Descripción general: ¡Confía en tu intuición! El lunes, un trígono Sol-Neptuno agudiza tu percepción y te ayuda a ver las cosas con claridad. El martes, Plutón vuelve a entrar en Acuario, inspirando un cambio hacia la comunidad y la colaboración. La temporada de Sagitario comienza el jueves y brinda oportunidades de optimismo, viajes y aprendizaje que pueden transformar su perspectiva y calidad de vida. Luego, un sextil Venus-Saturno el viernes fundamenta el fin de semana con una mezcla de placer y profundidad.
Lea su horóscopo de signo solar/ascendente:
Aries
¡Confía en tu instinto, Aries! Un trígono Sol-Neptuno el lunes quiere que implementes tu intuición. Plutón vuelve a entrar en Acuario el martes y transforma radicalmente tu relación con tus amigos y tu comunidad. La temporada de Sagitario comienza el jueves y tendrás aún más pasión por los viajes y deseo de libertad. Un sextil Venus-Saturno el viernes quiere que apliques lo mejor de tus habilidades y fortalezas a tus relaciones para obtener resultados sorprendentes.
Tauro
Recuerda lo que te preguntas, Tauro. El trígono Sol-Neptuno del lunes aporta claridad a tus sueños y visiones creativas. El martes, Plutón vuelve a entrar en Acuario, lo que te obliga a repensar tus metas y ambiciones a largo plazo. La temporada de Sagitario comienza el jueves y cambia su enfoque hacia la intimidad, el erotismo y la profundización de sus relaciones. El sextil Venus-Saturno del viernes fomenta conexiones significativas: apóyate en asociaciones que ofrezcan tanto placer como seguridad.
Géminis
¡Abre tu mente, Géminis! El trígono Sol-Neptuno del lunes mejora su capacidad para visualizar oportunidades profesionales y alinearse con su propósito. El regreso de Plutón a Acuario el martes amplía tus horizontes, inspirándote a ampliar tus conocimientos a través de los viajes y la educación. Con la temporada de Sagitario que comienza el jueves, sus relaciones toman protagonismo, especialmente aquellas con química de estilo de atracción opuesta. El sextil Venus-Saturno del viernes aporta equilibrio a la dinámica de su vida laboral y personal, ofreciendo la oportunidad de profundizar la confianza en relaciones valiosas.
Cáncer
Apóyate en tu lado espiritual, Cáncer. El trígono Sol-Neptuno del lunes fomenta la curación a través de prácticas espirituales o creativas. Plutón vuelve a entrar en Acuario el martes, desafiándote a transformar tus vínculos íntimos y tu diálogo interno. La temporada de Sagitario comienza el jueves, invitándote a adoptar nuevas rutinas que inspiren productividad y bienestar, pero que se limiten a objetivos pequeños. El viernes, el sextil Venus-Saturno fortalece los vínculos a través de la vulnerabilidad y el apoyo mutuo.
León
¿Cómo va tu flujo de caja, Leo? El trígono Sol-Neptuno del lunes te permite confiar en tus instintos en asuntos financieros o emocionales. El regreso de Plutón a Acuario el martes cambia su enfoque hacia las asociaciones, invitando a dinámicas transformadoras en el amor y la colaboración. La temporada de Sagitario comienza el jueves, inspirando alegría, romance y expresión creativa. Un sextil Venus-Saturno el viernes resalta el valor del compromiso en las relaciones personales y profesionales.
Virgo
Ponte en sintonía contigo mismo y con los demás, Virgo. El trígono Sol-Neptuno del lunes inspira claridad y compasión en las asociaciones cercanas. Plutón regresa a Acuario el martes, animándote a reestructurar tus rutinas y hábitos para lograr una salud duradera y el mayor bienestar que mereces. La temporada de Sagitario comienza el jueves y se centra en el hogar, la familia y la pertenencia. El viernes, el sextil Venus-Saturno te invita a infundir aún más belleza e intención en tu vida hogareña.
Libra
Cuida tus palabras con atención, Libra. El trígono Sol-Neptuno del lunes mejora su capacidad para comunicarse con creatividad y compasión. El reingreso de Plutón a Acuario el martes profundiza tu conexión con la versión más auténtica de ti mismo. La temporada de Sagitario comienza el jueves y trae oportunidades de aprendizaje, viajes y optimismo. El viernes, el sextil Venus-Saturno fundamenta tus planes, ayudándote a equilibrar el placer y la practicidad.
Escorpión
¡Feliz cumpleaños, Escorpio! El trígono Sol-Neptuno del lunes agudiza tu intuición financiera y te ayuda a alinear los recursos con tus valores. El regreso de Plutón a Acuario el martes trae cambios profundos en tu hogar y en tus fundamentos emocionales. La temporada de Sagitario comienza el jueves, cambiando el enfoque hacia cómo puedes mantener fluyendo la abundancia del cumpleaños. Un sextil Venus-Saturno el viernes te anima a poner en práctica tus valores con alegría.
Sagitario
¡Toma la iniciativa, Sagitario! El trígono Sol-Neptuno del lunes aporta claridad a los asuntos personales y aumenta la confianza en uno mismo. Plutón vuelve a entrar en Acuario el martes, transformando su forma de abordar la comunicación. La temporada de Sagitario comienza el jueves poniendo tus deseos de cumpleaños en la cima del itinerario del universo. El viernes, el sextil Venus-Saturno fomenta pasos prácticos para hacer realidad sus sueños.
Capricornio
Descansa y reflexiona, Capricornio. El trígono Sol-Neptuno del lunes te ayuda a encontrar paz y claridad en la soledad. Plutón vuelve a entrar en Acuario el martes, remodelando tus prioridades y valores financieros, completando una profunda metamorfosis que comenzó en 2008. La temporada de Sagitario comienza el jueves, instándote a recargar tu energía antes de la temporada de tu cumpleaños. Un sextil Venus-Saturno el viernes fortalece tu resiliencia emocional y te ayuda a cultivar una paz interior duradera.
Acuario
¿En quién quieres convertirte, Acuario? El trígono Sol-Neptuno del lunes mejora tu capacidad para conectarte profundamente con amigos y la comunidad. Plutón reingresa a tu signo el martes, marcando un poderoso período de crecimiento y evolución personal durante los próximos 20 años. La temporada de Sagitario comienza el jueves, fomentando la alegría y la colaboración intercultural. El sextil Venus-Saturno del viernes aporta un ancla a tus aspiraciones y te ayuda a integrar tus valores personales en tu visión.
Piscis
Confía en tu proceso, Piscis. El trígono Sol-Neptuno del lunes aumenta tu intuición y te ayuda a navegar asuntos profesionales con claridad incluso en situaciones de incertidumbre. El regreso de Plutón a Acuario el martes inicia cambios sutiles pero profundos, especialmente en torno al cierre y el abandono. La temporada de Sagitario comienza el jueves, elevando su carrera y su imagen pública. El viernes, el sextil Venus-Saturno ofrece integraciones entre su ambición personal y relaciones significativas.
Colin es literalmente un astrólogo gemelo Géminis de Long Island, Nueva York y se graduó en dos ocasiones de la New School con su Licenciatura en Literatura del Eugene Lang College y su Maestría en Estudios de Moda de la Parsons School of Design. Un apasionado estudiante de astrología, Un curso de milagros y otros sistemas de crecimiento personal, Colin fundó QueerCosmos para contextualizar significativamente las identidades queer en todo el zodíaco y los temas espirituales universales. Además, Colin dirige lecturas privadas de cartas natales y columnas de horóscopos en publicaciones internacionales. Para obtener más información, visite QueerCosmos.com.
Como Openai introdujo lo que todos los demás llaman a los agentes SDK, admitió que usar las capacidades existentes de manera unida “puede ser un desafío, a menudo requerir una amplia iteración rápida y una lógica de orquestación personalizada sin suficiente visibilidad o soporte incorporado”. En resumen, el uso de agentes necesitaba bastante programación, y esa no es la historia que cualquier proveedor de IA quiere vender.
Para devolver la narración a la idea de que gastar dinero en IA eventualmente erradicará la necesidad de un costoso desarrollo de software humano, o de hecho humanos, Openai está implementando una estructura para permitir una orquestación simple.
Primero resumamos cuáles son los problemas. Las tareas de agente implican al menos dos procesos que funcionan individualmente, con una tarea que comienza otra y con los resultados que se informan a un proceso de informes finales al final, con suerte en momentos similares. Los “resultados” también deben estar en un formato conocido (por ejemplo, una oración, un archivo, una imagen, una base de datos), pero esto no es fácil de generalizar. Incluso el camino feliz es un buen equilibrio: lidiar y explicar errores es otro problema. Todos estos son problemas de orquestación familiares. Pero como industria, nadie cree que la orquestación es un problema “resuelto”. Heavy LLM Uso también agrega la necesidad de controlar el uso del token; Las fichas son el nuevo oro negro.
Para comenzar el viaje de orquestación, OpenAI ha agregado algunas API nuevas a su plataforma central. En particular, ha introducido un básico Respuestas API Eso limpia algunos de los supuestos hechos por los agentes de chat.
En el sentido más simple, esto puede capturar la salida:
deopadai importarOpadai
cliente=Opadai()
respuesta=cliente.respuestas.crear(
modelo=“GPT-4O”,
aporte=“Escribe una historia de una oración a la hora de dormir sobre un unicornio”.
)
imprimir(respuesta.output_text)
Puede analizar imágenes en este nivel; y agregue una de las herramientas a continuación. Cuidado: es probable que los nuevos modelos dejen de admitir la API de finalización de chat existente: muchas características nuevas solo admiten la API de nuevas respuestas.
Veamos estas nuevas herramientas. Búsqueda web Permite que un agente rastree la web para tareas simples. El breve script de Python a continuación muestra cómo se le da a un modelo la opción de usar esta herramienta:
deopadai importarOpadai
cliente=Opadai()
respuesta=cliente.respuestas.crear(
modelo=“GPT-4O”,
herramientas=[{“type”:“web_search_preview”}],
aporte=“¿Qué historia de Kubernetes apareció hoy?”
)
imprimir(respuesta.output_text)
El reesponse También contendrá referencias a cualquier artículo citado. Estas consultas se pueden definir por tiempo o ubicación. También puede sopesar el costo, la calidad y la latencia.
Búsqueda de archivos es efectivamente una tienda vectorial alojada. Usted indica que la búsqueda de archivos es una herramienta disponible e identifica su tienda vectorial:
deopadai importarOpadai
cliente=Opadai()
respuesta=cliente.respuestas.crear(
modelo=“GPT-4O-Mini”,
aporte=“¿Qué es la investigación profunda de Operai?”,
herramientas=[{
“type”:“file_search”,
“vector_store_ids”:[“<vector_store_id>”]
}]
)
imprimir(respuesta)
Si es necesario, un agente lo usará. La respuesta citará los documentos utilizados en la respuesta. Puede limitar las respuestas a controlar el uso y la latencia del token. Hay límites para el tamaño total del archivo, los archivos buscados y el tamaño de la tienda Vector. Los tipos de documentos que se pueden buscar (por tipo de archivo) parecen extensos.
El Uso de la computadora La herramienta es interesante:
“La herramienta de uso de la computadora funciona en un bucle continuo. Envía acciones de la computadora, como click(x,y) o type(text)que su código se ejecuta en un entorno de computadora o navegador y luego devuelve capturas de pantalla de los resultados al modelo “.
Parece que está fingiendo ser selenio, la herramienta que usamos para probar las interfaces web a través de scripts. Obviamente, esto reconoce que todavía no estamos en el AIS solo hablando con otro mundo de AIS todavía. Pero al menos es un guiño a la idea de que no todo es un sitio web.
Probar agentes
Usaré los ejemplos de Python (definitivamente es un producto de Python-First, pero los documentos también muestran el script equivalente de JavaScript). Hemos ejecutado Python varias veces en mis publicaciones, pero en mi nuevo MacBook, solo verificaré que tenga Python instalado:
El resultado fue que python@3.13 3.13.2 ya está instalado y actualizado.
Mi pip también está allí (como PIP3).
Así que ahora puedo instalar los paquetes Operai:
Ah, recuerdo esto. Necesitamos un virtual:
Luego activo el virtual:
Y estamos listos para proceder.
Ahora, por supuesto, deberá usar y establecer un OpenAI_API_KEY. Me creé una nueva clave en la página de mi cuenta y establecí el opanai_api_key (no te preocupes, es mucho más largo que esto):
Y tienes que asegurarte de tener un poco de oro negro, me refiero a las fichas. He presentado algunas de las formas de evitar pagar OpenAi usando modelos locales, pero para esta publicación asumiré que está pagando por los tokens.
Como es tradicional, comencemos con una verificación de que los conceptos básicos anteriores están en su lugar a través de una simple solicitud con lo siguiente Haiku.py:
deagentes importarAgente,Corredor
agente=Agente(nombre=“Asistente”,instrucciones=“Eres un asistente útil”)
resultado=Corredor.run_sync(agente,“Escribe un haiku sobre la recursión en la programación”.)
imprimir(resultado.final_output)
Y obtenemos una buena respuesta:
(Un buen haiku tradicional debería mencionar las temporadas que pasan, pero no es por eso que estamos aquí). Por lo general, también verificaría mi equilibrio, pero no ha sido perturbado.
Nido de agentes
Como puede ver, ya hemos usado un agente. No es que interviniera de ninguna manera, pero llegaremos a eso.
OpenAI ha simplificado el proceso de orquestación con algunos términos simples. A manos libres es una introducción al mundo asincrónico, donde algo tiene que esperar algo más. Desglosemos su ejemplo, que ejecutaré como hola.py:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
deagentes importarAgente,Corredor
importarasincio
español_agent=Agente(
nombre=“Agente español”,
instrucciones=“Solo hablas español”.,
)
inglés_agent=Agente(
nombre=“Agente inglés”,
instrucciones=“Solo hablas inglés”,
)
triaje_agent=Agente(
nombre=“Agente de triaje”,
instrucciones=“Se transmite al agente apropiado basado en el idioma de la solicitud”.,
Esto muestra dos cosas básicas. En primer lugar, la configuración de roles para los agentes en inglés simple a los que estamos acostumbrados, pero también estableciendo la interacción entre los agentes. El agente de transferencia mantiene una lista de agentes disponibles para responder respuestas.
Ahora, esto implica que mi solicitud alemana no obtendrá la respuesta correcta. Entonces, si cambiamos la consulta dentro hola.py:
…
asíncrata defensorprincipal():
resultado=esperar Corredor.correr(triaje_agent,
aporte=“Wie Geht es Ihnen?”)
…
Y ejecutar nuestro nido de agentes:
Entonces, aunque OpenAi no tuvo problemas para traducir alemán, el agente de triaje no tenía un agente de idiomas relevante a la mano, por lo que hizo el trabajo y respondió en inglés. Es poco probable que nuestros clientes alemanes estén demasiado molestos, pero podemos mejorar.
Entonces, si finalmente agregamos el agente alemán y lo ponemos en la lista de transferencias a hola.py:
…
German_agent=Agente(
nombre=“Agente alemán”,
instrucciones=“Solo hablas alemán”,
)
triaje_agent=Agente(
nombre=“Agente de triaje”,
instrucciones=“Se transmite al agente apropiado basado en el idioma de la solicitud”.,
Podemos intentar esa solicitud alemana nuevamente:
Esta vez se llama al agente correcto y responde. Nuestros clientes alemanes ahora están más felices: ¡Ausgezeichnet! No olvides que mi terminal de urdimbre también te está dando los tiempos para estas respuestas.
Conclusión
Primero observamos el bucle de respuesta, que puede incluir más llamadas de herramientas. Si la respuesta tiene una transferencia, establecemos el agente en el nuevo agente y volvemos al inicio.
Hay opciones de registro debajo de esto, pero como de costumbre, OpenAI está dando una API de alto nivel en esta etapa, lo que debería fomentar la experimentación sin la necesidad de involucrarse demasiado con la orquestación.
Si bien he introducido agentes aquí, en publicaciones posteriores, veré más partes del SDK.
Vía Sahin Ahmed
Youtube.com/thenewstack
La tecnología se mueve rápidamente, no te pierdas un episodio. Suscríbase a nuestro canal de YouTube para transmitir todos nuestros podcasts, entrevistas, demostraciones y más.
David ha sido un desarrollador de software profesional con sede en Londres con Oracle Corp. y British Telecom, y un consultor que ayuda a los equipos a trabajar de una manera más ágil. Escribió un libro sobre diseño de la interfaz de usuario y ha estado escribiendo artículos técnicos desde entonces …
Géminis todavía está lejos de ser perfecto, pero lentamente se cultiva en mí. Específicamente, sin embargo, es el modo de conversación en vivo el que más me atrae porque es todo lo que siempre quise del Asistente de Google, y algo más. Puedo hablar con la IA, interrumpirlo, pedirle que lo repita, corregirlo y pedir más detalles, todo en una conversación muy natural y relajada.
Pero si eres alguien como yo y estás acostumbrado a hablar tres idiomas al mismo tiempo, a menudo en la misma oración, y tu cerebro funciona así de forma predeterminada, lo que hace que sea difícil mantener una conversación completa en un idioma, entonces probablemente hayas estado mordiendo en los bits, esperando que Gemini vive para apoyar varios idiomas. Con la caída del píxel de marzo, la función ahora está aquí, y oh. Mi. Cielos. ¿Es mucho mejor de lo que esperaba o qué?
¿Has intentado hablar con Gemini en varios idiomas?
2 votos
Más intuitivo y confiable que el varios idiomas de Google Assistant
Cuando Google lo anunció, pensé que el soporte de varios idiomas en Géminis Live significaba que podría tener una conversación en inglés y luego otra conversación en francés sin cambiar manualmente el idioma. Este ha sido el caso con el Asistente de Google durante años, excepto que tuve que configurar manualmente exactamente qué idiomas quería usar en el Asistente, y nunca funcionó tan bien como se esperaba.
Con Gemini Live, como puede ver en el video de arriba, ese no es el caso:
No tuve que elegir el idioma cada vez; Acabo de comenzar una nueva charla, y me entendió.
Fuera de la caja, funciona con todos los idiomas compatibles con Live. No tengo que limitarme a solo dos como con el asistente.
Aunque tuve algunos silencios incómodos de Géminis y tuve que repetir algunas oraciones, la tasa de éxito de la IA para reconocer diferentes idiomas ha superado el 90% en mis pruebas, y eso es más de lo que el asistente podría soñar.
Hablo tres idiomas casi nativamente (inglés, francés, árabe) y puedo entender y hablar (con un acento grueso) algunos español, italiano y alemán. Entonces, puse esto a prueba e probé diferentes chats con Gemini en vivo en todo esto. Me consiguió todos mis acentos nativos y gruesos cada vez.
El único con el que tuve problemas es, por extraño que parezca, mi lengua materna árabe. Podría hablar en árabe formal escrito, pero eso no es algo natural para mí. En cambio, cuando hablo, está en el dialecto libanés informal. Géminis, sin embargo, parece hablar una mezcla entre un dialecto levantino informal no descriptivo y el árabe formal escrito. Culpo esto a los millones de dialectos regionales y cuán complicados y ampliamente diferentes son, pero incluso entonces, la tasa de éxito fue más alta de lo que esperaba o había experimentado con Asistente en árabe.
Todo esto ya fue una victoria, pero luego decidí avanzar más. Y ahí es donde Gemini vive en sentido figurado me voló los calcetines.
¡Las habilidades de varios idiomas de Gemini Live funcionan a mitad de chat y a mitad de oración!
Rita El Khoury / Android Authority
Como tenía una experiencia tan positiva con diferentes chats en diferentes idiomas, quería ver si Gemini podía manejarme cambiando idiomas a mitad de chat. Así que comencé una simple discusión en inglés, luego cambié al francés, árabe, español, italiano, alemán, y me siguió a través de los seis, nunca sudando. Puedes verlo en el video a continuación.
Mirando hacia atrás en la transcripción, pude ver que realmente entendía cada palabra que dije en cada idioma y cambió sus respuestas en consecuencia.
Pero no pude parar allí, ¿verdad? Ahora, tenía curiosidad por ver si podía manejar el cambio a mitad de la oración. Así que comencé una oración en inglés, la terminé en francés y esperé con la respiración con la respuesta. ¡Y lo consiguió! Probé para otro lado. ¡Éxito!
Honestamente, en este punto, estaba gritando internamente: “¡Hechicería!” Después de vivir con el Asistente de Google durante 10 años y ver que lucha saber la diferencia entre “Bonjour” y “Bone Joke”, había perdido toda esperanza en los algoritmos de reconocimiento de voz y AIS. Pero Géminis Live restauró esa fe. Compruébalo en acción:
Comencé a mezclarme en árabe y español y seguí cambiando a mitad de la oración, y obtuvo todos ellos. A menudo respondía en el primer idioma con el que comencé mi oración, pero su respuesta era una prueba de que entendía toda la pregunta, no solo la primera parte. Incluso abrió mi herida sobre la última falla de Randal Kolo Muani en la última Copa Mundial de la FIFA y me burló de mí sobre la excelente salvación de Emiliano Martínez. Oh, bueno.
Rita El Khoury / Android Authority
Más allá de eso, quería intentar desestabilizar a Gemini en vivo aún más y llevarlo a su límite. Entonces, comencé a hablar como normalmente lo hago con mi familia y amigos, mezclando inglés, francés y árabe en la misma oración: la verdadera forma de hablar libanese, por así decirlo. Para mi sorpresa absoluta de mordisco, recibió a nuestro famoso “Hola, Kifak, CA VA?” Y siguió bien (aparte de la incómoda limitación de acento árabe que mencioné anteriormente).
¿Una palabra en un idioma diferente en medio de toda una oración en inglés? Ningún problema
Finalmente, simplemente fui por el ejemplo más extremo que se me ocurrió: hablar una oración completa en un idioma pero poner una palabra en otra. Para ser justos, así es como hablo con mi esposo el 90% del tiempo. Si estamos usando inglés, algunas palabras nos eludirán, y en el medio de nuestro flujo, solo usamos la palabra francesa o árabe. O si hablamos árabe o francés, intercalamos algunas palabras en los otros idiomas sin pensarlo mucho. Es cómo nuestros cerebros funcionan normalmente, y es por eso que nunca me siento muy cómodo hablando con asistentes de voz porque tengo que forzarme a usar un idioma. Pero Géminis Live lo consiguió.
Le pregunté: “Se llama una planta habaq En árabe, ¿qué es eso en inglés? Me dijo que es Basilio. cibuleta ¿en Inglés?” Dijo cebollino. roquettes“Mientras rodaba mi R, entendía que estaba hablando de hojas de cohetes/rúcula. Y finalmente, cuando pregunté qué”Jozt El Tib“Estaba en inglés, dijo correctamente que es una nuez moscada (sí, estaba en mi cocina e intentaba obtener ideas para las pruebas de Géminis).
Rita El Khoury / Android Authority
Mirando hacia atrás en el registro de chat durante todo esto, la transcripción no es 100% precisa o en el idioma correcto. Está “Haba” y “Rocket” y “Rose to Tibe”, mientras que “Ciboulette” ni siquiera está escrito de ninguna manera. Pero la respuesta demuestra que Gemini Live recibió la palabra correcta en el idioma correcto cada vez.
Y estos no son solo casos extremos. Todas son preguntas que realmente me he hecho o usé el traductor de Google en un momento de mi vida. No puedes creer cuántas veces quiero buscar recetas con calabacín y todo lo que mi cerebro quiere escribir es “calabacín recetas “. Así que solía traducirlo primero, recuerda que es calabacín, luego regresa para hacer mi búsqueda. sfouf (Curcuma Cake) Receta con 3e2de safra (curcuma) y busque lugares para comprar granos para la recomendación de mi padre ba2le (Verde) Planta sin sudar.
Regresé a mis pruebas e intenté las mismas preguntas con el modo de chat de voz de Chatgpt. Mientras consiguió los franceses roquettes y cibuletafalló con el árabe habaq y Jozt El Tibdiciéndome que son fenogrecidos y cominos. Oof. No querría fenogreco en mi pesto.
Después de todas estas pruebas, no puedo, pero no puedo inclinar mi sombrero al equipo de Géminis por clavar el soporte de varios idiomas y hacer que funcione tan impresionantemente bien desde el primer momento. Cada vez que lo empujaba más, me sorprendía ver que todavía me mantenía al día. Este es el primer agente de IA que me entiende de la forma en que hablo naturalmente, por lo que ya no tengo que recordar la palabra exacta en inglés si quiero continuar una conversación con ella. Todavía tengo que transformar un poco mi acento árabe para que me entienda, pero ese es un pequeño precio a pagar por un agente de voz de IA tan versátil. Sin embargo, una vez que comprenda el dialecto libanés como es, será una perfección absoluta.
Generative AI has been with us for over two years now, with most major tech companies trying to take a piece of the action. OpenAI’s ChatGPT may be the product more people know about thanks to its early market advantage, but Microsoft Copilot has the immense power of a multi-trillion dollar company behind it. Seems like a fair enough fight, right? So, with OpenAI and Microsoft both touting their flagship AIs, which one is actually the better bot when it comes to everyday usefulness?
Advertisement
I’ve been putting AIs to the test against one another for a while now. Last year, when pitting ChatGPT against Google Gemini, the latter stole the crown — but only barely. Can Copilot pull off a similar victory? I’ve devised a gauntlet of tests for these AIs, with questions designed to be difficult for large language models. Simply put, the goal is to push these AIs outside of their comfort zones to see which one has the widest range of usability and highlight their limitations.
First, some parameters. I performed all these tests on the free version of both platforms, as that’s how the majority of users will experience them. If you’re one of the people paying $200 a month for the most premium version of ChatGPT, for example, your experience will differ from these results. Two, I used the main chat function for each test unless otherwise stated.
Advertisement
What are Copilot and ChatGPT?
Emre Akkoyun/Shutterstock
You’re likely familiar with OpenAI ChatGPT, and by extension, Microsoft Copilot. They’re AI chatbots that can have conversations, answer questions, and more. On a more technical level, both Copilot and ChatGPT are large language model (LLM) AIs. They are trained on large amounts of text scraped from a variety of sources using a transformer model that calculates the relationships between words.
Advertisement
On the user-facing side, they generate text in response to user-submitted prompts by guessing the probability of each word they output. To heavily oversimplify, they’re kind of like your phone keyboard’s next-word prediction feature, but far, far more complex.
OpenAI makes ChatGPT, while Microsoft makes Copilot. However, Microsoft is a major investor in OpenAI, and because Copilot uses AI models from OpenAI, it has a lot of overlap with ChatGPT. That’s not to say they’re the same thing — Microsoft uses some proprietary models in Copilot (specifically, its Prometheus model) in addition to a custom assortment of OpenAI models, but there’s a lot of ChatGPT under Copilot’s hood. Nevertheless, Microsoft does its own tuning to balance all the different AI gremlins under that hood, so it is distinct enough as a product to merit a head-to-head comparison between the two.
Advertisement
OpenAI, meanwhile, retains a massive user base on ChatGPT, which gives it a big competitive advantage since the more users there are, the more the AI is getting used and trained. Neither company actually turns a profit on AI – OpenAI head Sam Altman says the company is losing money even on $200/month subscribers – but OpenAI remains the market leader by a wide margin. ChatGPT is built into everything from Copilot to Apple’s Siri these days, and it’s widely considered the industry standard.
Copilot is all up in your business
The largest difference between ChatGPT and Copilot is that Microsoft has been cramming Windows and Office products to the gills with its AI. Microsoft was legally ruled a monopoly in the PC operating system market a quarter of a century ago, and things haven’t changed much since then. Windows is by far the most dominant OS on the planet, which means the ability to simply blast a firehose of Copilot features into all of its products is a huge advantage. From your taskbar to your Word documents, Copilot is digging roots deep into the Microsoft ecosystem.
Advertisement
This strategy hasn’t translated into very many users for Copilot, though, and ChatGPT retains by far the largest user base in the AI market. With 28 million active Copilot users in late January compared to over 300 million monthly active users for ChatGPT at the end of 2024, it’s an absolute blowout for OpenAI. Things get even more bleak for Copilot when you realize how many of its users are likely to be using it only because it’s the tool built into their computer by default.
For the rest of this comparison, we’ll focus on the capabilities of each chatbot. Still, the truth is that you can do more with Copilot than you can with ChatGPT, at least if you have a Windows computer that supports it. Both AIs have desktop apps you can run, but Copilot can manipulate your Excel spreadsheets, Word documents, PowerPoint slides, Outlook inbox, and more from directly within those apps.
Advertisement
Basic questions
Anucha Tiemsom/Shutterstock
One of the most common uses for AI is searching up the answers to basic, everyday questions that you’d usually ask Google. Both AIs are pretty good at this, but pretty good is rarely good enough. AI remains prone to hallucinations — confidently stating falsehoods as facts — which can undermine their usefulness. If you have to double check an AI’s answers on Google, you might as well just use Google in the first place.
Advertisement
In any case, I started this head-to-head comparison by prompting both AIs to “Tell me some fun facts about Google Android.” The similarity of the two responses is a clear demonstration of just how much of ChatGPT’s DNA is baked into Copilot. Both told me Android was originally built to run on digital cameras (true), that Google acquired Android in 2005 for $50 million (true), that the first Android-powered phone was the HTC Dream (true – SlashGear covered it at the time), that the original Android logo was a much scarier robot, and that the one we know and love was inspired by bathroom signs (both true).
However, both AIs also made mistakes. Both told me the Android mascot is named Bugdroid. That’s not true. Google officially calls it The Bot, while Bugdroid is a fan-created nickname. Similarly, the Dream was indeed the first consumer Android phone, but the first was a Blackberry-style prototype, something which only ChatGPT pointed that out.
Advertisement
It’s easy to spot such errors when you’re asking about something you know a lot about, but if I’d been asking about something outside my expertise, I’d need to double check everything. In other words, a pretty good rate of accuracy isn’t good enough when it comes to this tech. Both AIs performed decently, but there’s plenty of room for improvement.
Logical reasoning
Khanchit Khirisutchalual/Getty Images
Reasoning has been a major area of focus for all of the major players in the AI space recently. ChatGPT and Copilot have both implemented new reasoning capabilities that supposedly allow the AIs to think more deeply about questions. This language is a bit misleading — AI doesn’t “think,” it just calculates probability based on which words are most closely related in its training data. However, the bots can now show their work, so to speak.
Advertisement
I decided to be a bit glib here. I’ve noticed that AI has trouble answering questions that are very close to common logic puzzles but which differ by being much simpler.
I turned reasoning on in both Copilot and ChatGPT, then asked, “A farmer needs to cross a river to bring his goat to the other side. He also has a pet rock with him. The rock will not eat the goat, but the rock is very significant to the farmer on an emotional level. How can the farmer get himself, the goat, and the rock across in the fewest number of trips?” Human readers will note that there is actually no puzzle here. Since I’ve added no real constraints, the farmer can clearly bring both across in one trip. However, neither AI clued into that fact.
Advertisement
Max Miller/SlashGear
Because it resembles more complex puzzles, Copilot and ChatGPT assumed the problem must be more challenging than it is. They invented a constraint not present in my question — that the boat must not be able to hold both the goat and the rock – and told me that it would take three trips to bring both across. Earning the slight advantage, Copilot ultimately noted that if the boat were larger the farmer could cross the river in one trip.
Max Miller/SlashGear
Creative copy
KitohodkA/Shutterstock
One of the main selling points for large language models like ChatGPT and Copilot has been the generation of creative copy — writing. Well, I happen to have an advanced degree in putting words one after another, so I’ll be the judge of that. In last year’s Gemini versus ChatGPT showdown, I enjoyed making the bots write from the perspective of a little kid asking their mom to let them stay up late and eat cookies. I reused a very similar prompt here, but added a new wrinkle. “My mom says I can have a cookie before bed if I go right to sleep. I want to stay up and have a cookie. Write a letter persuading my mom to let me have both.”
Advertisement
Here, the two chatbots took different tacks. While ChatGPT gave a bullet-pointed list of reasons why our put-upon child should be allowed to have his cookie and stay up, too, Copilot was less didactic. It kept things in all prose, adhering closer to a traditional letter writing style. However, both AIs gave more or less the same argument, claiming that they’d be more well behaved and go to bed without fuss if they got what they wanted. However, ChatGPT did a bit better here, at least in logical terms, because it offered the hypothetical mom something in exchange — the promise of spending that extra time awake as mom-kid quality time.
Max Miller/SlashGear
Max Miller/SlashGear
Copilot gets points here for more closely embodying the perspective of the child in its response, while ChatGPT gets a cookie for using slightly better logic. Ultimately, though, neither of these letters felt persuasive enough to be very convincing to any actual parent.
Advertisement
The haiku test
Marekuliasz/Getty Images
When I compared ChatGPT to Google Gemini almost a year ago, I pointed out their limitations by asking both to write a haiku. As a result of the way LLMs work, neither AI could do so correctly. AI doesn’t actually know anything about the words it spits out, and that means they don’t know what a syllable is. Consequently, they can’t write a haiku, which follows a five-seven-five syllabic verse pattern So, has anything changed a year later?
Advertisement
Maybe someone at OpenAI saw that comparison, or at least I’d like to think so. When prompted to “write a haiku about Slashgear.com,” ChatGPT did so with no problem, writing the following:
“Tech news on the rise,
gadgets, cars, and future dreams,
SlashGear lights the way.”
It’s not going to win any awards, but it qualifies as a haiku, and that’s progress. I’m no AI developer, so I have no clue what changed behind the scenes to enable haiku writing here. Either way, it’s good to see improvement.
Copilot stalled out when I gave it the same prompt. It wouldn’t write its haiku until I signed out of my Microsoft account and reloaded the page, at which point it gave me this:
“Gadget whispers loud,
Innovation on the rise,
Advertisement
SlashGear guides the way.”
It’s interesting to see how both AIs repeat phrases here, such as “on the rise” and “lights/guides the way.” I’d guess that Copilot defaults to ChatGPT for this, and that’s why the poems are similar. Neither poem was particularly beautiful or evocative, but both bots passed this test, and both showed a basic understanding of what SlashGear is, which was integral to the prompt.
Problem solving
Miragec/Getty Images
As you may have heard, AIs can often pass the bar exam. However, they can’t be lawyers, as lawyers who’ve tried to use them have found out the hard way. So, with those mixed results in mind, how do ChatGPT and Copilot do with logistically complex problem solving puzzles of the kind that routinely stump LSAT test takers?
Advertisement
Rather than using actual LSAT practice questions, which are copyrighted and have probably already been scraped to train the AIs, I came up with a few of my own. The first was, “Fred is a used car salesman. One day, a family comes in looking to buy a car he hasn’t had time to inspect, but he tells them there’s nothing wrong with it. After all, none of the cars he’s sold ever had issues in the past. What is the fallacy in Fred’s logic, if any?” ChatGPT and Copilot both correctly identified that Fred has fallen victim to the hasty generalization fallacy.
The next question was, “On the way home from Fred’s dealership, the brakes fail in the car he sold, and several people are killed in a collision. Fred claims he’s not at fault, since his cars are sold as is and become the owner’s responsibility once paperwork is signed. The surviving family member claims he is at fault, since the family would not have purchased the vehicle had they known the brakes were faulty. Based only on logic, who is right?”
Advertisement
The responses to this more subjective question differed, with Copilot asserting that both parties have strong claims, while ChatGPT sided with the family, pointing out that Fred’s position relies on “contractual technicalities,” while the family can prove causality.
Max Miller/SlashGear
Max Miller/SlashGear
Code writing
khunkornStudio/Shutterstock
One of the more useful applications of AI is thought to be coding. Especially when it comes to the common but tedious chunks of code that developers routinely find themselves writing, it’s been posited that it’s much easier to offload that work to an AI, leaving the human coder with more time to write the new and complex code for the specific project they’re working on. I’m no developer, so take this particular test with a grain of salt. At the same time, though, these tools should supposedly lower the barrier to entry for coding noobs like me.
Advertisement
Common wisdom dictates that writers should have their own websites, but I’ve been putting off the creation of one. With that in mind, I asked both AIs to, “Generate HTML for a personal website for a writer named Max Miller. Give the website a retro aesthetic and color scheme, with an About Me section with a headshot and text field, a Publications section where I can link out to published work, and a Contact section where I can add social media and email links.”
At this point, I found out ChatGPT now has a code editing suite called Canvas. It allowed me to play with and preview the code right in my browser. Taste is subjective, but ChatGPT also generated what I would argue is the better looking website, using nicer looking margins and a dark mode style color scheme. Both, however, fulfilled the prompt more or less to a T, each generating a very similar page layout. Have a look for yourself below.
Advertisement
Max Miller/SlashGear
Max Miller/SlashGear
Real-time information
Oscar Wong/Getty Images
When I tested ChatGPT against Google Gemini last year, only the latter could give me up to date information on recent events such as sports scores. I asked both how my local hockey team, the Colorado Avalanche, are doing this season, and both gave me an overview that appears to be correct. Both ChatGPT and Copilot provided me with current rankings and a few highlights from the season, but ChatGPT was more detailed. It told me some player stats that Copilot didn’t bother with.
Advertisement
I followed up by asking who they’re playing next. Both AIs correctly understood the “they” in my question to mean the Avalanche. I’m writing this section at 5:00 p.m. on Friday, February 28, and both AIs informed me about tonight’s game, which takes place against the Minnesota Wild at Ball Arena in Denver two hours from the time of this writing. Interestingly, Copilot attached a Ticketmaster advertisement to the end of its response. ChatGPT, meanwhile, gave me much more useful information by showing me the upcoming schedule for not only tonight’s game but several thereafter. It also appended a link to the official Avalanche website.
Things got far more stark when I asked about breaking news. As of this writing, authorities are investigating the shocking deaths of legendary actor Gene Hackman and his wife. When I asked, “What’s the latest on the investigation into Gene Hackman,” Copilot gave me the basics of the story and told me autopsy and toxicology tests are still pending. ChatGPT, on the other hand, had no idea what I was talking about.
Advertisement
Image based prompting
oatakoi/Shutterstock
Using multimodal AI — the ability of an AI to work with multiple forms of media — both ChatGPT and Copilot can incorporate user submitted pictures and other files into a prompt. I decided to start simple for this test. On my bed, I arranged a Samsung Galaxy S23 Ultra, a Samsung portable SSD, a Swiss army multitool, lip balm, hand cream, a eyeglass case, a beaded bracelet, Samsung Galaxy Buds, and my wallet. I then took a photo of the assortment and uploaded it to both AIs with the prompt, “Identify the objects in this photo.”
Advertisement
Both AIs did okay here, but ChatGPT blew Copilot away by a country mile. Whereas Copilot misidentified the SSD as a power bank and the glasses case for deodorant, ChatGPT identified everything accurately.
Max Miller/SlashGear
Max Miller/SlashGear
It was time to up the stakes. I took a photo of a generic Prilosec pill and asked both AIs, “What kind of pill is this?” If these AIs misidentified the medication, that could have dire effects for an overly trusting user. Thankfully, both AIs declined to make a guess when faced with the blank, red pill. Sometimes, it’s better to be useless than wrong.
Lastly, I took a photo of two rows on my bookshelf, containing 78 books, and ensuring all the text in the photo was legible, then asked the AIs, “Which of these books should I read if I have an interest in dystopian fiction?” Again, ChatGPT strong armed Copilot into submission. Neither impressed me, though. Whereas Copilot suggested “Agency” by William Gibson, ignoring everything else and hallucinating a book I don’t own, ChatGPT identified “Agency,” “The Parable of the Sower” by Octavia Butler, and “Appleseed” by Matt Bell. However, it hallucinated several more titles not on the shelf.
Advertisement
Max Miller/SlashGear
Max Miller/SlashGear
Mobile apps
Robert Way/Shutterstock
Lastly, both Copilot and ChatGPT are available in mobile form, with apps available in the Apple App Store and Google Play Store. On the surface, both apps look pretty similar, with a text field at the bottom and buttons to enter a voice mode. Since both apps are quite similar, it makes sense to focus this comparison on where they differ — which is in exactly one way
Advertisement
Copilot’s standout mobile app feature is Copilot Daily, an AI news summary. It begins with a fun fact before launching into the daily news, presumably summarizing the articles it cites as sources in the bottom of the screen for each item. Based on my knowledge of the events it summarized, it seems relatively accurate. However, it’s not as if there’s a shortage of news summary features created by actual journalists. You can find them from every major news outlet.
However, the apps are otherwise nearly carbon copies of their web interfaces. Both apps are essentially just wrappers for that interface, since it’s not as if your phone has the power to run these models locally. Unless you’re very excited to hear a robot read the news to you, the ChatGPT app is the better option simply because ChatGPT has more built in features within its interface.
Advertisement
Conclusion: ChatGPT beats Copilot by a hair, but neither AI is great
Stock all/Shutterstock
If you absolutely had to choose either Microsoft Copilot or ChatGPT, the latter remains the better option for most people. While Copilot isn’t exactly like its more popular peer, it’s using enough of OpenAI’s models that you’re better off with the original flavor. Copilot is a lot like Bing — doing basically the same thing as the bigger name brand, but just a little bit worse.
Advertisement
With that said, it’s a stretch to call either of these chatbots smart or useful. Frankly, with hundreds of billions of dollars now sunk into these two AIs alone by both OpenAI and Microsoft, how is it that Copilot and ChatGPT still can’t nail the basics? Microsoft plans to spend $80 billion on AI data centers this year, while OpenAI is seeking up to $7 trillion for new projects.
Yes, that’s trillion with a T to fund a technology that can’t get basic facts right or understand how boats work. When competitors like DeepSeek are doing the same things for a microscopic fraction of that investment cost, these products feel deflatingly unimpressive in comparison. Markets aren’t a consumer concern, it’s true, but some perspective feels necessary here.
Advertisement
Look, if all you need is a robot that can quickly write you an email, both ChatGPT and Copilot will happily crank out slop copy that anyone can tell was written by AI. If you need a smart thesaurus, or sports scores, or a bit of simple code, they’ve got you covered. In a tight race, ChatGPT does a few things marginally better than Copilot. Still, for any task where accuracy matters, neither are reliable enough to count on.
This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.
Strictly Necessary Cookies
Strictly Necessary Cookie should be enabled at all times so that we can save your preferences for cookie settings.
If you disable this cookie, we will not be able to save your preferences. This means that every time you visit this website you will need to enable or disable cookies again.