AI Update [08/06/2023]
Hi all(emaal),
Again lots of things happening in our precious AI world (these seem to get longer and longer, let me know if it’s getting too long), let's get right to it.
News
- In the field of medicine we had some great advancements. Diffusion models (the technique underlying lots of current-day image generation) can now also be used in reconstructing human vision from reading out an individual’s brain activity. A paralyzed Dutchman regained his ability to walk through a brain implant, where AI was used to translate brain signals to nerve signals, bridging the literal and figurative ‘gap’ between his brain and his legs. Deep learning also continues to help in the biology department, as it’s really good at finding very complex patterns and predicting how they would behave in practice, leading to scientists finding a new promising drug to combat antiobiotic-resistant bacteria (an enormous problem that’s getting larger day by day). Lastly, a new Large Language Model by Google named MedPaLM is getting so good that a majority of doctors prefer its feedback on questions compared to the feedback from other clinicians.
- In the creative sector Adobe and Microsoft are embracing AI generation. Adobe has started integrating AI generated ‘filling’ in Photoshop, and launched their new Firefly service if you want to generate entirely new images from scratch. Microsoft meanwhile has launched their Microsoft Designer service, and is integrating generative AI wherever they can in their Office suite. Both Adobe and Microsoft are implementing hidden watermarks wherever generative AI is used to help make clear is something is human made, or (partially) AI generated. You may also have seen this new interface for more fine-tuned control of generating images through ‘dot dragging’, it quickly garnered lots of attention. Generally, how to interface with these popular AI generation tools is a hot topic right now, as companies don’t want consumers to have to become prompt experts in order to use their products properly (which is also a reason why ‘learning to prompt’ may not be as important in the near future as it currently seems). Unreal Engine (a popular game creation platform) has furthermore showcased their impressive Metahuman software, allowing anyone to easily create photorealistic 3D representation of themselves (or others), even including real-time rendering of muscle movement, clothing deformation, and hair (link). Lastly, see this article by an editor who lost his most valuable customer to generative AI, and where he sees the silver lining – plus how the advertising sector is going wild on generative AI, likely impacting the work of many.
- Much is also happening in regulation space, with Big Tech saying they really want regulation, but Sam Altman (of OpenAI) showcasing how quickly they can change their mind by doing a 180 twice in the span of two days. The G7 also convened, agreeing that they want a standardization in the global regulation terms used when talking about AI (what does it constitute, what basic rules apply, etc.), figuring that out is now called the Hiroshima AI Process. The EU furthermore called into being a new regulatory body, the ‘European Centre for Algorithmic Transparency’, aimed at making the ‘black box models’ used in social media apps and big search platforms more transparent. Ah, and Meta got a wrist slap of 1.3B euro for not keeping data within the EU, guess that’s regulation. For a small overview of 6 international regulations on AI and their supposed effective: link.
- In the realm of ethics, there was this interesting blog on how different languages ‘tokenize’ unequally, basically meaning that you need much more text to say the same thing in certain non-English languages (Burmese for example needs 10x more text to say the same thing as English). This has wide-ranging effects, as there’s a limit to the amount of text that Large Language Models can take in, so certain languages will a.o. need to pay more to do the same thing in these new chatbots or get a much less effective version of the application – through no real fault of anyone. Furthermore, teenagers have been going wild trying to break the MyAI chatbot of Snapchat, both finding loopholes circumventing its safety programming and finding joy in bullying it (let’s reinforce that behaviour shall we); LLMs are very much biased and ‘human feedback training’ just makes it get a different bias; and people seem to be subtly influenced in their opinion through their interaction with chatbots. A lovely combination these 3 last points make.
- Robots! They are gradually becoming more advanced and capable. The biggest news comes from Meta, who has developed a generalized 'visual brain' for robotic tasks. This model can learn from human activities and simulations, then apply this knowledge to complete tasks in the real world, without needing to learn ‘by doing’. The Teslabot also keeps getting better due in part to human feedback, and the valuable aspect of Large Language Models is that they are making it easier for us to tell robots what to do and translate that to actions - check out Tidybot for an example. Amazon is also trying to enhance their Astro robot by equipping it with a Large Language Model but don’t expect much from that. Self-driving cars also seem to have at least a promising future in making our roads safer and more convenient in well-mapped urban environment (with stable climates), a general model for autonomous driving still seems like a gnarly problem though.
- Microsoft recently hosted their Build event, which is all about giving developers the latest scoop. There are a few interesting updates. First off, Cortana, Microsoft's previous voice assistant, is out. Now, they're introducing something called Windows Copilot, which will be part of Windows 11 and it'll be right there on your screen as a sidebar 'widget' of sorts (Microsoft is using this '_ Copilot' name for all their new features that are powered by Large Language Models, like in their Office products). There’s big potential there as a handy interface for your files (‘hey I am looking for that word file I wrote last week about this meeting with X’) but.. also quite a concern wrt privacy and safety. They've also announced Microsoft Fabric, a complete platform for working with data and analytics. Finally, they introduced Azure AI Studio, a user-friendly platform that will make it easier for businesses to create, test, keep track of, and use their own AI models.
- People working in the computer hardware industry are having a pretty good time right now. NVIDIA, a big name in graphics cards, is seeing a lot of success on the stock market. Not to mention, several Dutch chip-making companies like ASML, ASM, Besi, and NXP are also doing great. Everyone seems to be focusing on creating specialized chips that are tailored for specific tasks, especially for AI. These chips come in all sizes, from the tiny ones you might find in your smartphone to the massive ones used in supercomputers.
- For the leftover corporate news, Zoom has invested a lot in Anthropic, likely aiming to replicate Microsoft’s ambitions with their Teams Copilot feature. Neuralink, a company that's working on implanting tiny brain-machine interfaces, has received the green light from the FDA to start human trials. Naver, the South Korean equivalent to Google, is sharing their AI models with foreign governments outside the Western and English-speaking world. Meanwhile, Palantir, a company we discussed earlier that's building an AI Military Platform, is getting lots of 'positive' attention from various countries. Over in the banking sector, JPMorgan Chase is using Large Language Models to analyze statements from the US Federal Reserve in an attempt to predict interest rate changes, and they're planning to develop more than 30 similar models for similar purposes like predicting if the EU will raise interest rates. In a small whoopsies moment some information about OpenAI's future plans was accidentally leaked by Sam Altman, although it was quickly taken down.. nothing disappears on the internet however, so the GPT4 roadmap looks like: 1) a faster and cheaper GPT-4 model, 2) longer context windows for the AI to refer to, 3) a more finely-tuned API, and 4) multi-modal abilities (combining text, images, etc.) expected in 2024.
Thoughts
- I have a hunch that edge computing, which means running applications right where the data is, like on your phone or laptop, is the way of the future. AI models are becoming more efficient to train, even on single graphics cards that you can buy off the shelf, and they're also getting smaller. I wouldn't be surprised if we start seeing Large Language Models primarily running on local devices soon, only reaching out to their more powerful counterparts or other external tools when necessary and permitted. We might even start seeing 'AI-cores' (or as I like to call them, 'AI-hype-cores') being heavily marketed for both consumer and business devices like smartphones and laptops. The companies making the chips for these devices will likely be focusing on producing specialized products for this 'next stage'.
- There’s quite a lot of doom and gloom about AI nowadays (although, the narrative seems to return every so often). Personally I think it’s overblown to regard AI as critical of a threat as nuclear war like many tech companies and experts are calling it (I’d rather say rampant human greed is the general threat). Andrew Ng amongst others is also calling for a sobering conversation and I’d be inclined to join him in that regard.
- Although, let’s stay with one gloom shall we, namely security’s biggest problem with LLMs: prompt injection. It’s a problem we have no answer to, and that’s a biggy. Highly advise even non-tech folk to visit this link to read/listen to a security expert explaining it much better than I ever could.
Fun
To end on a positive note, I came across this Veritasium video about the history of the Micromouse tournament, and I found it fascinating (and hilarious). It’s a bit more on the nerdy-side of things, but even for non-tech folk it’s easy to understand as everything is explained very well. Here’s the link for anyone interested: link.
Leven is mooi!
Fabian Kok
___
Onderzoeker Lectoraat AI, Hogeschool Utrecht
Onderzoeker Responsible Applied Artificial InTelligence (RAAIT) programme
Responsible & Effective AI, Technology & Creativity