HNHacker News
TopNewBestAskShowJobs

fergal_reid

3,522 karma · joined September 13, 2010

@gmail.com: fergal.reid

https://www.linkedin.com/in/fergalreid VP of AI @Intercom PhD

Previous user id was 'feral'

Previously cofounder of Synference Predictive Analytics, applied Contextual Bandit / Reinforcement Learning tech, then PM of Optimizely Predictive Analytics team, following acquisition of Synference.

All views my own.

Coauthor of a paper on anonymity in Bitcoin, hence my interest in Bitcoin topics http://arxiv.org/abs/1107.4524

submissionscomments
fergal_reid··on What does copyright say about generative models?
I feel sympathy for the artists, and who knows what jobs AI will leave!

But I don't think the artists I'm seeing on Twitter understand that, even in their best case, where training on a copyright image is considered total infringement, this genie still isn't going back in the bottle.

Models are getting better and training more efficient all the time.

Soon you'll be able to train a model entirely on public domain content and have great generative art output.

It might be harder to express the style you want than just sharing a name of an artist in that style, but that won't hold things back.

With this worldview the discussion about whether training on something is infringing its copyright fast becomes irrelevant, for better or worse.

fergal_reid··on Building a Virtual Machine Inside ChatGPT
That's fair, its references for where chocolate is used are hallucinated (although it is the case that chocolate is used as a remedy in the books; I did check that independently!)

> it has no semantic comprehension

I'm not sure about this. I either don't think that's true, or I think we're going to need to start defining "semantics" in a circular way: as by definition something we conjecture humans can do that machines can't.

Now, I'm used to being the guy telling everyone ML models don't have semantic comprehension; but I don't think that's a sustainable position any more.

I'd say it has at least some degree of semantic comprehension.

It's got to have, in order to return the answer about Ron perhaps not getting sick.

It has to 'understand' that my query about getting sick refers to eating too much chocolate, and it has to understand enough to be uncertain of whether Ron will get sick because it thinks chocolate is a remedy for wizards.

That's quite a semantic thing to do. There's certainly no obvious or apparent way in which it's just doing some syntactic similarity.

Or, alternatively if we say that's just a super complex syntactic process, then I think that's like arguing there's no such thing as semantics, a bit like someone might argue there's no such thing as biology because it's all just physics lower down.

Yes, the model then goes too far and hallucinates a bunch of things, and that's a good example for never taking anything it says as authoritative.

But I think these examples are beyond certainly what I can identify as 'just' syntactic manipulation.

fergal_reid··on Building a Virtual Machine Inside ChatGPT
> There is no real abstraction.

I just tried to test it's ability to abstract directly, for your comment:

https://twitter.com/fergal_reid/status/1599191496372035584

I find it very hard to say this isn't doing abstraction.

I just made all those examples up, that's a new test I just created, so it didn't learn it elsewhere. I didn't cherrypick.

Isn't that doing out-of-sample abstraction?

That last sentence "species with 4 legs tend to be less intelligent and move at different speeds" is Abstraction or I don't know what Abstraction means.

(Not saying it's generally intelligent or anything; and I know it's just trained to predict symbols. But having a hard time with "...therefore it cannot abstract" in light of this sort of response.)

fergal_reid··on Building a Virtual Machine Inside ChatGPT
Well, Customer Support, which I build software for, for a start.

Existing bots are pretty good at answering simple informational queries. But the business has to do a lot of configuration work to make that happen, and that's a big ask.

Just the tech getting good enough to understand dialog more reliably, summarize, rephrase, ask for clarification, is a big deal.

Something like this or lambda configured for a particular business would likely handle a much larger percentage of their volume.

This is a big industry if you include phone, email, chat, all of which will eventually get bots; typically a single digit percentage of the workforce works in contact centers, stats like 4%.

It also seems like software development will be changed a lot.

Here's a related thought: how would you optimize your software stack to make it really easy for a tool like this to assist in?

Lots of dense self contained modules that can fit in token window, to maximize the systems ability to reason about them and alter them? Would that be worth it?

fergal_reid··on Building a Virtual Machine Inside ChatGPT
I've spent a couple of hours playing with ChatGPT since it was released and I am convinced people are not freaking out about it enough.

It is well beyond merely retrieving compressed knowledge; I see a lot of capability for synthesis and what I can only call 'understanding'.

I think a lot of people are getting tripped up by the fact it's trained to be very conservative in terms of what topics it will discuss. It's easy to think it's avoiding questions because it can't answer them. But often if you reframe the question you realize it is able to answer.

I also think there's an art to formulating your query. It's easy to get a bad response because your question is interpreted too literally, or didn't really say what you meant. Reframing can often get much better results.

It also appears bad at numbers/counting.

I think these are causing many people to underestimate it.

It's also easy to overestimate it sometimes - it often generates plausible but not quite right answers. You've got to read quite closely to make sure it's actually right.

But it's very good at a lot of things. Code synthesis. Summarization, Reformulation, explanation, information retrieval.

I think it's going to change whole industries.

It's also getting really hard to definitively say "well, AI definitely can't do X". X is getting small.

I think this is a watershed moment and people should be thinking hard about the implications.

This sample shocked me: https://mobile.twitter.com/fergal_reid/status/15981025960653...

I got it to write a poem, altered it step by step so that's clearly not something it will have seen before, and then asked it questions.

The bit where it explains why chocolate frogs mightn't make Ron from Harry Potter sick because of some wierd fact in that world, shocked me, because it's a really good retrieval of a very deep fact, relevant to a clearly out of sample input, and a good synthesis of that fact with reality.

It's amazing.

fergal_reid··on Parasite gives wolves what it takes to be pack leaders
That's misunderstanding brohee's post, which isn't talking about them leaving the pack but about them hunting individually before that time.

Which mightn't be what wolves do but that doesn't matter, it's a thought experiment.

fergal_reid··on Parasite gives wolves what it takes to be pack leaders
Havent read the paper, but your quote doesn't change the core point that it could be correlation.

E.g. perhaps aggression makes them more likely to eat infected animals and also more likely to challenge for leadership.

fergal_reid··on I stopped working on black hole information loss
I felt that this change reduced the quality of Veritasium, for me. I know it's just the titles, but it feels more bait and switch.

I understand why he did it, but wish he didn't. If I was a leader at YouTube I'd be pretty concerned about this - long term, it drives your platform in the direction of clickbait. Yeah there's benefits to that, but could also turn into a competitive weakness for YouTube. I think they should improve the recommender to make this unnecessary.

fergal_reid··on Entitlement issues (2009)
I don't agree with Gaiman here.

His basic argument is that the contract is you paid $10 for each book you got, so the author doesn't owe you the next book.

I disagree with this. First off, the fact that the fans are so upset is evidence that this wasn't _their_ understanding of the contract.

The thing is, they didn't just spend $10. They spent time getting invested in the story. This emotional investment is left unfulfilled if they don't get the implicitly promised ending.

Now, let's turn this around. Let's say fans do decide that their $10 strictly comes with no expectation of the next book coming.

Well then many fans wouldn't have bought the first books until the whole series is finished. It would likely become uneconomical to write long fantasy series. Writers like Martin would suffer.

It's easy for Martin or whoever to drop the ball on the implicit contract now, after becoming rich; but if this contract hadn't existed, authors like Martin would probably have desired to create it explicitly. If you believe that, then Gaiman's argument doesn't hold up.

And so not finishing a promised series is taking from that trust commons.

Of course, sometimes humans _can't_ finish something, and readers should accept that risk. The authors shouldn't be harassed.

But wanting them to follow through is not the same as thinking "they're your bitch".

fergal_reid··on Why babies cry in the first three months, how to tell them apart, and what to do
This may sound controversial, initially:

Consider getting a good pair of sound isolating earphones, or even ear defenders!

Not to neglect your child; but to reduce stress levels so you can problem solve effectively.

The cry of your baby is one of the most emotionally powerful sounds there is. You'll experience extreme stress if you can't calm them down quickly. Probably for good evolutionary reasons, so the parent doesn't ignore their child.

But sometimes there's a problem you just can't instantly solve.

If you find yourself getting really stressed, and losing your ability to act effectively, being able to put something on your ears and reduce the volume while you, e.g. change the nappy/diaper, can reduce stress and help you be a better parent.

I only had to do this a handful of times for each of our kids, but when it was helpful, it was really helpful.

fergal_reid··on Visual Information Theory (2015)
That's naive. Academia isn't some perfect system with perfect incentives.

There's plenty of incentive to dress something up as more complex and impressive and using more high powered machinary than needed, e.g. impressing reviewers, qualifying for a particular conference, convincing the funding body that your pet technique can actually be useful etc

I don't think you'd find many academics that dispute this, though hopefully they'd think a lot of people are trying to do good research.

It'd be like open source software if it was done as a hobby; a lot of academics get public funding.

fergal_reid··on When her best friend died, she rebuilt him using artificial intelligence
That seems a very specific/narrow point, not a general objection.

But anyway: I guess its possible human biological hardware does something difficult to simulate on a computer (at least one similar to our current architectures.)

Given a super advanced intelligence, though, I doubt adding hardware that does protein folding (or the protein folding analogue operation) fast - in biology if necessary - is going to be a major obstacle.

And that's if it can't come up with an algorithm that does an equivalent operation, that runs well on whatever hardware it already has.

fergal_reid··on When her best friend died, she rebuilt him using artificial intelligence
In fairness to Kurzweil, he isn't talking about a simple chatbot. That's a pretty big difference.

But beyond that, there's an argument that there's no clear distinction between a simulation of a person and the person.

If you want to get into this, start out by telling me the difference between A) a new sorting algorithm you wrote vs B) a simulation of the new sorting algorithm. In particular, why B isn't also running the same algorithm.

If that's not an objection, there's the question about "completely equivalent to the previously alive human being" - pragmatically, it seems impossible to me to fully recreate a 'completely equivalent' human from their writing/output; but, if you're willing to accept a 'so similar you cant tell', the idea is no longer as clearly ridiculous as it first seemed to me:

How big is the space of potential human minds?

Lets say you've an incredibly powerful intelligence: its got someone's DNA and their writings/output. And lets say it has a lot of data on humans, and has deduced a pretty good model of how their minds form, given their DNA and experiences. How much can it now deduce about how a deceased person's DNA was expressed, and the person they became, given their experiences? How small is the space of minds it could narrow in to, from that data? Can it find a small enough space that another human can't tell the difference?

How hard is it to model a human mind? I can make pretty good predictions about what people I know well will say in response to many situations.

But I don't know; I don't think we know enough about minds to say one way or the other.

← PreviousPage 2 of 2