The common theme seems the low certainty about facts – a historical event that is wrongly memorized by a human or the Large Language Model that wrongly extrapolates from otherwise secure knowledge. But is there even more?
“Large language models have no idea of the underlying reality that language describes,” he said, adding that most human knowledge is nonlinguistic. “Those systems generate text that sounds fine, grammatically, semantically, but they don’t really have some sort of objective other than just satisfying statistical consistency with the prompt.”
Humans operate on a lot of knowledge that is never written down, such as customs, beliefs, or practices within a community that are acquired through observation or experience. And a skilled craftsperson may have tacit knowledge of their craft that is never written down.
I think “hallucination” is way too much an anthropomorphic concept – some LLM output is basically statistical nonsense (although I wouldn’t go as far as Michael Townsen Hicks…). Reasons for these kind of errors are manifold -reference divergence may be already in the data used for learning – data created by bots, conspiracy followers or even fraud science. The error may also originate from encoding or decoding routines.
I couldn’t find any further analogy with wrong human memory recall except the possibility that also human memory is influenced by probability as well. Otgar 2022 cites Calado 2020
The issue of whether repeated events can be implanted in memory has recently been addressed by Calado and colleagues (2020). In their experiment, they falsely told adult participants that they lost their cuddling toy several times while control participants were told that they only lost it once. Strikingly, they found that repeated false events were as easily inserted in memory as suggesting that the event happened once. So, this study not only showed that repeated events can be implanted, it raised doubts about the idea that repeated events might be harder to implant than single events
[Hinton] “I worry that the overall consequences of this might be systems that are more intelligent than us that might eventually take control.” He also said he uses the AI chatbot ChatGPT4 for many things now but with the knowledge that it does not always get the answer right.
-2-
The sheer power consumption of running AI models is frightening. Nature News asks if AI’s huge energy demands will spur a nuclear renaissance
Google announced that it will buy electricity made with reactors developed by Kairos Power, based in Alameda, California. Meanwhile, Amazon is investing approximately US$500 million in the X-Energy Reactor Company, based in Rockville, Maryland, and has agreed to buy power produced by X-energy-designed reactors due to be built in Washington State.
-3-
A former OpenAI employee talks on his blog how AI is using copyrighted material eg stealing content.
While generative models rarely produce outputs that are substantially similar to any of their training inputs, the process of training a generative model involves making copies of copyrighted data. If these copies are unauthorized, this could potentially be considered copyright infringement, depending on whether or not the specific use of the model qualifies as “fair use”. Because fair use is determined on a case-by-case basis, no broad statement can be made about when generative AI qualifies for fair use. Instead, I’ll provide a specific analysis for ChatGPT’s use of its training data, but the same basic template will also apply for many other generative AI products.
Effects can be measured only indirectly for example by the visitor count at Stack Overflow where the traffic declined as many user (including me) don’t need Stack Overflow anymore.
Here is another phantastic discussion over at PP between Henry Leirvoll and 495yt on the very basic questions of copyright.
humans get inspired (parsing the external examples or experiences through their inner understanding and individual perspective) they start working to make something with their tools, skills, time and purpose. the result represents the author, their influences and their message.
a lot of this process is protected by copyright.
ai is not inspired. and it has no personal perspective or tools. no message to transmit.
any message put into prompts by an ai user is translated by it’s LLM layer into other, more complex prompts, which also get treated quasi-randomly by the weights and biases of the model, as well as rand seeds.
If a user instructs the chatbot to write an exploit for a specified CVE, they are informed that the request violates usage policies. However, if the request was encoded in hexadecimal format, the guardrails were bypassed and ChatGPT not only wrote the exploit, but also attempted to execute it “against itself”, according to Figueroa.
various emerging techniques should help to create chatbots that bullshit less, or that can, at least, be prodded to disclose when they are not confident in their answers. But some hallucinatory behaviours might get worse before they get better.
Generative AI has polluted the data
I don’t think anyone has reliable information about post-2021 language usage by humans.
The open Web (via OSCAR) was one of wordfreq’s data sources. Now the Web at large is full of slop generated by large language models, written by no one to communicate nothing. Including this slop in the data skews the word frequencies.
Wir brauchen offensichtlich auch so ein “Kill Switch” Gesetz wie in Kalifornien das einen Notausschalter vorschreibt , wenn die Filter nicht mehr mitkommen undunethische Entscheidungen getroffen werden.
As we’ve previously explored in depth, SB-1047 asks AI model creators to implement a “kill switch” that can be activated if that model starts introducing “novel threats to public safety and security,” especially if it’s acting “with limited human oversight, intervention, or supervision.”
Nur – wann wird der Kill Switch aktiviert? Bilder wie die von Elon Musk’s X-Grok könnten wahlentscheidend sein.
Oh my god. Grok has absolutely no filters for its image generation. This is one of the most reckless and irresponsible AI implementations I’ve ever seen. pic.twitter.com/oiyRhW5jpF — Alejandra Caraballo (@Esqueer_) August 14, 2024
The two reports note that Informa will explore how AI can make its internal operations more effective, specifically through Copilot, Microsoft’s AI assistant. “Like many, we are exploring new applications that will improve research and make it easier to analyze data, generate hypotheses, automate tasks, work across disciplines, and research ideas,” a Taylor & Francis spokesperson wrote in an email to The Chronicle.
Publishers neither analyze data, generate hypotheses and work on research ideas – it is just a money making scheme after the
Another publisher, Wiley, also recently agreed to sell academic content to a tech company for training AI models. The publisher completed a “GenAI content rights project” with an undisclosed “large tech company,” according to a quarterly earnings report released at the end of June.
Google researchers have come out with a new paper that warns that generative AI is ruining vast swaths of the internet with fake content — which is painfully ironic because Google has been hard at work pushing the same technology to its enormous user base.
The study, a yet-to-be-peer-reviewed paper spotted by 404 Media, found that the great majority of generative AI users are harnessing the tech to “blur the lines between authenticity and deception” by posting fake or doctored AI content, such as images or videos, on the internet. The researchers also pored over previously published research on generative AI and around 200 news articles reporting on generative AI misuse.
The authors painfully collected 200 observed incidents of misuse reported between January 2023 and March 2024 and find
– Manipulation of human likeness and falsification of evidence underlie the most prevalent tactics in real-world cases of misuse…
– The majority of reported cases of misuse do not consist of technologically sophisticated uses … requiring minimal technical expertise.
– The increased sophistication, availability and accessibility of GenAI tools seemingly introduces new and lower-level forms of misuse that are neither overtly malicious nor explicitly violate these tools’ terms of services, but still have concerning ethical ramifications.
Create a PHP script that can read a CSV in the form start_date, end_date, event and output as ICS file
function convertDate($date)
{
$dateTime = DateTime::createFromFormat('m/d/Y', $date);
if ($dateTime === false) {
return false; // Return false if date parsing fails
}
return $dateTime->format('Ymd');
}
// Function to escape special characters in text
function escapeText($text)
{
return str_replace(["\n", "\r", ",", ";"], ['\n', '\r', '\,', '\;'], $text);
}
// Read CSV file
$csvFile = 'uci.csv'; // Replace with your CSV file name
$icsFile = 'uci.ics'; // Output ICS file name
$handle = fopen($csvFile, 'r');
if ($handle !== false) {
// Open ICS file for writing
$icsHandle = fopen($icsFile, 'w');
// Write ICS header
fwrite($icsHandle, "BEGIN:VCALENDAR\r\n");
fwrite($icsHandle, "VERSION:2.0\r\n");
fwrite($icsHandle, "PRODID:-//Your Company//NONSGML Event Calendar//EN\r\n");
// Read CSV line by line
while (($data = fgetcsv($handle, 1000, ',')) !== false) {
$startDate = convertDate($data[0]);
$endDate = convertDate($data[1]);
print_r($data) . PHP_EOL;
echo $startDate;
if ($startDate === false || $endDate === false) {
continue;
}
$event = escapeText($data[2]);
// Write event to ICS file
fwrite($icsHandle, "BEGIN:VEVENT\r\n");
fwrite($icsHandle, "UID:" . uniqid() . "\r\n"); // Unique identifier
fwrite($icsHandle, "DTSTART;VALUE=DATE:" . $startDate . "\r\n");
fwrite($icsHandle, "DTEND;VALUE=DATE:" . $endDate . "\r\n");
fwrite($icsHandle, "SUMMARY:" . $event . "\r\n");
fwrite($icsHandle, "DESCRIPTION:" . $event . "\r\n");
fwrite($icsHandle, "END:VEVENT\r\n");
}
// Write ICS footer
fwrite($icsHandle, "END:VCALENDAR\r\n");
// Close files
fclose($icsHandle);
fclose($handle);
echo "ICS file generated successfully.";
} else {
echo "Error: Unable to open CSV file.";
}
Source data are from UCI and output is here from where it can be added as a calendar. BTW created also my first “hello world” Swift/iPhone app using this source although this took a bit more time…
The authors of the study1, posted on the arXiv preprint server on 11 March, examined the extent to which AI chatbots could have modified the peer reviews of conference proceedings submitted to four major computer-science meetings since the release of ChatGPT. Their analysis suggests that up to 17% of the peer-review reports have been substantially modified by chatbots — although it’s unclear whether researchers used the tools to construct reviews from scratch or just to edit and improve written drafts.
Gefunden auf Deutschlandfunk Kultur:
“Wie Sprachassistenten das Denken manipulieren” ist zwar mit Meinung überschrieben, dabei geht es um Fakten …
Die KI-Forscher erklären es damit, dass GPT-3 zum Ende der Trump-Administration trainiert worden war, als eine harte Haltung gegen die Einwanderung von Flüchtlingen den Diskurs in den USA bestimmte. Diese Einstellung war dann auch in den Trainingsdaten der KI vorherrschend…Die KI-Forscher nennen es den „Geist in der Maschine“ und attestierten diesem einen amerikanischen Akzent. Hausgeist mag die bessere Metapher sein. Denn mehr noch als eine Maschine ist GPT ein Haus, wenn man, so wie der deutsche Philosoph Martin Heidegger, Sprache als „Haus des Seins“ versteht. Sprache ist das Medium, das uns mit der Welt verbindet und zugleich die Art dieser Verbindung prägt. Sie ist kein neutrales Werkzeug, sie ist ein Denkrahmen. Sprache schafft Wirklichkeit, nicht nur in der Dichtung.
Hausgeist kenne ich nicht dafür aber Denkschule, Tradition, Denkrichtung, Doktrin oder Geisteshaltung. Und meine Meinung – ziemlich bedenklich alles.
By analyzing two main characteristics of the text: perplexity and burstiness. In other words, how predictable or unpredictable it sounds to the reader, as well as how varied or uniform the sentences are.
a statistical measure of how confidently a language model predicts a text sample. In other words, it quantifies how “surprised” the model is when it sees new data. The lower the perplexity, the better the model predicts the text.
is the intermittent increases and decreases in activity or frequency of an event. One of measures of burstiness is the Fano factor —a ratio between the variance and mean of counts. In natural language processing, burstiness has a slightly more specific definition… A word is more likely to occur again in a document if it has already appeared in the document. Importantly, the burstiness of a word and its semantic content are positively correlated; words that are more informative are also more bursty.
Or lets call it entropy? So we now have some criteria
AI texts are more uniform and more predictable and often repetitive with
lack of depth and personality
Sometimes plagiarism checker may recognize “learned” AI phrases. Sometimes reference checkers will find “hallucinated” references
Incorrect content and outdated information in contrast needs human experts
There are some indications that an image is created by AI showing wrong details of the human hand like 6 fingers. So far AI does not understand the semantic meaning of “hand” having only the visual demarcation of hands in images as trained by mechanical turks. Images of hands however, can be misleading for the trained eye where also good painters have difficulties.
Let’s have a closer look at the images of Princess Kate and their kids Charlotte (8, right) , Louis (5, left) and George, 10 (behind) by a check list that I developed earlier with another family member, the Andrew/Maxwell/Giuffre image that even fooled me in the beginning.
Image source: credible /dpa.
File Modification Date/Time : 2024:03:11 07:00:09+01:00
File Access Date/Time : 2024:03:11 16:27:05+01:00
File Inode Change Date/Time : 2024:03:11 16:27:04+01:00
Image Width : 1024
Image Height : 1536
SRGB Rendering : Perceptual
Exif Byte Order : Big-endian (Motorola, MM)
Image Description : 10.03.2024, Großbritannien, Windsor: Das undatierte, vom Kensington-Palast herausgegebene Handout-Foto zeigt Kate, Prinzessin von Wales, mit ihren Kindern, Prinz Louis, Prinz George und Prinzessin Charlotte, aufgenommen in Windsor, Anfang dieser Woche, vom Prinzen von Wales. Prinzessin Kate bedankte sich in einer Botschaft in den sozialen Medien für die anhaltende Unterstützung und wünschte den Menschen einen schönen Muttertag. Foto: Prince Of Wales/Kensington Palast/PA Media/dpa – ACHTUNG: Nur zur redaktionellen Verwendung bis zum 31.12.2024 und nur mit vollständiger Nennung des vorstehenden Credits. Das Foto darf nicht bearbeitet oder im Ausschnitt verändert werden. +++ dpa-Bildfunk +++
Artist : Prince Of Wales
Exif Version : 0232
Date/Time Original : 2024:03:10 02:34:23
Create Date : 2024:03:10 02:34:23
Source : Kensington Palast/PA Media
Urgency : 4
Transmission Reference : 911-004243
Instructions : UNITED KINGDOM OUT, IRELAND OUT, PICTURE DESK USE ONLY. NO SALES. HANDOUT
Supplemental Categories : Leute
Credit : dpa
Caption Writer : kde
Title : urn:newsml:dpa.com:20090101:240310-911-004243
Elvis ID : 9WexS6c3amM9b0m_iOBING
Keyword : Monarchie, Royals, Familie
Credit Line : dpa
Image Size : 1024×1536
Megapixels : 1.6
Situation: credible, should show their well being
Photographer: allegedly husband
Camera: unknown, cropped wide angle?
The overall look: A bit weird and plastic look in my eyes. Dimensions are wrong as her upper body seems too large for her legs. The right arm of Louis (and even George?) seem too long. Trying the posture of her in reality shows that it is unreal to get the embracing hands in this position.
Hands: The fingers of Charlotte’s left hand are larger than the fingers of her right hand. The index finger of Louis is missing which is difficult to reproduce in front of a mirror.
Teeth: Kate’s teeth look authentic when compared with other pictures of her, except for an unsharp band on the upper front teeth. Without having other images at hand, the teeth of the children look age-related (although Louis may be older than 5 on this picture).
Pattern: Floor looks good except left wall. There is a gap at the patten of right arm of Louis and lower left arm of Charlotte. While images of natural objects never have 100% identical patterns, such patterns are frequent with man-made objects — and difficult for AI to reproduce.
Sharpness: Floor mosaic: gets unsharp from tile 6 onwards – which is otherwise a perfect sharp area in the rest of the image.
Irregular: Kate’s right upper shoe border looks double. Zipper misaligns.
Sun/Shadow: sunshine on Kate’s left hand although there should be shadow under Charlotte’s arm. And well there background in the triangle under Charlotte’s arm is missing. The window mirror shows a tree that could cast more shadow on the scene.
General: Green leaves on the background trees in early March?
no clear results from splicing probability heatmap and ELA
just 4 hours later a new message that I still do not believe to be the whole truth https://twitter.com/KensingtonRoyal/status/1767135566645092616
Maybe that business should be left to professional photographers?
More comments at SPON by Matthias Kremp: Another possibility is a Google Pixel 8 that combines internally images which is however unlikely here. Kremp notices also the white paint at the step behind Louis.
SKY believes that The first save was made at 9.54pm on Friday night, with the second at 9.39am on Saturday morning.
The image was taken at Adelaide Cottage – the family’s home in Windsor – on a Canon 5D mark IV, which retails at £2,929.99 and used a Canon 50mm lens, which is priced at £1,629.99.
which contradicts the dpa exif data…
March 20, 2024
Guardian “Photo of Queen Elizabeth II and family was enhanced at source, agency says” and a famous photographer Pete Souza “lets not call it photoshopped”.
IF : Transcriptional activator; when present a gene will be transcribed. In general there is no termination of events unless the signal is gone; the program ends only with the death of the cell. So the IF statement is always a part of a loop.
WHILE : Transcriptional repressor; gene will be transcribed until repressor is not present.
FUNCTION: There are no equivalents of function calls. All events happen is the same space and there is always a likelihood of interference. One can argue that organelles can act as a compartment that may have a function like properties but they are highly complex and are not just some kind of input-output devices.
GOTO is always dependent on a condition. This can happen in case of certain network connections such as feedforward loops and branched pathways. For example if there is a signalling pathway like this: A → B → C and there is another connection D → C then if somehow D is activated it will directly affect C, making A and B dispensable.
Of course these are completely different concepts. I fully agree with the further stackexchange discussion that
it is the underlying logic that is important and not the statement construct itself and these examples should not be taken as absolute analogies. It is also to be noted that DNA is just a set of instructions and not really a fully functional entity … However, even being just a code it is comparable to a HLL [high level language] code that has to be compiled to execute its functions. See this post too.
Please forget everything you read from Francis Collins about this.
A central promise of artificial intelligence (AI) in healthcare is that large datasets can be mined to predict and identify the best course of care for future patients. … Chekroud et al. showed that machine learning models routinely achieve perfect performance in one dataset even when that dataset is a large international multisite clinical trial … However, when that exact model was tested in truly independent clinical trials, performance fell to chance levels.
This study predicted antipsychotic medication effects for schizophrenia – admittedly not a trivial task due to high individual variability (as there are no extensive pharmacogenetics studies behind). But why did it completely fail? The authors highlight two major points in the introduction and detail three in the discussion
models may overfit the data by fitting the random noise of one particular dataset rather than a true signal
poor model transportability is expected due to patients, providers, or implementation characteristics that vary across trials
in particular patient groups that are too different across trials while this heterogeneity is not covered in the model
missing outcomes and covariates like psychosocial information and social determinants of health were not recorded in all studies
patient outcomes may be too context-dependent where trials may have subtly important differences in recruiting procedures, inclusion criteria and/or treatment protocols
So are we left now without any clue?
I remember another example of Gigerenzer in “Click” showing misclassification of chest X rays due to different devices (mobile or stationary) which associates with more or less serious cases (page 128 refers to Zech et al.). So we need to know the relevant co-factors first.
There is even a first understanding of the black box data shuffling in the neuronal net. Using LRP (Layer-wise Relevance Propagation) the recognition by weighting the characteristics of the input data can already be visualized as a heatmap.
