Seasons

You continue:

and therefore we can only go by the best evidence we have at the time, accepting what we have as true until further notice but recognizing limits to that certainty...

I agree with all that!

But then you continue:

and saying there is no external reality and all things are equally true.

I just don't see why this follows. After all, if we are epistemically fallible, then how do we know that "there is no external reality and all things are equally true"? If you knew that all things were equally true, you'd know everything; surely any being aware of its epistemic fallibility would shy away from making such a completely general, universal claim.

The fact that bears are hard to hunt doesn't mean that there aren't any bears. The fact that truths are hard to find doesn't mean that there aren't any truths.

The fact that bears are hard to hunt doesn't mean that everything is a bear. the fact that truths are hard to find doesn't mean that everything is a truth.

ALL RIGHT YOU SNIVELING COWARDS!!! SAY SOMETHING!!!
DANCE! DANCE! [Fires concepts at your feet]

Interzone:

Sorry, I will now continue the discussion from messages 569 and 5670.

I was saying that the clever and knowledgeable aliens could figure out, by looking in detail at a well-preserved mouse, that it was a foraging animal adapted to a certain class of environments, even though they had no previous experience with terrestrial fauna. I hope you find this intuitively plausible.

Suppose one of them said, "No, it's now a forager, it's a paperweight!" How would the others argue with him? I presume, something like this:

"Well, Grzx, its being a paperweight would explain its size, shape, and weight. But look how much more would be explained if we take it to be a forager in environment X! It explains tens of thousands of chemical pathways! It explains tens of thousands of minute details about its physical structure! See, this thing here is a pump! [and so on].

So we could say that X is appropriate for doing Y in context C if and only if

(a) There is some significant probability that X would indeed do Y in context C, and

(b) Almost any change we make in X, larger than a certain very small size, will reduce the probability that X would indeed do Y in context C.

If the aliens examined a human being, they would be able to discover what it would be appropriate for.

Now, very roughly, I say that it is better when things are in contexts where it is very likely that they will do what it is appropriate for them to do. We may of course have to balance out the requirements of one thing against another's.

At any rate, this is a purely physical notion of goodness. Are you with me so far?

to all:

Actually: I got the Seasons forum going again, but now it seems I've chased everyone but Bev away by being so intense, so I'm bowing out for a few days.

Irina, aww, you and I will just take a break from the topic, unless someone else wants to argue with you about it. Think about Psi's Robot or have a drink at Dogh'ds.

Psi, Thanks for the video links. What are the search words am I looking for to google this type of bot? Do the makers release any info or have their own web page? If you have some mutilated accident victim inside, is it a cyborg like my mouse-brained Roomba? You've already told me the ethics boards for human research are easier to avoid in the UK, how far can that be pushed?

Wow. I'm affixing a post-it to my monitor never to get into any remotely philosophical-theological debates with Irina or Bev!

Thanks for the good thoughts/prayers. Kaye had to stay an extra night (so I did, and so James the cat had to take care of himself). Everybody's home and pretty much okay, just tired and glad to be home. And not be sleeping in a recliner.

The fun part was when Kaye rolled over her call thingy and simultaneously turned off the lights, turned on the TV, and called the nurse's desk. So many tubes and wires . . .

I'll be interested to watch the next Season.

I found the Titan info at Cyberstein http://www.cyberstein.co.uk/ They have one of those annoying flash thingies you have to allow and let load, but they make a cool bot so I will forgive them. I am only half kidding when I say I would like to see this combined with cybernetic research. I am Ironman (or woman, whateveer).

Clerk! Sorry it took me so long to say I am glad Kaye is doing better. Hope you are both hanging in there.

from Newcomers:
Perhaps each tentative output could be parsed, and not used unless it parses out grammatical. This would eliminate certain common uses of language, but it would have the advantage that, even at the beginning, the output made some kind of sense.

Yes - I think the linkgrammar parsing is integral to the whole process (and as I say, already in place.) I think there must be some sort of parsing in jabberwacky (unlike in Nick,) if only because he would have been intolerably random for the first 100,000 conversations or so!

Two concerns: I don't see much of a role for the individual botmaster here, and it seems as though every bot will ultimately turn out to be a sort of average of everyone else. Where is the particular personality of the particular bot going to manifest?

If the bot interface only connected to the learning bot, then yes - it would be effectively one bot, with a single emergent "personality" aggregated from the sum of its learning. But applied to just the xnones, the personality of a bot would diverge as other keyphrases were added by the botmaster. From my own experience, fewer than 1 in 20 of BJ's responses are xnones (and the vast majority of the other 19 are from a comparatively small subset of his keyphrases,) so individuality of expression is assured, assuming a reasonable level of development.

It would have the advantage that even a newborn bot would have an interesting and complex personality (albeit the same as all the other newborn bots, though improving and maturing over time,) - no more "I was only just born and can't speak very well yet". It could also be programmed to kick in instead of "There are no valid responses" or "goto not found/too many gotos in a row"-type errors.

I'm pretty sure the resources aren't there yet for an individual learning bot for each Forgebot, though with Moore's Law still merrily rolling along in the background, that may be only a matter of time. If each Forgebot had its own neural net, in addition to its keyphrases/responses then they would surely end up very different from each other.
They could still follow all other bots' conversations too, and benefit from the increase in raw data, but proportionally weight input from conversations they had (or were had by other bots belonging to the same botmaster/other bots in the same catagories/other bots with similar keyphrases) more heavily.

Perhaps the bot could search for analogues in its written-out part.

It might indeed be possible to use the keyphrase rankings to add another layer of weighting to the neural net. And also to use the bot characteristics in its setup (friendly/neutral/hostile etc.) to further tweak the process, and provide some further individuality.

Perhaps we should continue this in "Seasons", since this forum is for airing specific problems in writing bots for the forge as it is.

good idea

fewer than 1 in 20 of BJ's responses are xnones

Good point! Granted that xnones usually account for only a small part of output, it would still be odd if they expressed an average personality while the rest of the bot had a distinctive one of its own.

I suppose it would be advisable to check the output with any relevant memories/responses which factually conflicted with the putative xnone substitute, but a little such unpredictability in its output might be considered an integral part of a bot developing its own personality (and not just reflecting its maker's.)
Humans themselves are rarely perfectly consistent, after all

No matter how big we make case-based bots, they will never achieve consciousness - the future belongs to learning bots that can 'scale up' as the technology becomes available.
The Forge is certainly the best platform for bot development there currently is, but unless it grows to incorporate learning-based AI, it won't be in 10 years time. Because by then the learning bots will have the hardware resources to catch up and massively supersede CBR.