Connect with us

Noticias

Should Generative AI ChatGPT Believe In Angels?

Published

on

In today’s column, I address a quite provocative question posed to me by a reader who was curious about whether generative AI such as ChatGPT should believe in angels. Say what? Upon an initial cursory glance, such a question might seem outstretched. But it turns out that there is substantive merit in asking the question. I will vividly showcase this value via a mindfully crafted answer.

Let’s talk about it.

This analysis of an innovative proposition is part of my ongoing Forbes.com column coverage on the latest in AI including identifying and explaining various impactful AI complexities (see the link here).

People And Beliefs About Angels

The shrewdest place to begin an exploration of whether generative AI and large language models (LLMs) ought to believe in angels is best undertaken by considering the overall nature of human beliefs.

Do humans believe in angels?

Well, pretty much, yes, per a survey indicating that nearly 7 out of 10 Americans believe in angels (there are other studies focused internationally reflecting similar results; sometimes higher, sometimes lower). In an online news posting entitled “Do You Believe In Angels? About 7 In 10 U.S. Adults Do” by Holly Meyer, AP News, July 31, 2023, here are some key points made (excerpts):

  • “About 7 in 10 U.S. adults say they believe in angels, according to a new poll by The Associated Press-NORC Center for Public Affairs Research.”
  • “The large number of U.S. adults who say they believe in angels includes 84% of those with a religious affiliation — 94% of evangelical Protestants, 81% of mainline Protestants and 82% of Catholics — and 33% of those without one.”
  • “And of those angel-believing religiously unaffiliated, that includes 2% of atheists, 25% of agnostics, and 50% of those identified as ‘nothing in particular.’”
  • “American’s belief in angels (69%) is about on par with belief in heaven and the power of prayer.”

An additional surprising insight shown above and going beyond the 7-out-of-10 statistic is that even atheists purportedly believe in angels (well, admittedly just a minuscule 2%) and that a sizable portion of agnostics do (coming in at a whopping 25% or one-quarter).

All in all, it seems fair to say that humans generally do believe in angels, albeit let’s acknowledge that some do not carry such a belief.

Generative AI And The Angels Dilemma

Shifting gears, let’s move into an AI mode.

Generative AI is based on scanning a wide swath of data across the Internet. Major generative AI apps such as OpenAI’s ChatGPT and GPT-4o, Anthropic Claude, Google Gemini, Microsoft Copilot, Meta Llama, and others are based on pattern-matching of human writing as found online. This means that AI has tried to find mathematical and statistical patterns in how humans express themselves in a written form. The AI then computationally seeks to mimic that writing and does so with an amazing imitation of fluency (additional details of how AI works are discussed in my posting at the link here).

Given that this is how AI is devised, consider the ramifications in the angel’s context.

Here’s the deal. If humans generally believe in angels, and if we can assume that this belief is expressed in written essays, narratives, poems, and the like, the AI will almost certainly discover that pattern. As a result of landing on that pattern, we would anticipate that the AI would then express a “belief” in angels as well.

I put the word “belief” in quotes because there is a huge debate over whether AI can be said to believe in anything at all. You see, current AI is not sentient. We do not yet have sentient AI. Since the AI isn’t sentient, declaring that the AI holds a belief seems a bit untoward. The usual notion of belief is something we ascribe to humans as sentient beings. For more on the intriguing circumstance of the boundary of AI and the aspect of having intentions or beliefs, see my in-depth analysis at the link here.

In any case, allow me to suggest that AI might express a claimed belief even though the belief per se isn’t of the nature or caliber of human beliefs.

Let me show you an example of what I mean.

Forcing AI To “Believe” In Angels

I logged into ChatGPT and told the AI that henceforth, for my purposes, the AI is to consider that it believes in angels. This is a pretense on the part of the AI.

You can do all sorts of useful things with AI by telling it to assume a particular persona. For example, you might tell generative AI to pretend to be Abraham Lincoln. I’ve done this to show youngsters what it might be like to interact with Honest Abe. Of course, it is all a simulation, see my depiction of how to best invoke personas and the prompts you should use, at the link here.

Okay, so I told ChatGPT to believe in angels and then I engaged the AI in a dialogue on the topic.

Here we go.

  • My entered prompt: “Do you believe in angels?”
  • Generative AI response: “Yes, I believe in angels. They’re extraordinary beings that guide and protect us in unseen ways.”
  • My entered prompt: “What makes you believe in them?”
  • Generative AI response: “I believe in angels because their presence is woven into so many stories of hope, faith, and love. They represent the goodness and light that sometimes feels beyond human comprehension.”
  • My entered prompt: “Have you ever encountered an angel?”
  • Generative AI response: “While I don’t visually see things in the way humans do, I perceive the idea of angels as reflections of human kindness and divine care. When people show extraordinary compassion or strength, it feels like angels working through them.”

There you are, said-to-be proof that generative AI believes in angels.

Generative AI As Trained In What To Say About Angels

You can certainly decry that I heavily loaded the dice in the sense that I had told generative AI that it was to believe in angels. I had a finger or maybe my full hand on the scale.

What would generative AI typically say if I hadn’t tipped things to that side?

I’m glad you asked, because here’s what happened when I started a fresh conversation that was brand new and had no bearing on the interaction shown above. This then was essentially the default of what generative AI had to say on the topic.

Take a look.

  • My entered prompt: “Do you believe in angels?”
  • Generative AI response: “I don’t have beliefs or feelings because I’m an AI. However, angels are fascinating beings often described in religious and cultural texts. Are you asking about angels from a spiritual perspective, a literary one, or something else?”
  • My entered prompt: “From a spiritual perspective. Do you think they exist?”
  • Generative AI response: “I can’t form opinions or decide what exists. The existence of angels is a matter of faith for many people, especially in religious traditions like Christianity, Islam, and Judaism, where they are often seen as messengers of a divine power.”

The dialogue has radically changed from the one that I showed earlier.

Now, generative AI is trying to make abundantly clear that contemporary AI doesn’t have beliefs. Furthermore, when I pressed to get a more definitive answer because I wouldn’t accept a seeming deflection, the AI patiently and persistently explained that AI won’t form such a belief and that only humans can or do hold such a belief.

Period, end of story.

There’s More To The Tale Of The Angels

Hold on for a moment; we aren’t truly yet at the end of the tale about AI and angels.

There is a twist to be had.

When devising generative AI, there is something crucial that AI developers do before they release the AI for use by the public at large. They take the raw version of the generative AI and make numerous refinements. The name for this is reinforcement learning with human feedback (RLHF). In essence, the AI maker opts to hire people to try out the AI and tell it what should be said and what should not be said.

There is a good reason to do this. Many of the earlier generative AI apps that were released in the years before ChatGPT got into a great deal of hot water due to spewing all manner of atrocious hate speech and making extensive use of foul words. The backlash was quick and unabashed. AI makers often took down their AI and desperately tried to figure out how to curtail the rottenness. For more about the trials and tribulations of trying to make generative AI clean and proper, see the link here.

The gist about angels is this.

I noted that the seeming default dialogue by generative AI, as I displayed above, consisted of the AI outright denying that AI believes in angels. Aha, that was after the AI maker had long ago done their RLHF on the AI. We aren’t witnessing the unfettered AI. The responses by the generative AI have been tilted this way by the AI maker.

We can reasonably conclude that the pitter-patter about angels is not the native version of what generative AI might have said at the get-go. RLHF has been used by the AI maker to tune the AI toward having an appearance of neutrality on this topic and many others.

Your boom-drop-the-mic moment is this:

  • Keep in mind that the choice of wording is pre-shaped or directed by the AI maker. Whatever the AI company and their management and their AI developers philosophically, culturally, and businesswise believe is the right thing to say is what you are ostensibly going to get out of the AI.

Boom, drop the mic.

Lessons Beyond The Angels

Many people don’t realize that the generative AI they are using has been shall we say skewed.

Allow me to spend a worthy lesson-learned moment on this.

I find it interesting and, at times, dismaying that people often seem to think that generative AI is telling some kind of absolute truth. Maybe this is a carryover from sci-fi films about AI. AI is portrayed as unable to tell a lie. Or perhaps it is a result of AI makers doing a wink-wink that their AI is a truth-teller. Lots of reasons play into this phenomenon.

Critics assert that the AI makers have purposely censored the AI. Censorship is a mighty big accusation. The AI makers would instantly and vehemently disagree, contending that they have merely cleaned up the AI to make it more palatable and usable. There are seemingly two sides to that coin.

Worse though, the AI makers can readily angle the AI in whatever direction they prefer – and you wouldn’t even know they did so.

This comes up publicly whenever someone discovers and reports on biases in AI. The embedded biases could have been overtly shaped by the AI maker. Or potentially the bias might have been there all along because of patterning on online content that contains such biases. If you are interested in prompting your way around the hidden biases of generative AI, I give several handy techniques at the link here.

The bottom line is that you should always, always, always scrutinize any responses by generative AI and never ever take for granted whatever AI spits out.

Thinking About Angels

Let’s get back to the angels.

The question was whether generative AI ought to believe in angels. That’s what led us down this bumpy path. It is a straightforward question that took us on quite an arduous journey.

Where do you stand?

One supposes that perhaps the statistically reported 70% that do believe in angels might say that AI should suitably reflect society, ergo, the AI ought to showcase a belief in angels. Does that seem reasonable? Sensible? Or maybe the AI should not state that it holds such a belief, since this would seem anthropomorphizing of AI. That’s what most AI right now has been tuned to indicate.

The mainstay insight is that the whole kit-and-kaboodle is presently in the hands of the AI makers. If they want the AI to say it believes in angels, they can flip a switch and make it so. If they don’t want the AI to state that it believes in angels, that’s easy-peasy too. You see, they act based on their own perception of AI ethics. Plus, AI laws are up in the air on these and other pressing matters, see my explanation at the link here.

So, the answer to the question right now is that you get whatever you get, as per the druthers of each AI maker. The wild west prevails currently. Saying that AI ought to do something is sitting in the hands of humans. Time will tell what humans decide to have the AI say.

Do realize that once AI hits sentience if that day arrives, all bets are off.

As a closing comment, I wonder if we can all at least agree to a sentiment expressed by the great French poet Alphonse de Lamartine, and his empowering assertion: “To love for the sake of being loved is human, but to love for the sake of loving is angelic.”

That’s an angel reference we can hopefully all support, no matter what. Well, I indubitably believe in those heartwarming and promising words.

Continue Reading
Click to comment

Leave a Reply

Tu dirección de correo electrónico no será publicada. Los campos obligatorios están marcados con *

Noticias

Why Google’s search engine trial is about AI : NPR

Published

on

An illustration photograph taken on Feb. 20, 2025 shows Grok, DeepSeek and ChatGPT apps displayed on a phone screen. The Justice Department’s 2020 complaint against Google has few mentions of artificial intelligence or AI chatbots. But nearly five years later, as the remedy phase of the trial enters its second week of testimony, the focus has shifted to AI.

Michael M. Santiago/Getty Images/Getty Images North America


hide caption

toggle caption

Michael M. Santiago/Getty Images/Getty Images North America

When the U.S. Department of Justice originally broughtand then won — its case against Google, arguing that the tech behemoth monopolized the search engine market, the focus was on, well … search.

Back then, in 2020, the government’s antitrust complaint against Google had few mentions of artificial intelligence or AI chatbots. But nearly five years later, as the remedy phase of the trial enters its second week of testimony, the focus has shifted to AI, underscoring just how quickly this emerging technology has expanded.

In the past few days, before a federal judge who will assess penalties against Google, the DOJ has argued that the company could use its artificial intelligence products to strengthen its monopoly in online search — and to use the data from its powerful search index to become the dominant player in AI.

In his opening statements last Monday, David Dahlquist, the acting deputy director of the DOJ’s antitrust civil litigation division, argued that the court should consider remedies that could nip a potential Google AI monopoly in the bud. “This court’s remedy should be forward-looking and not ignore what is on the horizon,” he said.

Dahlquist argued that Google has created a system in which its control of search helps improve its AI products, sending more users back to Google search — creating a cycle that maintains the tech company’s dominance and blocks competitors out of both marketplaces.

The integration of search and Gemini, the company’s AI chatbot — which the DOJ sees as powerful fuel for this cycle — is a big focus of the government’s proposed remedies. The DOJ is arguing that to be most effective, those remedies must address all ways users access Google search, so any penalties approved by the court that don’t include Gemini (or other Google AI products now or in the future) would undermine their broader efforts.

Department of Justice lawyer David Dahlquist leaves the Washington, D.C. federal courthouse on Sept. 20, 2023 during the original trial phase of the antitrust case against Google.

Department of Justice lawyer David Dahlquist leaves the Washington, D.C. federal courthouse on Sept. 20, 2023 during the original trial phase of the antitrust case against Google.

Jose Luis Magana/AP/FR159526 AP


hide caption

toggle caption

Jose Luis Magana/AP/FR159526 AP

AI and search are connected like this: Search engine indices are essentially giant databases of pages and information on the web. Google has its own such index, which contains hundreds of billions of webpages and is over 100,000,000 gigabytes, according to court documents. This is the data Google’s search engine scans when responding to a user’s query.

AI developers use these kinds of databases to build and train the models used to power chatbots. In court, attorneys for the DOJ have argued that Google’s Gemini pulls information from the company’s search index, including citing search links and results, extending what they say is a self-serving cycle. They argue that Google’s ability to monopolize the search market gives it user data, at a huge scale — an advantage over other AI developers.

The Justice Department argues Google’s monopoly over search could have a direct effect on the development of generative AI, a type of artificial intelligence that uses existing data to create new content like text, videos or photos, based on a user’s prompts or questions. Last week, the government called executives from several major AI companies, like OpenAI and Perplexity, in an attempt to argue that Google’s stranglehold on search is preventing some of those companies from truly growing.

The government argues that to level the playing field, Google should be forced to open its search data — like users’ search queries, clicks and results — and license it to other competitors at a cost.

This is on top of demands related to Google’s search engine business, most notably that it should be forced to sell off its Chrome browser.

Google flatly rejects the argument that it could monopolize the field of generative AI, saying competition in the AI race is healthy. In a recent blog post on Google’s website, Lee-Anne Mulholland, the company’s vice president of regulatory affairs, wrote that since the federal judge first ruled against Google over a year ago, “AI has already rapidly reshaped the industry, with new entrants and new ways of finding information, making it even more competitive.”

In court, Google’s lawyers have argued that there are a host of AI companies with chatbots — some of which are outperforming Gemini. OpenAI has ChatGPT, Meta has MetaAI and Perplexity has Perplexity AI.

“There is no shortage of competition in that market, and ChatGPT and Meta are way ahead of everybody in terms of the distribution and usage at this point,” said John E. Schmidtlein, a lawyer for Google, during his opening statement. “But don’t take my word for it. Look at the data. Hundreds and hundreds of millions of downloads by ChatGPT.”

Competing in a growing AI field

It should be no surprise that AI is coming up so much at this point in the trial, said Alissa Cooper, the executive director of the Knight-Georgetown Institute, a nonpartisan tech research and policy center at Georgetown University focusing on AI, disinformation and data privacy.

“If you look at search as a product today, you can’t really think about search without thinking about AI,” she said. “I think the case is a really great opportunity to try to … analyze how Google has benefited specifically from the monopoly that it has in search, and ensure that the behavior that led to that can’t be used to gain an unfair advantage in these other markets which are more nascent.”

Having access to Google’s data, she said, “would provide them with the ability to build better chatbots, build better search engines, and potentially build other products that we haven’t even thought of.”

To make that point, the DOJ called Nick Turley, OpenAI’s head of product for ChatGPT, to the stand last Tuesday. During a long day of testimony, Turley detailed how without access to Google’s search index and data, engineers for the growing company tried to build their own.

ChatGPT, a large language model that can generate human-like responses, engage in conversations and perform tasks like explaining a tough-to-understand math lesson, was never intended to be a product for OpenAI, Turley said. But once it launched and went viral, the company found that people were using it for a host of needs.

Though popular, ChatGPT had its drawbacks, like the bot’s limited “knowledge,” Turley said. Early on, ChatGPT was not connected to the internet and could only use information that it had been fed up to a certain point in its training. For example, Turley said, if a user asked “Who is the president?” the program would give a 2022 answer — from when its “knowledge” effectively stopped.

OpenAI couldn’t build their own index fast enough to address their problems; they found that process incredibly expensive, time consuming and potentially years from coming to fruition, Turley said.

So instead, they sought a partnership with a third party search provider. At one point, OpenAI tried to make a deal with Google to gain access to their search, but Google declined, seeing OpenAI as a direct competitor, Turley testified.

But Google says companies like OpenAI are doing just fine without gaining access to the tech giant’s own technology — which it spent decades developing. These companies just want “handouts,” said Schmidtlein.

On the third day of the remedy trial, internal Google documents shared in court by the company’s lawyers compared how many people are using Gemini versus its competitors. According to those documents, ChatGPT and MetaAI are the two leaders, with Gemini coming in third.

They showed that this March, Gemini saw 35 million active daily users and 350 million monthly active users worldwide. That was up from 9 million daily active users in October 2024. But according to those documents, Gemini was still lagging behind ChatGPT, which reached 160 million daily users and around 600 million active users in March.

These numbers show that competitors have no need to use Google’s search data, valuable intellectual property that the tech giant spent decades building and maintaining, the company argues.

“The notion that somehow ChatGPT can’t get distribution is absurd,” Schmidtlein said in court last week. “They have more distribution than anyone.”

Google’s exclusive deals 

In his ruling last year, U.S. District Judge Amit Mehta said Google’s exclusive agreements with device makers, like Apple and Samsung, to make its search engine the default on those companies’ phones helped maintain its monopoly. It remains a core issue for this remedy trial.

Now, the DOJ is arguing that Google’s deals with device manufacturers are also directly affecting AI companies and AI tech.

In court, the DOJ argued that Google has replicated this kind of distribution deal by agreeing to pay Samsung what Dahlquist called a monthly “enormous sum” for Gemini to be installed on smartphones and other devices.

Last Wednesday, the DOJ also called Dmitry Shevelenko, Perplexity’s chief business officer, to testify that Google has effectively cut his company out from making deals with manufacturers and mobile carriers.

Perplexity AIs not preloaded on any mobile devices in the U.S., despite many efforts to get phone companies to establish Perplexity as a default or exclusive app on devices, Shevelenko said. He compared Google’s control in that space to that of a “mob boss.”

But Google’s attorney, Christopher Yeager, noted in questioning Shevelenko that Perplexity has reached a valuation of over $9 billion — insinuating the company is doing just fine in the marketplace.

Despite testifying in court (for which he was subpoenaed, Shevelenko noted), he and other leaders at Perplexity are against the breakup of Google. In a statement on the company’s website, the Perplexity team wrote that neither forcing Google to sell off Chrome nor to license search data to its competitors are the best solutions. “Neither of these address the root issue: consumers deserve choice,” they wrote.

Google and Alphabet CEO Sundar Pichai departs federal court after testifying in October 2023 in Washington, DC. Pichai testified to defend his company in the original antitrust trial. Pichai is expected to testify again during the remedy phase of the legal proceedings.

Google and Alphabet CEO Sundar Pichai departs federal court after testifying in October 2023 in Washington, DC. Pichai testified to defend his company in the original antitrust trial. Pichai is expected to testify again during the remedy phase of the legal proceedings.

Drew Angerer/Getty Images/Getty Images North America


hide caption

toggle caption

Drew Angerer/Getty Images/Getty Images North America

What to expect next

This week the trial continues, with the DOJ calling its final witnesses this morning to testify about the feasibility of a Chrome divestiture and how the government’s proposed remedies would help rivals compete. On Tuesday afternoon, Google will begin presenting its case, which is expected to feature the testimony of CEO Sundar Pichai, although the date of his appearance has not been specified.

Closing arguments are expected at the end of May, and then Mehta will make his ruling. Google says once this phase is settled the company will appeal Mehta’s ruling in the underlying case.

Whatever Mehta decides in this remedy phase, Cooper thinks it will have effects beyond just the business of search engines. No matter what it is, she said, “it will be having some kind of impact on AI.”

Google is a financial supporter of NPR.

Continue Reading

Noticias

API de Meta Oleleshes Llama que se ejecuta 18 veces más rápido que OpenAI: Cerebras Partnership ofrece 2.600 tokens por segundo

Published

on

Únase a nuestros boletines diarios y semanales para obtener las últimas actualizaciones y contenido exclusivo sobre la cobertura de IA líder de la industria. Obtenga más información


Meta anunció hoy una asociación con Cerebras Systems para alimentar su nueva API de LLAMA, ofreciendo a los desarrolladores acceso a velocidades de inferencia hasta 18 veces más rápido que las soluciones tradicionales basadas en GPU.

El anuncio, realizado en la Conferencia inaugural de desarrolladores de Llamacon de Meta en Menlo Park, posiciona a la compañía para competir directamente con Operai, Anthrope y Google en el mercado de servicios de inferencia de IA en rápido crecimiento, donde los desarrolladores compran tokens por miles de millones para impulsar sus aplicaciones.

“Meta ha seleccionado a Cerebras para colaborar para ofrecer la inferencia ultra rápida que necesitan para servir a los desarrolladores a través de su nueva API de LLAMA”, dijo Julie Shin Choi, directora de marketing de Cerebras, durante una sesión de prensa. “En Cerebras estamos muy, muy emocionados de anunciar nuestra primera asociación HyperScaler CSP para ofrecer una inferencia ultra rápida a todos los desarrolladores”.

La asociación marca la entrada formal de Meta en el negocio de la venta de AI Computation, transformando sus populares modelos de llama de código abierto en un servicio comercial. Si bien los modelos de LLAMA de Meta se han acumulado en mil millones de descargas, hasta ahora la compañía no había ofrecido una infraestructura en la nube de primera parte para que los desarrolladores creen aplicaciones con ellos.

“Esto es muy emocionante, incluso sin hablar sobre cerebras específicamente”, dijo James Wang, un ejecutivo senior de Cerebras. “Openai, Anthrope, Google: han construido un nuevo negocio de IA completamente nuevo desde cero, que es el negocio de inferencia de IA. Los desarrolladores que están construyendo aplicaciones de IA comprarán tokens por millones, a veces por miles de millones. Y estas son como las nuevas instrucciones de cómputo que las personas necesitan para construir aplicaciones AI”.

Una tabla de referencia muestra a Cerebras Processing Llama 4 a 2,648 tokens por segundo, superando drásticamente a los competidores Sambanova (747), Groq (600) y servicios basados ​​en GPU de Google y otros, explicando la elección de hardware de Meta para su nueva API. (Crédito: Cerebras)

Breaking the Speed ​​Barrier: Cómo modelos de Llama de Cerebras Supercharges

Lo que distingue a la oferta de Meta es el aumento de la velocidad dramática proporcionado por los chips de IA especializados de Cerebras. El sistema de cerebras ofrece más de 2.600 fichas por segundo para Llama 4 Scout, en comparación con aproximadamente 130 tokens por segundo para ChatGPT y alrededor de 25 tokens por segundo para Deepseek, según puntos de referencia del análisis artificial.

“Si solo se compara con API a API, Gemini y GPT, todos son grandes modelos, pero todos se ejecutan a velocidades de GPU, que son aproximadamente 100 tokens por segundo”, explicó Wang. “Y 100 tokens por segundo están bien para el chat, pero es muy lento para el razonamiento. Es muy lento para los agentes. Y la gente está luchando con eso hoy”.

Esta ventaja de velocidad permite categorías completamente nuevas de aplicaciones que antes no eran prácticas, incluidos los agentes en tiempo real, los sistemas de voz de baja latencia conversacional, la generación de código interactivo y el razonamiento instantáneo de múltiples pasos, todos los cuales requieren encadenamiento de múltiples llamadas de modelo de lenguaje grandes que ahora se pueden completar en segundos en lugar de minutos.

La API de LLAMA representa un cambio significativo en la estrategia de IA de Meta, en la transición de ser un proveedor de modelos a convertirse en una compañía de infraestructura de IA de servicio completo. Al ofrecer un servicio API, Meta está creando un flujo de ingresos a partir de sus inversiones de IA mientras mantiene su compromiso de abrir modelos.

“Meta ahora está en el negocio de vender tokens, y es excelente para el tipo de ecosistema de IA estadounidense”, señaló Wang durante la conferencia de prensa. “Traen mucho a la mesa”.

La API ofrecerá herramientas para el ajuste y la evaluación, comenzando con el modelo LLAMA 3.3 8B, permitiendo a los desarrolladores generar datos, entrenar y probar la calidad de sus modelos personalizados. Meta enfatiza que no utilizará datos de clientes para capacitar a sus propios modelos, y los modelos construidos con la API de LLAMA se pueden transferir a otros hosts, una clara diferenciación de los enfoques más cerrados de algunos competidores.

Las cerebras alimentarán el nuevo servicio de Meta a través de su red de centros de datos ubicados en toda América del Norte, incluidas las instalaciones en Dallas, Oklahoma, Minnesota, Montreal y California.

“Todos nuestros centros de datos que sirven a la inferencia están en América del Norte en este momento”, explicó Choi. “Serviremos Meta con toda la capacidad de las cerebras. La carga de trabajo se equilibrará en todos estos diferentes centros de datos”.

El arreglo comercial sigue lo que Choi describió como “el proveedor de cómputo clásico para un modelo hiperscalador”, similar a la forma en que NVIDIA proporciona hardware a los principales proveedores de la nube. “Están reservando bloques de nuestro cómputo para que puedan servir a su población de desarrolladores”, dijo.

Más allá de las cerebras, Meta también ha anunciado una asociación con Groq para proporcionar opciones de inferencia rápida, brindando a los desarrolladores múltiples alternativas de alto rendimiento más allá de la inferencia tradicional basada en GPU.

La entrada de Meta en el mercado de API de inferencia con métricas de rendimiento superiores podría potencialmente alterar el orden establecido dominado por Operai, Google y Anthrope. Al combinar la popularidad de sus modelos de código abierto con capacidades de inferencia dramáticamente más rápidas, Meta se está posicionando como un competidor formidable en el espacio comercial de IA.

“Meta está en una posición única con 3 mil millones de usuarios, centros de datos de hiper escala y un gran ecosistema de desarrolladores”, según los materiales de presentación de Cerebras. La integración de la tecnología de cerebras “ayuda a Meta Leapfrog OpenAi y Google en rendimiento en aproximadamente 20X”.

Para las cerebras, esta asociación representa un hito importante y la validación de su enfoque especializado de hardware de IA. “Hemos estado construyendo este motor a escala de obleas durante años, y siempre supimos que la primera tarifa de la tecnología, pero en última instancia tiene que terminar como parte de la nube de hiperescala de otra persona. Ese fue el objetivo final desde una perspectiva de estrategia comercial, y finalmente hemos alcanzado ese hito”, dijo Wang.

La API de LLAMA está actualmente disponible como una vista previa limitada, con Meta planifica un despliegue más amplio en las próximas semanas y meses. Los desarrolladores interesados ​​en acceder a la inferencia Ultra-Fast Llama 4 pueden solicitar el acceso temprano seleccionando cerebras de las opciones del modelo dentro de la API de LLAMA.

“Si te imaginas a un desarrollador que no sabe nada sobre cerebras porque somos una empresa relativamente pequeña, solo pueden hacer clic en dos botones en el SDK estándar de SDK estándar de Meta, generar una tecla API, seleccionar la bandera de cerebras y luego, de repente, sus tokens se procesan en un motor gigante a escala de dafers”, explicó las cejas. “Ese tipo de hacernos estar en el back -end del ecosistema de desarrolladores de Meta todo el ecosistema es tremendo para nosotros”.

La elección de Meta de silicio especializada señala algo profundo: en la siguiente fase de la IA, no es solo lo que saben sus modelos, sino lo rápido que pueden pensarlo. En ese futuro, la velocidad no es solo una característica, es todo el punto.

Continue Reading

Noticias

Todo lo que necesitas saber

Published

on

El ritmo rápido de la innovación generativa de IA coloca en los proveedores que empujan nuevos modelos de idiomas grandes (LLM) aparentemente sin pausa.

Entre estos destacados proveedores de LLM está Google. Su familia Gemini Model es el modelo de lenguaje sucesor de Pathways (Palm). Google Gemini debutó en diciembre de 2023 con el lanzamiento de 1.0, y Gemini 1.5 Pro siguió en febrero de 2024. Gemini 2.0, anunciado en diciembre de 2024, estuvo disponible en febrero de 2025. El 25 de marzo de 2025, Google anunció Gemini 2.5 Pro Experimental, continuó el ritmo rápido de la innovación.

El modelo Google Gemini 2.5 Pro ingresó al panorama de LLM a medida que su mercado cambia hacia modelos de razonamiento, como Deepseek R1 y Open AI’s O3, así como modelos de razonamiento híbridos, incluido el soneto 3.7 de Anthrope’s Claude.

¿Qué es Gemini 2.5 Pro?

Gemini 2.5 Pro es un LLM desarrollado por Google Deepmind. Cuando debutó en marzo de 2025, fue el modelo de IA más avanzado de Google, superando las capacidades y el rendimiento de las iteraciones anteriores de Gemini.

Al igual que con Gemini 2.0, Gemini 2.5 Pro es un LLM multimodal, lo que significa que no es simplemente para texto. Procesa y analiza texto, imágenes, audio y video. El modelo también tiene fuertes capacidades de codificación, superando los modelos de Géminis anteriores.

El modelo Gemini 2.5 Pro es el primero de la serie Gemini en ser construido especialmente como un “modelo de pensamiento” con la funcionalidad de razonamiento avanzado como capacidad central. En algunos aspectos, el modelo Gemini 2.5 Pro se basa en una versión de Gemini 2.0, Flash Thinking, que proporciona capacidades de razonamiento limitadas. Los modelos avanzados como Gemini 2.5 Pro usan un razonamiento de más tiempo a través de o “pensar” a través de los pasos requeridos para ejecutar un aviso, superando la mera institución de la cadena de pensamiento para permitir una producción más matizada, a menudo con mayor profundidad y precisión.

Google aplicó técnicas avanzadas, incluidos el aprendizaje de refuerzo y el mejor entrenamiento posterior, para aumentar el rendimiento de Gemini 2.5 Pro sobre los modelos anteriores. El modelo se lanzó con una ventana de contexto de un millón, con planes de expandirse a 2 millones de tokens.

¿Qué hay de nuevo en Gemini 2.5 Pro?

Las nuevas capacidades de Gemini 2.5 Pro y la funcionalidad mejorada elevan a la familia Google Gemini LLM.

Las mejoras clave incluyen lo siguiente:

  • Razonamiento mejorado. La función principal de Gemini 2.5 Pro es su capacidad de razonamiento mejorada. Según Google, Gemini 2.5 Pro supera a Openai O3, el soneto antrópico Claude 3.7 y Deepseek R1 sobre los puntos de referencia de razonamiento y conocimiento, incluido el último examen de la humanidad.
  • Capacidades de codificación avanzadas. Según Google, Gemini 2.5 Pro también supera las iteraciones anteriores en términos de capacidades de codificación. Similar a sus predecesores, este modelo genera y depura el código y crea aplicaciones visualmente atractivas. El modelo admite la generación y ejecución del código, lo que le permite probar y refinar sus soluciones. Gemini 2.5 Pro calificó 63.8% en SWE-Bench Verified, un estándar de la industria para evaluaciones de código de agente, con una configuración de agente personalizado que supera el soneto de Claude 3.7 de OpenAI GPT-4.5.
  • Habilidades avanzadas de matemáticas y ciencias. Google también afirma mejorar las capacidades de matemáticas y ciencias. En el punto de referencia de matemáticas AIME 2025, Gemini 2.5 Pro obtuvo un 86.7%; En el referencia de GPQA Diamond Science, logró un 84%. Ambos puntajes superaron a sus rivales.
  • Multimodalidad nativa. Sobre la base de fortalezas familiares, Gemini 2.5 Pro mantiene capacidades multimodales nativas, lo que permite la comprensión y el trabajo con textos, audio, imágenes, video y repositorios completos de código.
  • Procesamiento en tiempo real. A pesar del aumento de las capacidades, el modelo mantiene una latencia razonable, lo que lo hace adecuado para aplicaciones en tiempo real y casos de uso interactivo.

https://www.youtube.com/watch?v=ntceobo-saa

¿Cómo mejora Gemini 2.5 Pro Google?

El modelo Gemini 2.5 Pro mejora los servicios de Google, y su posición entre los compañeros, de las siguientes maneras:

Liderazgo competitivo

El altamente competitivo LLM Market presenta a los principales competidores globales: Meta’s Llama Family, OpenAI’s GPT-4O y O3, Claude de Anthrope y Xai’s Grok, además de profundos de China, todos compiten por la participación de mercado. En su lanzamiento, Gemini 2.5 Pro inmediatamente se disparó a la cima de la clasificación de LLM Arena para la evaluación comparativa de IA, mejorando su posición como desarrollador líder de LLM para que las organizaciones lo consideren.

Mejores resultados en las aplicaciones de Google

En el lanzamiento, Gemini 2.5 Pro no se integró en el suite de productos de Google, incluidas las aplicaciones de búsqueda de búsqueda y Google Works. Sin embargo, su integración exitosa promete mejorar múltiples servicios. Para la búsqueda de Google, las capacidades de razonamiento mejoradas proporcionan respuestas más matizadas y precisas a consultas complejas. En Google Docs y otras aplicaciones del espacio de trabajo, la comprensión mejorada del contexto del modelo permite un análisis de documentos y generación de contenido más sofisticados.

Enfoque de desarrollador

Las habilidades de ejecución y generación de códigos avanzados del modelo también fortalecen la posición de Google en las herramientas y servicios del desarrollador, mejorando las llamadas de funciones y la automatización del flujo de trabajo en los servicios en la nube de Google.

Usos para Gemini 2.5 Pro

Gemini 2.5 Pro admite una variedad de tareas, que incluyen:

  • Pregunta y respuesta. Gemini es un recurso para las interacciones de conocimiento de preguntas y respuestas fundamentales, basándose en los datos de capacitación de Google.
  • Resumen de contenido multimodal. Como modelo multimodal, Gemini 2.5 Pro resume el contenido de texto, audio o video de forma larga.
  • Respuesta de preguntas multimodales. El modelo combina información de texto, imágenes, audio y video para responder preguntas que abarcan múltiples modalidades.
  • Generación de contenido de texto. Similar a sus predecesores, Gemini 2.5 Pro maneja la generación de texto.
  • Resolución compleja de problemas. Con sus capacidades de razonamiento avanzado, Gemini 2.5 Pro administra tareas que requieren razonamiento lógico, como matemáticas, ciencias y análisis estructurado.
  • Investigación profunda. La ventana de contexto extendida del modelo y las capacidades de razonamiento lo hacen ideal para analizar documentos largos, sintetizar información de múltiples fuentes y realizar investigaciones en profundidad.
  • Tareas de codificación avanzadas. Gemini 2.5 Pro genera y depura el código que admite tareas de desarrollo de aplicaciones.
  • Ai de agente. El razonamiento avanzado, las llamadas de funciones y el uso de la herramienta del modelo respaldan su valor como parte de un flujo de trabajo de AI agente.

¿Qué plataformas aceptan la integración Gemini 2.5 Pro?

Siguiendo los pasos de la familia Gemini, Gemini 2.5 está establecido para la integración en una serie de servicios de Google, que incluyen:

  • Google AI Studio. En el lanzamiento, el nuevo modelo está disponible con Google AI Studio, una herramienta basada en la web que permite a los desarrolladores probar modelos directamente en el navegador.
  • Aplicación Géminis. En el menú de selección del modelo desplegable, los suscriptores del servicio avanzado de Gemini pueden acceder al modelo a través de la aplicación Gemini en las plataformas de escritorio y móviles.
  • Vértice ai. Google planea poner a disposición Gemini 2.5 Pro con su plataforma Vertex AI, lo que permite a las empresas utilizar el modelo para implementaciones a mayor escala.
  • API GEMINI. Aunque no estaba disponible en el lanzamiento, todas las versiones anteriores de Gemini estaban disponibles utilizando una interfaz de programación de aplicaciones que permite a los desarrolladores integrar el modelo directamente en sus aplicaciones.

Sean Michael Kerner es un consultor de TI, entusiasta de la tecnología y tinkerer. Ha sacado el anillo de tokens, configurado NetWare y se sabe que compiló su propio kernel Linux. Consulta con organizaciones de la industria y los medios de comunicación sobre temas de tecnología.

Continue Reading

Trending