Posts by Paul Kinsler

OpenAI says its latest model is less likely to beat around the bush

Re: I always wondered if

Hmm. Well I suppose if they generate word sequences according to some probability distribution, they should be able to report on the on the likelihood of any given output.

However, I imagine that for any long sequence this is probability likely to be tiny (given the large number of plausible/similar outputs), but if you used a logarithmic measure, and adjusted for the sequence length, and perhaps other things, I guess you could be given a sense of how "generic" the output was...

Anthropic writes 23,000-word 'constitution' for Claude, suggests it may have feelings

Re: Not really ...

Whilst at a bit of at a tangent to your "understanding" comment, some might find the following interesting:

https://zenodo.org/records/18231172

What is reasoning anyway? A closer look at reasoning in LLMs

U.Hahn,

There is a remarkable degree of polarisation in current debate about the capacities of Large Language Models (LLMs). One example of this is the debate about reasoning. Some researchers see ample evidence of reasoning in these systems, while others maintain that these systems do not reason at all. This paper seeks to shed light on this debate by examining the divergent uses of the term reasoning across different disciplines. It provides a simple clarificatory framework for talking about behaviour that highlights key dimensions of variation in how ‘reasoning’ is used across psychology, philosophy and AI. This highlights not just the extent to which researchers are talking past each other, but also that common inferences about model capability that accompany classification decisions are, in fact, far less compelling than they might seem.

Welcome to America - now show us your last five years of social media posts

Re: 3.434359 x 10^168186 corners is *rather* more than I was expecting.

But at least we might be grateful that it's fewer than TREE(3) corners.

Irish Excel whiz sheets all over the competition in Vegas showdown

"but I do any serious modeling in Excel"

I'm curious to know what sort of models this "serious modeling" referred to by Diarmuid Early involves.

Aviation delays ease as airlines complete Airbus software rollback

Re: The mystery wrinkles

Well, indeed; but the event described in the article appears to have occurred on 30 October, a day characterised by an unremarkable Kp, low solar proton flux, and no flares as evidenced by significant X-ray flux. One can, for example, scroll back and zoom out at ...

https://spaceweather.knmi.nl/viewer/

... to compare the two dates in question.

This aviation event may indeed have been a solar-activity related occurrence - spaceweather.knmi is far from comprehensive - but there seems no reason that I can see that justifies claiming it occurred during any kind of "solar storm" (although any corrections and additional evidence, are, of course, welcome).

Re: The mystery wrinkles

I was also a bit mystified; as far as I could tell from the CNN report the incident was c 18:48UT on the 30th (from 13:48EST), although and (as you say) nothing shows on space weather live. Perhaps it was a radio-only flare, and thus would not appear on the GOES X-ray flux?

TryHackMe races to add women to Christmas cyber challenge roster after backlash

Re: So they did try

It's thing that when you try to increase diversity on some sort of panel, committee, or whatever by inviting the under-represented, it somewhat tends to be from the same smallish pool of "obvious" candidates from the pool of the under-represented who get invited. However, these are people who are very possibly quite busy already, and so maybe are starting to decline some of the invites, however much they might -- ordinarily -- be inclined to "fly the flag", as it were, by attending. This also happens with the over-represented, but since there are many more of them, it's very much less of an issue.

The solution, of course, is to try to widen the pool of "the under-represented" candidates and not just plump straight for those with the highest profile.

Linux admin hated downtime so much he schlepped a live UPS during office move

Re: Checkpoints

I've manually coded in checkpointing at various times over my career for long-running simulations -- "long-running" for me being a week or more; but I've done some that took months to run :-/

Ubuntu 25.10's Rusty sudo holes quickly welded shut

Re: sendmail.cf

I editted a sendmail.cf precisely once; to remove an open relay that was there by default. I think the instructions were essentially "change just this one line", so that is all I did.

Some time later, I was asked in an interview if I had ever editted a sendmail.cf; so I told them the above. But I have never been sure, subsequently, if the question was intended as a trap (which I had then presumably failed), or a test of knowledge/experience. *shrug*

First stellar Coronal Mass Ejection detected beyond our Sun

It's locked and I couldn't find a preprint.

Since it's in Nature, and there was a press release, I expect it was under embargo.

FWIW, the detected event occurred May 2016; this result was obtained by mining old/existing data. Not that this was announced first off; you have to skim/read all the way past the references and into the "Methods" section to find this out.

Boffins: cloud computing's on-demand biz model is failing us

I don't know what the [...] boffins get up to,

It's likely to vary. For some, it will indeed be analysing a giant chunk of just-harvested data, or running very specific simulations, and so might well be occasional and compute intensive. But at the other extreme, there is continuous processing of incoming data in real-time (or near real time) ... as those in space weather, e.g. doing CME modelling/ forecasting, based on satellite or radio-telescope outputs.

Gullible bots struggle to distinguish between facts and beliefs

"how anthropomorphized all this research comes off"

Indeed. But perhaps since these models are touted as "AI", it is of some interest exactly to treat them as if they were, and see what an analysis reveals.

And possibly, with an anthro/psych framing like this, results might be more easily taken on board by those who either believe the "AI" claim, or at least who are inclined to treat their interactions with them in that way -- because that is how things *seem* to them.

The Chinese Box and Turing Test: AI has no intelligence at all

... the perfect name for LLMs

Hence:

https://thebullshitmachines.com/

BOFH: Saving the planet, one falsified metric at a time

Re: I think the river that runs through it was also called the Thames for a while?

Tis the Waihou now (or, rather, the Waihou _again_, I would imagine)

How do you solve a problem like Discovery?

Easy

Rename Texas as "Virginia", and Virginia as "Texas".

Solved :-)

Brit boffins teach fusion plasma some manners with 3D magnetic field

Re: We have to stop using Heat Engines to produce energy

I'm half inclined to put on a thermodynamics hat, and say that anything that does work is, in some sense, a heat engine. :-)

OpenAI releases bot-tom feeding browser with ChatGPT built in

Re: "There are only 10 types of people : those who understand binary and those who don't"

Or:

There are only 10 types of people : those who understand binary, those who don't, and those who understand Gray Code.

Square Kilometre Array is so sensitive, its datacenter needs two Faraday cages to stop RF leaks

Re: focused radiowaves?

It's SKA, not EISCAT-3D

:-)

Re: What are they going to do about people showing up on site with phones, watches, and ...

At a guess, detect them, with extraordinary sensitivity and precision, from quite far away :-)

Classic Psion fan releases proof-of-concept language server for OPL

Re: and one of those Sharp pocket computer things?

A Zaurus? I still have mine, although it's only use now is an alarm clock.

Two wrongs don’t make a copyright

Re: entertaining advertisement could be part of a narrative left hanging

This is (or at least was) a thing that used to be done sometimes on linear tv - the ad for a thing would change week-to-week or whatever, advancing some small narrative, and it indeed seemed to engage some viewers. But it probably relied mostly on the inertia of those stuck - or prepared to wait - in front of the tv whilst waiting for the ad break to finish (I mean, how many cups of tea might you actually need in an hour?), and what with all the time shifting, catchup, bingeing, and skipping capabilities now I'm not sure it could manage anything like the same traction.

Mysterious X-37B spaceplane flies again, this time carrying a quantum GPS alternative

take out our GPS for that part of the world

However, note that GPS-style GNSS satellites (GPG, Glonass, Galilleo, Baidou) - with the exception of a couple of Baidou ones - are not geostationary; so individual satellites do not cover just some specific part of the world; although there are a few regional services.

CIO made a dangerous mistake and ordered his security team to implement it

if he waited a year for new hardware ...

... he could then use the year to come up with a better and more intensive calculation that would *still* take a year of computer time, even if on faster hardware, and with optimized algorithms. :-)

IBM, NASA cook up AI model to predict solar tantrums

Re: Trained?

11 years. But then it might be that that is all the useful data they have, or that the solar cycle stage is not especially relevant to the characteristics of the cues in that data.

A Linux alternative? Debian/Hurd shows microkernel Unix dream is alive

Re: Microkernels have a *much* smaller attack surface.

I would have thought that it would be better to say that microkernels move most (or a lot of) the attack surface into userland; which might indeed solve or reduce the many problems introduced by that attack surface, but not -- I presume -- all of them.

Generative AI isn't just a matter of life and death. It's far more important than that

"If you want a picture of the future, imagine a boot stamping on a human face — forever."

I find it hard to believe you passed up the opportunity to say: "If you want a picture of the future, imagine a bot stamping on a human face — forever."

Some users report their Firefox browser is scoffing CPU power

Re: Too much like Firefox ...

Well, I'm not sure why one might expect anything else, given it's a firefox variant.

FWIW, if you run slackware, slackbuilds.org has a slackbuild that repackages the official binaries; so flatpak isn't an intrinsic requirement.

Hanging up: AOL to pull the plug on its dial-up service after 36 years

Re: I might just

Well, if you've got two, you can probably get one to dial the other? (cables permitting)

NASA boss calls for nuclear reactor on the Moon

Re: right now our power requirements on the Moon are nil.

OK, but if you want a deliverable lunar reactor for when you need it to appear, some design and testing will have to have been done in advance ... such as in this project. And if you expect to need a reactor to power your moonbase, you might even want to have it in place and running reliably for a while *before* it becomes a critical part of your infrastructure.

And in the meantime, some physicists will -- most likely -- be along shortly with innumerable projects for a wide variety of sensors, telescopes, and the like, all of which might benefit from a convenient power source. Might be tricky to get your robot to plug in the necessary powerboard/extension cable though. :-)

Behold the wood-block wonder of the Kilopixel display

invented steampunk.

Alternatively: K.W. Jeter

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

Quote: "Morlock Night is a science fiction novel by American writer K. W. Jeter. It was published in 1979. In a letter to Locus Magazine in April 1987, Jeter coined the word "steampunk" to describe it and other novels by James Blaylock and Tim Powers. "

Japan discovers object out beyond Pluto that rewrites the Planet 9 theory

Re: Two months a week and five years a month

I might hazard a guess that having spotted it, they were then able to go back through older data and spot it there also.

Ex-OpenAI engineer pulls the curtain back on a chaotic hot mess

Re: Actually looks like a great place to work at

As a general point, and irrespective of whether-or-not the subject is LLMs, many apparently "unsolvable" problems can be solved, if only you work out how - may be this is a moment of genius, or maybe some better mathematics or technology that overcomes the roadblock, and sometimes - like in some research - a sheer bloody-minded but systematic iteration through a vast array of combinations. And, as a case in point, computerised/robotic labs have made such brute-force efforts much more tractable.

It might be that LLM's are a true dead end, or it might be that they are not a true dead-end, but have some rare but valuable use-cases. But if finding out how to engineer those use-cases is hard or unlikely (as you suggested with your "dead end" remark), then IMO trying a wide and diverse mixture of strategies is probably more likely to be successful at finding that unexpected valuable use, than is focusing in on just a few decided at the top level.

But which does not necessarily mean that OpenAI are using their money wisely, or have their balance right; just that they might not be entirely wrong in their approach.

Scientists spot massive black hole collision that defies current theories

Re: not from inside the hole

Although see e.g.

Electromagnetic Energy Extraction from Kerr Black Holes: Ab-Initio Calculations

Meringolo et al

The possibility of extracting energy from a rotating black hole via the Blandford-Znajek mechanism represents a cornerstone of relativistic astrophysics. We present general-relativistic collisionless kinetic simulations of Kerr black-hole magnetospheres covering a wide range in the black-hole spin. Considering a classical split-monopole magnetic field, we can reproduce with these ab-initio calculations the force-free electrodynamics of rotating black holes and measure the power of the jet launched as a function of the spin. The Blandford-Znajek luminosity we find is in very good agreement with analytic calculations and compatible with general-relativistic magnetohydrodynamics simulations via a simple rescaling. These results provide strong evidence of the robustness of the Blandford-Znajek mechanism and accurate estimates of the electromagnetic luminosity to be expected in those scenarios involving rotating black holes across the mass scale.

https://arxiv.org/abs/2507.08942

Junior developer's code worked in tests, destroyed data in production

Re: I'm pretty sure the en dashes are purely ...

I did think that they also rather nicely illustrated the oncoming confusion.

Xlibre forks to the rescue – but Kubuntu gives X11 the boot

Re: [X11] not being able to deliver what people needed

I am not sure that overclaiming with "what people needed" really helps your argument. It should be pretty clear by now that many of the people who post here are perfectly happy with what X11 has delivered and still delivers (me included). Thus at best your statement has to become:

"X11 has a long history of not being able to deliver what some people needed".

Perhaps that "some" should be "many"; although where you might find reliable stats one way or another is an interesting question.

Broadly, however, just like with systemd, the choice of wayland vs X11 is not being actively made by anything like a majority of linux users; who most likely just make do more-or-less happily with whatever their distro-of-choice devs happen to prefer, and in all likelihood really haven't given the issue much thought (unlike us variously partisan commentards here).

SpaceX's Starship explodes again ... while still on the ground

Re: playing KSP.

If a Kerbal enthusiast, you might enjoy reading this:

"Large Language Models as Autonomous Spacecraft Operators in Kerbal Space Program", Carrasco etal.

https://arxiv.org/abs/2505.19896

Boffins found self-improving AI sometimes cheated

Aiiiiii

At a tangent, but since I was reading a couple of interrelated discussions on ways of thinking about AI/LLMs yesterday, I thought others here might be interested in them:

http://bactra.org/weblog/feral-library-card-catalogs.html

https://crookedtimber.org/2023/07/03/shoggoths-amongst-us/

https://www.science.org/doi/10.1126/science.adt9819

Datacenters near Heathrow seemingly stay up as substation fire closes airport

Re: [Optical fibres] It's not as if one has to inject photons at 400keV just ...

You might find it educational to read up on the subject: coupling light into an optical fibre is not necessarily trivial, and optical fibre will not usefully transmit any frequency of light you might wish to send (notably due to absorption, inconvenient dispersive properties, or indeed various other things).

Ubuntu 25.10 plans to swap GNU coreutils for Rust

Re: Speed comparisons

Indeed. I've got some ongoing hourly processing that regularly throws a couple of million lines of data through various stages often involving multiple instances of grep/sed/awk/sort etc; any sort of significant speed penalty would not be welcome.

Dash to Panel maintainer quits after donations drive becomes dash to disaster

I'd put a button on the about page for donations.

I think it would best to *start* the project off with a small, visible, but nonetheless tasteful and polite request for donations (or even just noting that they might eventually be requested). That way everyone will just automatically have to get used to it by default, and so regular users will not be shocked and/or annoyed by one suddenly appearing a few years down the track. :-)

Techie pulled an all-nighter that one mistake turned into an all-weekender

Re: This is the reason

"mc" exists on my Slackware 15.0, for whatever reason (I have almost never used it myself).

Since it seems fairly innocuous to me, so can anyone enlighten me as to what the problem is with it?

A last look at the Living Computers museum before collection heads to auction

My thesis is on 9-track magnetic tape.

I also have my thesis on some old tape format, but it doesn't matter that I can't read it -- I have had the file in my work archive copied from machine to machine to machine as I moved around.

The problem is, in fact, not how its been stored, but that it was created by the then version of MS Word. Fortunately I still have the nicely bound hardcopy version ... not that I need to read it very often.

Dark mode might be burning more juice than you think

Shocker

On the BBC page there is a link to a pdf. In it both OLED and LCD displays are mentioned, and the difference between OLED and LCD noted.

However, their single test device is in fact a Macbook Pro with an LCD display.

As a final remark, this seems to be a report which is really more about addressing user-expectations, rather than about technology differences. As in e.g. a user might have heard "dark mode saves power", but not twigged to the fact that it isn't true for their LCD display, especially once they perhaps have also put the brightness up a bit.

Copilot+ PCs? Customers just aren't buying it – yet

a "bullshit simulator".

Just saw this elsewhere ...

https://thebullshitmachines.com/

(a humanities course about how to learn and work and thrive in an AI world.)

Windows 10's demise nears, but Linux is forever

Re: My last major interface change was going from Blackbox to Fluxbox...

I think mine was twm to fvwm

:-D

Can AWS really fix AI hallucination? We talk to head of Automated Reasoning Byron Cook

Re: treat each citation it finds as a single unique token

They are presumably not infallible, but many scientific journals now automatically check the citations in submitted papers, and raise a query if they cannot find an authoritative match.

Thus you might imagine a scheme where any "citation" detected and tokenised, could also be tagged as validated, or as unvalidated; and if used, reported as such.

Unlikely to be infallible, and only workable in specific cases, sure -- but still an improvement. But you probably wouldn't want an LLM to do the validating :-)

cases to cite

It seems to me the training process, at the very least, needs to treat each citation it finds as a single unique token, rather than just another miscellaneous collection of characters or words. Then it would at least only generate actual citations, rather than merely some text that resembles a citation. And it might even manage to put them - sometimes - in a correct context, but I don't think you could rely on it - the reasoning behind why authors cite a thing is not always clear - it can range from some-generic-backgound, all the way down to a-specific-result-on-page-something.

UK ICO not happy with Google's plans to allow device fingerprinting

Re: On Android, GPS data can be spoofed. Search "android mock locations"

Hmm. Recently Google used raw GNSS data to run a study on android phones without explicit permission:

https://www.nature.com/articles/s41586-024-08072-x

If they can yank raw code phases and pseudodistances, what makes you think a position spoofing app is a bulletproof solution?

We told Post Office about system problems at the highest level, Fujitsu tells Horizon Inquiry

Re: an organised litany of lies

Only a minor point of pedantry, for which I apologise, but the exact phrasing used by Peter Mahon was -

"orchestrated litany of lies"

Huawei handed 2,596,148,429,267,413,
814,265,248,164,610,048 IPv6 addresses

It's also annoyingly vague about what a particle is

For these kinds of order-of-magnitude estimates it doesn't matter very much whether you count a neutron as one particle or its three constituent quarks (or, for the old school types, a proton plus an electron). The number is to get an idea of the scale, not to get a value which is supposed to "correct".