Oh Go on

Tom Reading, Tech Leave a Comment

More on AlphaGo. Such respect to Sedol for choosing to play the last match against what he views as the stronger incarnation of the machine. It didn’t help, of course.

Fan Hui, the 3-time European champion has improved from number 600-ish in the world to be in the 300s while playing AlphaGo as part of its training. It’s been emotional, mostly sad, but he has found some beautiful moments.

Fabulous long Facebook reflections from Eliezer Yudkowsky. My take-aways (but it’s all worth reading).

  • AlphaGo is superhuman with bugs, not near-human.
  • Optimised strategies may look stupid, be strange edges of probability space and feel alien to us. We might have to get multiple AlphaGos to help us explain the “meaning” of moves.
  • We can’t necessarily see AlphaGo moves as powerful as we can’t see the consequences because we are working with too small a probability space.
  • “…when you’ve been been placed in an adversarial relation to something smarter than you, you don’t always know that you’ve lost, or that anything is even wrong, until the end.”
  • “AI is either overwhelmingly stupider or overwhelmingly smarter than you.” There’s not much space for human-level competence.

More on how DeepMind works. We should all start understanding this, as best we can. Suleyman thinks it is too early to be talking about AGI rights: many people “know how difficult it is to get these things to anything.”

AlphaGo and cyborg rats

Tom Personal, Reading, Science, Tech Leave a Comment

So much sci-fi coming to earth this last week.

AlphaGo beat Lee Sedol and convincingly. It’s important because the game was thought to be too “big” for traditional computer strategies based on simulating playing the game forward a few rounds and seeing what happens (like chess programs often do). Go has 10^761 possible games compared to the estimated 10^120 for chess.

For me the most astonishing thing about AlphaGo is that it was not designed to play Go. It is a generic learning engine that was trained on 30m Go positions from public databases and then played itself across 50 computers to reinforce its learnings. It has improved steadily over time and now plays “a little strange, but a very strong player, a real person”. Lots more in this Nature article. Lee Sedol won a remarkable game four where a stunning move from the 9p seemed to break the computer’s learning and caused it to play weak move after weak move leading to an eventual resignation from AlphaGo.

Just imagine what a pattern engine could do when applied to other endeavours for example the law (pattern: winning vs losing cases) and risk management (pattern: successful vs failed contracts), medical image diagnosis (pattern: life-years saved vs lost). Sadly, the Go learning can’t easily be transferred to another field of endeavour: AlphaGo is now a Go specialist and nothing else.

Even so, the speed at which these artificial general intelligences learn outpaces humans by orders of magnitude. We’re toast.

Separately, scientists have shown that rat cyborgs are better at solving mazes than rats left to their own devices. Maybe that’s our way back in?

Go. Gone?

Tom Reading, Tech Leave a Comment

Everyone expected computer domination of Go to be “ten years” away. It looks strongly like it’s nearly with us. A European master was defeated 5-0 and a match up with a world champion match is due in March.

Once again, the human player described the computer as “a wall”: it doesn’t make mistakes and doesn’t spend too much time on particular moves. It’s widely thought that Kasparov cracked under the pressure Deep Blue exerted in the rematch that he lost.

I wonder if humans will be able to find a winning strategy against this kind of pressure: it’s certainly possible. It seems as though human chess players haven’t quite given up yet.

Source: Digital intuition : Nature News & Comment

The horse has always bolted

Tom General Leave a Comment

Two articles covering different aspects of the fact that policy always has to follow developments in the world and is always rushing to keep up. This often leads to terrible law-making.

On privacy,  a postulated axiom:

the defense of privacy follows, and never precedes, the emergence of new technologies for the exposure of secrets

Lots of interest in the article: the discussion as to whether e.g. wire-tapping is enforced testimony is one I haven’t encountered before. I wonder if we should we all have some version of canary out there (“I have never been asked to provide my clients’ data to law enforcement or other government agencies”)?

On genetic engineering: we’ve been doing it for so long using more-or-less test-tubey methods that these hopefully discrete edits (hornless Holsteins, fast-growing salmon, non-malaria-carrying mosquitoes) shouldn’t be a cause for concern. Yes, we need to monitor very closely how they develop to see if there are any unexpected consequences, but we shouldn’t ban them up front.

We’re going to see a stream of edited animals coming through because it’s so easy


Man up

Tom Personal, View 6 Comments

Thanks Vancouver, you were beautiful and efficient. But clearly I was not butch enough for you. That is some ride down to Seattle. 

Now let’s hope for an easy border crossing.