Connect with us

Noticias

Google Gemini: Everything you need to know about the generative AI models

Published

on

Google’s trying to make waves with Gemini, its flagship suite of generative AI models, apps, and services. But what’s Gemini? How can you use it? And how does it stack up to other generative AI tools such as OpenAI’s ChatGPT, Meta’s Llama, and Microsoft’s Copilot?

To make it easier to keep up with the latest Gemini developments, we’ve put together this handy guide, which we’ll keep updated as new Gemini models, features, and news about Google’s plans for Gemini are released.

What is Gemini?

Gemini is Google’s long-promised, next-gen generative AI model family. Developed by Google’s AI research labs DeepMind and Google Research, it comes in four flavors:

  • Gemini Ultra, a very large model.
  • Gemini Pro, a large model – though smaller than Ultra. The latest version, Gemini 2.0 Pro Experimental, is Google’s flagship.
  • Gemini Flash, a speedier, “distilled” version of Pro. It also comes in a slightly smaller and faster version, called Gemini Flash-Lite, and a version with reasoning capabilities, called Gemini Flash Thinking Experimental.
  • Gemini Nano, two small models: Nano-1 and the slightly more capable Nano-2, which is meant to run offline

All Gemini models were trained to be natively multimodal — that is, able to work with and analyze more than just text. Google says they were pre-trained and fine-tuned on a variety of public, proprietary, and licensed audio, images, and videos; a set of codebases; and text in different languages.

This sets Gemini apart from models such as Google’s own LaMDA, which was trained exclusively on text data. LaMDA can’t understand or generate anything beyond text (e.g., essays, emails, and so on), but that isn’t necessarily the case with Gemini models.

We’ll note here that the ethics and legality of training models on public data, in some cases without the data owners’ knowledge or consent, are murky. Google has an AI indemnification policy to shield certain Google Cloud customers from lawsuits should they face them, but this policy contains carve-outs. Proceed with caution — particularly if you’re intending on using Gemini commercially.

What’s the difference between the Gemini apps and Gemini models?

Gemini is separate and distinct from the Gemini apps on the web and mobile (formerly Bard).

The Gemini apps are clients that connect to various Gemini models and layer a chatbot-like interface on top. Think of them as front ends for Google’s generative AI, analogous to ChatGPT and Anthropic’s Claude family of apps.

Image Credits:Google

Gemini on the web lives here. On Android, the Gemini app replaces the existing Google Assistant app. And on iOS, the Google and Google Search apps serve as that platform’s Gemini clients.

On Android, it also recently became possible to bring up the Gemini overlay on top of any app to ask questions about what’s on the screen (e.g., a YouTube video). Just press and hold a supported smartphone’s power button or say, “Hey Google”; you’ll see the overlay pop up.

Gemini apps can accept images as well as voice commands and text — including files like PDFs and soon videos, either uploaded or imported from Google Drive — and generate images. As you’d expect, conversations with Gemini apps on mobile carry over to Gemini on the web and vice versa if you’re signed in to the same Google Account in both places.

Gemini Advanced

The Gemini apps aren’t the only means of recruiting Gemini models’ assistance with tasks. Slowly but surely, Gemini-imbued features are making their way into staple Google apps and services like Gmail and Google Docs.

To take advantage of most of these, you’ll need the Google One AI Premium Plan. Technically a part of Google One, the AI Premium Plan costs $20 and provides access to Gemini in Google Workspace apps like Docs, Maps, Slides, Sheets, Drive, and Meet. It also enables what Google calls Gemini Advanced, which brings the company’s more sophisticated Gemini models to the Gemini apps.

Gemini Advanced users get extras here and there, too, like priority access to new features, the ability to run and edit Python code directly in Gemini, and a larger “context window.” Gemini Advanced can remember the content of — and reason across — roughly 750,000 words in a conversation (or 1,500 pages of documents). That’s compared to the 24,000 words (or 48 pages) the vanilla Gemini app can handle.

Screenshot of a Google Gemini commercial
Image Credits:Google

Gemini Advanced also gives users access to Google’s Deep Research feature, which uses “advanced reasoning” and “long context capabilities” to generate research briefs. After you prompt the chatbot, it creates a multi-step research plan, asks you to approve it, and then Gemini takes a few minutes to search the web and generate an extensive report based on your query. It’s meant to answer more complex questions such as, “Can you help me redesign my kitchen?”

Google also offers Gemini Advanced users a memory feature, that allows the chatbot to use your old conversations with Gemini as context for your current conversation. Gemini Advanced users also get increased usage for NotebookLM, the company’s product that turns PDFs into AI-generated podcasts.

Gemini Advanced users also get access to Google’s experimental version of Gemini 2.0 Pro, the company’s flagship model that’s optimized for difficult coding and math problems.

Another Gemini Advanced exclusive is trip planning in Google Search, which creates custom travel itineraries from prompts. Taking into account things like flight times (from emails in a user’s Gmail inbox), meal preferences, and information about local attractions (from Google Search and Maps data), as well as the distances between those attractions, Gemini will generate an itinerary that updates automatically to reflect any changes. 

Gemini across Google services is also available to corporate customers through two plans, Gemini Business (an add-on for Google Workspace) and Gemini Enterprise. Gemini Business costs as low as $6 per user per month, while Gemini Enterprise — which adds meeting note-taking and translated captions as well as document classification and labeling — is generally more expensive, but is priced based on a business’s needs. (Both plans require an annual commitment.)

In Gmail, Gemini lives in a side panel that can write emails and summarize message threads. You’ll find the same panel in Docs, where it helps you write and refine your content and brainstorm new ideas. Gemini in Slides generates slides and custom images. And Gemini in Google Sheets tracks and organizes data, creating tables and formulas.

Google’s AI chatbot recently came to Maps, where Gemini can summarize reviews about coffee shops or offer recommendations about how to spend a day visiting a foreign city.

Gemini’s reach extends to Drive as well, where it can summarize files and folders and give quick facts about a project. In Meet, meanwhile, Gemini translates captions into additional languages.

Gemini in Gmail
Image Credits:Google

Gemini recently came to Google’s Chrome browser in the form of an AI writing tool. You can use it to write something completely new or rewrite existing text; Google says it’ll consider the web page you’re on to make recommendations.

Elsewhere, you’ll find hints of Gemini in Google’s database products, cloud security tools, and app development platforms (including Firebase and Project IDX), as well as in apps like Google Photos (where Gemini handles natural language search queries), YouTube (where it helps brainstorm video ideas), and the NotebookLM note-taking assistant.

Code Assist (formerly Duet AI for Developers), Google’s suite of AI-powered assistance tools for code completion and generation, is offloading heavy computational lifting to Gemini. So are Google’s security products underpinned by Gemini, like Gemini in Threat Intelligence, which can analyze large portions of potentially malicious code and let users perform natural language searches for ongoing threats or indicators of compromise.

Gemini extensions and Gems

Announced at Google I/O 2024, Gemini Advanced users can create Gems, custom chatbots powered by Gemini models. Gems can be generated from natural language descriptions — for example, “You’re my running coach. Give me a daily running plan” — and shared with others or kept private.

Gems are available on desktop and mobile in 150 countries and most languages. Eventually, they’ll be able to tap an expanded set of integrations with Google services, including Google Calendar, Tasks, Keep, and YouTube Music, to complete custom tasks.

Gemini Gems
Image Credits:Google

Speaking of integrations, the Gemini apps on the web and mobile can tap into Google services via what Google calls “Gemini extensions.” Gemini today integrates with Google Drive, Gmail, and YouTube to respond to queries such as “Could you summarize my last three emails?” Later this year, Gemini will be able to take additional actions with Google Calendar, Keep, Tasks, YouTube Music and Utilities, the Android-exclusive apps that control on-device features like timers and alarms, media controls, the flashlight, volume, Wi-Fi, Bluetooth, and so on.

Gemini Live in-depth voice chats

An experience called Gemini Live allows users to have “in-depth” voice chats with Gemini. It’s available in the Gemini apps on mobile and the Pixel Buds Pro 2, where it can be accessed even when your phone’s locked.

With Gemini Live enabled, you can interrupt Gemini while the chatbot’s speaking (in one of several new voices) to ask a clarifying question, and it’ll adapt to your speech patterns in real time. At some point, Gemini is supposed to gain visual understanding, allowing it to see and respond to your surroundings, either via photos or video captured by your smartphones’ cameras.

Gemini Live
Image Credits:Google

Live is also designed to serve as a virtual coach of sorts, helping you rehearse for events, brainstorm ideas, and so on. For instance, Live can suggest which skills to highlight in an upcoming job or internship interview, and it can give public speaking advice.

You can read our review of Gemini Live here. Spoiler alert: We think the feature has a ways to go before it’s super useful — but it’s early days, admittedly.

Image generation via Imagen 3

Gemini users can generate artwork and images using Google’s built-in Imagen 3 model.

Google says that Imagen 3 can more accurately understand the text prompts that it translates into images versus its predecessor, Imagen 2, and is more “creative and detailed” in its generations. In addition, the model produces fewer artifacts and visual errors (at least according to Google), and is the best Imagen model yet for rendering text.

Google Imagen 3
A sample from Imagen 3.Image Credits:Google

Back in February 2024, Google was forced to pause Gemini’s ability to generate images of people after users complained of historical inaccuracies. But in August, the company reintroduced people generation for certain users, specifically English-language users signed up for one of Google’s paid Gemini plans (e.g., Gemini Advanced) as part of a pilot program.

Gemini for teens

In June, Google introduced a teen-focused Gemini experience, allowing students to sign up via their Google Workspace for Education school accounts.

The teen-focused Gemini has “additional policies and safeguards,” including a tailored onboarding process and an “AI literacy guide” to (as Google phrases it) “help teens use AI responsibly.” Otherwise, it’s nearly identical to the standard Gemini experience, down to the “double check” feature that looks across the web to see if Gemini’s responses are accurate.

Gemini in smart home devices

A growing number of Google-made devices tap Gemini for enhanced functionality, from the Google TV Streamer to the Pixel 9 and 9 Pro to the newest Nest Learning Thermostat.

On the Google TV Streamer, Gemini uses your preferences to curate content suggestions across your subscriptions and summarize reviews and even whole seasons of TV.

Google TV Streamer set up
Image Credits:Google

On the latest Nest thermostat (as well as Nest speakers, cameras, and smart displays), Gemini will soon bolster Google Assistant’s conversational and analytic capabilities.

Subscribers to Google’s Nest Aware plan later this year will get a preview of new Gemini-powered experiences like AI descriptions for Nest camera footage, natural language video search and recommended automations. Nest cameras will understand what’s happening in real-time video feeds (e.g., when a dog’s digging in the garden), while the companion Google Home app will surface videos and create device automations given a description (e.g., “Did the kids leave their bikes in the driveway?,” “Have my Nest thermostat turn on the heating when I get home from work every Tuesday”).

Google Gemini in smart home
Gemini will soon be able to summarize security camera footage from Nest devices.Image Credits:Google

Also later this year, Google Assistant will get a few upgrades on Nest-branded and other smart home devices to make conversations feel more natural. Improved voices are on the way, in addition to the ability to ask follow-up questions and “[more] easily go back and forth.”

What can the Gemini models do?

Because Gemini models are multimodal, they can perform a range of multimodal tasks, from transcribing speech to captioning images and videos in real time. Many of these capabilities have reached the product stage (as alluded to in the previous section), and Google is promising much more in the not-too-distant future.

Of course, it’s a bit hard to take the company at its word. Google seriously underdelivered with the original Bard launch. More recently, it ruffled feathers with a video purporting to show Gemini’s capabilities that was more or less aspirational — not live.

Also, Google offers no fix for some of the underlying problems with generative AI tech today, like its encoded biases and tendency to make things up (i.e., hallucinate). Neither do its rivals, but it’s something to keep in mind when considering using or paying for Gemini.

Assuming for the purposes of this article that Google is being truthful with its recent claims, here’s what the different tiers of Gemini can do now and what they’ll be able to do once they reach their full potential:

What you can do with Gemini Ultra

Google says that Gemini Ultra — thanks to its multimodality — can be used to help with things like physics homework, solving problems step-by-step on a worksheet, and pointing out possible mistakes in already filled-in answers.

However, we haven’t seen much of Gemini Ultra in recent months. The model does not appear in the Gemini app, and isn’t listed on Google Gemini’s API pricing page. However, that doesn’t mean Google won’t bring Gemini Ultra back to the forefront of its offerings in the future.

Ultra can also be applied to tasks such as identifying scientific papers relevant to a problem, Google says. The model can extract information from several papers, for instance, and update a chart from one by generating the formulas necessary to re-create the chart with more timely data.

Gemini Ultra technically supports image generation. But that capability hasn’t made its way into the productized version of the model yet — perhaps because the mechanism is more complex than how apps such as ChatGPT generate images. Rather than feed prompts to an image generator (like DALL-E 3, in ChatGPT’s case), Gemini outputs images “natively,” without an intermediary step.

Ultra is available as an API through Vertex AI, Google’s fully managed AI dev platform, and AI Studio, Google’s web-based tool for app and platform developers.

Gemini Pro’s capabilities

Google says that its latest Pro model, Gemini 2.0 Pro, is its best model yet for coding performance and complex prompts. It’s currently available as an experimental version, meaning it can have unexpected issues.

Gemini 2.0 Pro outperforms its predecessor, Gemini 1.5 Pro, in benchmarks measuring coding, reasoning, math, and factual accuracy. The model can take in up to 1.4 million words, two hours of video, or 22 hours of audio and can reason across or answer questions about that data (more or less).

However, Gemini 1.5 Pro still powers Google’s Deep Research feature.

Gemini 2.0 Pro works alongside a feature called code execution, released in June alongside Gemini 1.5 Pro, which aims to reduce bugs in code that the model generates by iteratively refining that code over several steps. (Code execution also supports Gemini Flash.)

Within Vertex AI, developers can customize Gemini Pro to specific contexts and use cases via a fine-tuning or “grounding” process. For example, Pro (along with other Gemini models) can be instructed to use data from third-party providers like Moody’s, Thomson Reuters, ZoomInfo and MSCI, or source information from corporate datasets or Google Search instead of its wider knowledge bank. Gemini Pro can also be connected to external, third-party APIs to perform particular actions, like automating a back-office workflow.

AI Studio offers templates for creating structured chat prompts with Pro. Developers can control the model’s creative range and provide examples to give tone and style instructions — and also tune Pro’s safety settings.

Vertex AI Agent Builder lets people build Gemini-powered “agents” within Vertex AI. For example, a company could create an agent that analyzes previous marketing campaigns to understand a brand style and then apply that knowledge to help generate new ideas consistent with the style. 

Gemini Flash is lighter but packs a punch

Google calls Gemini 2.0 Flash its AI model for the agentic era. The model can natively generate images and audio, in addition to text, and can use tools like Google Search and interact with external APIs.

The 2.0 Flash model is faster than Gemini’s previous generation of models and even outperforms some of the larger Gemini 1.5 models on benchmarks measuring coding and image analysis. You can try Gemini 2.0 Flash in the Gemini web or mobile app, and through Google’s AI developer platforms.

In December, Google released a “thinking” version of Gemini 2.0 Flash that’s capable of “reasoning,” in which the AI model takes a few seconds to work backwards through a problem before it gives an answer.

In February, Google made Gemini 2.0 Flash thinking available in the Gemini app. The same month, Google also released a smaller version called Gemini 2.0 Flash-Lite. The company says this model outperforms its Gemini 1.5 Flash model, but runs at the same price and speed.

An offshoot of Gemini Pro that’s small and efficient, built for narrow, high-frequency generative AI workloads, Flash is multimodal like Gemini Pro, meaning it can analyze audio, video, images, and text (but it can only generate text). Google says that Flash is particularly well-suited for tasks like summarization and chat apps, plus image and video captioning and data extraction from long documents and tables.

Devs using Flash and Pro can optionally leverage context caching, which lets them store large amounts of information (e.g., a knowledge base or database of research papers) in a cache that Gemini models can quickly and relatively cheaply access. Context caching is an additional fee on top of other Gemini model usage fees, however.

Gemini Nano can run on your phone

Gemini Nano is a much smaller version of the Gemini Pro and Ultra models, and it’s efficient enough to run directly on (some) devices instead of sending the task to a server somewhere. So far, Nano powers a couple of features on the Pixel 8 Pro, Pixel 8, Pixel 9 Pro, Pixel 9 and Samsung Galaxy S24, including Summarize in Recorder and Smart Reply in Gboard.

The Recorder app, which lets users push a button to record and transcribe audio, includes a Gemini-powered summary of recorded conversations, interviews, presentations, and other audio snippets. Users get summaries even if they don’t have a signal or Wi-Fi connection — and in a nod to privacy, no data leaves their phone in process.

Image Credits:Google

Nano is also in Gboard, Google’s keyboard replacement. There, it powers a feature called Smart Reply, which helps to suggest the next thing you’ll want to say when having a conversation in a messaging app such as WhatsApp.

In the Google Messages app on supported devices, Nano drives Magic Compose, which can craft messages in styles like “excited,” “formal,” and “lyrical.”

Google says that a future version of Android will tap Nano to alert users to potential scams during calls. The new weather app on Pixel phones uses Gemini Nano to generate tailored weather reports. And TalkBack, Google’s accessibility service, employs Nano to create aural descriptions of objects for low-vision and blind users.

How much do the Gemini models cost?

Gemini 1.5 Pro, 1.5 Flash, 2.0 Flash, and 2.0 Flash-Lite are available through Google’s Gemini API for building apps and services — all with free options. But the free options impose usage limits and leave out certain features, like context caching and batching.

Gemini models are otherwise pay-as-you-go. Here’s the base pricing — not including add-ons like context caching — as of September 2024:

  • Gemini 1.5 Pro: $1.25 per 1 million input tokens (for prompts up to 128K tokens) or $2.50 per 1 million input tokens (for prompts longer than 128K tokens); $5 per 1 million output tokens (for prompts up to 128K tokens) or $10 per 1 million output tokens (for prompts longer than 128K tokens)
  • Gemini 1.5 Flash: 7.5 cents per 1 million input tokens (for prompts up to 128K tokens), 15 cents per 1 million input tokens (for prompts longer than 128K tokens), 30 cents per 1 million output tokens (for prompts up to 128K tokens), 60 cents per 1 million output tokens (for prompts longer than 128K tokens)
  • Gemini 2.0 Flash: 10 cents per 1 million input tokens, 40 cents per 1 million output tokens. For audio specifically, it costs 70 center per 1 million input tokens, and also 40 centers per 1 million output tokens.
  • Gemini 2.0 Flash-Lite: 7.5 cents per 1 million input tokens, 30 cents per 1 million output tokens.

Tokens are subdivided bits of raw data, like the syllables “fan,” “tas,” and “tic” in the word “fantastic”; 1 million tokens is equivalent to about 700,000 words. Input refers to tokens fed into the model, while output refers to tokens that the model generates.

2.0 Pro pricing has yet to be announced, and Nano is still in early access.

What’s the latest on Project Astra?

Project Astra is Google DeepMind’s effort to create AI-powered apps and “agents” for real-time, multimodal understanding. In demos, Google has shown how the AI model can simultaneously process live video and audio. Google released an app version of Project Astra to a small number of trusted testers in December but has no plans for a broader release right now.

The company would like to put Project Astra in a pair of smart glasses. Google also gave a prototype of some glasses with Project Astra and augmented reality capabilities to a few trusted testers in December. However, there’s not a clear product at this time, and it’s unclear when Google would actually release something like this.

Project Astra is still just that, a project, and not a product. However, the demos of Astra reveal what Google would like its AI products to do in the future.

Is Gemini coming to the iPhone?

It might. 

Apple has said that it’s in talks to put Gemini and other third-party models to use for a number of features in its Apple Intelligence suite. Following a keynote presentation at WWDC 2024, Apple SVP Craig Federighi confirmed plans to work with models, including Gemini, but he didn’t divulge any additional details.

This post was originally published February 16, 2024, and is updated regularly.

Continue Reading
Click to comment

Leave a Reply

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

Noticias

Lo que significa para el futuro chatgpt

Published

on

Después de anunciar la gran reestructuración de OpenAI con fines de lucro que todos esperaban, OpenAi acaba de presentar una nueva iniciativa de IA que podría hacer de ChatGPT la opción de IA predeterminada para los países democráticos y darle al chatbot una ventaja importante sobre los competidores. Operai lo llama “OpenAi para países”, esencialmente la versión internacional de Stargate.

Project Stargate es el plan de OpenAI para invertir $ 500 mil millones en infraestructura de IA en los próximos años, con ayuda y apoyo del gobierno de los Estados Unidos. Operai está trabajando con Microsoft, SoftBank y otros en el proyecto. Stargate le dará a OpenAI el poder de cómputo masivo que necesita para el futuro de ChatGPT.

Operai quiere replicar ese tipo de asociación con otros gobiernos de todo el mundo que están interesados ​​en sus propios proyectos similares a Stargate. La iniciativa Operai for Country se lanzará con 10 países no identificados, donde OpenAI construirá infraestructura ChatGPT adaptada a las necesidades de cada mercado, con la bendición del gobierno de los Estados Unidos.

Operai anunció el Proyecto Operai for Country en una publicación de blog, presentándolo como una forma de hacer que la IA segura y democrática sea el valor predeterminado, en lugar de la IA construida por regímenes autoritarios:

Queremos ayudar a estos países, y en el proceso, difundir la IA democrática, lo que significa el desarrollo, el uso y el despliegue de IA que protege e incorpora principios democráticos de larga data. Ejemplos de esto incluyen la libertad para que las personas elijan cómo trabajan y dirigen la IA, la prevención del uso gubernamental de la IA para acumular el control y un mercado libre que garantiza la libre competencia. Todas estas cosas contribuyen a una amplia distribución de los beneficios de la IA, desalientan la concentración de poder y ayudan a avanzar en nuestra misión. Del mismo modo, creemos que asociarse estrechamente con el gobierno de los Estados Unidos es la mejor manera de avanzar en la IA democrática.

Como usuario de ChatGPT desde hace mucho tiempo, estoy de acuerdo con todo eso. No usaría software de IA como Deepseek, sin importar cuán avanzado, porque incluye barandillas de censura alineadas con los valores del gobierno chino. También envía todos sus datos a China.

De hecho, si Deepseek no se hubiera convertido en una alternativa tan grande a ChatGPT, no estoy seguro de que Operai hubiera avanzado con esta iniciativa tan fácilmente. Deepseek ofrece un fuerte rival de chatgpt que es de uso gratuito. Puede instalar la versión de código abierto en su computadora para evitar la censura y el intercambio de datos con China. Pero aún se origina en un país no democrático, y eso es parte de lo que impulsa la campaña demócrata de IA de OpenAI.

Dejando a un lado la política, OpenAi para países será una asociación comercial entre Operai e internacionales. Estos acuerdos implicarán la construcción de infraestructura de IA localmente, por lo que los datos de ChatGPT de los usuarios en esas regiones pueden almacenarse dentro de sus fronteras.

Operai también creará versiones personalizadas de CHATGPT para los ciudadanos de cada país. No está claro si estas versiones coincidirán con el CHATGPT disponible internacionalmente, pero esperaría que lo hagan:

Esto ayudará a ofrecer una mejor atención médica y educación, servicios públicos más eficientes y más. Esto será AI de, por y para las necesidades de cada país en particular, localizado en su idioma y para su cultura y respetar futuros estándares globales.

Operai también dice que continuará construyendo modelos de IA seguros que respeten los valores democráticos y los derechos humanos. OpenAi para países incluirá recaudar fondos nacionales de inicio para apoyar “ecosistemas nacionales de IA saludables”.

Todo esto sucedería a través de “colaboraciones de infraestructura formalizadas y en coordinación con el gobierno de los Estados Unidos”.

Todo eso suena muy bien en teoría. En la práctica, podría darle a OpenAi un nivel de poder e influencia que hace que sea casi imposible competir.

Una de las principales prioridades de OpenAI en este momento es asegurar la energía de cálculo y la energía necesarias para desarrollar AGI. Eso requiere capital, por lo que está creando un brazo con fines de lucro. Operai también quiere monetizar el chatgpt. No hay nada de malo en eso.

OpenAi para países ayuda en todos los frentes. La financiación del gobierno para la infraestructura permitirá que OpenAi siga construyendo mejores modelos de chatgpt en su camino hacia una IA más avanzada.

Pero las asociaciones también brindarán ventajas significativas a OpenAI y posiciones dominantes en el mercado, potencialmente sofocando la competencia en el espacio de IA.

Si un gobierno democrático invierte en infraestructura similar a Stargate, ChatGPT podría convertirse en la IA predeterminada en industrias como la atención médica y la educación. Ese gobierno querrá ver un retorno de su inversión en AI entregando beneficios inmediatos a sus ciudadanos.

A su vez, esos ciudadanos se acostumbrarán a confiar en ChatGPT para tareas personales de IA, especialmente si también es lo que usan en el trabajo.

Si bien Operai será un socio, también se beneficiará de las instituciones gubernamentales que utilizan modelos CHATGPT personalizados y de personas que pueden suscribirse a versiones pagas de la herramienta.

La IA democrática es una base sólida para la iniciativa Operai para los países, pero también podría llevar a OpenAi a convertirse en un monopolio de IA. Hemos visto un dominio similar antes con compañías como Microsoft y Google, y eso sucedió sin apoyo del gobierno. Un monopolio de IA sería mucho más grave que uno que involucre a Internet Explorer o Google Search.

Digo eso como suscriptor de chatgpt más que no tiene planes de irse. Pero tampoco quiero que ninguna compañía de IA domine el espacio. Los países que exploran la infraestructura de estilo Stargate pueden querer apoyar el desarrollo local de IA antes de recurrir a OpenAI para países.

El ejemplo más cercano que tenemos en este momento es una asociación multimillonaria entre Operai y SoftBank en Japón. Los dos crearon una empresa conjunta local llamada SB OpenAi Japón para administrar productos de IA con alimentación de chatgpt que SoftBank se extenderá entre sus negocios. Ese es un acuerdo privado, pero puede ofrecer una idea de cómo puede operar OpenAi para los países.

Continue Reading

Noticias

Rampant AI Cheating Is Ruining Education Alarmingly Fast

Published

on

Illustration: New York Magazine

Chungin “Roy” Lee stepped onto Columbia University’s campus this past fall and, by his own admission, proceeded to use generative artificial intelligence to cheat on nearly every assignment. As a computer-science major, he depended on AI for his introductory programming classes: “I’d just dump the prompt into ChatGPT and hand in whatever it spat out.” By his rough math, AI wrote 80 percent of every essay he turned in. “At the end, I’d put on the finishing touches. I’d just insert 20 percent of my humanity, my voice, into it,” Lee told me recently.

Lee was born in South Korea and grew up outside Atlanta, where his parents run a college-prep consulting business. He said he was admitted to Harvard early in his senior year of high school, but the university rescinded its offer after he was suspended for sneaking out during an overnight field trip before graduation. A year later, he applied to 26 schools; he didn’t get into any of them. So he spent the next year at a community college, before transferring to Columbia. (His personal essay, which turned his winding road to higher education into a parable for his ambition to build companies, was written with help from ChatGPT.) When he started at Columbia as a sophomore this past September, he didn’t worry much about academics or his GPA. “Most assignments in college are not relevant,” he told me. “They’re hackable by AI, and I just had no interest in doing them.” While other new students fretted over the university’s rigorous core curriculum, described by the school as “intellectually expansive” and “personally transformative,” Lee used AI to breeze through with minimal effort. When I asked him why he had gone through so much trouble to get to an Ivy League university only to off-load all of the learning to a robot, he said, “It’s the best place to meet your co-founder and your wife.”

By the end of his first semester, Lee checked off one of those boxes. He met a co-founder, Neel Shanmugam, a junior in the school of engineering, and together they developed a series of potential start-ups: a dating app just for Columbia students, a sales tool for liquor distributors, and a note-taking app. None of them took off. Then Lee had an idea. As a coder, he had spent some 600 miserable hours on LeetCode, a training platform that prepares coders to answer the algorithmic riddles tech companies ask job and internship candidates during interviews. Lee, like many young developers, found the riddles tedious and mostly irrelevant to the work coders might actually do on the job. What was the point? What if they built a program that hid AI from browsers during remote job interviews so that interviewees could cheat their way through instead?

In February, Lee and Shanmugam launched a tool that did just that. Interview Coder’s website featured a banner that read F*CK LEETCODE. Lee posted a video of himself on YouTube using it to cheat his way through an internship interview with Amazon. (He actually got the internship, but turned it down.) A month later, Lee was called into Columbia’s academic-integrity office. The school put him on disciplinary probation after a committee found him guilty of “advertising a link to a cheating tool” and “providing students with the knowledge to access this tool and use it how they see fit,” according to the committee’s report.

Lee thought it absurd that Columbia, which had a partnership with ChatGPT’s parent company, OpenAI, would punish him for innovating with AI. Although Columbia’s policy on AI is similar to that of many other universities’ — students are prohibited from using it unless their professor explicitly permits them to do so, either on a class-by-class or case-by-case basis — Lee said he doesn’t know a single student at the school who isn’t using AI to cheat. To be clear, Lee doesn’t think this is a bad thing. “I think we are years — or months, probably — away from a world where nobody thinks using AI for homework is considered cheating,” he said.

In January 2023, just two months after OpenAI launched ChatGPT, a survey of 1,000 college students found that nearly 90 percent of them had used the chatbot to help with homework assignments. In its first year of existence, ChatGPT’s total monthly visits steadily increased month-over-month until June, when schools let out for the summer. (That wasn’t an anomaly: Traffic dipped again over the summer in 2024.) Professors and teaching assistants increasingly found themselves staring at essays filled with clunky, robotic phrasing that, though grammatically flawless, didn’t sound quite like a college student — or even a human. Two and a half years later, students at large state schools, the Ivies, liberal-arts schools in New England, universities abroad, professional schools, and community colleges are relying on AI to ease their way through every facet of their education. Generative-AI chatbots — ChatGPT but also Google’s Gemini, Anthropic’s Claude, Microsoft’s Copilot, and others — take their notes during class, devise their study guides and practice tests, summarize novels and textbooks, and brainstorm, outline, and draft their essays. STEM students are using AI to automate their research and data analyses and to sail through dense coding and debugging assignments. “College is just how well I can use ChatGPT at this point,” a student in Utah recently captioned a video of herself copy-and-pasting a chapter from her Genocide and Mass Atrocity textbook into ChatGPT.

Sarah, a freshman at Wilfrid Laurier University in Ontario, said she first used ChatGPT to cheat during the spring semester of her final year of high school. (Sarah’s name, like those of other current students in this article, has been changed for privacy.) After getting acquainted with the chatbot, Sarah used it for all her classes: Indigenous studies, law, English, and a “hippie farming class” called Green Industries. “My grades were amazing,” she said. “It changed my life.” Sarah continued to use AI when she started college this past fall. Why wouldn’t she? Rarely did she sit in class and not see other students’ laptops open to ChatGPT. Toward the end of the semester, she began to think she might be dependent on the website. She already considered herself addicted to TikTok, Instagram, Snapchat, and Reddit, where she writes under the username maybeimnotsmart. “I spend so much time on TikTok,” she said. “Hours and hours, until my eyes start hurting, which makes it hard to plan and do my schoolwork. With ChatGPT, I can write an essay in two hours that normally takes 12.”

Teachers have tried AI-proofing assignments, returning to Blue Books or switching to oral exams. Brian Patrick Green, a tech-ethics scholar at Santa Clara University, immediately stopped assigning essays after he tried ChatGPT for the first time. Less than three months later, teaching a course called Ethics and Artificial Intelligence, he figured a low-stakes reading reflection would be safe — surely no one would dare use ChatGPT to write something personal. But one of his students turned in a reflection with robotic language and awkward phrasing that Green knew was AI-generated. A philosophy professor across the country at the University of Arkansas at Little Rock caught students in her Ethics and Technology class using AI to respond to the prompt “Briefly introduce yourself and say what you’re hoping to get out of this class.”

It isn’t as if cheating is new. But now, as one student put it, “the ceiling has been blown off.” Who could resist a tool that makes every assignment easier with seemingly no consequences? After spending the better part of the past two years grading AI-generated papers, Troy Jollimore, a poet, philosopher, and Cal State Chico ethics professor, has concerns. “Massive numbers of students are going to emerge from university with degrees, and into the workforce, who are essentially illiterate,” he said. “Both in the literal sense and in the sense of being historically illiterate and having no knowledge of their own culture, much less anyone else’s.” That future may arrive sooner than expected when you consider what a short window college really is. Already, roughly half of all undergrads have never experienced college without easy access to generative AI. “We’re talking about an entire generation of learning perhaps significantly undermined here,” said Green, the Santa Clara tech ethicist. “It’s short-circuiting the learning process, and it’s happening fast.”

Before OpenAI released ChatGPT in November 2022, cheating had already reached a sort of zenith. At the time, many college students had finished high school remotely, largely unsupervised, and with access to tools like Chegg and Course Hero. These companies advertised themselves as vast online libraries of textbooks and course materials but, in reality, were cheating multi-tools. For $15.95 a month, Chegg promised answers to homework questions in as little as 30 minutes, 24/7, from the 150,000 experts with advanced degrees it employed, mostly in India. When ChatGPT launched, students were primed for a tool that was faster, more capable.

But school administrators were stymied. There would be no way to enforce an all-out ChatGPT ban, so most adopted an ad hoc approach, leaving it up to professors to decide whether to allow students to use AI. Some universities welcomed it, partnering with developers, rolling out their own chatbots to help students register for classes, or launching new classes, certificate programs, and majors focused on generative AI. But regulation remained difficult. How much AI help was acceptable? Should students be able to have a dialogue with AI to get ideas but not ask it to write the actual sentences?

These days, professors will often state their policy on their syllabi — allowing AI, for example, as long as students cite it as if it were any other source, or permitting it for conceptual help only, or requiring students to provide receipts of their dialogue with a chatbot. Students often interpret those instructions as guidelines rather than hard rules. Sometimes they will cheat on their homework without even knowing — or knowing exactly how much — they are violating university policy when they ask a chatbot to clean up a draft or find a relevant study to cite. Wendy, a freshman finance major at one of the city’s top universities, told me that she is against using AI. Or, she clarified, “I’m against copy-and-pasting. I’m against cheating and plagiarism. All of that. It’s against the student handbook.” Then she described, step-by-step, how on a recent Friday at 8 a.m., she called up an AI platform to help her write a four-to-five-page essay due two hours later.

Whenever Wendy uses AI to write an essay (which is to say, whenever she writes an essay), she follows three steps. Step one: “I say, ‘I’m a first-year college student. I’m taking this English class.’” Otherwise, Wendy said, “it will give you a very advanced, very complicated writing style, and you don’t want that.” Step two: Wendy provides some background on the class she’s taking before copy-and-pasting her professor’s instructions into the chatbot. Step three: “Then I ask, ‘According to the prompt, can you please provide me an outline or an organization to give me a structure so that I can follow and write my essay?’ It then gives me an outline, introduction, topic sentences, paragraph one, paragraph two, paragraph three.” Sometimes, Wendy asks for a bullet list of ideas to support or refute a given argument: “I have difficulty with organization, and this makes it really easy for me to follow.”

Once the chatbot had outlined Wendy’s essay, providing her with a list of topic sentences and bullet points of ideas, all she had to do was fill it in. Wendy delivered a tidy five-page paper at an acceptably tardy 10:17 a.m. When I asked her how she did on the assignment, she said she got a good grade. “I really like writing,” she said, sounding strangely nostalgic for her high-school English class — the last time she wrote an essay unassisted. “Honestly,” she continued, “I think there is beauty in trying to plan your essay. You learn a lot. You have to think, Oh, what can I write in this paragraph? Or What should my thesis be? ” But she’d rather get good grades. “An essay with ChatGPT, it’s like it just gives you straight up what you have to follow. You just don’t really have to think that much.”

I asked Wendy if I could read the paper she turned in, and when I opened the document, I was surprised to see the topic: critical pedagogy, the philosophy of education pioneered by Paulo Freire. The philosophy examines the influence of social and political forces on learning and classroom dynamics. Her opening line: “To what extent is schooling hindering students’ cognitive ability to think critically?” Later, I asked Wendy if she recognized the irony in using AI to write not just a paper on critical pedagogy but one that argues learning is what “makes us truly human.” She wasn’t sure what to make of the question. “I use AI a lot. Like, every day,” she said. “And I do believe it could take away that critical-thinking part. But it’s just — now that we rely on it, we can’t really imagine living without it.”

Most of the writing professors I spoke to told me that it’s abundantly clear when their students use AI. Sometimes there’s a smoothness to the language, a flattened syntax; other times, it’s clumsy and mechanical. The arguments are too evenhanded — counterpoints tend to be presented just as rigorously as the paper’s central thesis. Words like multifaceted and context pop up more than they might normally. On occasion, the evidence is more obvious, as when last year a teacher reported reading a paper that opened with “As an AI, I have been programmed …” Usually, though, the evidence is more subtle, which makes nailing an AI plagiarist harder than identifying the deed. Some professors have resorted to deploying so-called Trojan horses, sticking strange phrases, in small white text, in between the paragraphs of an essay prompt. (The idea is that this would theoretically prompt ChatGPT to insert a non sequitur into the essay.) Students at Santa Clara recently found the word broccoli hidden in a professor’s assignment. Last fall, a professor at the University of Oklahoma sneaked the phrases “mention Finland” and “mention Dua Lipa” in his. A student discovered his trap and warned her classmates about it on TikTok. “It does work sometimes,” said Jollimore, the Cal State Chico professor. “I’ve used ‘How would Aristotle answer this?’ when we hadn’t read Aristotle. But I’ve also used absurd ones and they didn’t notice that there was this crazy thing in their paper, meaning these are people who not only didn’t write the paper but also didn’t read their own paper before submitting it.”

Still, while professors may think they are good at detecting AI-generated writing, studies have found they’re actually not. One, published in June 2024, used fake student profiles to slip 100 percent AI-generated work into professors’ grading piles at a U.K. university. The professors failed to flag 97 percent. It doesn’t help that since ChatGPT’s launch, AI’s capacity to write human-sounding essays has only gotten better. Which is why universities have enlisted AI detectors like Turnitin, which uses AI to recognize patterns in AI-generated text. After evaluating a block of text, detectors provide a percentage score that indicates the alleged likelihood it was AI-generated. Students talk about professors who are rumored to have certain thresholds (25 percent, say) above which an essay might be flagged as an honor-code violation. But I couldn’t find a single professor — at large state schools or small private schools, elite or otherwise — who admitted to enforcing such a policy. Most seemed resigned to the belief that AI detectors don’t work. It’s true that different AI detectors have vastly different success rates, and there is a lot of conflicting data. While some claim to have less than a one percent false-positive rate, studies have shown they trigger more false positives for essays written by neurodivergent students and students who speak English as a second language. Turnitin’s chief product officer, Annie Chechitelli, told me that the product is tuned to err on the side of caution, more inclined to trigger a false negative than a false positive so that teachers don’t wrongly accuse students of plagiarism. I fed Wendy’s essay through a free AI detector, ZeroGPT, and it came back as 11.74 AI-generated, which seemed low given that AI, at the very least, had generated her central arguments. I then fed a chunk of text from the Book of Genesis into ZeroGPT and it came back as 93.33 percent AI-generated.

There are, of course, plenty of simple ways to fool both professors and detectors. After using AI to produce an essay, students can always rewrite it in their own voice or add typos. Or they can ask AI to do that for them: One student on TikTok said her preferred prompt is “Write it as a college freshman who is a li’l dumb.” Students can also launder AI-generated paragraphs through other AIs, some of which advertise the “authenticity” of their outputs or allow students to upload their past essays to train the AI in their voice. “They’re really good at manipulating the systems. You put a prompt in ChatGPT, then put the output into another AI system, then put it into another AI system. At that point, if you put it into an AI-detection system, it decreases the percentage of AI used every time,” said Eric, a sophomore at Stanford.

Most professors have come to the conclusion that stopping rampant AI abuse would require more than simply policing individual cases and would likely mean overhauling the education system to consider students more holistically. “Cheating correlates with mental health, well-being, sleep exhaustion, anxiety, depression, belonging,” said Denise Pope, a senior lecturer at Stanford and one of the world’s leading student-engagement researchers.

Many teachers now seem to be in a state of despair. In the fall, Sam Williams was a teaching assistant for a writing-intensive class on music and social change at the University of Iowa that, officially, didn’t allow students to use AI at all. Williams enjoyed reading and grading the class’s first assignment: a personal essay that asked the students to write about their own music tastes. Then, on the second assignment, an essay on the New Orleans jazz era (1890 to 1920), many of his students’ writing styles changed drastically. Worse were the ridiculous factual errors. Multiple essays contained entire paragraphs on Elvis Presley (born in 1935). “I literally told my class, ‘Hey, don’t use AI. But if you’re going to cheat, you have to cheat in a way that’s intelligent. You can’t just copy exactly what it spits out,’” Williams said.

Williams knew most of the students in this general-education class were not destined to be writers, but he thought the work of getting from a blank page to a few semi-coherent pages was, above all else, a lesson in effort. In that sense, most of his students utterly failed. “They’re using AI because it’s a simple solution and it’s an easy way for them not to put in time writing essays. And I get it, because I hated writing essays when I was in school,” Williams said. “But now, whenever they encounter a little bit of difficulty, instead of fighting their way through that and growing from it, they retreat to something that makes it a lot easier for them.”

By November, Williams estimated that at least half of his students were using AI to write their papers. Attempts at accountability were pointless. Williams had no faith in AI detectors, and the professor teaching the class instructed him not to fail individual papers, even the clearly AI-smoothed ones. “Every time I brought it up with the professor, I got the sense he was underestimating the power of ChatGPT, and the departmental stance was, ‘Well, it’s a slippery slope, and we can’t really prove they’re using AI,’” Williams said. “I was told to grade based on what the essay would’ve gotten if it were a ‘true attempt at a paper.’ So I was grading people on their ability to use ChatGPT.”

The “true attempt at a paper” policy ruined Williams’s grading scale. If he gave a solid paper that was obviously written with AI a B, what should he give a paper written by someone who actually wrote their own paper but submitted, in his words, “a barely literate essay”? The confusion was enough to sour Williams on education as a whole. By the end of the semester, he was so disillusioned that he decided to drop out of graduate school altogether. “We’re in a new generation, a new time, and I just don’t think that’s what I want to do,” he said.

Jollimore, who has been teaching writing for more than two decades, is now convinced that the humanities, and writing in particular, are quickly becoming an anachronistic art elective like basket-weaving. “Every time I talk to a colleague about this, the same thing comes up: retirement. When can I retire? When can I get out of this? That’s what we’re all thinking now,” he said. “This is not what we signed up for.” Williams, and other educators I spoke to, described AI’s takeover as a full-blown existential crisis. “The students kind of recognize that the system is broken and that there’s not really a point in doing this. Maybe the original meaning of these assignments has been lost or is not being communicated to them well.”

He worries about the long-term consequences of passively allowing 18-year-olds to decide whether to actively engage with their assignments. Would it accelerate the widening soft-skills gap in the workplace? If students rely on AI for their education, what skills would they even bring to the workplace? Lakshya Jain, a computer-science lecturer at the University of California, Berkeley, has been using those questions in an attempt to reason with his students. “If you’re handing in AI work,” he tells them, “you’re not actually anything different than a human assistant to an artificial-intelligence engine, and that makes you very easily replaceable. Why would anyone keep you around?” That’s not theoretical: The COO of a tech research firm recently asked Jain why he needed programmers any longer.

The ideal of college as a place of intellectual growth, where students engage with deep, profound ideas, was gone long before ChatGPT. The combination of high costs and a winner-takes-all economy had already made it feel transactional, a means to an end. (In a recent survey, Deloitte found that just over half of college graduates believe their education was worth the tens of thousands of dollars it costs a year, compared with 76 percent of trade-school graduates.) In a way, the speed and ease with which AI proved itself able to do college-level work simply exposed the rot at the core. “How can we expect them to grasp what education means when we, as educators, haven’t begun to undo the years of cognitive and spiritual damage inflicted by a society that treats schooling as a means to a high-paying job, maybe some social status, but nothing more?” Jollimore wrote in a recent essay. “Or, worse, to see it as bearing no value at all, as if it were a kind of confidence trick, an elaborate sham?”

It’s not just the students: Multiple AI platforms now offer tools to leave AI-generated feedback on students’ essays. Which raises the possibility that AIs are now evaluating AI-generated papers, reducing the entire academic exercise to a conversation between two robots — or maybe even just one.

It’ll be years before we can fully account for what all of this is doing to students’ brains. Some early research shows that when students off-load cognitive duties onto chatbots, their capacity for memory, problem-solving, and creativity could suffer. Multiple studies published within the past year have linked AI usage with a deterioration in critical-thinking skills; one found the effect to be more pronounced in younger participants. In February, Microsoft and Carnegie Mellon University published a study that found a person’s confidence in generative AI correlates with reduced critical-thinking effort. The net effect seems, if not quite Wall-E, at least a dramatic reorganization of a person’s efforts and abilities, away from high-effort inquiry and fact-gathering and toward integration and verification. This is all especially unnerving if you add in the reality that AI is imperfect — it might rely on something that is factually inaccurate or just make something up entirely — with the ruinous effect social media has had on Gen Z’s ability to tell fact from fiction. The problem may be much larger than generative AI. The so-called Flynn effect refers to the consistent rise in IQ scores from generation to generation going back to at least the 1930s. That rise started to slow, and in some cases reverse, around 2006. “The greatest worry in these times of generative AI is not that it may compromise human creativity or intelligence,” Robert Sternberg, a psychology professor at Cornell University, told The Guardian, “but that it already has.”

Students are worrying about this, even if they’re not willing or able to give up the chatbots that are making their lives exponentially easier. Daniel, a computer-science major at the University of Florida, told me he remembers the first time he tried ChatGPT vividly. He marched down the hall to his high-school computer-science teacher’s classroom, he said, and whipped out his Chromebook to show him. “I was like, ‘Dude, you have to see this!’ My dad can look back on Steve Jobs’s iPhone keynote and think, Yeah, that was a big moment. That’s what it was like for me, looking at something that I would go on to use every day for the rest of my life.”

AI has made Daniel more curious; he likes that whenever he has a question, he can quickly access a thorough answer. But when he uses AI for homework, he often wonders, If I took the time to learn that, instead of just finding it out, would I have learned a lot more? At school, he asks ChatGPT to make sure his essays are polished and grammatically correct, to write the first few paragraphs of his essays when he’s short on time, to handle the grunt work in his coding classes, to cut basically all cuttable corners. Sometimes, he knows his use of AI is a clear violation of student conduct, but most of the time it feels like he’s in a gray area. “I don’t think anyone calls seeing a tutor cheating, right? But what happens when a tutor starts writing lines of your paper for you?” he said.

Recently, Mark, a freshman math major at the University of Chicago, admitted to a friend that he had used ChatGPT more than usual to help him code one of his assignments. His friend offered a somewhat comforting metaphor: “You can be a contractor building a house and use all these power tools, but at the end of the day, the house won’t be there without you.” Still, Mark said, “it’s just really hard to judge. Is this my work?” I asked Daniel a hypothetical to try to understand where he thought his work began and AI’s ended: Would he be upset if he caught a romantic partner sending him an AI-generated poem? “I guess the question is what is the value proposition of the thing you’re given? Is it that they created it? Or is the value of the thing itself?” he said. “In the past, giving someone a letter usually did both things.” These days, he sends handwritten notes — after he has drafted them using ChatGPT.

“Language is the mother, not the handmaiden, of thought,” wrote Duke professor Orin Starn in a recent column titled “My Losing Battle Against AI Cheating,” citing a quote often attributed to W. H. Auden. But it’s not just writing that develops critical thinking. “Learning math is working on your ability to systematically go through a process to solve a problem. Even if you’re not going to use algebra or trigonometry or calculus in your career, you’re going to use those skills to keep track of what’s up and what’s down when things don’t make sense,” said Michael Johnson, an associate provost at Texas A&M University. Adolescents benefit from structured adversity, whether it’s algebra or chores. They build self-esteem and work ethic. It’s why the social psychologist Jonathan Haidt has argued for the importance of children learning to do hard things, something that technology is making infinitely easier to avoid. Sam Altman, OpenAI’s CEO, has tended to brush off concerns about AI use in academia as shortsighted, describing ChatGPT as merely “a calculator for words” and saying the definition of cheating needs to evolve. “Writing a paper the old-fashioned way is not going to be the thing,” Altman, a Stanford dropout, said last year. But speaking before the Senate’s oversight committee on technology in 2023, he confessed his own reservations: “I worry that as the models get better and better, the users can have sort of less and less of their own discriminating process.” OpenAI hasn’t been shy about marketing to college students. It recently made ChatGPT Plus, normally a $20-per-month subscription, free to them during finals. (OpenAI contends that students and teachers need to be taught how to use it responsibly, pointing to the ChatGPT Edu product it sells to academic institutions.)

In late March, Columbia suspended Lee after he posted details about his disciplinary hearing on X. He has no plans to go back to school and has no desire to work for a big-tech company, either. Lee explained to me that by showing the world AI could be used to cheat during a remote job interview, he had pushed the tech industry to evolve the same way AI was forcing higher education to evolve. “Every technological innovation has caused humanity to sit back and think about what work is actually useful,” he said. “There might have been people complaining about machinery replacing blacksmiths in, like, the 1600s or 1800s, but now it’s just accepted that it’s useless to learn how to blacksmith.”

Lee has already moved on from hacking interviews. In April, he and Shanmugam launched Cluely, which scans a user’s computer screen and listens to its audio in order to provide AI feedback and answers to questions in real time without prompting. “We built Cluely so you never have to think alone again,” the company’s manifesto reads. This time, Lee attempted a viral launch with a $140,000 scripted advertisement in which a young software engineer, played by Lee, uses Cluely installed on his glasses to lie his way through a first date with an older woman. When the date starts going south, Cluely suggests Lee “reference her art” and provides a script for him to follow. “I saw your profile and the painting with the tulips. You are the most gorgeous girl ever,” Lee reads off his glasses, which rescues his chances with her.

Before launching Cluely, Lee and Shanmugam raised $5.3 million from investors, which allowed them to hire two coders, friends Lee met in community college (no job interviews or LeetCode riddles were necessary), and move to San Francisco. When we spoke a few days after Cluely’s launch, Lee was at his Realtor’s office and about to get the keys to his new workspace. He was running Cluely on his computer as we spoke. While Cluely can’t yet deliver real-time answers through people’s glasses, the idea is that someday soon it’ll run on a wearable device, seeing, hearing, and reacting to everything in your environment. “Then, eventually, it’s just in your brain,” Lee said matter-of-factly. For now, Lee hopes people will use Cluely to continue AI’s siege on education. “We’re going to target the digital LSATs; digital GREs; all campus assignments, quizzes, and tests,” he said. “It will enable you to cheat on pretty much everything.”

Continue Reading

Noticias

Korl lanza una plataforma que orquestan agentes de IA de OpenAi, Gemini y Anthrope para hipercustomizar la mensajería de los clientes

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


Es un enigma: los equipos de clientes tienen más datos de los que pueden comenzar a usar, desde las notas de Salesforce, los boletos JIRA, los paneles de proyectos, los documentos de Google, pero tienen dificultades para reunirlo todo al elaborar mensajes de clientes que realmente resuenan.

Las herramientas existentes a menudo dependen de plantillas o diapositivas genéricas y no pueden proporcionar una imagen completa de viajes de clientes, hojas de ruta, objetivos del proyecto y objetivos comerciales.

Korl, una startup lanzada hoy, espera superar estos desafíos con una nueva plataforma que funciona en múltiples sistemas para ayudar a crear comunicaciones altamente personalizadas. La herramienta multimodal múltiple utiliza una mezcla de modelos de OpenAI, Géminis y antrópico para obtener datos y contextualizar los datos.

“Los ingenieros tienen herramientas de IA potentes, pero los equipos orientados al cliente están atrapados con soluciones poco profundas y desconectadas”, dijo Berit Hoffmann, CEO y cofundador de Korl, a VentureBeat en una entrevista exclusiva. “La innovación central de Korl se basa en nuestras tuberías avanzadas de múltiples agentes diseñados para construir el contexto del cliente y el producto que carecen las herramientas genéricas de presentación”.

Creación de materiales de cliente personalizados a través de una vista de múltiples fuentes

Los agentes de AI de Korl agregan información de diferentes sistemas, como la documentación de ingeniería de JIRA, contornos de Google Docs, diseños de Figma y datos de proyectos de Salesforce, para construir una vista de múltiples fuentes.

Por ejemplo, una vez que un cliente conecta a Korl con JIRA, su agente estudia las capacidades de productos existentes y planificadas para descubrir cómo mapear datos e importar nuevas capacidades de productos, explicó Hoffmann. La plataforma coincide con los datos del producto con la información del cliente, como el historial de uso, las prioridades comerciales y la etapa del ciclo de vida, que llena los vacíos con el uso de la IA.

“Los agentes de datos de Korl recopilan, enriquecen y estructuran diversos conjuntos de datos de fuentes internas y datos públicos externas”, dijo Hoffmann.

Luego, la plataforma genera automáticamente revisiones comerciales trimestrales (QBR) personalizadas, lanzamientos de renovación, presentaciones a medida y otros materiales para su uso en hitos importantes del cliente.

Hoffmann dijo que el diferenciador central de la compañía es su capacidad para ofrecer “materiales pulidos listos para el cliente”, como diapositivas, narraciones y correos electrónicos, “en lugar de simplemente análisis o ideas crudas”.

“Creemos que esto ofrece un nivel de valor operativo que los equipos orientados al cliente necesitan hoy dadas las presiones para hacer más con menos”, dijo.

Cambiar entre OpenAi, Géminis, Anthrope, basado en el rendimiento

Korl orquesta un “conjunto de modelos” en OpenAi, Gemini y Anthrope, seleccionando el mejor modelo para el trabajo en el momento basado en la velocidad, la precisión y el costo, explicó Hoffmann. Korl necesita realizar tareas complejas y diversas (narraciones matizadas, computación de datos, imágenes), por lo que cada caso de uso coincide con el modelo más desempeñado. La compañía ha implementado “mecanismos sofisticados de respaldo” para mitigar las fallas; Al principio, observaron altas tasas de falla al confiar en un solo proveedor, informó Hoffman.

La startup desarrolló un aplazamiento de automóviles patentado para manejar diversos esquemas de datos empresariales en JIRA, Salesforce y otros sistemas. La plataforma se asigna automáticamente a los campos relevantes en Korl.

“En lugar de solo una coincidencia semántica o de nombre de campo, nuestro enfoque evalúa factores adicionales como la escasez de datos para obtener y predecir coincidencias de campo”, dijo Hoffmann.

Para acelerar el proceso, Korl combina modelos de baja latencia y alto rendimiento (como GPT-4O para respuestas rápidas de construcción de contexto) con modelos analíticos más profundos (Claude 3.7 para comunicaciones más complejas y orientadas al cliente).

“Esto garantiza que optimizemos para la mejor experiencia del usuario final, haciendo compensaciones basadas en el contexto entre inmediatez y precisión”, explicó Hoffmann.

Debido a que “la seguridad es primordial”, Korl busca garantías de privacidad de grado empresarial de los proveedores para garantizar que los datos del cliente estén excluidos de los conjuntos de datos de capacitación. Hoffmann señaló que su orquestación múltiple y contextual, lo que impulsa adicional limita la exposición inadvertida y las fugas de datos.

Lidiar con datos que son ‘demasiado desordenados’ o ‘incompletos’

Hoffman señaló que, al principio, Korl escuchó de los clientes que les preocupaba que sus datos fueran “demasiado desordenados” o “incompletos” para ser aprovechados. En respuesta, la compañía construyó tuberías para comprender las relaciones de los objetos comerciales y llenar los vacíos, como cómo posicionar las características externamente o cómo alinear los valores en torno a los resultados deseados.

“Nuestro agente de presentación es lo que aprovecha esos datos para generar diapositivas de clientes y pista de conversación [guide conversations with potential customers or leads] dinámicamente cuando sea necesario ”, dijo Hoffmann.

También dijo que Korl presenta “verdadera multimodalidad”. La plataforma no es solo extraer datos de varias fuentes; Está interpretando diferentes tipos de información, como texto, datos estructurados, gráficos o diagramas.

“El paso crítico es ir más allá de los datos sin procesar para responder: ¿Qué historia cuenta este gráfico? ¿Cuáles son las implicaciones más profundas aquí, y realmente resonarán con este cliente específico?”, Dijo. “Hemos creado nuestro proceso para realizar esa diligencia debida crucial, asegurando que la producción no sea solo datos agregados, sino un contenido genuinamente rico entregado con un contexto significativo”.

Dos de los competidores cercanos de Korl incluyen Gainsight y Clari; Sin embargo, Hoffmann dijo que Korl se diferencia al incorporar un contexto profundo de productos y hoja de ruta. Las estrategias efectivas de renovación y expansión del cliente requieren una comprensión profunda de lo que hace un producto, y esto debe combinarse con el análisis de los datos y el comportamiento del cliente.

Además, Hoffmann dijo que Korl aborda dos “deficiencias fundamentales” de las plataformas existentes: contexto comercial profundo y precisión de la marca. Los agentes de Korl recopilan el contexto comercial de múltiples sistemas. “Sin esta inteligencia integral de datos, las cubiertas automatizadas carecen de valor comercial estratégico”, dijo.

Cuando se trata de la marca, la tecnología patentada de Korl extrae y replica las pautas de los materiales existentes.

Reducir el tiempo de preparación de la cubierta de ‘varias horas a minutos’

Las primeras indicaciones sugieren que Korl puede desbloquear al menos una mejora de 1 punto en la retención de ingresos netos (NRR) para las compañías de software del mercado medio, dijo Hoffmann. Esto se debe a que descubre el valor del producto previamente no realizado y facilita la comunicación de los clientes antes de que se conviertan o toman decisiones de renovación o expansión.

La plataforma también mejora la eficiencia, reduciendo el tiempo de preparación de la plataforma para cada llamada del cliente de “varias horas a minutos”, según Hoffman.

Los primeros clientes incluyen la plataforma de construcción de habilidades Datacamp y Gifting y Direct Mail Company Sendoso.

“Abordan un desafío crítico y pasado por alto: con demasiada frecuencia, las características del producto se lanzan mientras que los equipos de mercado (GTM) no están preparados para venderlas, apoyarlas o comunicarlas de manera efectiva”, dijo Amir Younes, director de clientes de Sendoso. “Con la IA de Korl, [go-to-market] La habilitación de GTM y la creación de activos podrían estar a solo un clic de distancia, sin agregar sobrecarga para los equipos de I + D “.

Korl ingresó hoy al mercado con $ 5 millones en fondos iniciales en una ronda co-liderada por Mac Venture Capital y subrayado VC, con la participación de Perceptive Ventures y Diane Greene (fundador de VMware y ex CEO de Google Cloud).

Continue Reading

Trending