The Skeptic AI Enthusiast #29

The current tech landscape seen critically by an AI veteran

Illustration for this week’s post: “The Missing Piece: Symbolic AI’s Role in Solving Generative AI Hurdles” made with Meta AI

My take on the news

• Google announced advanced voice chat:
  At its recent Made by Google event, Google announced that all its new phones include the Gemini voice chat, called Gemini Live. I checked it out from the event, and I can tell you it’s darn good! Gone are the days when artificial voices had that charming robotic tone. There is nuance, flow, intonation, and even emotion (Okay, simulated emotion, given that AI can’t feel a thing).
  The main takeaway is that Google has caught up with OpenAI, which recently announced an “advanced voice mode” for its premium users.
  Voice interaction could become the dominant mode of interaction with phones and AI in general… except for the privacy loss. As I published here a few weeks ago, I patented a “silent voice” device so you can talk to your phone, but nobody will hear anything, keeping our precious privacy.
  

• Chinese maglev vacuum train:
  You know that hyperloop trains are dead after both Elon Musk and Sir Branson pulled the plug on them, but this won’t stop the Chinese from advancing their own version. Their “T-flight” train, which is levitated over the ground by powerful magnets and travels inside a tube with reduced air pressure, has been recently tested. They recently achieved a top speed of over 600 kph in an extremely short track (just 2km), which is a lot compared to non-maglev trains like the French TGV (“just” 300 kph), but it’s still far from its intended speed of 4,000 kph (it’s not a typo).
  Funded by the Chinese government and without the profitability restrictions of private Western companies, I’d bet the Chinese will succeed. Not next year, not in five years, but I guess they’d be happy achieving this in a lifetime. They play the long game.
  

• Google's robot plays table tennis:
  Recently, Google posted the research article “Achieving Human-Level Competitive Robot Table Tennis,” which describes how a robotic arm learned to play table tennis at a human level. The hardware is pretty basic, but the modular software architecture described in the paper is very interesting.
  What I don’t really agree with is the goal of building robots that can compete with humans in the physical world. You see, it’s not too hard to build a robot that hits the ball ten times harder than a human and, thus, wins almost every match. Of course, such a robot would have “superhuman” capabilities, but what is the point?
  I remember a popular robotic competition called “Robocup,” where I participated in the software category with decent results (that is a way to say that we were eventually eliminated). The goal of the robotic category was “to compete at the level of the best human soccer teams,” which is ridiculous. Tell me, if you are a soccer player, and you have to dispute a ball with the head of a steel robot, would you dare to do it? I wouldn’t.
  

This week’s quote

The antidote to hype is to focus on concrete value.
—Eric Siegel, Machine learning expert, YouTuber (Big Think channel), entrepreneur.

I couldn’t agree more! I have written several times about why the hype about AI isn’t what’s important about it but the concrete problems it solves. Recently, I read an article by a consultant who complained about many clients who asked about how to bring AI to solve their problems; after closely looking at the problems, most of them didn’t require AI at all.

AI can and will bring value to enterprises and users, but this will need carefully designed products with specific use cases and specific value for the users.

This week’s resource

This page provides an accessible explanation of how the deep-learning transformer architecture works. It’s a very polished explanation that includes graphics and even animations, so I expect you to find it useful.

As you may know, the transformer architecture became the founding block of the LLMs’ revolution, and it all started with Google researchers' publication of the “Attention is all you need” article.

What is…?

…a cognitive architecture.

In short, it is a type of call used in a specific LLM application.

As explained in this post by LangChain, there are different ways of making calls to an LLM, from a single call and single response, to an indefinitely long series of calls. This is illustrated in the figure:

You can see the “levels” of autonomy in an LLM-based system, from conventional code (no call at all to an LLM) to the “autonomous” level. Each level below is one step towards agentic autonomy.

I think this conceptualization is nice to distinguish the different levels of complexity in LLM use, but perhaps I’m not convinced about using the “cognitive” term. There seems to be nothing “cognitive” in the way LLMs are called, in my opinion. But hey, once technical terms get sticky, it’s hard to change them. That’s exactly what happened with the “Artificial Intelligence” term…

Blog piece highlights

Yesterday, I published my blog post, “The Missing Piece: Symbolic AI’s Role in Solving Generative AI Hurdles.” Its highlights are the following:

• Current LLM-based AI systems are not close to being perfect; they still suffer from hallucinations, factual errors, and other problems.

• As I show in a simple example, LLMs struggle even with basic arithmetic and fail to explain their own errors.

• One problem with LLMs is that they have been trained to “sound correct” instead of “being correct.”

• When I started as an AI researcher in the mid-’80s, two camps were disputing AI's dominance: symbolic AI and Neural Networks.

• I spent my entire research career in the symbolic AI camp. Was it a good idea?

• Given the hurdles of LLMs, many have considered combining them with symbolic AI.

• There are currently two ways of doing so: one is to call the symbolic part from the LLM, and the other one is to feed the LLM with knowledge accessible to the symbolic part.

• The Wolfram plugin and RAG based on a Knowledge Graph have been the most successful combinations so far.

• The article explains step-by-step how both methods are carried out.

• There are other combination ideas, like using a “mixture of experts” architecture.

• The combination of symbolic AI and LLMs could be very promising in the future.

By the way, the figure at the top of this post is the one I originally intended to include in my blog piece, but the publication (Towards Data Science) doesn’t accept AI figures, only real-life photos. I guess they will have a hard time guessing whether the photo is real or not with some incredibly believable “photos” made by Midjourney or Flux.

Here is the end of the free preview of “The Skeptic AI Enthusiast.” Please become a paying subscriber for more exclusive content, including a “Friend Link” to the mentioned article and links for some curated articles on Medium.