and you will find out that it totally does. However, being able to take a look at such a large list actually made me a little bit concerned about any of these protocols actually becoming dominant and taking over in practice. 

Let’s take this post in a different, more meta direction. Ok, here are all the spilled beans on that one, I ended up making an executive decision on how to manage my writing backlog. It was a super disruptive and spur of the moment decision at this point in the process. I went ahead and at the week 128 issue (which is this post) pasted into the backlog the new list of 49 items from a Google Keep note. That happened as a result of a really productive day that happened last week. During one of the vacation days at the beach recently, I started making a list of topics I would like to spend some time either researching or writing about. 

It turns out that list ended up including 49 total items. With the addition of those 49 items, my backlog is now tracking out to week 218. Switching things up like that means I’m going to struggle with updating the next five weeks of forward looking items on a bunch of blocks of content, but that is a solvable problem in terms of editing and writing. It’s just a bit annoying and somewhat time consuming vs. creating any really ongoing problematic situation. That does mean that starting in July you will start to get the benefit of this new backlog and strictly speaking you are going to end up waiting about a year to return to the previously structured program content. 

While we are zoomed out and considering some of the more meta implications of things I’ll note that I have kept the podcast part of this effort going. Readership numbers indicate that it is a smaller portion of things compared to the standard text based link clicking. About 20% of the ongoing traffic for this writing project seems to be related to the audio based podcast. The recording process is now pretty streamlined and I have it as a part of my overall creative weekend routine. These weekly blocks of writing won’t grow long enough to make recording the audio for them problematic in terms of a time commitment. 

Footnotes:

[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=%22ai+security%22&btnG= 

[2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=ai+security&btnG= 

[3] https://www.theverge.com/2023/4/20/23689570/activitypub-protocol-standard-social-network 

[4] https://en.wikipedia.org/wiki/Comparison_of_software_and_protocols_for_distributed_social_networking

What’s next for The Lindahl Letter? 

* Week 129: How do you use Colab in a generative way?

* Week 130: Build captain fractal using Colab

* Week 131: Bulk imagine improvement scripting

* Week 132: Synthetic data notebooks

* Week 133: Automated survey methods

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Things have been happening with Google DeepMind. A lot of news has been dropping since I started crafting this post. I don’t want to write about the news of the moment on this one. The context and analysis within this post is going to try to be deeper than a base reaction to that recent news. The Verge (and a ton of other people) reported that, “Google’s big AI push will combine Brain and DeepMind into one team” [1]. Apparently, some drama exists around those two organizations being combined. We will almost certainly see some books about this from the principal players. Let’s hope they turn out as good as the 2020 memoir “Uncanny Valley” by Anna Wiener. Some of them won’t be that dynamic, but they should be. 

I spent some time reading content from DeepMind. They have a solid research page [2]. It highlights some interesting things including AlphaFold, WaveNet, and AlphaGo [3][4][5]. They have a lot of content out on GitHub in a variety of repositories [6]. They have over 190 repositories with a variety of code you could go take a look at. You may even notice that they have signed the “Lethal Autonomous Weapons Pledge” [7][8].

Some good conversation is occurring about what happens when AI hallucinates, bias bounties, new bug bounties, and how red teams work [9]. You could also find out a little more about how the GitLab AI-based security feature helps to scan codebases for vulnerabilities [10]. You could go check out the Google Cloud Security Podcast EP52 Securing AI with DeepMind CISO Vijay Bolina from February 14, 2022 [11].

You really have to consider that even the best pair programming with assistive technology could introduce security vulnerabilities [12]. We could probably spend an entire week looking at the medical artificial intelligence system or more clearly named Med-PaLM 2 that was released [13]. They also shared an interesting method to discover faster matrix multiplication algorithms which I thought was rather interesting [14]. 

Don’t worry I did go out and grab a couple (really 4) scholarly papers to share with you this week:

Beattie, C., Leibo, J. Z., Teplyashin, D., Ward, T., Wainwright, M., Küttler, H., ... & Petersen, S. (2016). Deepmind lab. arXiv preprint arXiv:1612.03801. https://arxiv.org/pdf/1612.03801.pdf

Powles, J., & Hodson, H. (2017). Google DeepMind and healthcare in an age of algorithms. Health and technology, 7(4), 351-367. https://link.springer.com/article/10.1007/s12553-017-0179-1

Tassa, Y., Doron, Y., Muldal, A., Erez, T., Li, Y., Casas, D. D. L., ... & Riedmiller, M. (2018). Deepmind control suite. arXiv preprint arXiv:1801.00690. https://arxiv.org/pdf/1801.00690 

Evans, R., & Gao, J. (2016). Deepmind ai reduces google data centre cooling bill by 40%. DeepMind blog, 20, 158. https://www.deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40 

Footnotes:

[1] https://www.theverge.com/2023/4/20/23691468/google-ai-deepmind-brain-merger 

[2] https://www.deepmind.com/research

[3] https://www.deepmind.com/research/highlighted-research/alphafold

[4] https://www.deepmind.com/research/highlighted-research/wavenet 

[5] https://www.deepmind.com/research/highlighted-research/alphago

[6] https://github.com/deepmind 

[7] https://www.deepmind.com/safety-and-ethics

[8] https://futureoflife.org/open-letter/lethal-autonomous-weapons-pledge/?cn-reloaded=1

[9] https://www.theregister.com/2023/04/26/is_your_ai_hallucinating/

[10] https://www.spiceworks.com/tech/artificial-intelligence/news/gitlab-launches-ai-security-feature/

[11] https://cloud.withgoogle.com/cloudsecurity/podcast/ep52-securing-ai-with-deepmind-ciso/

[12] https://www.techtarget.com/searchsoftwarequality/news/252523049/Developers-beware-AI-pair-programming-comes-with-pitfalls

[13] https://www.verdict.co.uk/google-lauches-new-medtech-ai-system-amid-calls-for-greater-data-security/

[14] https://venturebeat.com/ai/top-5-stories-of-the-week-deepmind-and-openai-advancements-intels-plan-for-gpus-microsofts-zero-day-flaws/

What’s next for The Lindahl Letter? 

* Week 128: Democratizing AI system security

* Week 129: Snapchat Security

* Week 130: Generative model security

* Week 131: Profiling Microsoft Azure security

* Week 132: Engineering discipline security

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Something is going to have to give here pretty soon for Hugging Face or they are going to get left behind by the extended model, plugin driven, and now internet connected GPT models that some of the biggest platforms are working to take mainstream. We are seeing the real time development creating extensibility into actions from companies like OpenAI and their partner Microsoft that are going to be a foundational groundwork for how connectivity works within these models. This ecosystem is going to be something so highly proprietary and interconnected between a set of foundational companies that no ability to directly open source a competitor is going to exist. Part of that will be due to the payment models that are going to get setup as a foundational services layer starts to get setup.

People have been really excited about Hugging Face for some time now. Since 2016 the private company Hugging Face has worked to share machine learning content with the world. You can easily get to some of the training courses they have set up on NLP and Deep RL [1]. They are pretty good training courses. People have really dug into models and spaces from Hugging Face to help democratize NLP models. At one point, everybody was checking in with what Hugging Face was up to and they made a very large space in the AI and ML space. You can easily go out to Google Scholar and see a bunch of academic papers that reference Hugging Face either as two words or sometimes one word [2]. I spent some time looking for papers with a decent number of citations. Here are 5 that were selected:

Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A. M. (2019). Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. https://arxiv.org/pdf/1910.03771

Jiang, W., Synovic, N., Hyatt, M., Schorlemmer, T. R., Sethi, R., Lu, Y. H., ... & Davis, J. C. (2023). An empirical study of pre-trained model reuse in the hugging face deep learning model registry. arXiv preprint arXiv:2303.02552. https://arxiv.org/pdf/2303.02552

Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., ... & Rush, A. M. (2020, October). Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations (pp. 38-45). https://aclanthology.org/2020.emnlp-demos.6.pdf 

Pfeiffer, J., Rücklé, A., Poth, C., Kamath, A., Vulić, I., Ruder, S., ... & Gurevych, I. (2020). Adapterhub: A framework for adapting transformers. arXiv preprint arXiv:2007.07779. https://arxiv.org/pdf/2007.07779 

Zhang, Y., Sun, S., Galley, M., Chen, Y. C., Brockett, C., Gao, X., ... & Dolan, B. (2019). Dialogpt: Large-scale generative pre-training for conversational response generation. arXiv preprint arXiv:1911.00536. https://arxiv.org/pdf/1911.00536.pdf%7D. 

You can get a feel for what the Hugging Face community has been working on by taking a look at their GitHub repository [3]. They have transformers, datasets, diffusers, and tools for a variety of things that people find useful. I would describe Hugging Face as a very hands-on community where you can dig in and be a part of what is going on within the space. That is the distinctive difference from some of the other larger corporations that do open source things. Most of the corporate AI labs will produce and distribute things when they are ready. Iterative tool releases and a vibrant community full of contributions are popping up all over. We are also seeing a newer trend where companies like OpenAI are releasing models and other technology via API without open sourcing the content. This pay to play API model shields the intellectual property better, but in this space people are rapidly learning from innovations and working to build alternatives.  

Here are a couple of YouTube videos including one from Hugging Face:

You can dig into the security features offered by Hugging Face [4]. It says clearly on the security page that they are SOC2 Type 1 certified with a link to the AICPA page [5]. With all sources where you might download code and use it you are going to want to understand the code, be ready to keep it current, and of course be prepared to take rapid action. They are even trying to compete with ChatGPT and share more open models for that type of effort [6]. I spent some time looking around to try to find some deep assessments of the security profile for Hugging Face and have not really found what I was looking for yet.

Footnotes:

[1] https://huggingface.co/learn/nlp-course/chapter1/1 

[2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=%22hugging+face%22&btnG= 

[3] https://github.com/huggingface 

[4] https://huggingface.co/docs/hub/security 

[5] https://us.aicpa.org/interestareas/frc/assuranceadvisoryservices/aicpasoc1report.html 

[6] https://venturebeat.com/ai/hugging-face-launches-open-source-version-of-chatgpt-in-bid-to-battle-openai/ 

What’s next for The Lindahl Letter? 

* Week 127: Profiling DeepMind Security

* Week 128: Democratizing AI system security

* Week 129: Snapchat Security

* Week 130: Generative model security

* Week 131: Profiling Microsoft Azure security

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Security professionals the world over are concerned about what people like some of Samsung’s engineers might be doing with OpenAI’s ChatGPT services [1]. Let’s back up just a bit from that very real breaking news security reality and make sure to set the stage. Instead of starting with a complete corporate history about OpenAI, you knew that I would go out and search Google Scholar to find some of the key OpenAI related scholarly works [2]. Alternatively, you could surf over to the OpenAI website and find the section they have for over 100 papers and just download the PDFs from the source [3]. You could download papers like the much maligned “GPT-4 Technical Report” that was lengthy at over 100 pages, but did not go into the model mechanics people were interested in understanding [4]. Other papers are hosted on that page like the Dall-E-2 paper “Hierarchical Text-Conditional Image Generation with CLIP Latents” [5]. It is a lot of great content and you will get a sense that at one point OpenAI was sharing and building a great legacy of published content that helped people really understand where artificial intelligence was going. They also did some really awesome work building agents that competed in pretty complex video games against world class players. That part of the equation is what caught my attention and really pulled me into trying to understand OpenAI. 

Here are 5 interesting papers with a solid number of citations:

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). Gpts are gpts: An early look at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130. https://arxiv.org/pdf/2303.10130.pdf 

Berner, C., Brockman, G., Chan, B., Cheung, V., Dębiak, P., Dennison, C., ... & Zhang, S. (2019). Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680. https://arxiv.org/pdf/1912.06680.pdf 

Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016). Openai gym. arXiv preprint arXiv:1606.01540. https://arxiv.org/pdf/1606.01540.pdf

Gawłowicz, P., & Zubow, A. (2019, November). Ns-3 meets openai gym: The playground for machine learning in networking research. In Proceedings of the 22nd International ACM Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (pp. 113-120). https://www2.informatik.hu-berlin.de/~zubow/gawlowicz19_mswim.pdf 

Zamora, I., Lopez, N. G., Vilches, V. M., & Cordero, A. H. (2016). Extending the openai gym for robotics: a toolkit for reinforcement learning using ros and gazebo. arXiv preprint arXiv:1608.05742. https://arxiv.org/pdf/1608.05742.pdf

Now that we have covered some of the decently cited scholarly content it is probably good to zoom out from that space into the broader media coverage of OpenAI. Most of it is about the risk, reward, and fear that this type of generative AI fosters. You will see that security is not at the forefront of these discussions. 

From TheVerge “OpenAI co-founder on company’s past approach to openly sharing research: ‘We were wrong’”. Ok. Where do you go from there…https://www.theverge.com/2023/3/15/23640180/openai-gpt-4-launch-closed-research-ilya-sutskever-interview

You can hear Greg Brockman who was an OpenAI cofounder talk about the astonishing potential during a 30 minutes TED talk.

The CEO of OpenAI Sam Altman sat down with ABC News for 20 minutes here:

You can see a long form interview with Lex Fridman and Sam Altman here:

They have an OpenAI YouTube channel that gets a lot of views. They used to share some pretty interesting videos.

Ok at this point in our journey here, we have covered OpenAI in general and looked at what is being created. Now it’s probably best to try to explain the drama that surrounds the company as quickly as possible. Oddly, it’s not a drama focused on the technology being created that certainly has entered the public mind. Elon Musk was an initial board member of OpenAI back in 2015 [6]. It used to be a non-profit and things have changed. That is where the real drama exists about the company. Microsoft now has a multi-billion dollar investment in OpenAI [7]. You can read the Microsoft official blog where they detail that they have invested in OpenAI three times including 2019, 2021, and 2023 and is now the exclusive cloud provider [7]. You might be thinking how did Elon Musk invest or more accurately donate 100 million to OpenAI which now is clearly partnered with Microsoft [9]. It’s a confusing plot twist for sure and a lot of media coverage exists trying to explain how that exactly happened [10]. I’m sure we will learn more about what exactly happened in the next couple of years.

Let’s circle back to the security realities of using things like OpenAI’s ChatGPT system. People in this case are going to use the system to input things to the prompt. Some of that input might very well be intellectual property owned distinctly by somebody. You could ask a model to create more content in an author’s style or to extend the story of characters. It’s also entirely possible that somebody might feed in code for review or ask the model to produce code. 

Footnotes:

[1] https://www.securitymagazine.com/articles/99220-cybersecurity-leaders-reflect-on-samsung-chatgpt-incidents 

[2] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=openai&btnG= 

[3] https://openai.com/research 

[4] https://cdn.openai.com/papers/gpt-4.pdf 

[5] https://cdn.openai.com/papers/dall-e-2.pdf 

[6] https://finance.yahoo.com/news/openai-cofounder-elon-musk-said-142815403.html 

[7] https://www.theverge.com/2023/1/23/23567448/microsoft-openai-partnership-extension-ai 

[8] https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/ 

[9] https://fortune.com/2023/03/16/elon-musk-openai-non-profit-switch-30b-market-cap-for-profit-after-donation-chatgpt/ 

[10] https://www.businessinsider.com/elon-musk-defends-role-in-openai-chat-gpt-microsoft-2023-2 

What’s next for The Lindahl Letter? 

* Week 126: Profiling Hugging Face Security

* Week 127: Profiling DeepMind Security

* Week 128: Democratizing AI system security

* Week 129: Snapchat Security

* Week 130: Generative model security

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

It's all happening. I'm pivoting to AI + Security as the central topic for The Lindahl Letter. The next weekly modules I'm going to create are all going to be pivoting to a clear security based focus each week. Content for this year of The Lindahl Letter was packaged into posts encompassing weeks 105 to 156 or 52 total blocks of content. Think year divisible by week with production spread evenly. This is the 3rd year of that type of writing effort. This is week 124 in that sequence and I have yelled pivot and brought in 11 new topics that will follow the profiles of OpenAI, Hugging Face, and DeepMind. That content pivot pretty much involved highlighting a section of the backlog and just moving it beyond week 156. I’m going to keep on working to balance the topics together to bring forward the best possible set of content for this year. Getting the best possible blocks of content brought together as a research project is essential to my work moving forward. 

It’s possible that I will go back and rework some of the content from weeks 105 to 123 to include more of a security focus. Obviously, that won’t change the previously released podcast audio or published content, but it would update the content that gets put into a manuscript at the end of the year. That content gets edited and revised during the course of the publishing process so it is never exactly the same as the content that was published throughout the year in weekly installments. While I certainly strive to produce high quality prose my ability to edit to perfection on a weekly basis remains suspect. Editing for continuity and general cohesiveness is a different element within that large of a manuscript compared to producing a weekly installment. Putting content into an eBook and ultimately a hardback or paperback format means that it needs to be free of grammarian enraging distractions.

During my initial cut at introducing security related topics into the backlog 11 topics jumped out at me and got included. My backlog editing will be pretty extreme for the next couple of weeks to really get the most out of this epic content based pivot. Right now we are at the precipice of some very serious questions about how AI will impact society. Seriously, AI and society are very alive within the public mind at the moment. Understanding how we secure, interact over time, and ultimately establish security within the watershed event of modern AI development will be more important than my rather casual walk into AI in general. You will find that this will be a good pivot that ends up providing more depth and context to what is happening currently. Each of the new installments will be written with a degree of perspective that helps prevent them from being passing observation installments. Writing pure reaction content is not where I’m trying as that provides less of an advisor type function and more of an ephemeral of the moment observation.

You may well be aware that I’m still struggling with the idea that all the content that gets produced as a part of my weekly research notes could be replicated by ChatGPT in minutes. This is the 124th installment and OpenAI’s ChatGPT could easily provide alternative versions of the content I have created. Understanding how to bring larger themes together and diligently searching for the best academic articles is not something that ChatGPT does well at the moment. It’s entirely possible that the combination of content selection and prompt engineering could change the potential output. I’m pretty sure somebody will work to better adapt ChatGPT to the review of complex academic work. Having the right data sources selected for a model will be a key element of how things work moving forward. You could take the papers from my independent study ML syllabus and get a decent set of inputs for a model. However, would you add that as a layer to a current model like ChatGPT 3.5 or 4.0 or will you have focused layers for the models to consider in sequence or as a refinement step along the way. All of those things are up in the air and we will learn more about where they will ultimately end up. 

As a general editorial decision here and now I may consult and work with ChatGPT, but the output in terms of writing posts for the rest of the year will just include my words. For me the novelty of including some examples of the differences between what I would write and what OpenAI’s ChatGPT would create has lost a certain degree of interest for me. Maybe the novelty or the thrill of it is gone. It took a couple months for that to happen, but now I’m going to focus on providing the best research note I can each week. Even this note focusing my content going forward signals an important change in my research trajectory and was worth devoting a week of content creation to building and sharing. 

Links and thoughts:

What’s next for The Lindahl Letter? 

* Week 125: Profiling OpenAI Security

* Week 126: Profiling Hugging Face security

* Week 127: Profiling DeepMind security

* Week 128: Democratizing AI systems security

* Week 129: Snapchat Security

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Substack Week 123: We are wholesale oversubscribed on AI related content

Thank you for tuning in to this audio only podcast presentation. This is week 123 of The Lindahl Letter publication. A new edition arrives every Friday. This week the topic under consideration for The Lindahl Letter is, “We are wholesale oversubscribed on AI related content.”

You may have noticed that the topic for the post this week changed. I called an audible and wrote a different piece of content.

All that AI content has hit a real saturation point. We are wholesale oversubscribed on AI related content and we have not even really crossed into the situation I’m concerned about where people are using the content to flood the open internet with synthetic content. Generally given the current state of the AI excitement level and some of the fear, apprehension, and concern a lot of people are writing or commenting. All that commenting has built into a very real saturation point where AI content is literally everywhere. Sam Altman of OpenAI recently noted in a Wired article that, “the Age of Giant AI Models Is Already Over” [1]. The initial release date of ChatGPT was November 30, 2022 so the cycle on this one was no more than a 6 month explosion of hype content. You can go out to Google Trends and pretty quickly get a sense of just how fast ChatGPT spun up in December of 2022 [2]. If you went out and added a comparison term within Google Trends of auto-gpt or AutoGPT, then you would get a sense of just how much more popular ChatGPT happens to be in the wild [3]. Keep in mind that since April 11, 2023 the searches for AutoGPT have taken off exponentially [4]. 

Major things are happening in terms of ChatGPT, AutoGPT, and people expanding how agents are able to cooperate. A paper from Park et al. that was just published on April 7, 2023 shows some very interesting use cases for simulating behavior using generative agents.  

Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv preprint arXiv:2304.03442. https://arxiv.org/pdf/2304.03442.pdf 

Rewind a little bit and I thought the key Stanford University paper would be that 214 page multi author one on foundation models.

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., ... & Liang, P. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258. https://arxiv.org/pdf/2108.07258.pdf 

You can get a sense from the demonstration of capability and use that occurred in that paper on generative agents just a couple years after scholars circled the wagons on foundation models just how fast the field is changing. Things are changing so fast that Sam Altman above noted that the age of giant models might have arrived and be complete between the publication of those two papers. It can make conducting research a very interesting thing to contemplate. Content at the bleeding edge of AI technology today could very well encounter a context, vibeshift, or even meaning change within months. Taking that into consideration I’m going to recognize that we are approaching the halfway point on the production of this year’s Substack content. To that end, I’m starting to ponder how to improve things and really make things better in terms of the content that is being produced. 

Links and thoughts:

Footnotes:

[1] https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/

[2] https://trends.google.com/trends/explore?geo=US&q=chatgpt&hl=en 

[3] https://trends.google.com/trends/explore?date=now%207-d&geo=US&q=chatgpt,autogpt&hl=en 

[4] https://trends.google.com/trends/explore?date=2023-03-18%202023-04-18&geo=US&q=autogpt&hl=en 

What’s next for The Lindahl Letter? 

* Week 124: Profiling OpenAI 

* Week 125: Profiling Hugging Face (open and collaborative machine learning)

* Week 126: Profiling Deep Mind

* Week 127: Democratizing AI systems

* Week 128: Building dreaming into AI systems

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

I spent some time over the last couple of weeks looking for scholarly articles that reference AI as a service (AIaaS) [1]. Then I realized that things are changing so quickly in the landscape of AI these days that none of those articles even consider Auto-GPT [2]. Unless you have been on vacation and hiding from AI news recently you have been hearing about the advent of this plugin, multiagent, and task targeted implementation of a GPT model called Auto-GPT. People are using it to do things. You can see that that GitHub repository already has 50,000 stars and the demo was from March 30, 2023. In terms of open source code projects this one has taken off with rocket boosters. Right now the project has been forked almost 8,000 times. I’m trying to find a version that can be run in a jupyter notebook on Colab. This space is rapidly changing and will probably be changing rapidly over the next 6 months.

What did the OpenAI ChatGPT 4.0 model spit out on this one when prompted?

AI can be both a platform and a service, depending on how it is implemented and used. The distinction between the two is not mutually exclusive, as AI applications often span across various domains and industries. Let's explore each concept:

* AI as a Platform: An AI platform provides the infrastructure, tools, and libraries necessary for developing and deploying AI applications. These platforms enable developers and data scientists to build, train, and manage AI models without having to start from scratch. Examples of AI platforms include TensorFlow, PyTorch, and OpenAI's GPT series. In this context, AI serves as a foundational technology upon which various solutions can be built.

* AI as a Service (AIaaS): AIaaS refers to the delivery of AI capabilities as a service, usually through APIs or web services. Companies that offer AIaaS provide access to pre-built AI models, which clients can use without having to develop their own AI capabilities. This allows organizations to leverage AI technology without investing in infrastructure or expertise. Examples of AIaaS include Google Cloud AI, IBM Watson, and Amazon Web Services' machine learning services.

In conclusion, AI can function both as a platform and a service, depending on its implementation and usage. Many companies and developers use AI platforms to create AI services, which can then be integrated into various applications and business processes.

I thought it might be fun to follow up and prompt to see who was the best at AIaaS… 

It is difficult to definitively say who the "best" AI-as-a-Service (AIaaS) provider is, as different providers excel in different areas and offer varying services that cater to specific use cases. However, some of the most prominent and reputable AIaaS providers include:

* Google Cloud AI: Google offers a comprehensive suite of AI services, including natural language processing, computer vision, and machine learning. Its offerings, such as Google Cloud Vision API, Dialogflow, and AutoML, are widely used and respected for their accuracy and ease of use.

* Amazon Web Services (AWS): AWS provides a range of AI services, including machine learning, computer vision, and natural language processing. Some popular offerings are Amazon Rekognition, Amazon Lex, and Amazon SageMaker.

* IBM Watson: IBM's AIaaS platform, Watson, offers various services, including natural language processing, machine learning, and chatbot development. Watson Assistant and Watson Discovery are among its popular services.

* Microsoft Azure: Azure offers a suite of AI services, including machine learning, natural language processing, and computer vision. Key offerings include Azure Cognitive Services, Azure Machine Learning, and Azure Bot Service.

The "best" AIaaS provider for your specific needs will depend on several factors, such as the required AI capabilities, pricing, ease of integration, and the level of customization needed. It's essential to research and evaluate each provider based on your unique requirements to make an informed decision.

Links and thoughts:

Top 4 Tweets of the week:

https://twitter.com/ericschmidt/status/1659619610063249409

https://twitter.com/CaseyNewton/status/1659579798857871360

https://twitter.com/robmay/status/1659610763005399042

https://twitter.com/DynamicWebPaige/status/1659075284891738112

Footnotes:

[1] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=AIaaS&btnG= 

[2] https://github.com/Torantulino/Auto-GPT 

What’s next for The Lindahl Letter? 

* Week 123: Considering open source AI

* Week 124: Profiling OpenAI 

* Week 125: Profiling Hugging Face (open and collaborative machine learning)

* Week 126: Profiling Deep Mind

* Week 127: Democratizing AI systems

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

My initial take on writing an independent study based syllabus for applied AI was to find the best collection of freely available scholarly papers that somebody could read as an onramp to beginning to understand the field. That I think is a solid approach to helping somebody get going within a space that is very complex and full of content. It’s a space that is perpetually adding more content than any one person could possibly read or consume. Before you take that approach it is important to understand that one definitive textbook does exist. You certainly could go out and read it. 

Russell, S. J. (2010). Artificial intelligence a modern approach. Pearson Education, Inc.. 

You could find the first edition, second edition, or third edition for sale on eBay or somewhere else if you wanted a physical copy of the book. The book is currently in a 4th edition run, but I don’t have a copy of that edition yet. It’s used by over 1,500 schools so a lot of copies exist out in the wild [1]. The authors Stewart Russell and Peter Norvig have shared a PDF of the bibliography for that weighty tome of AI insights as well [2]. Even with 35 pages of bibliography nobody with the name Lindahl made the cut. On a side note you can find the name Schmidhuber included twice if that sort of thing is important to you. 

Let’s reset for a second here. If you are brand new to the field of AI or want to read a textbook based introduction, then you should seriously consider buying a copy of the aforementioned textbook. That is a really great way to start which has worked for tens of thousands of people. My approach here is going to be a little bit unorthodox, but it works for me. My last run at this type of effort was, “An independent study based introduction to machine learning syllabus for 2022” and you can find it out on Google Scholar [3]. This outline will be the basis of a similar type of work that will end up getting crafted in Overleaf and shared out to the world. 

Searching for just pure introductions to artificial intelligence is really hit or miss. A lot of different introductions to various fields exist. In this case, I’m trying to zoom out a little more into a larger evaluation of content instead of focusing on any one field. Nothing I ran into during my search had the number of citations or impact of the Russel and Norvig textbook. I’m going to endeavor to structure and organize 70+ articles into a syllabus. To give you an idea of the kind of things that are going to get pulled together here are 5 different papers. 

Oke, S. A. (2008). A literature review on artificial intelligence. International journal of information and management sciences, 19(4), 535-570. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=f4dfcfe3f132b1079d054e2db29adc063fab0007 

Contreras, I., & Vehi, J. (2018). Artificial intelligence for diabetes management and decision support: literature review. Journal of medical Internet research, 20(5), e10775. https://www.sciencedirect.com/science/article/pii/S0004370218305988/pdfft?md5=ec6948d3f66efe5e57d1336a54d1604d&pid=1-s2.0-S0004370218305988-main.pdf 

Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., & Aerts, H. J. (2018). Artificial intelligence in radiology. Nature Reviews Cancer, 18(8), 500-510. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6268174/ 

Ghahramani, Z. (2015). Probabilistic machine learning and artificial intelligence. Nature, 521(7553), 452-459. https://www.repository.cam.ac.uk/bitstream/handle/1810/248538/Ghahramani%25202015%2520Nature.pdf?sequence=1 

Jiang, F., Jiang, Y., Zhi, H., Dong, Y., Li, H., Ma, S., ... & Wang, Y. (2017). Artificial intelligence in healthcare: past, present and future. Stroke and vascular neurology, 2(4). https://svn.bmj.com/content/svnbmj/2/4/230.full.pdf 

Obviously, working out of a textbook format will help provide a better structured introduction and that is why I plan on mentioning that in the paper and I have noted it here. On a scholarly basis I understand the need to dive directly into source papers and really get a sense of what is happening within the applied side of AI research in the last 10 years. That is why I want to spend some time preparing the independent study based syllabus for others as well as it being a good solid effort for me to learn more rapidly with an extreme degree of focus. 

What did ChatGPT 4.0 model come up with?

An independent study applied AI syllabus will provide you with the necessary knowledge and skills to develop practical AI solutions. Below is a suggested 14-week syllabus to guide you through the essential topics and resources.

* Week 1: Introduction to AI and Machine Learning

* History of AI

* Overview of machine learning, deep learning, and reinforcement learning

* Common machine learning algorithms

* Week 2: Python for AI and Machine Learning

* Basic Python programming

* NumPy, pandas, and matplotlib

* Setting up a Python environment for machine learning

* Week 3: Data Preprocessing and Exploration

* Data cleaning and transformation

* Handling missing data and outliers

* Feature engineering and selection

* Week 4: Supervised Learning - Regression

* Linear regression

* Ridge and Lasso regression

* Decision trees and random forests for regression

* Week 5: Supervised Learning - Classification

* Logistic regression

* k-Nearest Neighbors (k-NN)

* Support vector machines (SVM)

* Week 6: Unsupervised Learning

* k-Means clustering

* Hierarchical clustering

* Principal component analysis (PCA)

* Week 7: Neural Networks and Deep Learning

* Introduction to neural networks

* Activation functions and backpropagation

* Convolutional neural networks (CNNs)

* Week 8: Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM)

* Introduction to RNNs and LSTMs

* Time series forecasting

* Natural language processing (NLP) applications

* Week 9: Reinforcement Learning

* Introduction to reinforcement learning

* Q-learning and Deep Q-Networks (DQNs)

* Policy gradients and actor-critic methods

* Week 10: Transfer Learning and Pretrained Models

* Introduction to transfer learning

* Fine-tuning pretrained models

* Pretrained models for NLP and computer vision

* Week 11: Natural Language Processing

* Word embeddings and Word2Vec

* Text classification and sentiment analysis

* Text summarization and translation

* Week 12: Computer Vision

* Object detection and recognition

* Semantic segmentation

* Image-to-image translation

* Week 13: Generative Models

* Introduction to generative models

* Variational autoencoders (VAEs)

* Generative adversarial networks (GANs)

* Week 14: Ethics and AI Deployment

* Ethical considerations in AI

* Bias and fairness in machine learning

* Deploying AI models

Key resources to study these topics include popular textbooks, online courses, blogs, and tutorials. Some notable resources include:

* "Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow" by Aurélien Géron

* "Deep Learning" by Ian Goodfellow, Yoshua Bengio, and Aaron Courville

* "Reinforcement Learning: An Introduction" by Richard S. Sutton and Andrew G. Barto

* Online courses like Coursera's "Machine Learning" by Andrew Ng or fast.ai's "Practical Deep Learning for Coders"

* AI research papers from sources like arXiv.org, NeurIPS, and ICML

* By following this syllabus, you will gain a solid foundation in applied AI and be well-equipped to pursue more advanced topics or specific application domains.

Links and thoughts:

Top Tweets of the week:

https://twitter.com/ijustine/status/1657505930563448833

https://twitter.com/CaseyNewton/status/1657062682229633024

Footnotes:

[1] https://aima.cs.berkeley.edu/adoptions.html 

[2] https://aima.cs.berkeley.edu/Bibliography.pdf 

[3] https://scholar.google.com/scholar?hl=en&as_sdt=0%2C6&q=An+independent+study+based+introduction+to+machine+learning+syllabus+for+2022&btnG= 

What’s next for The Lindahl Letter? 

* Week 122: Will AI be a platform or a service?

* Week 123: Considering open source AI

* Week 124: Profiling OpenAI 

* Week 125: Profiling Hugging Face (open and collaborative machine learning)

* Week 126: Profiling Deep Mind

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the year ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com

Right now at the start of 2023, I would probably highlight 3 AI trends: generative models, automation, and legislation. Before we get into those specific topics let’s zoom out for just a second and look at two different reports you could read to get a sense of what is going on right now. One of the great places to start would be with the recently released 2023 AI Index report from the Institute for Human Centered AI. 

Nestor Maslej, Loredana Fattorini, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Helen Ngo, Juan Carlos Niebles, Vanessa Parli, Yoav Shoham, Russell Wald, Jack Clark, and Raymond Perrault, “The AI Index 2023 Annual Report,” AI Index Steering Committee, Institute for Human-Centered AI, Stanford University, Stanford, CA, April 2023. https://aiindex.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf 

You can look around the website they have setup here:

https://hai.stanford.edu/research/ai-index-2023

If the 386 page PDF seems like a lot of content to consume, then feel free to read the slimed down 2023 state of AI in 14 charts that they also produced

https://hai.stanford.edu/news/2023-state-ai-14-charts 

The other interesting report that I read recently was the Google Cloud 2023 Data and AI Trends Report which you can download by giving them your email. That step seemed annoying so I’m going to leave the link to the report below, but it might default back to the information page

https://services.google.com/fh/files/misc/data_and_ai_trends.pdf 

Those two reports aside, let’s jump into the 3 topics that will comprise the bulk of this missive. 

Generative models - generative models are probably the biggest thing happening at the start of 2023. We are setting chat based and image based models like GPT-4 and DALL-E 2 from OpenAI [1]. One of the things spinning out of the generative models are the low code, no code, and copilot efforts. This is one of the areas where you will find a lot of coverage in the media as the generative models are being used as search engines and the image generating efforts are so easy to use people have gone thumbnail happy using them recently [2]. 

Automation - We are going to see a tremendous amount of better automation out of the new technology that is available [3]. I’m tempted to mention that I consider a lot of the copilot efforts to be a part of early automation. Those copilot efforts will move from being inline assistive to probably a more prompt based delivery. That type of request and receiving copiloting is what I would consider a form of automation. 

Legislation - Countries are wholesale banning ChatGPT [4]. A lot of legislation will probably come into being related to trying to ensure safety [5]. I’m not going to spend a ton of time on this trend as it's going to unfold in a pretty predictable way.

You might find that this week the large language model from OpenAI has produced a better post than what I crafted. To that end, let's find out now what the ChatGPT model GPT-4 produces after being prompted about this one…

As of April 2023, the top trends in AI include the following:

* Federated Learning: This approach allows AI models to be trained on decentralized data while maintaining privacy. It has gained prominence as a way to harness the potential of data from various sources without compromising user privacy.

* AI Ethics and Regulation: As AI systems become more integrated into everyday life, there is an increasing focus on the ethical implications and regulatory measures required to ensure AI is used responsibly.

* Explainable AI (XAI): With the growing complexity of AI models, there is a rising demand for models that provide interpretable and understandable explanations for their predictions or decisions.

* AI in Healthcare: AI applications in healthcare are expanding rapidly, from diagnosis and treatment recommendations to drug discovery and personalized medicine.

* AI for Climate Change: AI techniques are being used to tackle climate change by optimizing energy consumption, forecasting extreme weather events, and modeling climate patterns.

* Edge AI: As IoT devices proliferate, there is a growing interest in running AI models on-device (at the edge) rather than relying on cloud-based processing. This approach can improve latency, privacy, and energy efficiency.

* AI for Cybersecurity: AI is being deployed to enhance cybersecurity measures by detecting and preventing potential threats and attacks using advanced pattern recognition and anomaly detection.

* Reinforcement Learning (RL): RL, a type of machine learning that focuses on decision-making, continues to gain traction as researchers explore its potential applications in various fields, including robotics, finance, and gaming.

* Natural Language Processing (NLP): NLP techniques continue to advance, enabling more sophisticated AI applications in areas like language translation, sentiment analysis, and chatbot development.

* AI and Robotics: AI is playing a key role in the development of advanced robotics systems, allowing for improved autonomy, decision-making, and adaptability in various domains, including manufacturing, logistics, and personal assistance.

These trends are expected to evolve as AI research and development continues to advance, addressing the opportunities and challenges associated with the integration of AI into various aspects of human life.

Links and thoughts:

Footnotes:

[1] https://openai.com/

[2] https://www.forbes.com/sites/ganeskesari/2022/12/22/ai-trends-for-2023-industry-experts-and-chatgpt-ai-make-their-predictions/

[3] https://venturebeat.com/programming-development/3-key-trends-for-2023-low-code-no-code-generative-ai-and-the-evolution-of-programming/ 

[4] https://www.cnbc.com/2023/04/04/italy-has-banned-chatgpt-heres-what-other-countries-are-doing.html 

[5] https://apnews.com/article/joe-biden-artificial-intelligence-science-technology-chatgpt-6948df344041ef1e794d199595bf69e9 

What’s next for The Lindahl Letter? 

* Week 121: Considering an independent study applied AI syllabus

* Week 122: Will AI be a platform or a service?

* Week 123: Considering open source AI

* Week 124: Profiling OpenAI 

* Week 125: Profiling Hugging Face (open and collaborative machine learning)

If you enjoyed this content, then please take a moment and share it with a friend. If you are new to The Lindahl Letter, then please consider subscribing. New editions arrive every Friday. Thank you and enjoy the week ahead.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.nelsx.com