alan-crowe · HackerTrans

alan-crowe

no profile record

Submissions

Do "Runge Spikes" offer a better metaphor for hallucinations by LLMs?

8 pointsby alan-crowevorig jaar0 comments

Is Bayesian probability individual, situational, or transcendental...

cawtech.freeserve.co.uk

2 pointsby alan-crowe15 jaar geleden0 comments

Saving forums from themselves with shared hierarchical white lists

cawtech.demon.co.uk

39 pointsby alan-crowe17 jaar geleden26 comments

Ghost Benefits

cawtech.demon.co.uk

2 pointsby alan-crowe18 jaar geleden0 comments

comments

4 uur geledendiscuss

People with serious illnesses long to get on to a clinical trial, hoping to get the latest and best treatment, even before it is approved. But it is not really like that. Only one phase III drug in ten really helps. And what of the other nine? Derek Lowe doesn't go into the downside, so I'm left to guess that eight out of ten do neither good nor harm and one in ten get their trial stopped in a panic when they turn out to be harmful. By all means volunteer for a clinical trial as an altruistic act, hoping to advance science and confer a benefit on those who come after. With such a high failure rate, the expected benefit is likely net zero.

2 maanden geledendiscuss

There is a folk phenomenon, called The Hum

https://www.bbc.co.uk/news/magazine-13752688

https://www.reddit.com/r/todayilearned/comments/xlsdk5/til_t...

so my guess is that LLMs see The Hum in their training data, and then put the word "hum" in their output. Since humans occupy varied, small media bubbles, many haven't encountered text talking about The Hum at all. The LLM's use of the word "hum" then stands out as excessive and a tell. And a mysterious one!

2 maanden geledendiscuss

As a child, I noticed that the proofs of mathematical theorems were esoteric knowledge, known only to a few adults. I struggled to follow even the simplest proofs, and hoped that one day I might learn to create a proof or two of my own. This was not only a high aspiration, but a dangerous one. I saw no reason why certain knowledge of a true fact would be accessible to humans via proof. Any-one who embarked on the quest to find a proof risked embarking on a doomed quest to find a non-existent proof.

For me Gödel's completeness theorem is the miracle. Every valid statement has a proof. Amazing!

Aim a little higher, every true statement, and there might not be a proof. It is no surprise to me that this is true. It is a big surprise to me that Gödel was able to prove it; ordinary proofs are hard to find, and proofs of the limits of provability presumably even more deeply hidden.

Non-standard models of arithmetic are weird. Theorems that are true of the standard model of arithmetic and false in some non-standard model must surely be convoluted and obscure. The first order version of the Peano axioms nail down the integers, not perfectly, but very well. Restricting one-self to theorems that are true in all models of them, even the weird, non-standard ones, feels like a very minor restriction. Gödel's completeness theorem raises the possibility of writing a computer program to find a proof of every theorem that isn't convoluted and obscure. Gödel completeness theorem is the really big deal.

Except it isn't. That computer program turns out to be one of those wretched tree search ones that soon bogs down. The real problem turns out to be the combinatorial explosion inherent in unstructured search through the Herbrand universe. One needs Unification and one needs a still missing ingredient to give search a sense of direction. The interesting questions are about the "sense of direction" that lets us find some of the deeply hidden proofs that do exist. Will LLM's help? The answer will be interesting, either way.

5 maanden geledendiscuss

In the UK, there are tax breaks for renting out your spare room https://www.gov.uk/rent-room-in-your-home/the-rent-a-room-sc...

I've done this for much of my life. It has always worked out OK. Perhaps because I've always picked a nice middle class person as my lodger. Even the mentally ill homeless person I rented my spare room to was a nice middle class mentally ill homeless person and a friend of a friend, who recovered from his reactive depression and moved on as anticipated.

If there is a university nearby, there are probably graduate students looking for a cheap place to stay while they finish writing up their PhD thesis. Since graduate students are pretty much guaranteed to be respectable, you may need to ask for less than the market rent to grab one, but you don't mention being short of money. They will have their own concerns about you, but your dog will vouch for you.

You don't mention having a spare bedroom that you could rent out, I'm just guessing that "working remote" implies living somewhere with affordable housing, and that you have extra space. If you live in a one bedroom flat, maybe sell it or rent it out entirely, and take the other side of the deal, becoming a lodger yourself.

That is not an option for me. I'm trapped by my stuff: grand piano, Boxford CUD lathe, too many maths books. I cannot vouch for that option, but it works for my lodger, so it works for some people. (I'm over sixty and he is older than me.)

I fear that my comment is a little "off" and not quite appropriate, but in my defense, it is responsive to "I sometimes panic when it's been too long since I've seen another person."

7 maanden geledendiscuss

My "importance of privacy" story:

I get my gas and electricity from Scottish Power. Recently a rival company, Ovo Energy made a clerical error and sent me a bill, leading to a dispute. The front line of defence against this kind of dispute is that the bills give the serial numbers of the meters. The bill from Scottish Power gives the same meter serial numbers that are embossed on the front of my meters, and is therefore valid. The bill from Ovo Energy gives different serial numbers and is therefore in error.

Picture though the internal processes in Ovo Energy. A second clerk is tasked with attending to the problem. He has a choice. He can change the address to agree with the meter serial numbers, correcting the error. Or he can change the meter serial numbers to those for my address, compounding the error.

Since the meter serial numbers are confidential, to me and Scottish Power, Ovo Energy does not have the second option; they do not know the serial numbers (which are long, like a credit card number, not just 1,2,3,...). Thus the clerical error gets corrected, or just left, but not compounded.

My guess is that confidential information, (such as meter serial numbers, credit card numbers, and account numbers), are the front like of defence against both clerical error and fraud based on impersonation. It is a rather weak defence, but it is light weight, and seems to how much of billing and billing disputes work.

We all have lots to hide: the confidential information that the system needs us to keep confidential to stop clerical errors from compounding.

7 maanden geledendiscuss

I'm still using fvwm2

    $ pkg info fvwm
    fvwm-2.6.9_4
    Name           : fvwm
    Version        : 2.6.9_4
    Installed on   : Mon Dec  8 02:01:51 2025 GMT
    Origin         : x11-wm/fvwm2
    Architecture   : FreeBSD:15:amd64

Very happy with it :-)

10 maanden geledendiscuss

On a 19 x 19 board the ranks are traditionally determined by the handicap needed to give an even game. So a 14 kyu would give 20 - 14 = 6 stones to a 20 kyu. 20 kyu is a rank that often sees rapid improvement, as the basic ideas "click" just through play and experience. You might be stronger than that now.

Handicap stones give a greater advantage on smaller boards, but 19 x 19 is the standard size. I've not seen any specific guidance for smaller boards.

10 maanden geledendiscuss

I'm on the other side of this, meeting a friend two or three times a month to play Go and giving him a three stone handicap on a 13 by 13 board.

Sometimes I play a move with a huge, but hidden threat behind it. If he plays elsewhere instead of answering locally, I get to play a clever sequence and capture some stones. I could just wait for the blunder and win. Instead I give a quick lesson in tactics: here is my plan, if you want to play elsewhere, your move needs to have an even bigger threat behind it.

He is learning, and now I face my clever moves being player against me. This makes it harder for me to win (it is about 50:50 with the handicap), but also more fun for me to play.

You could ask your spouses brother for a "teaching game" or a larger handicap, or a bit of both.

10 maanden geledendiscuss

You seem to be telling a tale of two industries, sports-betting and advertising. The sports-betting industry lobbies for reduced regulation, while the advertising industry sits on its hands, ticking along providing advertising for other businesses.

Eventually the sports-betting industry wins its lobbying campaign. But what did they win? The different sports-betting companies are in fierce and unprofitable competition, while the advertising industry walks off with all the profits.

That tale is at its most ironic when the advertising industry is consolidated and makes monopoly profits on advertising on behalf of sports-betting companies.

10 maanden geledendiscuss

It is tricky to boycott Amazon because when I search for a product using Google, I get lots of links to Amazon, and not much else. So I look at the Amazon pages to get search terms and then type those in to other search engines, such as Bing, Brave, or Yandex, and keep going to find online shops. Since Yandex is Russian, I add my country code to my search terms. I also try adding my city name and sometimes find a bricks-and-mortar shop close enough.

12 maanden geledendiscuss

LLMs have demonstrated that "intelligence" is a broad umbrella term that covers a variety of very different things.

Think about this story https://news.ycombinator.com/item?id=44845442

Med-Gemini is clearly intelligent, but equally clearly it is an inhuman intelligence with different failure modes from human intelligence.

If we say Med-Gemini is not intelligent, we will end up having to concede that actually it is intelligent. And the danger of this concession is that we will under-estimate how different it is from human intelligence and then get caught out by inhuman failures.

12 maanden geledendiscuss

What I found most interesting in the article was the frank acknowledgement by J B Rhine that Walter Levy, director of the Institute for Parapsychology was committing fraud.

I view research into parapsychology with a background assumption that I understand the motivation. Reductionist materialism suggests that death is the end. Parapsychology may imply the existence of souls, a spiritual realm, life after death, and maybe the chance of heaven. What wonderful balm for existential angst!

Does this make me doubt research into parapsychology? No, quite the opposite. A researcher can only soothe the pain of existential angst with genuine results. If they fake the research, they know that they faked it, and it provides no consolation. My take is that I cannot dismiss positive results in parapsychology as fraud.

Grifters are real. It is possible in principle that researchers in parapsychology are faking results with an eye to making money selling pills that "boost your ESP". I think that I am able to spot and dismiss grifts without difficulty, and that it is not what is at issue here.

But now I learn that I'm wrong. Walter Levy's results are fake. I have to flip from modus ponens to modus tollens. Instead of saying "I understand the motivation, therefore Levy's results are genuine", I have to say "Levy's results are fake, therefore I don't understand the motivation".

I'm in a pickle. A central principle of how I understand the world is that fraud is motivated. No motive, no fraud. Now what? Since there is motiveless fraud, much of what I thought I knew about the world is on shaky ground.

12 maanden geledendiscuss

I've been thinking of the issue like this: a traditional approach to symbolic artificial intelligence frames problem solving as a search of a tree. Do you go depth first, dodging infinite branches? Or breadth first (do you have enough memory?). What about iterative deepening?

Sometimes the solution is in the tree, but it is too deep, and one runs out of time before it is found.

Statistical based learning could act as a branch predictor. Sometimes guiding the search to go very deep in the right place and find the hidden solution. Sometimes guiding the search to go very deep in the wrong place; one runs out of time as usual.

Notice the strength of hybrid approach. One isn't accepting the probably correct answer of the statistical part. It is only a guide, and if the answer is found, and the symbolic part of the software is correct, the answer will be reliable.

I think this is already being done with maths problems. The LLM is writing proof attempts in Lean. But Lean is traditional symbolic AI. If the LLM can come up with a proof that Lean approves, then it really has a proof. From your comment I learn that Stockfish has already got something like this to work very well.

vorig jaardiscuss

We may be underestimating the effort that goes into cleaning up LLM messes. LLMs learn to program from code bases written by humans. Not just written by humans, maintained by humans. So the bugs that humans spot and remove are under-represented in the training data. Meanwhile, the bugs that evade human skill at debugging lurk indefinitely and are over-represented in the training data.

We have created tools to write code with bugs that humans have difficulty spotting. Worse, we estimate the quality of the code that our new tools produce on the basis that they are inhuman and have no special skill at writing bugs that we cannot spot, despite the nature of their training data.

vorig jaardiscuss

Nuclear powered ships and submarines.

https://en.wikipedia.org/wiki/Nuclear_marine_propulsion

The attraction is that high enrichment permits a smaller reactor and less frequent refuelling. US submarines go over 96% enrichment.

Bunker fuel is in the news. Cheap oil for marine diesel engines is both polluting in the traditional sulfur etc sense and in the modern CO2 global warming sense. There are occasional attempts at wind powered sailing ships, or at least sail assisted ships, to reduce the burning of bunker fuel. Nuclear power for shipping is currently out of fashion, but is a legitimate topic for research.

vorig jaardiscuss

I've no idea. I just want to mourn the utter failure of my 2001 prediction that telepresence would displace business air travel. http://web.archive.org/web/20160305121400/http://www.cawtech...

vorig jaardiscuss

LLMs have taught us that there is more than one kind of intelligence. They are definitely intelligent, but not the specific kind of intelligence we were hoping for. We get to be wise after the event and move the goal-posts. It is not so much that they have the wrong kind of intelligence, as that we never suspected the variety of possible forms of intelligence. We are clearly making progress towards ASI, but we don't know how distant the goal really is, because we don't know what the goal actually is.

Fusion is much better understood. We are not going to create "the wrong kind of fusion" and have to come up with a new plan.

vorig jaardiscuss

There is an interesting phenomenon with polynomial interpolation called Runge Spikes. I think "Runge Spikes" offers a better metaphor than "hallucination" and argue the point: https://news.ycombinator.com/item?id=43612517

vorig jaardiscuss

I've suggested Runge Spiking as an alternative, non-anthropomorphic metaphor

https://news.ycombinator.com/item?id=43612517

vorig jaardiscuss

I was expecting ancient history, the story of the Turing Institute in Glasgow. Established 1983. Dissolved 1994. https://en.wikipedia.org/wiki/Turing_Institute

The article is about the second try, commencing 2014.