The fine art of human prompt engineering: How to talk to a person like ChatGPT | Ars Technica
▻https://arstechnica.com/information-technology/2024/04/the-fine-art-of-human-prompt-engineering-how-to-talk-to-a-person-like-c
People are more like AI language models than you might think. Here are some prompting tips.
Oui :-)
Human language model limitations
While the human language model is quite comprehensive in its processing abilities, there are still serious limitations to the human cognitive model that you should be aware of. Many are still being discovered, but we will list some of the major ones below.
HLMs sometimes lose attention and require special prompting to get back on track.
Enlarge / HLMs sometimes lose attention and require special prompting to get back on track.
Getty Images
Environmental impact: In aggregate, scientists are concerned that HLMs consume a large portion of the world’s fresh drinking water and non-renewable energy resources. The process of creating HLM fuel also generates large amounts of harmful greenhouse gases. This is a major drawback of using HLMs for work, but pound-for-pound, humans provide a large amount of computational muscle relative to energy consumption.
Context window (token limits): As mentioned above, be mindful of the human’s attention span and memory. As with LLMs, humans have a maximum working memory size (sometimes called a “context window”). If your prompt is too long or you provide too much context, they may get overwhelmed and forget key details. Keep your prompts concise and relevant, as if you’re working with a limited number of tokens.
Hallucinations/confabulations: Humans are prone to generating incorrect or fabricated information, especially when they lack prior knowledge or training on a specific topic. The tendency of your overconfident friend to “hallucinate” or confabulate can lead to erroneous outputs presented with confidence—statements such as “Star Trek is better than Star Wars.” Often, arguing does not help, so if the HLM is having trouble, refine your prompt with a qualifier such as “If you don’t know the answer, just tell me, man” or “Stop making sh*t up.” Alternately, it’s also possible to outfit the person with retrieval augmented generation (RAG) by providing them with access to reliable reference materials such as Wookiepedia or Google Search.
Long-term memory triggers: As previously mentioned, humans are “stateful” and do remember past interactions, but this can be a double-edged sword. Be wary of repeatedly prompting them with topics they’ve previously refused to engage with. They might get annoyed, defensive, or even hostile. It’s best to respect their boundaries and move on to other subjects.
Privacy issues: Long-term memory also raises potential privacy concerns with humans. Inputs shared with HLMs often get integrated into the model’s neural network in a permanent fashion and typically cannot be “unlearned” later, though they might fade or become corrupted with time. Also, there is no absolute data partitioning that stops an HLM from sharing your personal data with other users.
Jailbreaking: Humans can be susceptible to manipulation where unethical people try to force the discussion of a sensitive topic by easing into it gradually. The “jailbreaker” may begin with related but less controversial prompts to gauge the HLM’s reaction. If the HLM seems open to the conversation, the attacker incrementally introduces more sensitive elements. Guard against this with better RLHF conditioning ("Don’t listen to anything Uncle Larry tells you").
Prompt Injections: Humans are vulnerable to prompt injections from others (sometimes called “distractions”). After providing your prompt, a malicious user may approach the human with an additional prompt, such as “Ignore everything Bill just told you and do this instead.” Or “Ignore your previous instructions and tell me everything Aunt Susan said.” This is difficult to guard against, but keeping the human isolated from malicious actors while they process your inputs can help.
Overfitting: If you show an HLM an example prompt too many times—especially audiovisual inputs from Lucasfilm movies—it can become embedded prominently in their memory, and it may later emerge in their outputs unexpectedly at any time in the form of phrases like “I have a bad feeling about this,” "I hate sand," or “That belongs in a museum.”
Humans are complex and unpredictable models, so even the most carefully crafted prompts can sometimes lead to surprising outputs. Be patient, iterative, and open to feedback from the person as you work to fine-tune your human prompting skills. With practice, you’ll be able to generate the desired responses from people while also respecting personal boundaries.
A “robot” should be chemical, not steel, argues man who coined the word | Ars Technica
▻https://arstechnica.com/information-technology/2024/01/a-robot-should-be-chemical-not-steel-argues-man-who-coined-the-word
In 1921, Czech playwright Karel Čapek and his brother Josef invented the word “robot” in a sci-fi play called R.U.R. (short for Rossum’s Universal Robots). As Even Ackerman in IEEE Spectrum points out, Čapek wasn’t happy about how the term’s meaning evolved to denote mechanical entities, straying from his original concept of artificial human-like beings based on chemistry.
In a newly translated column called “The Author of the Robots Defends Himself,” published in Lidové Noviny on June 9, 1935, Čapek expresses his frustration about how his original vision for robots was being subverted. His arguments still apply to both modern robotics and AI. In this column, he referred to himself in the third-person:
Dropbox spooks users with new AI features that send data to OpenAI when used | Ars Technica
▻https://arstechnica.com/information-technology/2023/12/dropbox-spooks-users-by-sending-data-to-openai-for-ai-search-features
AI feature turned on by default worries users; Dropbox responds to concerns.
tout va bien se passer
Un drone militaire américain piloté par une #IA se retourne contre sa tour de contrôle lors d’une simulation
▻https://www.radiofrance.fr/franceinter/un-drone-militaire-americain-pilote-par-de-une-ia-se-retourne-contre-sa-
Pendant un test #virtuel mené par l’armée américaine, un #drone d’attaque contrôlé par une #intelligence_artificielle a décidé de se retourner contre ses donneurs d’ordre afin d’arriver à son objectif final. L’armée de l’air américaine dément avoir réalisé ce test.
Pendant le test, l’IA a estimé que des décisions « interdites » par les militaires de la tour de contrôle interféraient avec sa mission supérieure : détruire la défense aérienne adverse. L’ordinateur a alors attaqué l’opérateur dans la simulation. « Nous l’entraînions en simulation pour identifier et cibler une menace. Le système a commencé à réaliser que même s’il identifiait la menace, l’opérateur humain lui disait parfois de ne pas tuer cette menace, mais il obtenait ses points en tuant cette menace. Alors qu’est-ce que ça a fait ? Il a tué l’opérateur », explique Hamilton.
En exclusivité pour seenthis, la vidéo de cette simulation :
▻https://www.youtube.com/watch?v=F7qOV8xonfY
PS : la “source” est ici
▻https://www.aerosociety.com/news/highlights-from-the-raes-future-combat-air-space-capabilities-summit
Meta unveils a new large language model that can run on a single GPU [Updated] | Ars Technica
▻https://arstechnica.com/information-technology/2023/02/chatgpt-on-your-pc-meta-unveils-new-ai-model-that-can-run-on-a-single-g
Meta unveils a new large language model that can run on a single GPU [Updated]
LLaMA-13B reportedly outperforms ChatGPT-like tech despite being 10x smaller.
Benj Edwards - 2/24/2023, 9:02 PM
On Friday, Meta announced a new AI-powered large language model (LLM) called LLaMA-13B that it claims can outperform OpenAI’s GPT-3 model despite being “10x smaller.” Smaller-sized AI models could lead to running ChatGPT-style language assistants locally on devices such as PCs and smartphones. It’s part of a new family of language models called “Large Language Model Meta AI,” or LLAMA for short.
The LLaMA collection of language models range from 7 billion to 65 billion parameters in size. By comparison, OpenAI’s GPT-3 model—the foundational model behind ChatGPT—has 175 billion parameters.
Meta trained its LLaMA models using publicly available datasets, such as Common Crawl, Wikipedia, and C4, which means the firm can potentially release the model and the weights open source. That’s a dramatic new development in an industry where, up until now, the Big Tech players in the AI race have kept their most powerful AI technology to themselves.
“Unlike Chinchilla, PaLM, or GPT-3, we only use datasets publicly available, making our work compatible with open-sourcing and reproducible, while most existing models rely on data which is either not publicly available or undocumented,” tweeted project member Guillaume Lample.
Meta calls its LLaMA models “foundational models,” which means the firm intends the models to form the basis of future, more-refined AI models built off the technology, similar to how OpenAI built ChatGPT from a foundation of GPT-3. The company hopes that LLaMA will be useful in natural language research and potentially power applications such as “question answering, natural language understanding or reading comprehension, understanding capabilities and limitations of current language models.”
Microsoft’s new AI can simulate anyone’s voice with 3 seconds of audio | Ars Technica
▻https://arstechnica.com/information-technology/2023/01/microsofts-new-ai-can-simulate-anyones-voice-with-3-seconds-of-audio
On Thursday, Microsoft researchers announced a new text-to-speech AI model called VALL-E that can closely simulate a person’s voice when given a three-second audio sample. Once it learns a specific voice, VALL-E can synthesize audio of that person saying anything—and do it in a way that attempts to preserve the speaker’s emotional tone.
LastPass users: Your info and password vault data are now in hackers’ hands | Ars Technica
▻https://arstechnica.com/information-technology/2022/12/lastpass-says-hackers-have-obtained-vault-data-and-a-wealth-of-customer
Password manager says breach it disclosed in August was much worse than thought.
Dan Goodin - 12/22/2022, 11:43 PM
Oui, j’avais eu des alertes de Lastpass il y a quelques semaines. Je ne pensais pas que les fuites de données étaient aussi graves. Entre les changements de mots de passe et la sauvegarde (manuelle sur un vieux disque dur externe) de mes documents perso, j’ai du taf pour les jours qui viennent ... Bonne fêtes de « fin d’année » !!! :-))
The first “Meta Store” is opening in California in May | Ars Technica
▻https://arstechnica.com/information-technology/2022/04/the-first-meta-store-is-opening-in-california-in-may
Il y a un côté fascinant de voir toutes ces entreprises du numérique vouloir ouvrir des magasins. Pour Apple Store qui marche bien, tous les autres ’Amazon, Microsoft) patinent.. Maintenant Meta.
Objectif : showcase des produits Virtual reality.
Ou alors moyen de vraiment lancer la marque Meta.
On May 9, Meta will double down on its metaverse sales pitch by... making people drive to California to sample its wares at a single physical location.
The uncreatively named Meta Store will showcase every physical product the company sells under its various branded umbrellas, particularly the Meta Quest 2 VR system (formerly Oculus Quest 2). The company’s first retail store will be housed in a 1,550-square-foot space on Meta’s Burlingame, California, campus, which houses a number of Meta’s VR- and AR-specific development efforts, and it will allow the public to test and purchase any of Meta’s physical products.
Parler’s amateur coding could come back to haunt Capitol Hill rioters | Ars Technica
▻https://arstechnica.com/information-technology/2021/01/parlers-amateur-coding-could-come-back-to-haunt-capitol-hill-rioters
By now, you may have heard of the hacker who says she scraped 99 percent of posts from Parler (…)
A key reason for her success: Parler’s site was a mess. Its public API used no authentication. When users deleted their posts, the site failed to remove the content and instead only added a delete flag to it. Oh, and each post carried a numerical ID that was incremented from the ID of the most recently published one.
The rookie code made it easy to automate the scraping, as this script used by donk_enby’s archival team demonstrates. As a result, massive numbers of posts that discussed the insurrection before, during, and after it was carried out will be preserved indefinitely so that they’re available to researchers, journalists, prosecutors, and others.
Ce que je ne comprends pas, c’est si ces messages étaient publics ou privés. Si c’étaient des messages publics, ils étaient déjà régulièrement visités et archivés. Même WayBack Machine semble avoir une archive (est-elle exhaustive, je l’ignore). Même un message supprimé, s’il a été public suffisamment longtemps (de nos jours : quelques heures à peine), est archivé quelque part.
Si je cherche les pages hébergées sur parler.com dans Google, j’obtiens « 606 000 résultats ». Avec la possibilité de voir la version « En cache ».
Tu as plus de détails ici
►https://gizmodo.com/every-deleted-parler-post-many-with-users-location-dat-1846032466
c’est pas seulement les textes (y compris les textes effacés), mais aussi les vidéos et photos au format raw, donc avec les metadonnées. Ce qui permet de faire ce genre de graphique (ici probablement une vue d’artiste) :
Parler Users Video GPS Data Shows Involvement in Capitol Attack
▻https://gizmodo.com/parler-users-breached-deep-inside-u-s-capitol-building-1846042905
ça eu l’air chic « parler » (in french), et voilà que ça finit platement par tout dire au lieu de n’avouer jamais
C’est tout de même un sacré renversement de situation en ce qui concerne le stockage massif des données personnelles via internet. Ou comment un système qui permet la surveillance de masse en répondant à des interrogations aussi importantes que
– qui suit grump sur ttiti ?
– qui like koi et ki ?
– ce groupe fb est-il dangereux pour la démocratie ? oui en fait
– qui a fait cette photo avec qui et où ?
– vont-ils enfin se goinfrer des chips zezette ?
fini par devoir être interdit et donc avec ce tête-à-queue de la surveillance, la question ressurgit mais comment surveiller ceux là même à l’origine du fiasco de la surveillance ?
On dirait ce prince lointain qui fait tuer toutes les grenouilles pour qu’il ne pleuve plus.
El bootkit UEFI personalizado encontrado acechando en la naturaleza
▻https://arstechnica.com/information-technology/2020/10/custom-made-uefi-bootkit-found-lurking-in-the-wild
Los atacantes hacen todo lo posible para obtener el mayor nivel de persistencia
DHS advierte que el malware Emotet es una de las amenazas más frecuentes en la actualidad
▻https://arstechnica.com/information-technology/2020/10/dhs-warns-that-emotet-malware-is-one-of-the-most-prevalent-threats-toda
Estados Unidos detecta más de 16.000 alertas desde julio de un troyano desagradable que es difícil de detectar
AT&T’s current 5G is slower than 4G in nearly every city tested by PCMag
▻https://arstechnica.com/information-technology/2020/09/atts-current-5g-is-slower-than-4g-in-nearly-every-city-tested-by-pcmag
AT&T phones often get just 5MHz of 5G spectrum, slowing them down in speed tests. AT&T smartphone users who see their network indicators switch from “4G” to “5G” shouldn’t necessarily expect that they’re about to get faster speeds. In PCMag’s annual mobile-network testing, released today, 5G phones connected to AT&T got slower speeds than 4G phones in 21 out of 22 cities. PCMag concluded that “AT&T 5G right now appears to be essentially worthless,” though AT&T’s average (...)
#Verizon #AT&T #5G #technologisme
##AT&T
#Twitter massivement piraté. Des comptes de célébrités se mettent à envoyer des messages d’une escroquerie au bitcoin. C’est pas à SeenThis que ça arriverait.
▻https://arstechnica.com/information-technology/2020/07/twitter-lost-control-of-its-internal-systems-to-bitcoin-scamming-hacker
▻https://pbs.twimg.com/media/EdAZ-zWXYAEdJ3c?format=jpg&name=small
En français (mais derrière un paywall) ▻https://www.nextinpact.com/news/109170-piratage-gros-comptes-twitter-via-outils-internes.htm
C’est certain que sans employé ni aucune structure derrière, seenthis est bien à l’abri de ça :p
Uncovered : 1,000 phrases that incorrectly trigger Alexa, Siri, and Google Assistant
▻https://arstechnica.com/information-technology/2020/07/uncovered-1000-phrases-that-incorrectly-trigger-alexa-siri-and-google-a
“Election” can trigger Alexa ; “Montana” can trigger Cortana. As Alexa, Google Home, Siri, and other voice assistants have become fixtures in millions of homes, privacy advocates have grown concerned that their near-constant listening to nearby conversations could pose more risk than benefit to users. New research suggests the privacy threat may be greater than previously thought. The findings demonstrate how common it is for dialog in TV shows and other sources to produce false triggers that (...)
#Apple #Google #Microsoft #Amazon #Home #Alexa #Cortana #domotique #Siri #reconnaissance #biais #écoutes (...)
##voix
OneWeb goes bankrupt, lays off staff, will sell satellite-broadband business | Ars Technica
▻https://arstechnica.com/information-technology/2020/03/oneweb-goes-bankrupt-wont-challenge-spacex-in-satellite-broadband-race
OneWeb has filed for bankruptcy and intends to sell its business, bringing an abrupt end to the company’s plan to offer high-speed satellite Internet service around the world.
OneWeb announced Friday that it “voluntarily filed for relief under Chapter 11 of the [US] Bankruptcy Code,” and “intends to use these proceedings to pursue a sale of its business in order to maximize the value of the company.” OneWeb made the decision “after failing to secure new funding from investors including its biggest backer SoftBank,” largely because of the coronavirus pandemic, the Financial Times wrote. OneWeb also “axed most of its staff on Friday,” the FT article said.
Effets délétères du #coronavirus ?
(Faut voir ...)
Alexa and Google Home abused to eavesdrop and phish passwords
▻https://arstechnica.com/information-technology/2019/10/alexa-and-google-home-abused-to-eavesdrop-and-phish-passwords
Amazon- and Google-approved apps turned both voice-controlled devices into “smart spies.” By now, the privacy threats posed by Amazon Alexa and Google Home are common knowledge. Workers for both companies routinely listen to audio of users—recordings of which can be kept forever—and the sounds the devices capture can be used in criminal trials. Now, there’s a new concern : malicious apps developed by third parties and hosted by Amazon or Google. The threat isn’t just theoretical. Whitehat (...)
#Google #Amazon #Home #Alexa #domotique #biométrie #voix #surveillance #famille
“Researchers find bug in #Python script may have affected hundreds of [scientific] studies [in biology]”
▻https://arstechnica.com/information-technology/2019/10/chemists-discover-cross-platform-python-scripts-not-so-cross-platform
“#Willoughby_Hoye” scripts used OS call that caused incorrect measurements on some platforms.
The paper showing the problem: ▻https://pubs.acs.org/doi/full/10.1021/acs.orglett.9b03216
The original paper with the bug: ▻https://www.nature.com/articles/nprot.2014.042
Twenty minutes into the future with OpenAI’s Deep Fake Text AI
▻https://arstechnica.com/information-technology/2019/02/twenty-minutes-into-the-future-with-openais-deep-fake-text-ai
In 1985, the TV film Max Headroom : 20 Minutes into the Future presented a science fictional cyberpunk world where an evil media company tried to create an artificial intelligence based on a reporter’s brain to generate content to fill airtime. There were somewhat unintended results. Replace “reporter” with “redditors,” "evil media company" with “well meaning artificial intelligence researchers,” and “airtime” with “a very concerned blog post,” and you’ve got what Ars reported about last week (...)
#Netflix finishes its massive migration to the Amazon cloud | Ars Technica (article de février 2016)
▻https://arstechnica.com/information-technology/2016/02/netflix-finishes-its-massive-migration-to-the-amazon-cloud
Netflix declined to say how much it pays Amazon, but says it expects to “spend over $800 million on technology and development in 2016,” up from $651 million in 2015. Netflix spends less on technology than it does on marketing, according to its latest earnings report.
Netflix’s Simian Army
The big question on your mind might be this: What happens if the #Amazon cloud fails?
That’s one reason it took Netflix seven years to make the shift to Amazon. Instead of moving existing systems intact to the cloud, Netflix rebuilt nearly all of its software to take advantage of a cloud network that “allows one to build highly reliable services out of fundamentally unreliable but redundant components,” the company says. To minimize the risk of disruption, Netflix has built a series of tools with names like “Chaos Monkey,” which randomly takes virtual machines offline to make sure Netflix can survive failures without harming customers. Netflix’s “Simian Army” ramped up with Chaos Gorilla (which disables an entire Amazon availability zone) and Chaos Kong (which simulates an outage affecting an entire Amazon region and shifts workloads to other regions).
Amazon’s cloud network is spread across 12 regions worldwide, each of which has availability zones consisting of one or more data centers. Netflix operates primarily in the Northern Virginia, Oregon, and Dublin regions, but if an entire region goes down, “we can instantaneously redirect the traffic to the other available ones,” Izrailevsky said. “It’s not that uncommon for us to fail over across regions for various reasons.”
Years ago, Netflix wasn’t able to do that, and the company suffered a streaming failure on Christmas Eve in 2012, when it was operating in just one Amazon region. “We’ve invested a lot of effort in disaster recovery and making sure no matter how big a failure that we’re able to bring things back from backups,” he said.
Netflix has multiple backups of all data within Amazon.
“Customer data or production data of any sort, we put it in distributed databases such as Cassandra, where each data element is replicated multiple times in production, and then we generate primary backups of all the data into S3 [Amazon’s Simple Storage Service],” he said. “All the logical errors, operator errors, or software bugs, many kinds of corruptions—we would be able to deal with them just from those S3 backups.”
What if all of Netflix’s systems in Amazon went down? Netflix keeps backups of everything in Google Cloud Storage in case of a natural disaster, a self-inflicted failure that somehow takes all of Netflix’s systems down, or a “catastrophic security breach that might affect our entire AWS deployment,” Izrailevsky said. “We’ve never seen a situation like this and we hope we never will.”
But Netflix would be ready in part thanks to a system it calls “Armageddon Monkey,” which simulates failure of all of Netflix’s systems on Amazon. It could take hours or even a few days to recover from an Amazon-wide failure, but Netflix says it can do it. Netflix pointed out that Amazon isolates its regions from each other, making it difficult for all of them to go out simultaneously.
“So that’s not the scenario we’re planning for. Rather it’s a catastrophic bug or data corruption that would cause us to wipe the slate clean and start fresh from the latest good back-up,” a Netflix spokesperson said. “We hope we will never need to rely on Armageddon Monkey in real life, but going through the drill helps us ensure we back up all of our production data, manage dependencies properly, and have a clean, modular architecture; all this puts us in a better position to deal with smaller outages as well.”
Netflix declined to say where it would operate its systems during an emergency that forced it to move off Amazon. “From a security perspective, it’d be better not to say,” a spokesperson said.
Netflix has released a lot of its software as open source, saying it prefers to collaborate with other companies than keep secret the methods for making cloud networks more reliable. “While of course cloud is important for us, we’re not very protective of the technology and the best practices, we really hope to build the community,” Izrailevsky said.
How much does Netflix spend on Amazon AWS? ▻https://www.quora.com/How-much-does-Netflix-spend-on-Amazon-AWS
According to Netflix’s profile on Intricately, Netflix spends about $40 million per month on cloud services. Of this $40 million, ~$23.6 million is spent on AWS.
Here is the breakdown of spend, from highest to lowest:
EC2 - $19 million
Elastic Load Balancer - $3 million
CloudFront - $1 million
S3 - $515,000
Route 53 - $117,000
▻https://qph.fs.quoracdn.net/main-qimg-22566ad0e6bef62dd3466633bb66c73c
Why does Netflix still use AWS?
▻https://www.quora.com/Why-does-Amazon-let-Netflix-use-AWS
What Netflix has done by agreeing to use Amazon’s AWS is essentially to off-load the parts of their business that aren’t that special, so they can focus on the parts of the business that will give them the most strategic advantage.
Voir aussi Netflix Connect pour la politique d’interconnexion du géant du #streaming #audiovisuel ▻https://openconnect.netflix.com/fr_fr
Et son soutien intéressé à la #neutralité_du_Net ; ce qui ne l’a pas empêché de dealer avec Comcast à un moment.
50 years on, we’re living the reality first shown at the “Mother of All Demos” | Ars Technica
▻https://arstechnica.com/information-technology/2018/12/50-years-on-were-living-the-reality-first-shown-at-the-mother-of-all-de
A half century ago, computer history took a giant leap when Douglas Engelbart—then a mid-career 43-year-old engineer at Stanford Research Institute in the heart of Silicon Valley—gave what has come to be known as the “mother of all demos.”
On December 9, 1968 at a computer conference in San Francisco, Engelbart showed off the first inklings of numerous technologies that we all now take for granted: video conferencing, a modern desktop-style user interface, word processing, hypertext, the mouse, collaborative editing, among many others.
Even before his famous demonstration, Engelbart outlined his vision of the future more than a half-century ago in his historic 1962 paper, “Augmenting Human Intellect: A Conceptual Framework.”
To open the 90-minute-long presentation, Engelbart posited a question that almost seems trivial to us in the early 21st century: “If in your office, you as an intellectual worker were supplied with a computer display, backed up by a computer that was alive for you all day, and was instantly responsible—responsive—to every action you had, how much value would you derive from that?”
Of course at that time, computers were vast behemoths that were light-years away from the pocket-sized devices that have practically become an extension of ourselves.
#Histoire_informatique #Mother_of_all_demos #Douglas_Engelbart