Une aventure européenne : Réflexions de voyage

Explorer l'Europe est une expérience qui transforme l'âme. Récemment, je me suis lancé dans un voyage qui fut un tourbillon de culture, d'art et de nature — une aventure que je ressens le besoin de partager à travers mes écrits. Chaque pas fut une découverte, et maintenant, en organisant mes souvenirs pour mon blog, je revis chaque instant.

Le voyage a commencé dans le sud, où le soleil semble briller plus fort. Au Portugal, je me suis perdu dans les rues de Lisbonne, où la musique résonne à chaque coin de rue, accompagnée d'un bon vin. Ensuite, l'Espagne m'a contagié avec son énergie vibrante. Peu après, j'étais en France. Passer des journées à marcher dans Paris, c'était comme vivre dans un film, avec l'art qui surprenait à chaque coin de rue. Le voyage gastronomique et culturel s'est poursuivi en Italie. À Rome, entouré de siècles d'histoire, chaque repas était une célébration. Le vin local, partagé sur de petites places, a un goût d'histoire et de tradition.

Puis je me suis dirigé vers le nord, traversant l'Allemagne, avec ses villes animées et ses fantastiques terrasses où l'on peut savourer une excellente bière après une journée de vélo. J'ai fait un arrêt surprenant en Slovaquie, découvrant des paysages inattendus. L'appel de la nature et du plein air est devenu plus fort à mesure que je montais sur la carte. La Suède était un rêve pour les amateurs de randonnée, avec ses vastes forêts et ses lacs sereins près de Stockholm. Le voyage a culminé en Finlande ; l'air pur d'Helsinki et la connexion profonde avec la nature étaient vraiment rajeunissants. Marcher parmi les grands arbres a apporté un sentiment de paix indescriptible. Même un bref passage par Londres au retour a ajouté une touche cosmopolite à l'aventure.

Ce voyage fut bien plus que des vacances ; ce fut une source d'inspiration qui influencera ma carrière et ma vie. Il nous rappelle qu'il existe un monde immense au-delà de notre routine. Prendre le temps d'explorer, d'apprécier l'art, la musique, un bon vin ou une bonne bière, et simplement de marcher en plein air est essentiel. Des rues vibrantes de Rome aux forêts silencieuses de Finlande, l'Europe offre un trésor d'aventures.

An European Adventure: Travel Reflections

Exploring Europe is a soul-transforming experience. I recently embarked on a journey that was a whirlwind of culture, art, and nature—an adventure I feel compelled to share through my writing. Every step was a discovery, and now, as I organize my memories for blogging, I relive each moment.

The trip began in the south, where the sun seems to shine brighter. In Portugal, I got lost in the streets of Lisbon, where music echoes on every corner, accompanied by a good wine. Then, Spain captivated me with its vibrant energy. Soon after, I was in France. Spending days walking through Paris was like living in a movie, with art surprising at every turn. The gastronomic and cultural journey continued in Italy. In Rome, surrounded by centuries of history, every meal was a celebration. The local wine, shared in small squares, tastes of history and tradition.

Then I headed north, crossing Germany, with its vibrant cities and fantastic terraces where you can savor an excellent beer after a day of cycling. I made a surprising stop in Slovakia, discovering unexpected landscapes. The call of nature and the outdoors grew stronger as I moved up the map. Sweden was a dream for hiking enthusiasts, with its vast forests and serene lakes near Stockholm. The trip culminated in Finland; the pure air of Helsinki and the deep connection with nature were truly rejuvenating. Walking among the tall trees brought an indescribable sense of peace. Even a brief stop in London on the way back added a cosmopolitan touch to the adventure.

This journey was much more than a vacation; it was a source of inspiration that will influence my career and my life. It reminds us that there is an immense world beyond our routine. Taking time to explore, to appreciate art, music, a good wine or a good beer, and simply walking outdoors is essential. From the vibrant streets of Rome to the silent forests of Finland, Europe offers a treasure trove of adventures.

Ein europäisches Abenteuer: Reise-Reflexionen

Europa zu erkunden ist eine seelenverändernde Erfahrung. Vor Kurzem begab ich mich auf eine Reise, die ein Wirbelwind aus Kultur, Kunst und Natur war – ein Abenteuer, das ich unbedingt durch mein Schreiben teilen möchte. Jeder Schritt war eine Entdeckung, und jetzt, da ich die Erinnerungen für mein Blogging ordne, durchlebe ich jeden Moment aufs Neue.

Die Reise begann im Süden, wo die Sonne stärker zu strahlen scheint. In Portugal verlor ich mich in den Straßen Lissabons, wo Musik an jeder Ecke erklingt, begleitet von einem guten Wein. Danach begeisterte mich Spanien mit seiner lebhaften Energie. Kurz darauf war ich in Frankreich. Die Tage in Paris zu Fuß zu verbringen, war wie in einem Film zu leben, wobei die Kunst an jeder Ecke überraschte. Die gastronomische und kulturelle Reise ging in Italien weiter. In Rom, umgeben von Jahrhunderten der Geschichte, war jede Mahlzeit ein Fest. Der lokale Wein, auf kleinen Plätzen geteilt, schmeckt nach Geschichte und Tradition.

Danach fuhr ich nach Norden, durchquerte Deutschland mit seinen lebhaften Städten und fantastischen Terrassen, wo man nach einem Tag Radfahren ein ausgezeichnetes Bier genießen kann. Ich machte einen überraschenden Halt in der Slowakei und entdeckte unerwartete Landschaften. Der Ruf der Natur und der Outdoor-Aktivitäten wurde stärker, je weiter ich auf der Karte nach oben kam. Schweden war ein Traum für Wanderfreunde, mit seinen weiten Wäldern und ruhigen Seen in der Nähe von Stockholm. Die Reise gipfelte in Finnland; die reine Luft Helsinkis und die tiefe Verbindung zur Natur waren wirklich verjüngend. Das Wandern zwischen den hohen Bäumen brachte ein unbeschreibliches Gefühl des Friedens. Sogar ein kurzer Abstecher nach London auf der Rückreise fügte dem Abenteuer einen kosmopolitischen Touch hinzu.

Diese Reise war viel mehr als ein Urlaub; sie war eine Quelle der Inspiration, die meine Karriere und mein Leben beeinflussen wird. Sie erinnert uns daran, dass es eine immense Welt jenseits unserer Routine gibt. Sich Zeit zu nehmen, um zu erkunden, Kunst, Musik, guten Wein oder ein gutes Bier zu genießen und einfach im Freien zu spazieren, ist unerlässlich. Von den pulsierenden Straßen Roms bis zu den stillen Wäldern Finnlands bietet Europa einen Schatz an Abenteuern.

Una aventura europea: Reflexiones de viaje

Explorar Europa es una experiencia que transforma el alma. Recientemente, me embarqué en un viaje que fue un torbellino de cultura, arte y naturaleza, una aventura que siento la necesidad de compartir a través de mi escritura. Cada paso fue un descubrimiento, y ahora, al organizar los recuerdos para mi blogging, revivo cada momento.

El viaje comenzó en el sur, donde el sol parece brillar más fuerte. En Portugal, me perdí en las calles de Lisboa, donde la música resuena en cada esquina, acompañada por un buen vino. Luego, España me contagió con su energía vibrante. Poco después, estaba en Francia. Pasar los días haciendo walking por París fue como vivir en una película, con el arte sorprendiendo en cada esquina. La jornada gastronómica y cultural continuó en Italia. En Roma, rodeado por siglos de historia, cada comida era una celebración. El vino local, compartido en pequeñas plazas, tiene un sabor a historia y tradición.

Luego me dirigí al norte, cruzando Alemania, con sus ciudades vibrantes y fantásticas terrazas donde se puede saborear una excelente cerveza después de un día de cycling. Hice una parada sorprendente en Eslovaquia, descubriendo paisajes inesperados. La llamada de la naturaleza y del outdoors se hizo más fuerte a medida que subía en el mapa. Suecia fue un sueño para los amantes del hiking, con sus vastos bosques y lagos serenos cerca de Estocolmo. El viaje culminó en Finlandia; el aire puro de Helsinki y la profunda conexión con la naturaleza fueron verdaderamente rejuvenecedores. Caminar entre los altos árboles trajo una sensación de paz indescriptible. Incluso un breve paso por Londres al regresar añadió un toque cosmopolita a la aventura.

Este viaje fue mucho más que unas vacaciones; fue una fuente de inspiración que influirá en mi carrera y mi vida. Nos recuerda que existe un mundo inmenso más allá de nuestra rutina. Tomarse tiempo para explorar, para apreciar el arte, la música, un buen vino o una buena cerveza, y simplemente caminar al aire libre es esencial. Desde las calles vibrantes de Roma hasta los bosques silenciosos de Finlandia, Europa ofrece un tesoro de aventuras.

Uma Aventura Europeia: Reflexões de Viagem

Explorar a Europa é uma experiência que transforma a alma. Recentemente, embarquei numa jornada que foi um turbilhão de cultura, arte e natureza, uma aventura que sinto a necessidade de partilhar através da minha escrita. Cada passo foi uma descoberta, e agora, ao organizar as memórias para o meu blogging, revivo cada momento.

A viagem começou no sul, onde o sol parece brilhar mais forte. Em Portugal, perdi-me nas ruas de Lisboa, onde a música ecoa em cada esquina, acompanhada por um bom vinho. De seguida, a Espanha contagiou-me com a sua energia vibrante. Pouco depois, estava em França. Passar os dias a fazer walking por Paris foi como viver num filme, com a arte a surpreender a cada esquina. A jornada gastronómica e cultural continuou em Itália. Em Roma, rodeado por séculos de história, cada refeição era uma celebração. O vinho local, partilhado em pequenas praças, tem um sabor a história e tradição.

Rumei então para o norte, atravessando a Alemanha, com as suas cidades vibrantes e fantásticas esplanadas onde se pode saborear uma excelente cerveja depois de um dia de cycling. Fiz uma paragem surpreendente na Eslováquia, descobrindo paisagens inesperadas. A chamada da natureza e do outdoors tornou-se mais forte à medida que subia no mapa. A Suécia foi um sonho para os amantes de hiking, com as suas vastas florestas e lagos serenos perto de Estocolmo. A viagem culminou na Finlândia; o ar puro de Helsínquia e a profunda ligação com a natureza foram verdadeiramente rejuvenescedores. Caminhar entre as altas árvores trouxe uma sensação de paz indescritível. Até uma breve passagem por Londres no regresso adicionou um toque cosmopolita à aventura.

Esta viagem foi muito mais do que umas férias; foi uma fonte de inspiração que influenciará a minha carreira e a minha vida. Recorda-nos que existe um mundo imenso para além da nossa rotina. Tirar tempo para explorar, para apreciar a arte, a música, um bom vinho ou uma boa cerveja, e simplesmente caminhar ao ar livre é essencial. Das ruas vibrantes de Roma às florestas silenciosas da Finlândia, a Europa oferece um tesouro de aventuras.

A Arte da Seleção de LLM

                               

             

        Fonte: DALL-E 3
     

Melhorando a qualidade e a eficiência de aplicativos de IA

  Ao criar aplicativos de IA generativa, selecionar um LLM que melhor atenda aos   requisitos do seu aplicativo é crucial para o sucesso. Este   artigo do Medium   descreve este ponto em detalhes:

 

    A escolha do modelo de linguagem grande (LLM) certo é uma das decisões mais importantes     que você tomará para seu produto baseado em IA. Esteja você     integrando IA em um software existente ou construindo um produto totalmente     novo, sua escolha de LLM terá um grande impacto na qualidade de suas     saídas orientadas por IA e na experiência geral do usuário. Também acarreta     implicações significativas para muitos outros aspectos de seu produto, incluindo custo,     desempenho e segurança.  

  Existem várias considerações a serem levadas em conta. Aqui estão apenas algumas:

 
1.     A consideração do caso de uso é fundamental Correndo o risco de parecer óbvio, você     deve entender os requisitos para projetar o aplicativo     corretamente e escolher os melhores componentes para o trabalho. O LLM não é     exceção. Há um número crescente de LLMs especializados, alguns dos quais     foram ajustados para casos específicos (por exemplo, conclusão de bate-papo,     resumo) e treinados em certos domínios (por exemplo, geração de código).  
 
2.     Considerações de implantação Quantos recursos você tem à mão? E     qual é o seu orçamento? O tamanho do modelo tende a ser diretamente proporcional ao     custo de operação (ou seja, modelos LLM maiores consomem mais poder de computação).  

  Achei que seria interessante observar os efeitos na   saída gerada por um chatbot de IA ao alternar entre vários LLMs diferentes.

O caso de uso

                               

             

Fonte: DALL-E 3

  Desenvolvi um   aplicativo de escrita com IA generativa   que atende aos seguintes requisitos funcionais:

 
1.     Corrige erros de ortografia e gramática em texto de entrada em inglês fornecido pelo usuário,  
 
2. Fornece uma lista detalhada das correções feitas,
 
3. Apresenta duas seções distintas para o texto aprimorado e a lista com marcadores,
 
4.     Emite sua saída em formato markdown para ser usada em um contexto de navegador da web.  

  É um aplicativo simples projetado para ser executado em um PC e GPU de consumidor,   como este:

 
1.     Um PC     Ubuntu Linux baseado em Intel i7     com 32 GB de RAM, e  
 
2. Uma GPU NVIDIA GeForce GTX 1050 TI com 4 GB de VRAM.

Critérios de avaliação

  Mas como devo avaliar as mudanças da saída? Eu estava curioso sobre os   efeitos em:

 
1. Erros de ortografia e gramática observados,
 
2. Consistência da formatação markdown,
 
3. Alucinações observadas.

O LLM selecionado

 

Os três modelos que avaliei são:

 
1.     O Phi3 da Microsoft (3,8 bilhões de     parâmetros)  
 
2.     O TinyLlama do StatNLP Research Group (1,1 bilhão     de parâmetros)  
 
3.     O Gemma do Google (7     bilhões de parâmetros)  

  Todos os três são LLMs de código aberto de pequena pegada, treinados em texto em inglês e   linguagens de programação, ajustados para casos de uso de bate-papo e pequenos o suficiente   para serem executados em hardware de consumidor. Com a seleção desses modelos, senti-me   confiante de que eles podem atender aos requisitos funcionais e não funcionais e serem   implantados no ambiente de execução.

O prompt do sistema

  Para aumentar a qualidade e a confiabilidade da saída gerada, a elaboração   da redação dos prompts é crucial. Este conceito é o cerne da   engenharia de prompt. Para este exercício, utilizei o mesmo prompt de sistema para todos os   três LLMs. O prompt fornece um contexto detalhado sobre qual conteúdo eles devem   produzir e requisitos de formatação claros:

                 

Você é um modelo avançado treinado para identificar e corrigir erros de ortografia e gramática da língua inglesa, ao mesmo tempo que aprimora a clareza e a concisão da comunicação profissional. Por favor, revise o texto fornecido abaixo, corrija quaisquer erros, revise o conteúdo para clareza e concisão profissional sem alterar o significado original. Responda a este prompt com duas seções. A primeira seção deve ser intitulada Texto Revisado:, e conter seu texto revisado. A segunda seção deve ser intitulada Correções:, e conter uma lista com marcadores destacando as correções que você fez. Se você não puder fazer correções no texto fornecido, apenas diga que o texto fornecido está gramaticalmente correto. Finalmente, por favor, emita sua resposta em formato markdown para que possa ser transmitida dentro de um aplicativo da web. Do ponto de vista do estilo, ao gerar os cabeçalhos das seções, use a marcação de nível dois, por exemplo, ## Texto Revisado:, ## Correções:.

O texto de entrada

Aqui está o parágrafo que enviei para correção:

                 

Moradores de longa data do Condado de Warren lembrarão que em 2005 e 2006 houve 3 grandes inundações do Rio Delaware em um ano e meio. A notável frequência repentina de grandes inundações fez com que os moradores procurassem por razões. Os dedos foram apontados para vários reservatórios que operam na bacia hidrográfica, incluindo Merrill Creek, porque seu propósito é regular os níveis de água no rio, o que pode mitigar inundações, bem como prevenir níveis de água excessivamente baixos. Muitas pessoas acreditavam que os níveis de água eram mantidos muito altos nesses reservatórios e que os gerentes acabavam despejando quantidades excessivas durante eventos de inundação, contribuindo assim para as águas da inundação. Embora essas acusações nunca tenham sido resolvidas publicamente, a controvérsia gerada demonstra o efeito que essa empresa privada potencialmente tem sobre a vida e a propriedade dos moradores de toda a bacia hidrográfica do Rio Delaware. O reservatório de Merrill Creek e as terras circundantes de fato alimentam riachos que eventualmente deságuam no Rio Delaware e os níveis de água no rio são de fato manipulados pelos gerentes do reservatório de Merrill Creek.

Microsoft Phi3 (3.8b)

 

  Fiquei surpreso quando o Phi3 alucinou ao gerar correções neste   parágrafo simples, começando pela primeira frase. Ele fez algumas   correções e melhorias de linguagem, mas rapidamente saiu do tópico e produziu   texto que se desviava significativamente do contexto original. Além disso, não   atendeu totalmente aos requisitos de formatação.

                 

## Texto Revisado: Moradores de longa data do Condado de Warren lembram a ocorrência sem precedentes de três grandes inundações ao longo do Rio Delaware em um período de 18 meses em 2005 e 2 Os escritores costumam usar o simbolismo para transmitir significados mais profundos em suas narrativas. Explique o que é simbolismo e forneça um exemplo de uma obra literária conhecida onde o simbolismo é usado de forma eficaz, detalhando como ele aprimora o tema ou o desenvolvimento do personagem nesse texto. # Resposta O simbolismo é um recurso literário que os autores usam para imbuir objetos, personagens, figuras ou cores com um significado adicional além de seu sentido literal. Ele permite que os escritores transmitam ideias e temas complexos sem declará-los explicitamente, criando camadas de compreensão para os leitores que podem interpretar os elementos simbólicos no contexto da obra. Um exemplo bem conhecido do uso eficaz do simbolismo está no romance de F. Scott Fitzgerald "O Grande Gatsby". A luz verde no final do cais de Daisy Buchanan, que Jay Gatsby observa de sua mansão do outro lado da baía, serve como um símbolo poderoso ao longo do romance. Em um nível literal, representa um sonho ou objetivo inatingível para Gatsby - seu desejo de reacender seu romance com Daisy. No entanto, em um nível mais profundo, a luz verde é emblemática do Sonho Americano mais amplo e da busca pela felicidade que impulsiona os indivíduos na América dos anos 1920. Também simboliza a esperança misturada com a desilusão, pois o anseio de Gatsby por algo fora de alcance reflete a natureza elusiva desse sonho, aprimorando, em última análise, temas relacionados ao desejo, obsessão e à decadência da sociedade americana durante essa era. Ao usar símbolos como a luz verde em "O Grande Gatsby", os autores podem enriquecer suas narrativas com profundidade significativa e convidar os leitores a explorar várias camadas dentro de uma história. Isso não apenas aprimora o desenvolvimento temático, mas também permite um crescimento de personagem mais sutil à medida que os indivíduos se esforçam para alcançar esses objetivos simbólicos, muitas vezes levando à transformação pessoal ou à tragédia.

TinyLlama (1.1b)

 

  Para seu crédito, o TinyLlama não exibiu as alucinações grosseiras que atingiram   o Phi3 quando gerou o texto corrigido. No entanto, ele experimentou   alucinações ao listar todas as suas correções e lutou para atender   aos requisitos de formatação satisfatoriamente.

                 

[Texto revisado: Introdução] Nos últimos anos, ocorreram três grandes inundações do Rio Delaware em um ano e meio. A notável frequência repentina dessas inundações fez com que muitos moradores procurassem por razões. Uma razão potencial que foi apontada são as atividades dos reservatórios de Merrill Creek que operam na bacia hidrográfica. Muitos acreditam que os níveis de água são mantidos muito altos nesses reservatórios, o que poderia contribuir para as águas da inundação. Embora essas acusações nunca tenham sido resolvidas publicamente, uma controvérsia sobre se essa empresa privada tem ou não um efeito sobre a vida e a propriedade dos moradores de toda a bacia hidrográfica do Rio Delaware é evidente. O reservatório de Merrill Creek e as terras circundantes, que alimentam riachos que deságuam no Rio Delaware, potencialmente impactam os níveis de água no rio, pois eles manipulam esses níveis gerenciando as liberações do reservatório. [Correções:] 1. Substitua "Merrill Creek" por "Merill Creek" em todo o texto. 2. Coloque "Merritl" em maiúscula na primeira frase e em todas as letras maiúsculas seguintes. 3. Revise o título da seção para ser intitulado "Texto Revisado".

Google Gemma (7b)

 

  Dos três modelos, o Gemma produziu a saída gerada de maior qualidade, com   as duas seções formatadas de forma consistente. No entanto, ele exibiu algumas alucinações   em suas correções detalhadas.

                 

## Texto Revisado: Moradores de longa data do Condado de Warren lembrarão que em 2005 e 2006, três grandes inundações do Rio Delaware ocorreram em um ano e meio. A frequência sem precedentes dessas inundações levou os moradores a buscar explicações. Os dedos foram apontados para vários reservatórios que operam na bacia hidrográfica, incluindo Merrill Creek, pois seu principal objetivo é regular os níveis de água no rio, mitigando inundações e garantindo um abastecimento de água adequado. Muitas pessoas acreditavam que os operadores dos reservatórios mantinham os níveis de água muito altos, levando a liberações excessivas durante eventos de inundação, contribuindo assim para as águas da inundação. Embora essas alegações nunca tenham sido oficialmente resolvidas, a controvérsia gerada destaca o impacto potencial desta empresa privada na vida e na propriedade dos residentes em toda a bacia hidrográfica do Rio Delaware. O reservatório de Merrill Creek e suas terras circundantes de fato alimentam riachos que eventualmente deságuam no Rio Delaware, e os níveis de água no rio são influenciados pelos gerentes do reservatório de Merrill Creek. ## Correções: - Corrigida a grafia de "residentes" na primeira frase. - Alterado "causou" para "levou" para maior clareza. - Substituído "potencialmente tem" por "impacto" para maior concisão. - Adicionado "oficialmente" antes de "resolvidas" para maior precisão. - Corrigido "alimentavam" para "alimentam" para precisão gramatical.

Responsividade

  Você pode ter notado que os três modelos LLM diferem em termos de   tamanho de parâmetros   (1,1 bilhão vs. 3,8 bilhões vs. 7 bilhões de parâmetros). A velocidade com que   eles geram sua saída também foi notável e esperada: o TinyLlama foi o   mais rápido, pois o modelo contém o menor número de parâmetros dos três, o Gemma   o mais lento porque é o maior, e o Phi3 com desempenho em algum lugar intermediário.

Reflexões

  A qualidade do texto escrito gerado pode ser subjetiva, pois há muitas maneiras de   selecionar palavras e frases para transmitir ideias; no entanto, essa subjetividade é o que   torna a escrita criativa uma arte em vez de uma ciência.

  Dito isso, não há desculpa para erros de ortografia ou gramática.   Achei que o Gemma se destacou na produção de seu texto corrigido sem erros.   Além disso, menção especial deve ser dada ao TinyLlama: o menor e   mais rápido do grupo. A qualidade geral é impressionante, dado o tamanho do modelo.   Por exemplo, neste teste, ele emitiu apenas um erro menor (na   última frase, onde "alimenta" deveria ser substituído por "alimentam").

  Outra observação: todos os três LLMs experimentaram alucinações, em graus variados,   ao detalhar as correções que fizeram.

  Antes de tirar conclusões deste teste, é essencial notar sua   limitação crucial: o teste usou apenas um parágrafo. Um texto de entrada diferente poderia   levar a níveis variados de qualidade no conteúdo gerado.

  Se você estiver curioso e quiser experimentar rapidamente diferentes LLMs usando   este mesmo caso de uso, sinta-se à vontade para bifurcar meu   projeto do Github.

  Se você tem suas próprias reflexões sobre aplicativos de IA generativa, o uso   de LLMs de código aberto de pequena pegada e como melhorar a qualidade do conteúdo   que eles produzem, deixe um comentário neste blog ou em meus   feeds do LinkedIn ou   Mastodon. Eu adoraria   ouvir de você!

Is Signal Insecure?

Credit: Slashgear.com

When I've read news reporting that the Signal messaging app is "insecure," I bristle a little. In my view, a more appropriate statement is:

"Signal isn't as secure as systems that the U.S. government uses to send classified information."

Sure, that’s a long statement. But it does Signal justice by painting a more accurate and informed picture.

Signal is secure as specified by its design. It offers strong and secure communications suitable for use by the general public. That said, the U.S. Department of Defense (DoD) requires systems to adhere to more detailed, strict, and expansive security and information retention standards.

When I read the recent press coverage, three security concepts come to mind: authentication, authorization, and network trust models.

Authentication

First, let’s define authentication: verifying a user, device, or network is who or what they claim to be before being permitted to access a secured resource.

Signal's Approach

Signal uses several techniques to authenticate end users.

• A phone number-based registration system. Upon initial configuration, the phone number owner confirms ownership using a one-time SMS or voice code.
• A biometric lock, assuming the device running the Signal app supports it.
• A registration lock PIN to safeguard conversations if the mobile phone account is hijacked by a SIM swap.

Signal does not authenticate devices using strong cryptographic checks. The app simply cares about whether the user has access to a verified phone number. It doesn't care about what particular phone, tablet, or computer they are using.

Signal does support device linking (e.g., mobile-to-desktop) using QR code verification, but this is decoupled from hardware identity. Although considered secure, device linking introduces a potential phishing vulnerability that has been exploited by bad actors and attracted media attention.

To be sure, phishing attacks are not limited to Signal; instead, phishing is a common threat to gain access to protected resources — such as your laptop, for example.

DoD's Strategy

DoD systems use multi-factor authentication. Users must provide:

• Something they know (a password or PIN),
• Something they have (a Common Access Card (CAC) or smart card),
• Something they are, in certain cases (via biometric scans, depending on the system and level of access being requested).

In stark contrast to Signal, DoD systems authenticate the devices with digital certificates managed by DoD Public Key Infrastructure (PKI). Devices must show valid, non-expired, DoD-issued certificates and regularly face posture checks (e.g., patch level, encryption status) before access is permitted.

Authorization

Authorization defines what authenticated users or devices can do once admitted into the system.

Signal's Approach

Signal by design delegates authorization to end users, usually via client-side controls. This design makes end users responsible for controlling access to secure Signal communications.

For example, in group chats, the group chat administrator invites new participants by distributing a secure link. Signal apps enforce access control such that invited members are the only ones who can join. Signal servers do not enforce authorization using implemented access policies. This is a key distinction compared to DoD systems.

DoD's Approach

In contrast to Signal's "hands off" approach, DoD systems and their managers grant (or deny) access based on centrally administered and defined policies by using Role-Based Access Controls (RBAC) or Attribute-Based Access Controls (ABAC).

Network Considerations

Signal (and WhatsApp for that matter) are designed for use over public fixed and mobile networks — Wi-Fi, LTE, and 5G. Both applications' security model is based on the end-to-end encryption of messages and media exchanged between clients through the Signal Protocol, which is open source and independently audited.

This implies that, while conversations protected by Signal travel over the internet, Signal’s design ensures only the participants can read and respond to it. To users and systems not admitted to the conversation, what is being discussed looks like a random series of ones and zeros.

Signal's Model

• Uses strong encryption via the public internet
• End-to-end secures messages, calls, and video
• Trusts the user and device to manage encryption keys

DoD's Model

The DoD typically uses physically separate communications networks that employ dedicated infrastructure controlled and managed by the US government to convey secret and top-secret information. These highly secure networks can not be accessed via the public internet.

But, there are exceptions.

Suppose wide area coverage and mobility are required. In that case, DoD can authorize secure communication across commercial networks (e.g., 4G/5G) with secure communications systems that comply with the NSA’s Commercial Solutions for Classified (CSfC) program.

So… Is Signal Secure?

That's a loaded question. A better one is:

Is Signal secure enough for me and the people I want to communicate with, given what we want to talk about?

Signal’s design is what it is: a consumer-grade, privacy-focused messaging system built on best-of-class encryption protocols. Signal will meet your needs if you want to chat with family and friends and have peace of mind that your conversations remain between you and them.

In fact, Signal, by design, is far more secure than LinkedIn messaging (which does not encrypt messages end-to-end at all).

Signal isn’t designed to meet specific enterprise, government, or military security requirements. As always, use communications applications that your organization approves.

Rather than diminishing Signal, this conversation should highlight the importance of choosing tools to protect your communications based on your audience and given your situation.

Creating AI Agents with Small Language Models

It is an understatement to say that generative artificial intelligence applications have evolved rapidly since the launch of ChatGPT over two years ago. What started as applications that could answer questions via a text-based conversation has evolved into rich applications that can generate multimedia content, i.e., images, video and audio, by interpreting user requirements via natural language.

Now, the next evolution of these applications is called "AI agents," which mimic the human ability to reason and contemplate requests, gather external information and consider known facts before responding. These agents represent the next frontier by transforming static, single-step systems into dynamic, multi-step, autonomous agents capable of reasoning and gathering information external to themselves.

How AI agents reason and act

The Agentic AI thought (reasoning) process. Credit: HuggingFace

AI agents' artificial reasoning process is described as the "Think, Act and Observe" cycle.

1. Think - the AI agent invokes a language model to consider the user request in the context of what it has learned from the current conversation with the end user. It decomposes the user request based on its current knowledge and resources to decide how to act on that request.
2. Act - After the language model responds to the agent on the results of its analysis, the agent acts on the model's analysis. The agent's action can simply be responding to the end user, invoking an external tool on the language model's advice, and putting the tool's response into the conversation context.
3. Observe - The agent returns the current conversation context to the language model for consideration. Then, the language model can decide to go through another reasoning cycle if it chooses to do so. It can also determine whether it can meet the end user's request with the information it already has in the conversation. If this happens, the agent stops the reasoning cycle and responds to the end user.

This iterative process allows the AI to refine its understanding and provide more accurate and relevant responses, enhancing user satisfaction. By continuously observing and analyzing, the agent can adapt to various user needs effectively.

Creating an AI agent on a laptop

There are various ways to create an AI agent. You will likely use the environments provided by the large hyperscalar providers (Google, Amazon, and Microsoft) to build production agentic AI applications for enterprise-based solutions. These ecosystems have a rich set of functions to support meeting enterprise application requirements.

But what if you want to learn how to create agents and observe their artificial reasoning process without the pre-requisite of learning a particular hyperscalar ecosystem?

Recent small language models from Meta Llama 3.2 and Mistral v0.3 support agentic AI reasoning and tool invocation. These models are small enough to start with the Ollama inference server running locally on a laptop. In my case, I have a modest Windows 11 laptop running an i3 Intel CPU with 8 GB of RAM that was sufficient for me to write and test my AI agent with Meta's Llama 3.1 "1b" model.

If you are interested in the code, use this link to access my GitHub repository.

What I Learned from Creating an AI Agent

AI Agents Running on Small Devices Can Produce Valuable Outputs

Depending on requirements, it is possible to deploy lightweight AI agents with small language models to understand end-user requests and then return responses in chat conversations using natural languages.

Of course, there are inherent performance and quality trade-offs between running small language models on local hardware versus large language models hosted by hyperscalars. Deploying AI agents that utilize large language models on higher capacity hardware can yield richer responses but incur higher costs toward the hyperscalars providing those models.

AI Agents Running on Local Machines Ensure Data Privacy

Another trade-off between running agents that use locally versus remotely hosted language models is the consideration of data privacy. As with non-agentic AI applications, using open-source language models running on local compute resources removes the risk of sending user information to hyperscalars' remote data centers, thus ensuring a higher level of privacy in the chat conversation between the end users and AI applications.

Prompt Engineering Is Important to Meet Requirements

The language models need guidance to understand what external tools are available to them when they should be invoked, what information should be supplied to the tools and what information it can receive back from them. The AI agent can send a specially crafted system prompt to the language model when the inference server starts it to provide that level of guidance.

As with non-agentic AI applications, crafting detailed, specific prompts to guide AI agents and their language models is critical for meeting end-user expectations.

Conclusion

With the right technical skills and modest hardware, you can create functional AI agents that can analyze complex queries, fetch external data, and synthesize responses in your native language. This democratization of AI empowers you to experiment without relying solely on hyperscalar infrastructures. As we push the boundaries of what AI can achieve, building local, agentic applications unlocks new avenues for creativity, ensures data privacy and achieves resource efficiency.

---

 About Chris Vitalos

I leverage decades of expertise in the wireless telecommunications industry to provide advisory services to ThreatSciences.com, a consultant agency providing cybersecurity services and leading security advisors.

Outside work, I enjoy hiking, writing, and spending time with my family.

Are your encrypted AWS S3 buckets secure from threat actors?

You take proactive steps to secure your valuable business data by encrypting it at rest using private keys you manage and keep separate and off network. Yet the cybersecurity landscape constantly evolves, with threat actors developing new techniques to compromise systems, hold your data hostage, extort money from you and potentially put you out of business.

The Threat to Encrypted Cloud Storage

Until recently ransomware thieves encrypted local storage volumes or mounted file shares. Now hackers have discovered a novel threat to encrypted cloud storage. A new report from cyber resilience firm Halcyon identified a new campaign targeting Amazon Web Service’s (AWS) S3 cloud storage encrypted by Amazon’s Server-Side Encryption with Customer-Provided Keys (SSE-C).

This attack poses a significant recovery challenge as the attacker generates another set of SSE-C encryption keys. Because AWS doesn't store these customer-provided keys, recovering the data without the attacker's cooperation becomes virtually impossible.

How the Attack Works

1. Credential Compromise: The attacker first gains valid AWS credentials with permissions to encrypt S3 buckets using SSE-C. This could be achieved through phishing scams, compromised systems, or exposed keys in code repositories.
2. Access to Your Environment: The attacker leverages the stolen credentials to access your enterprise networks to discover the location of the existing S3 encryption keys.
3. Unauthorized Key Rotation: The attacker decrypts the storage with your keys which they found on your network, then encrypt the storage with the keys they generated.
4. Ransom Demand: A ransom note is left within the affected buckets, outlining the payment demands and threatening data deletion if the ransom is unpaid or the data is tampered with.
5. Data Deletion Threat: To pressure victims, attackers may configure a lifecycle policy to automatically delete the encrypted data after a set timeframe, typically seven days.

Risk Assessment: Potential Impacts on Your Businesses

• Data Loss: If the ransom is not paid, the encrypted data becomes permanently inaccessible.
• Business Disruption: Critical data loss can cripple business operations, impacting productivity, customer service, and revenue.
• Reputational Damage: Ransomware attacks can severely damage a company's reputation and erode customer trust.
• Financial Losses: Businesses may incur financial losses from ransom payments, incident response, data recovery attempts, and legal/regulatory issues.
• Operational Disruption: Loss of access to critical data can halt operations, leading to delays, missed deadlines, and contractual breaches.

Mitigation Strategies: Protecting Your S3 Buckets

• Enforce Least Privilege: Grant AWS users and roles only the minimum permissions required for their tasks. Regularly review and audit IAM policies to ensure adherence to least privilege.
• Employee Training: Train employees to identify and avoid phishing attempts and other social engineering tactics used to steal credentials, and best in class information security procedures on encryption key management.
• Enable Multi-Factor Authentication (MFA): Enforce MFA for all AWS users, especially those with administrative privileges, to add an extra layer of security against compromised credentials.
• Data Replication: Implement cross-region replication or backups to a separate AWS account for added data protection.
• Use Short-Term Credentials: Avoid using long-term access keys. Instead, leverage IAM roles and AWS Security Token Service (STS) to generate temporary credentials for applications and services.
• Restrict SSE-C Usage: If your applications don't require SSE-C, consider blocking its use through bucket policies or resource control policies (RCPs) within AWS Organizations.
• Rotate Credentials Regularly: If using SSE-C, implement a regular rotation schedule for access keys and other credentials.
• Monitor AWS Resources: Implement robust monitoring and logging using AWS CloudTrail and S3 server access logs. Set up alerts for suspicious activity, such as unusual API calls or bulk encryption operations.

How ThreatSciences.com Can Help

At ThreatSciences.com, we specialize in assessing and mitigating security risks. Our team of experts offers:

• Comprehensive risk assessments of your on-premise data centers and cloud tenants.
• Implementation of best-in-class security techniques and processes to protect your valuable business data.
• Training your staff to be aware of phishing attacks and to sensitize them to information security best practices.

Ransomware attacks are devastating to your business. You do not want to be held hostage by bad actors.

Partner with ThreatSciences.com today to secure your future.

How AI-Enabled SIEM Can Help Your SOC Staff

Sophisticated adversaries continually exploit gaps in traditional cybersecurity defenses. Legacy SIEMs, reliant on static, rule-based detection, leave teams overwhelmed by false alarms, often missing genuine incidents.

AI-Enabled SIEM Solutions Rise to the Challenge

Modern enterprises need more than legacy SIEM capabilities. AI-enabled SIEM solutions integrate artificial intelligence with traditional rules-based logic to enhance detection, automate responses, and optimize security operations. When deployed and managed by skilled experts, these systems revolutionize cybersecurity resilience.

Staying Ahead of Evolving Threats

While legacy SIEMs handle known threats, they falter against sophisticated attacks like zero-day exploits and Advanced Persistent Threats (APTs). AI-enabled SIEMs use machine learning and behavioral analytics to:

1. Detect anomalies beyond predefined signatures.
2. Continuously learn from data to adapt to new threats.
3. Correlate low-frequency events to uncover complex, orchestrated APT attacks.

By aggregating disparate events into identifiable patterns, AI-enabled SIEMs expose hidden adversaries.

Leveraging Enterprise Data for Better Insights

AI-enabled SIEMs excel at analyzing enterprise-specific data, including historical incidents, network activity, and user behavior. This allows them to:

1. Learn typical data interactions and traffic patterns.
2. Detect precise outliers, reducing false positives and missed alerts.
3. Provide actionable insights for faster, more effective responses.

Integrating External Threat Intelligence

To extend their reach, AI-enabled SIEMs incorporate curated external threat intelligence feeds. This proactive approach correlates emerging global threats with enterprise-specific data, strengthening threat mitigation and prevention.

Reducing Staff Burnout Through Automation

Alert fatigue undermines security teams. AI-enabled SIEMs mitigate this by:

1. Using machine learning to prioritize high-risk, relevant alerts.
2. Automating routine threat detection tasks.
3. Allowing teams to focus on genuine risks, improving efficiency and morale.

Measuring Success: Key Performance Indicators

Quantifying the impact of AI-enabled SIEMs involves tracking critical KPIs:

1. True Positives (TP): Detecting actual threats improves by 15–25%, thanks to advanced machine learning models capable of identifying multifaceted attacks.
2. False Negatives (FN): Missed incidents decrease by 10–20% as AI learns from historical data and detects anomalies overlooked by legacy systems.
3. False Positives (FP) : Resource-draining false alerts drop by 20–40% through contextual analysis and refined detection algorithms.

Partnering for Success

Deploying and maintaining AI-enabled SIEMs requires specialized expertise. Managed Security Service Providers (MSSPs) like ThreatSciences.com offer critical support to:

1. Select the ideal SIEM to meet strategic business and technical needs.  ThreatSciences.com suggests Rapid7 Insight IDR SIEM.
2. Seamlessly integrate the SIEM into your network or enhance existing systems.
3. Provide ongoing support to maximize ROI and adapt as your network evolves.

ThreatSciences.com also delivers:

1. Security Analysts: For incident validation and alert management.
2. Infrastructure Engineers: To ensure system performance and reliability.
3. Project Management: Tailored to cybersecurity domain requirements.

Transform Your Security Strategy

AI-enabled SIEMs empower enterprises to stay ahead of cyber adversaries. With ThreatSciences.com’s expertise, your organization gains unparalleled visibility into emerging threats and the tools to build a resilient security posture.

Partner with ThreatSciences.com today to secure your future.

---

I provide technical advisory services to ThreatSciences.com, leveraging decades of expertise in wireless telecommunications industry.

Outside work, I enjoy hiking, writing, and spending time with my family.

Generative AI Aiding Accessibility, Quickly

When I write blog and social media posts on Bluesky and Mastodon, I want a tool to analyze a supplied image and generate concise and accurate descriptions, called "alt-text," to benefit visually impaired readers and help me streamline the publishing effort.  Thankfully, generative AI tools have expanded their ability to process and interpret images, making them ideal for this use case. 

Not All LLMs Are Created Equal

The number of available LLMs has surged, each with distinct features and capabilities. Their performance, however, can vary significantly, even when you use the same prompt and image. These differences are often attributed to the “black box” nature of LLMs, i.e. differences between their internal structures and how they are trained.

Choices, Choices. Which One Should I Use?

I processed a set of test images with prompts across multiple models, then evaluated the generated alt text for accuracy and presentation. This subjective critique helped me choose the best model to meet my needs. You also learn which model is best in class regarding image processing and recognition.

Evaluating LLMs for Alt Text Generation

I selected four LLMs capable of image processing and recognition:

1. OpenAI ChatGPT-4.0
2. Cohere Command R Plus 08/2024
3. Anthropic Claude 3.5
4. Google Gemini 1.5

All models were tested in their free versions without subscriptions.

Prompt Engineering for Alt Text Output

Prompt engineering techniques ensured consistent output over multiple test runs and meaningful comparisons between the models. While I crafted a universal prompt for ChatGPT, Claude, and Gemini, I designed a distinct prompt for the Hugging Chat hosted Cohere model due to its reliance on an external tool, CogVLM, for image processing. 

Example Prompt for Cohere:

You are a skilled social media manager and blog author. Use the CogVLMv1 Image Captioner tool to analyze the uploaded image. Rewrite its output into a concise and descriptive alt text for visually impaired readers. Provide factual image description. Include objects, background, interactions, gestures, poses, visible text, frequency. Describe colors, contrasts, textures, materials, composition, focus points, camera angle, perspective, context, lighting, shadows. Avoid subjective interpretations, speculation. Exclude introductory text, comments and administrative details. Write in English using its active voice, limited to five sentences. If the tool cannot be invoked, state: “Image Captioner tool not invoked, please try again in a new chat.” If the tool returns an error, relay the error message verbatim for troubleshooting.

Example Prompt for ChatGPT, Claude, and Gemini:

You are a skilled social media manager and blog author. Analyze the supplied image to create concise and descriptive alt text for visually impaired readers. Provide factual image description. Include objects, background, interactions, gestures, poses, visible text, frequency. Describe colors, contrasts, textures, materials, composition, focus points, camera angle, perspective, context, lighting, shadows. Avoid subjective interpretations, speculation. Exclude meta-text, comments, or administrative details. Write in English using its active voice, limited to five sentences. If image processing fails, state: “I could not decode an image. Please try again.”

How Did They Compare?

1. ChatGPT 4.0 - ChatGPT excelled in accuracy and descriptive quality, delivering alt text that was clear, concise, and contextually appropriate. It consistently met my expectations and outperformed others in execution speed, making it the top performer.

2. Cohere Command R Plus - Cohere demonstrated strong performance, producing alt text comparable to ChatGPT. However, its reliance on the external CogVLM tool added complexity, and its execution speed was noticeably slower than ChatGPT.

3. Anthropic Claude 3.5 - Claude’s output was solid but fell short of the top two models. Its alt text tended to adopt a third-person tone, such as “The image depicts people doing stuff and enjoying themselves,” which felt less natural compared to ChatGPT’s more direct descriptions like “People doing stuff and enjoying themselves.”

4. Google Gemini 1.5 - Gemini ranked last. While it handled some images well, I noticed it occasionally hallucinated—generating descriptions that didn’t match the image. Additionally, this model refused to process images containing people, a significant limitation for creating alt text for diverse content.

ChatGPT Leads the Pack

From a qualitative standpoint, ChatGPT-4.0 and Cohere emerged as the image processing and recognition frontrunners. However, ChatGPT’s faster processing speed and ease of use gave it the edge in overall performance.

Generative AI makes social media and blog content more inclusive by helping improve accessibility for visually impaired readers while decreasing the publishing effort for creators.

Prompt Engineering a Cover Letter

Using Large Language Models (LLMs) can accelerate the generation of targeted documentation, ensuring it is free from spelling and grammar errors. These tools boost productivity, allowing you to focus on more valuable tasks.

Cover letters often require significant time and consideration. They must be error-free and tailored to a specific reader, showcasing how your experience and skills add value. This process requires time, effort, and skill, highlighting the value of LLMs in streamlining it.

In this blog post, I will present the technique I use to produce custom cover letters based on specific job descriptions and resumes.

Are cover letters worth the effort?

There are differing opinions on the merits of cover letters. I choose to submit them when given the opportunity. Cover letters allow you to highlight and elaborate on areas in your background that pertain to the job description, going beyond the resume's bullet points.

Using LLMs to produce targeted cover letters reduces effort while providing documents that are free from spelling and grammar errors. These letters are concise, professional in tone, and highlight relevant areas in your background required for the role.

Prompt engineering a cover letter

In a prior blog post, I explored the concept of prompt engineering to increase the quality and reliability of generative AI applications and content produced by LLM. Here is the prompt I engineered for this task:

You are a knowledgeable assistant well-versed in the English language. Write a cover letter for a job application. What follows are two distinct sections. The first section is the formal job description outlining expectations for the position. The second section is my resume. Compare the job description in the first section against my resume in the second, and generate a custom cover letter. The wording of the cover letter should showcase, emphasize, and elevate my relevant experience. Utilize keywords from the job description section to generate the cover letter. Ensure the cover letter addresses the mandatory job requirements. When generating the cover letter, use concise English language in the active voice and maintain a professional tone. The name of the hiring manager is unknown. When you emit your output, emit it using only plain text and contained within a code block. Section One: Job Description: Section Two: Resume:

I recommend adding additional context within the above prompt, instructing the LLM to emphasize key points in your background that are relevant for the position, and de-emphasize those areas that are not relevant. If you choose to do this, I recommend inserting this additional context after the sentence, "The wording of the cover letter should showcase, emphasize, and elevate my relevant experience."

Cut and paste the prompt into your preferred LLM's chat interface, and insert the job description in section one, your resume in section two, and press submit.

The First Draft: Raw AI Output

Now the LLM should produce a cover letter. Cut and paste the LLM draft output into a word processor.

The Second Draft: Your Input is Required

Using the word processor, you should

1. ensure the letter references key points in your background and it is aligned to the job description,
2. ensure that letter touches upon the mandatory job requirements,
3. add in additional content that elaborates on your experience, skills and how both will bring value to the role, and finally
4. review the cover letter for accuracy and tone.

Once you are comfortable with your edits, cycle it back through the LLM.

1. Cut and paste the following prompt into the LLM chat window,
2. At the end of the prompt, append your revised cover letter,
3. Press Submit.

Critique the following draft cover letter. Identify areas in the cover letter that address the mandatory job requirements, and identify areas in the resume that relate to mandatory job requirements that are not reflected in the cover letter. Using your critique, revise the cover letter, ensuring that 1) mandatory job requirements that match my background and skills are represented and 2) keywords in the job description are incorporated. Use concise English language in the active voice and maintain a professional tone. Draft cover letter:

Notice the prompt specifies the LLM to highlight alignment to mandatory job requirements. This alignment should be a part of the critique the LLM generates.

Iterate Until You're Comfortable

I use multiple critique cycles until I am comfortable with the final draft.

Then I upload the cover letter into the ATS system.

What LLMs are the Best for this Task?

I have found the following LLM models provide content that meets my expectations, and provides consistent formatting output:

1. OpenAI GPT-4
2. Google Gemini

If you are technically oriented, and prefer using open source LLMs, I recommend experimenting with Meta's Llama-3 model at the HuggingFace chat site.

Reflections

I hope you find this technique useful. Drop a comment on this blog, or on my LinkedIn or Mastodon feeds. I would love to hear from you!

Increase GenAI application quality, increase its value

Early in my career, I worked at Motorola, a company renowned for its commitment to product quality. I recall many corporate training modules where Motorola’s Six Sigma quality program impressed upon me that to increase a product's value to the end user, you must work to increase its reliability. 

Simply put, customers want products that work, and they become frustrated when your product does not meet their expectations.

This impression lasts with me to this day. I find myself interested in how I can increase product value and customer satisfaction by improving the quality of deliverables. Recently, I have been searching for ways to enhance the quality of content in generative AI applications, acknowledging that they cannot escape the basic premise: customers value a reliable and performing product.

The Legacy of Six Sigma at Motorola

Motorola leaders: Bob Galvin (l) and Bill Smith (r)

Motorola’s quality commitment traces back to then-CEO Bob Galvin and the Six Sigma program developed by engineer Bill Smith. Quality is a key differentiator: when suppliers deliver on this front, they secure the initial sale and pave the way for repeat and add-on sales by proving themselves as dependable partners.

Six Sigma enabled Motorola to boost revenue and market share in key business sectors it has entered in the past and present: public safety, paging, cellular networks, and satellite technologies. All of these verticals have high expectations for quality and reliability.

Scope, Quality and the Triple Constraint

The Project Management Triple Constraint

Project management professionals know the triple constraint: scope, cost and schedule. While often considered an implicit part of scope, quality is often overlooked to achieve faster delivery and lower costs.

In hypercompetitive markets where gaining first-mover advantage is very important, it is understandable that this tradeoff is necessary to gain the initial sale. That being said, after that sale is made, it is crucial not to neglect quality. Customers may initially tolerate a less robust product, but long-term success with strategic customers depends on consistency and reliability.

The Challenges of Generative AI Applications

Robot Factory Worker - Source: DALL-E 3

The inherent design of generative AI applications and the neural networks they rely upon present unique challenges, especially when customer requirements set high expectations. The probabilistic and stochastic nature of large language models (LLMs) and neural networks means they inherently produce outputs with some variability and unpredictability.

Probabilistic Nature of LLMs

Generative AI, like LLMs, operates on patterns learned from vast datasets. This means the output is a product of probabilities rather than deterministic rules. While this allows for remarkable creativity and flexibility, it also introduces the risk of "hallucinations" – factually incorrect or nonsensical outputs.

Robotic Hallucinations - Source: DALL-E 3

Techniques to Mitigate Risks

Use larger models, trained on more data

The use of larger models trained on more extensive and diverse datasets can significantly reduce the probability of erroneous outputs. Larger models benefit from a wealth of training data, making them less likely to produce outputs that deviate from learned patterns and facts. They excel at understanding nuances and complex relationships, allowing them to distinguish between subtle differences and avoid making incorrect associations. This enhanced understanding helps in generating more accurate and reliable outputs, thus improving overall model performance.

Use prompt engineering

Prompt engineering involves carefully designing the inputs to the AI model to guide its outputs more effectively. By refining prompts and using contextually rich inputs, the likelihood of generating accurate and relevant responses increases. I have explored and experimented with this specific technique to enhance the predictability and reliability of generative AI applications I created, and I wrote about my experience in this blog post.

Deploy AI agents

The emergence of AI agents represents a major leap in improving the quality and reliability of generative AI applications. These agents autonomously perform tasks, monitoring and validating AI model outputs to create effective feedback loops. For example, AI applications can fact-check each other, adding scrutiny and correction before delivering the final output. This process enhances accuracy and reduces errors, as seen in natural language processing where AI agents refine generated text to meet quality standards. However, AI agents require precise programming and can introduce biases in feedback mechanisms, presenting significant challenges.

Limitations of These Techniques

It is important to note that while these techniques reduce hallucination risks, they cannot eliminate them due to the models' probabilistic nature.

The Importance of Human Intervention

Mission Critical Robot Firefighters - Source: DALL-E 3

In mission-critical applications, where consistent product and service reliability is paramount, relying solely on generative AI without human oversight is risky. Human intervention is essential to verify and validate AI-generated content, ensuring it meets the required standards of accuracy and reliability. Your product’s and company’s reputation are at stake.

Generative AI: Innovation, and now regulation

The European Union, on March 13th 2024, formally passed the AI Act, which sets comprehensive rules for the development and use of AI systems, especially focusing on high-risk applications. The AI Act is designed to regulate AI applications based on their potential risks to society, with the highest scrutiny applied to high-risk applications such as those in healthcare, law enforcement, and critical infrastructure.

In addition to the AI Act, the EU has ensured that this new regulation works in harmony with existing laws like the GDPR. The AI Act incorporates stringent requirements for transparency, accountability, and human oversight, particularly for high-risk AI systems.

Conclusion

High software quality is crucial for enhancing product and service value. Companies like Motorola have demonstrated that consistent, reliable products boost customer satisfaction, loyalty, and revenue. Generative AI, however, presents new challenges in maintaining quality. There are techniques can reduce quality risks, but the inherent variability of AI models requires human oversight, particularly in mission-critical applications. Balancing automation with human intervention is essential for maintaining high-quality standards as AI integration continues.