Friday Squid Blogging: Japanese Divers Video Giant Squid - Source: www.schneier.com

Comments

not important • May 9, 2025 6:37 PM

Visit the Arctic vault holding back-ups of great works
https://www.bbc.com/news/articles/c7vnyn17p57o

=At the back of the chamber, another large metal box contains GitHub’s Code Vault.

The software developer has archived hundreds of reels of open source code here, which are the building blocks underpinning computer operating systems, software, websites and apps.

Programming languages, AI tools, and every active public repository on its platform, written by its 150 million users, are also stored here.

“It’s incredibly important for humanity to secure the future of software, it’s become so critical to our day to day lives,” Githhub’s chief operating officer, Kyle Daigle tells the BBC.

His firm has explored a variety of long-term storage solutions, he said, and there are challenges. “Some of our existing mechanisms can be stored for a very long time, but you need technology to read them.”

At Piql’s headquarters in southern Norway, data files are encoded onto photosensitive film.

“Data is a sequence of bits and bytes,” explains senior product developer, Alexey Mantsev, as film ran through a spool at his fingertips.

“We convert the sequence of the bits which come from our clients data into images. Every image [or frame] is about eight million pixels.”

Once these images are exposed and developed, the processed film appears grey, but viewed more closely, it’s similar to a mass of tiny QR codes.

The information can’t be deleted or changed, and is easily retrievable explains Mr Mantsev.

“We can scan it back, and decode the data just the same way as reading data from a hard drive, but we will be reading data from the film.”

One key question arising with long-term storage methods, is whether people will understand what has been preserved and how to recover it, centuries into the future.

… a team of scientists from the University of Southhampton have created a so-called 5D memory crystal, which has saved a record of the human genome.=

Clive Robinson • May 10, 2025 12:47 PM

Trying to fix AI hallucinations are making them more frequent.

For various reasons current AI LLM and ML systems produce significant amounts of erroneous output. Call it “soft bullshit” or “hallucination” is as far as quantity is concerned just a matter of taste, but the repeated occurrence tempo is apparently increasing as attempts to reduce it do not.

Is the finding of a round up of available reports by Nee Scientists,

https://www.newscientist.com/article/2479545-ai-hallucinations-are-getting-worse-and-theyre-here-to-stay/

The prognosis is that these errors are inherent in the way LLMs fundamentally work, thus hallucination or soft bullshit they are going to stack up more and more because any “fix” is just going to make things worse.

I can take a guess as to why this might be and there are several reasons that come to mind.

The first consideration is fact the training and update input information is basically shoveled in from the swamp that the Internet etc is, without any discernment is certainly not going to help.

Because like it or not the LLM is in reality little more than a complex filter that is designed by the square of the input information error statistics. Which might be great for producing a smooth curve for the descent algorithms but…

Look at the comments from Emily Bender at the University of Washington,

‘These models work by repeatedly answering the question of “what is a likely next word” to formulate answers to prompts, and so they aren’t processing information in the usual sense of trying to understand what information is available in a body of text’

She at least understands the issue of the failing of training data input statistics. And goes on to make a further couple of points with,

“‘Hallucination’ as a term is doubly problematic,” says Bender. “On the one hand, it suggests that incorrect outputs are an aberration, perhaps one that can be mitigated, whereas the rest of the time the systems are grounded, reliable and trustworthy. On the other hand, it functions to anthropomorphise the machines – hallucination refers to perceiving something that is not there [and] large language models do not perceive anything.”

Pointing out firstly language statistics for next word prediction are never going to be a sensible way of storing, searching, or assessing factual information. Thus all such LLM use is going to produce “soft bullshit” and is never going to be “unbiased” as far as usage requirements… Secondly making the output sound eloquent in human terms is going to cause hallucinations in the people using the LLM, we politely call “anthropomorphization”.

Let’s be honest if next doors cat started spouting “Shakespeare sonnets” you would be impressed, but honestly would you use it for health / legal / relationship advise?

If you think you would can I interest you in these equipment racks under a bridge?

After all who does not love a troll…

But the last point in the article goes to Emily Bender again with a hint at a 1980’s movie,

But the best move may be to completely avoid relying on AI chatbots to provide factual information

In short “Don’t play” (or pay for that matter).

Clive Robinson • May 10, 2025 1:44 PM

@ ALL,

From time to time I mention that “ICT Sec” has foundations in the standard infrastructure we all rely on. That is the supply of power, water, and communications.

I also point out that when “power” goes down so does water and communications shortly there after. Worse it causes a latch-up effect where both power and water need communications to be brought back up in a reasonable time period.

So anything that effects power in a major way is going to be a serious event. Around 60million in South West Europe discovered this awkward fact a few days back.

But that was due as far as we can tell due to “safety systems” tripping and cascading without actual damage.

However we know from the Victorian era that the earth can get hit by significant solar “Coronal Mass Ejections”(CMEs) the worst recorded being the Carrington Event of Sep 1859,

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

A time when there was little infrastructure and electrical communications was by hand telegraphy and chemical batteries.

A century and a half later, the situation is very different in certain Western Cities survival will not be possible if the power stays out for more than a week because not only will water not come in nor will food or the power to cook it, oh and sewage won’t get removed.

As a “little” hint as to what that means have a read up on a refuse removal stoppage due to strike action and political stupidity. That started at the begining of the year and effects over a million people and is now oft called the “Birmingham bin strike” the photos are quite graphic but unsurprising.

So imagine what will happen if the power grid goes down and can not be brought back up for months if not a year or more?

Well a full on CME / Carrington Event is very likely to “destroy the grid” if it is above ground.

The probability of a CME doing this is fairly obviously related to the frequency of CMEs and of their effective power and duration.

All of which are apparently getting worse according to this Scientific American article,

https://www.scientificamerican.com/article/stronger-solar-activity-cycles-may-be-in-the-suns-forecast/

As the article points out, humans have short memories and thus poor long term planning, and… as far as solar activity is concerned we’ve had it quiet for most of the past decade or so.

So consider what that means as far as “maintenance costs” and “shareholder value” and executive bonuses…

Subscribe to comments on this entry