Since I had the pleasure to listen to Jacopo Tagliabue’s presentations during my MA – when he was already a PhD candidate, I was struck. It was an important experience for me. It proved to me that intellectual achievements can be explained in a plain, clear, and passionate way. Encountering him during my own studies was a discovery I just tried to reiterate here in this interview. So, Jacopo – thank you for all your achievements. I hope they can be inspiring to many other Italian and international readers as well. I believe we need to speak more about hard-working, insightful people because, after all, the world can be wonderful if we are able to believe in it. Then, it is with my distinct pleasure to publish the interview on Scuola Filosofica – for those who don’t know it yet, is one of the leading philosophical blogs in Italy. In the name of Scuola Filosofica Team, our readers, and myself (Dr Giangiuseppe Pili), Jacopo: thank you!
[1] Dear Dr Jacopo Tagliabue, it is a distinct pleasure and a real honor to be your interviewer, thank you for your interest and availability. I hope you will enjoy the interview! So, thank you for joining Scuola Filosofica – Philosophical School! Let’s start with an introductory question: how would you present yourself to Scuola Filosofica readers?
I am Jacopo! Educated in several acronyms across the globe (UNISR, SFI, MIT), I’m now Lead A.I. Scientist for Coveo, after the acquisition of my own A.I. startup, Tooso. I failed the Turing Test once, but that was many friends ago.
[2] Let’s start with your own experience. You studied widely different topics from model theory, computability, and Artificial Intelligence. You are working at COVEO as a leading scientist. First, what was the path that leads you through philosophy to artificial intelligence and machine learning?
I read a book – Godel, Escher, Bach – in high school and decided logic and language should be the focus of my university studies. It was difficult to choose between math and philosophy, but I ultimately picked philosophy after seeing the multi-disciplinary offer of UNISR. I started coding during my B.A. internship and picked it up quite quickly – to be honest, I still suck at coding compared to my peers and colleagues, but I learned enough to make myself useful. In a sense, I was doing A.I. before A.I. was cool, so I’m glad the world has finally caught up with the idea that teaching language to machines is fun!
[3] Let’s go deeper into your vision of artificial intelligence. You work directly at the forefront of a cutting-edge field in an impressive, well-known firm. Indeed, your impressive scientific achievements are coupled with your striking working knowledge of AI. One of the reasons I hope you would have accepted to be interviewed was indeed this unusual blend between the abstract research and direct access to the work in the field. In addition, in COVEO’s homepage, I found this interesting inspirational mission: “Provide effortless tailored journeys with the Coveo Experience Intelligence Platform.” Where the stress is indeed on the word “Experience Intelligence”. So, in your own view, what is intelligence? Even Alan Turing stated that it was impossible – probably, to define it. But after 70 years of research and years of direct experience, what is your vision of it?
Yep, I am not going to attempt to solve a problem Alan declined to solve! While intelligence is impossible to define, I can touch upon something that I find to be a key component of intelligence, and this is that is the ability to quickly learn/adapt/generalize in new situation: I never played paddle tennis, but it took me 10 mins to start having fun since I was able to re-use all my explicit and implicit knowledge of tennis to bootstrap my learning curve – how do we build machines that can do that?
[4] Let’s come to the future… starting from the past. Frankenstein was the creature “born” by biochemistry, and electrodynamic, the sciences of the future for the XIX century. Then, it was thought that those two sciences could have led human beings to replicate a body and a mind entirely similar to ours. But, in the end, we are still waiting for Frankenstein. Then, psychology and neurosciences emerged, which told us that we were near to understand human intelligence. Inside the neurosciences, AI emerged as the leading science able to replicate human intelligence. After twenty years or so of relative stagnation of AI, today Machine Learning and Deep Learning brought us again into the conception that, after all, it is only a matter of time to shake the hand to a machine. Is it really just a matter of time? Is an AlphaZero-like program able to bring us to solving general intelligence?
It is a matter of time in the sense that I haven’t come across a convincing argument (only a LOT of unconvincing ones!) to support the idea that human behavior is not, in principle, reproducible by a machine. However, it is not a matter of time if we think of 10 years ahead. Too many pieces of the puzzle are still missing, a reality often highlighted by other A.I. researchers (e.g. Melanie Mitchell from SFI). What we learned from the last 10 years is not so much how to build smart machines, but that many “narrow” tasks that we thought required intelligence, really just require statistical pattern recognition. This obviously does not in any way diminish the incredible progress of computer science – but I think it’s important to jointly consider scientific achievements like the DeepMind Atari paper and industry-grade engineering systems built out of deep learning tools.
[5] In a recent interview you gave to professor Giuseppe Primiero – The Reasoner dec. 2019, you stated that you were and always had been interested in language. So, what do you think of the Turing Test today? First, are at least some machines able to pass it? Are we near to it? Why natural language seems so difficult to crack for a pure computational machine – if this is indeed the case?
Not that I know of. Language is very hard as it is potentially unbounded due to recursion: even if you tried, you cannot “memorize” English, so you need to generalize the underlying rules that produce an infinite number of sentences from finite atoms. We made some decent progress on syntactic rules, but we are still very far from a computational understanding of semantics, that is, how the meaning of a phrase emerges from the meaning of its constituents. Truth is, the two most popular perspectives on meaning (dense vectors vs. model-theoretic object) are very difficult to merge.
[6] Recently, AlphaZero simply changed chess. Indeed, for the first time, a machine is teaching Great Masters new ways to conceive their own field (chess). Personally, I super-enjoy AlphaZero games because they are very different from our way of playing so much that a good chess player once told me that AlphaZero is how a Martian would play – that is an entity far from our own knowledge and experience (and culture) would play. Do you think that AlphaZero constitutes a revolution even beyond chess?
There is a nice documentary on DeepMind playing Go, and it really highlights all these issues in a nice way: the fight between human intelligence and “alien” intelligence, and how much we can learn about ourselves by observing our inferences through the eyes of the machine. Scientifically, the most famous DeepMind challenges take place in the context of “reinforcement learning” i.e. a system learning through constant feedback from the environment. This works well in video-games, when you can “die” virtually thousands of times without consequences, but it doesn’t work quite as well in reality.
[7] Maybe neurosciences and AI still didn’t change the world as much as it could have been thought in the first stance. However, how do you see the future in this respect? For instance, some scientists think that one day will be able to download our minds into a computer. Personally, I wouldn’t like it just because my laptop is not very good at escaping from a burning house… What is your thought? And what about the singularity vision of AI?
As of today, your body is not very good at escaping from a fire / surviving a car accident, and your brain cannot copy itself when needed… so I definitely think that mind uploading would be kind of useful! As stressed above, I don’t see any reason, in principle, why our functional layer could not be copied “somewhere else”, but for now the discussion around the possibility of doing so is largely speculative: there are pressing problems caused by A.I. even today, well before the singularity. That said, I always liked deep discussions on “existential threats” – scenarios that are maybe unlikely today but potentially very dangerous to mankind – so I am glad that some serious scholars are thinking about the implications of radical technological progress.
[8] Let’s come to the current cyber threats. The public debate seems to be heavily orientated between who believes that cyber threats are everywhere and unstoppable, and who believes that we are far from seeing a cyber Pearl Harbor. The literature on cyberwar is as old as thirty years now and everything still looks possible. It looks as if the cyber-things are able to achieve everything we desire, and at the same time, it looks as if we are still dealing with long-lasting problems, from security to economic uncertainties – to put aside even more controversial issues considered by the recent pandemic. So, how do you see AI applied to cybersecurity?
By delegating a great deal of our societal processes to the “digital world”, we are obviously exposing ourselves to new vulnerabilities, but this also presents huge opportunities. As a point to prevent panic, the “old world” had a lot of vulnerabilities as well, but for other reasons (say, some complex patterns of global incentives) disruptive events like 9/11 were still rare. On top of that, technological progress can solve problems at scale: sure, pandemics in a globalized world are a big issue, but without technology (cell phones, Zoom, etc.) it would have been impossible to achieve any level of coordination/productivity. We definitely need to stay vigilant, but the future won’t be that bad if we pay attention.
[9] I want to conclude coming back on your own experience. How was your transition between academia and new activity in the market? This can be of inspiration for many of our young readers!
All in all, it was pretty smooth. Thanks to my internship, I picked up coding; from there, my somewhat serendipitous academic background turned out to be a great asset in industry, where technical skills need to be matched by “soft skills” and a general ability to see the “big picture”. I have been very lucky in having great mentors in my life! Every time I find myself learning something new, I still refer to a good Batman motto: “training is nothing, will is everything”.
[10] How would you define yourself in one sentence?
Whereof one cannot speak…
[11] May you give us five keywords that you would like to be linked to you?
Artificial Intelligence, NLP, Cognitive Sciences, Entrepreneurship
Be First to Comment