Tim at AGI MIT 2018

MIT is holding a series of free public lectures on "AGI" in 2018. They have a web page about their lectures here:

It looks a lot like a reimagining of their 2017 lectures on self-driving cars:

...though with a much larger number of famous invited speakers.

The instructor, Lex Fridman says he will post the videos on his YouTube channel.

I attended a number of the lectures in the series, I also asked questions at the end of every lecture I attended.

2018-01-22 - Lex Fridman - Artificial General Intelligence

I attended the first lecture. It was crowded, with somewhere around 200 people in attendance. Most of them appeared to be MIT students.

This was an introduction to the topic and to the course presented by Lex Fridman, the course instructor.

According to the slides there will be other lectures in 2018 - beyond those currently advertised on the web site. Lex Fridman mentioned subsequent planned lectures in February and March and said that he hoped to keep the course alive throughout the year. Here's his slide about the extended course schedule:


A video of the talk is now available here:

2018-01-23 - Josh Tenenbaum: Computational Cognitive Science

This was a great talk by Josh. Josh also had great slides. They were crammed with stimulating data and there were a lot of them.

I got to ask a question at the end and had a nice chat with audience members after the talk. Here is Josh giving a somewhat similar talk. Some of the slides are also the same.

Josh described the limitations of modern pattern recognition systems, giving examples, and argued that much more would be needed to create advanced machine intelligence. In particular, Josh says, it involves modeling the world, understanding what we see, imagining things, problem solving and planning.

While there's something to this perspective, I felt that Josh repeatedly and systematically undersold the potential of connectionist networks and modern neural networks.


Attendees 5 minutes before the start of Josh's talk. A lot more people showed up after that.

A video of the talk is now available here:

I asked a question at the end of the talk at 1:15:38

2018-01-24 - Ray Kurzweil: How to Create a Mind

I didn't attend this talk. A video of it is now available here:

2018-01-25 - Lisa Feldman Barrett: Emotion Creation

I didn't have high expectations for this lecture, but it was even worse than I had feared. Lisa promoted the concept of "social construction of emotion", apparently ignoring the evidence for cross-cultural common emotional states.

Another of Lisa's points was that intelligent machines would need some kind of bodies if they are to have human-like emotions. Meh.

Overall, I felt that the lecture suffered from poor presentation and was just weak.

I started to ask a question at the end of the lecture, but Lisa took issue with my premises and I was told that I was asking the "wrong question". Apparently, I was falsely assuming an association between hormones, neurotransmitters and emotional states. I guess that if you believe that emotions are socially constructed than any biochemical basis for them is a big no-no. Pfff. I never got to ask my question because the speaker and I were just not on the same page.

For me, the best part of the experience was a discussion with a fellow attendee named "Eli". I talked to Eli for about 20 minutes on the train ride home. It was a bit of a frustrating discussion because we digressed quite a lot and didn't seem to agree on very much. However, Eli did at least appear to know some stuff about machine intelligence - unlike this speaker.

A video of this talk is now available here:

2018-01-26 - Nate Derbinsky: Cognitive Modeling

Nate's talk was about cognitive architecture. It was quite nice and interesting. About a third of the talk was about SOAR.

I found myself wondering whether the modular cognitive architectures he described would be better off being implemented as multi-agent systems. Instead of rules for accessing short and long term memory stores, agents could consult with a library. ISTM that a multi-agent system approach would be more flexible.

My other issue with the talk was that Nate didn't really do a very good job of convincing me that systems much like the ones he described would be particularly important in the future.

Here is the video:

2018-01-29 - Andrej Karpathy: Deep Learning

I didn't attend this talk. For a long time it was listed as an upcoming video. However now there is no longer any trace of the talk on the MIT web site. It looks as though the video for this talk probably didn't work out for some reason. However, I notice that there's a recent lecture by Andrej Karpathy on YouTube here.

2018-01-29 - Stephen Wolfram: Knowledge-Based Programming

Here, Stephen Wolfram showed off his Mathematica / Wolfram Language product on stage. The presentation was interesting and accessible. Its knowledge-based approach reminded me of "Cyc".

A video of the talk is now available here:

Unfortunately, the video didn't manage to capture the screen during Wolfram's presentation, and a considerable amount of action took place there.

I asked the an audience question after this talk at 1:03:47. Wolfram gave a lengthy response.

2018-01-30 - Richard Moyes: AI Safety and Autonomous Weapon Systems

I didn't attend this talk. Richard Moyes is the managing director of Article 36. Their web site says:

Article 36 is a UK-based not-for-profit organisation working to prevent the unintended, unnecessary or unacceptable harm caused by certain weapons. Article 36 undertakes research, policy and advocacy and promotes civil society partnerships to respond to harm caused by existing weapons and to build a stronger framework to prevent harm as weapons are used or developed in the future.

I gather from attendees that this talk was quite negative about the topic - a "please slow down" talk. Here is the video:

2018-01-31 - Marc Raibert: Robotics

Mark is the CEO of Boston Dynamics. A recent version of spot mini was on stage, and we got to see a lot of unreleased footage from Boston Dynamics video archives. I shot some video of the robot in action. Most of the audience seemed to have the same idea. Here is one of my videos:

Marc Raibert shows up about 1 minute into this video. Here's the video of his lecture:

This talk was pretty well attended. There were around 300 people in it, I estimate. Many of them did not find anywhere to sit and stood around the edges of the room.

Lex also posted a much better video of the robot at the talk. You can see me in the audience if you know where to look.


Marc Raibert's audience.

2018-02-01 - Ilya Sutskever: Deep Reinforcement Learning

Ilya Sutskever was a bit challenging to understand initially. However as the talk progressed, I was able to grasp what he was talking about a bit better. The talk was interesting and Ilya Sutskever seemed sincere. He answered questions honestly. He would frequently answer with: I don't think we really know the answer to that yet, but I have some opinions. Saying when he did not know the answer was quite refreshing.

Here is Ilya Sutskever's talk:


Ilya Sutskever's audience.

2018-02-02 - Lex Fridman - Human-Centered Artificial Intelligence

Lex gave a teaser for his proposed Human-Centered Artificial Intelligence course. There's a web site for this at: https://hcai.mit.edu/.

Lex proposed that a symbiotic relationship between humans and machines would be productive and positive. I asked a rather sceptical audience question, suggesting that: symbiotic relationships can become unstable if there is too large a power imbalance between the parties - and it pays for one party to eat the other rather than trade with them - so wouldn't it be better to try and unite and fuse with them rather than attempt to coexist in symbiosis? Lex gave an extended reply.

No video for this talk was ever mentioned on the web site. Not did any video subsequently appear.

Attendees all got free T-shirts. Here is me wearing mine mine:


Me in my MIT AGI t-shirt.

Lex also announced that videos of all the lectures would be posted free online. Most of the talks were indeed subsequently posted, though two still seem to be missing.

Lex has also posting some related videos - which he describes as "conversations":

Tim Tyler | Contact