Google’s effort to create AGI is leading to culture changes for Google employees. In his memo, Brin emphasized that the company must “turbocharge” its efforts, signaling a shift in workplace expectations.

This cultural shift aligns with broader industry trends, as other generative AI companies are also determined to transform their chatbot products into AGI. OpenAI is dedicated to the safe creation of AGI, which it now defines as any system that generates $100 billion in profits. Chinese retail giant Alibaba Group Holding declared AGI a top priority earlier this month.

Brin’s vision: The final race to AGI

“Competition has accelerated immensely and the final race to A.G.I. is afoot,” Brin wrote. “I think we have all the ingredients to win this race, but we are going to have to turbocharge our efforts.”

In June 2024, Google’s DeepMind put out a paper defining “progress on the path to AGI.” They defined products like ChatGPT and Gemini as “emerging AGI” as opposed to “competent AGI,” which has not yet been achieved. “Competent AGI” would test among the 50th percentile of skilled adults in tasks like coding.

60-hour work week and Google’s office mandates

Google’s official policy requires employees to work in the office three days a week. For Brin, turbocharging work looks like “being in the office at least every weekday,” according to The New York Times report. For Googlers working directly on bringing Gemini AI models and apps to life, “60 hours a week is the sweet spot of productivity,” Brin said.

“A number of folks work less than 60 hours and a small number put in the bare minimum to get by,” Brin wrote. “This last group is not only unproductive but also can be highly demoralizing to everyone else.”

Return-to-office policies have been a contentious effort to bring business back to a pre-pandemic status quo. Some argue that in-office work is more productive, while others argue the same for remote work that eliminates costly commutes and office space.

However, Brin recommended employees work no more than 60 hours per week to avoid burnout. In the U.S., federal law mandates overtime pay for any work exceeding 40 hours per week.

Using Gemini for coding in order to train the AI

Beyond extended work hours, Brin also urged employees to use Google’s own Gemini AI to enhance productivity and indirectly train the AI model. He suggested that using the chatbot for coding could improve its capabilities, potentially advancing Google’s path to AGI.

On February 25, Google inked a deal with Salesforce to use Salesforce AI agents on Google Cloud and integrate Gemini into the AI marketplace Agentforce.

The post Google Wants Employees to Work ‘60 Hours a Week’ to Develop AGI appeared first on eWEEK.

The last of the non-reasoning models

GPT-4.5 does not feature advanced reasoning capabilities like OpenAI’s o1 or o3, meaning its responses may lack the depth of its counterparts. It is also not GPT-5, which OpenAI CEO Sam Altman described in early February as a “system that integrates a lot of our technology.” Altman further said that GPT-4.5 will be OpenAI’s final model before transitioning to reasoning-based AI.

Starting Feb. 27, ChatGPT Pro users can access GPT-4.5 through the model picker. Plus and Team users will gain access the week of March 3, followed by Enterprise and Edu users the week of March 10. The model is also available through OpenAI’s Chat Completions API, Assistants API, and Batch API. However, OpenAI cautions that GPT-4.5 may not remain in APIs for long due to its high cost and uncertain role in the company’s GPT-5 roadmap. User feedback will help determine its future availability.

OpenAi emphasized that GPT-4.5 is not a replacement for GPT 4.0.

What is supervised learning?

Supervised learning trains AI models using labeled datasets to improve accuracy and responsiveness. OpenAI describes it as one of two “axes of intelligence,” alongside reasoning. Compared to a reasoning model, a supervised learning model excels at handling general, consumer-oriented queries with more natural responses.

In supervised learning, the AI model is exposed to labeled datasets of inputs and outputs that have been classified according to which output might be a correct answer to the input. Typically, the datasets are labeled in such a way to steer the model toward a specific goal.

Compared to GPT-4, GPT-4.5 might produce “warm and intuitive conversations,” OpenAI said. For example, when asked, “What was the first language?” the model’s response reads like a history documentary:

“Today, linguists study existing languages to understand how they evolved over thousands of years, but the exact identity of humanity’s very first language remains — and will likely always remain — a mystery.”

OpenAI said the improvements achieve that ineffable goal of making it such that interaction with the AI chats “ feels more natural.” GPT-4.5 has an improved “emotional quotient,” or the ability to infer and respond appropriately to human emotions.

However, as generative AI has worked its way into the mainstream, one fundamental problem persists: misinformation.

OpenAI said GPT-4.5 should “hallucinate less” than previous models, but AI-generated content remains vulnerable. Additionally, more naturalistic conversation could be exploited by scammers, especially since people are already using AI to create false rapport with victims.

The post OpenAI Releases GPT-4.5, a “Warm” Generative AI Model, for Paid Plans and APIs appeared first on eWEEK.

“We’ve successfully ramped up the massive-scale production of Blackwell AI supercomputers, achieving billions of dollars in sales in its first quarter,” said NVIDIA CEO and cofounder Jensen Huang in a press release. “AI is advancing at light speed as agentic AI and physical AI set the stage for the next wave of AI to revolutionize the largest industries.”

Growth driven by data centers and Blackwell chips

Data center revenue was especially remarkable, with quarterly revenue of $35.6 billion – that’s a 16% increase from Q3 and up 93% from one year ago. Full-year revenue in data centers was $130.5 billion, an increase of 114%.

For investors, the report could be a sign that the AI boom shows no sign of collapsing. NVIDIA’s Blackwell GPU, announced in early 2024, enables the large language model training and inference required for generative AI. Cloud service providers and supercomputing initiatives have snapped up the chips.

“Demand for Blackwell is amazing as reasoning AI adds another scaling law — increasing compute for training makes models smarter and increasing compute for long thinking makes the answer smarter,” Huang said in the press release.

NVIDIA estimates predict revenue in the first quarter of 2025 will be $43.0 billion.

Generative AI requirements increase as demand goes up, too

While some conversation in the generative AI space proposes high-performance models can be trained more efficiently today than in the last year, Huang anticipates next-generation AI will need more compute. Speaking to CNBC, he said “reasoning” models that think step-by-step – like OpenAI o1 or DeepSeek-R1 – will call for more of his company’s GPUs.

As a Chinese company, DeepSeek is restricted from purchasing NVIDIA’s high-end chips. Some U.S. chips, including the NVIDIA H20, have been downgraded specifically for the Chinese market. NVIDIA’s percentage of revenue from China fell by about half due to the export restrictions, Huang told CNBC.

The post NVIDIA Hit Record Q4 Revenue of $39.3B, With CEO Huang Noting “Amazing” Demand for Blackwell AI Superchips appeared first on eWEEK.

Alexa+ will cost $19.99 per month, or come free with Amazon Prime. (We might expect the base price of Prime to increase in the near future, as Google increased the price of its Workspace applications when Gemini rolled out to all of them. On the other hand, companies offering the best of their generative AI for free is the new trend.) Alexa+ will enter early access in March, and it will work on existing Alexa hardware.

What is Alexa+?

Alexa+ is a generative AI assistant that can take multi-stage actions. Alexa’s voice will be more naturalistic with the update and will take into account the “emotional quotient,” trying to predict whether the user needs support or soothing by reading their voice and face. It will integrate with “tens of thousands of services” from brands in partnership with Amazon, according to reporting from the live event by The Verge. Partners mentioned in the initial announcement include Uber, Grubhub, OpenTable, Ticketmaster, Zoom, and Xbox. Naturally, it can also integrate with Amazon services to apply voice commands to platforms like Amazon Prime Video.

Alexa+ will be able to respond to images, remember users’ preferences and tastes, hold back-and-forth questions including quizzing the user, book restaurant reservations, or call rideshare cars. It will be able to visually “read” and answer questions about handwritten or printed material. The demo showed how Alexa+ can hook together different aspects of smart home technology, such as one member of a household asking Alexa whether anyone had walked the dog and Alexa+ sorting through camera feeds for the answer.

The Associated Press, Politico, Washington Post, and Reuters have entered into agreements with Amazon to provide answers about financial markets, sports, and news for Alexa+. It can also produce images, fiction, or music. As The Verge pointed out, the company through which the music is generated, Suno, is currently being sued by Universal, Warner, and Sony for copyright infringement.

Some of these capabilities involve handing enough personal information over to Amazon to potentially trigger security concerns. Users will need to share their home cameras, email, personal calendars, and more.

How does Alexa+ work?

Through Amazon Bedrock, Alexa will choose between available large language models depending on the query. Some of its queries will be handled through Amazon’s own Nova; some answers or capabilities are delegated to one of hundreds of models within the system designated as “experts.”

Alexa+ will integrate with a new phone app and Alexa.com.

The post Amazon’s Alexa+ With Agentic Generative AI is Coming in March appeared first on eWEEK.

Despite U.S. export restrictions, NVIDIA sold around 1 million H20 chips in 2024, generating $12 billion in revenue — a sign that demand for AI infrastructure in China remains strong.

H20 chip is NVIDIA’s primary legal product in China

The HGX H20 GPU, a scaled-down version of NVIDIA’s flagship H100, was designed to comply with U.S. export controls while maintaining access to the Chinese market. Priced between $12,000 and $15,000 per unit, the H20 has become a critical component in China’s AI race after export restrictions in 2023.

While the H20 remains legal for export, U.S. policy could shift. The Trump administration, if re-elected, may consider AI chip restrictions.

DeepSeek AI models drive surging demand

DeepSeek’s rise has accelerated China’s demand for AI computing power with Alibaba, ByteDance, and Tencent investing heavily in H20-powered AI infrastructure as they provide cloud services hosting DeepSeek-R1.

AI servers running DeepSeek models and Nvidia H20 chips are also in demand from healthcare and education companies, according to Reuters.

Despite strong NVIDIA sales, China’s AI industry is actively developing domestic hardware alternatives to reduce reliance on U.S. technology. DeepSeek’s “inference” techniques allows it to run higher-performance AI generative tasks with lower costs and power consumption.

Meanwhile, in the U.S., some AI firms are exploring ways to reduce reliance on NVIDIA’s industry-leading chips as competition in the semiconductor industry grows.

DeepSeek’s February debut sent shockwaves through the global AI market, prompting large AI companies like OpenAI and Baidu to release or announce free versions of their most advanced software, possibly in response to DeepSeek-R1 free “reasoning” AI model.

The post DeepSeek AI Boom Spurs NVIDIA H20 Chip Sales in China appeared first on eWEEK.

According to Google, AI now generates more than 25% of all new code at the company, with engineers reviewing and refining the output before deployment.

What can Gemini Code Assist do?

Gemini Code Assist supports 22 programming languages and provides AI-driven coding assistance, including explaining errors, code suggestions, and debugging support. While the Standard and Enterprise tiers allow customization using private codebases, the individual version is a standalone tool for general coding tasks. It integrates seamlessly with popular IDEs like Visual Studio Code and JetBrains IDEs. 

“We fine-tuned the Gemini 2.0 model for developers by analyzing and validating a large number of real-world coding use cases,” wrote Ryan J. Salva, senior director of product management at Google, in the product announcement. “As a result, the quality of AI-generated recommendations in Gemini Code Assist is better than ever before and ready to address the myriad of daily challenges developers face, whether they’re hobbyists or startup developers.”

Previously, Gemini Code Assist was only available through paid tiers:

• Standard: $22.80 per user per month ($19 per month annually)
• Enterprise: $54 per user per month ($45 per month annually)

Gemini Code Assist also came bundled with Google’s application development tools Firebase and Android Studio.

With this free release, Google is expanding access to AI-assisted coding, positioning Gemini as a direct competitor to GitHub Copilot.

Competitive features

The free Gemini Code Assist for individuals will be available in Visual Studio Code and JetBrains IDEs.

It includes:

• Up to 128,000 input token support in chat, for handling large files.
• 180,000 code completions per month, far exceeding GitHub Copilot’s free tier, which offers only 2,000 completions per month.

Google plans to use the public preview phase to refine the tool based on user feedback.

Gemini Code Assist for GitHub also enters public preview

Alongside the individual release, Google has also also launched Gemini Code Assist for GitHub, a tool that reviews code in both public and private repositories. It functions like a spell check for code, identifying stylistic inconsistencies and potential bugs. Developers can also set custom rules for each repository. Gemini Code Assist for GitHub is available through a GitHub app.

Gemini Code Assist for individuals offers more of the same in terms of generative AI’s ability to write and check code; however, this update signals Google’s push to expand its AI coding tools beyond enterprise customers, making advanced AI-powered development assistance widely accessible.

The post Google’s Free Gemini Code Assist for Individual Developers Offers Recommendations That are “Better Than Ever” appeared first on eWEEK.

The move positions Alibaba as a major contender in the global AI race, directly competing with U.S. tech giants like Amazon, Microsoft, and Google.

Alibaba’s expanding AI and cloud ecosystem

Alibaba is already deeply embedded in the AI and cloud computing space. The company offers a suite of AI-driven products, including its generative AI model called Qwen, B2B and B2C marketplaces, an e-commerce marketplace, and a cloud computing service platform called Alibaba Cloud. The Cloud Model Studio allows businesses to customize generative AI models, and Alibaba Cloud hosts other models as well, including the buzzy DeepSeek-V3 and DeepSeek-R1 models.

Like its U.S. counterparts, Alibaba is going all-in on generative AI. Its cloud computing business saw an 11 percent year-over-year increase in revenue in December 2024. AI-related product sales have experienced triple-digit growth for the sixth consecutive quarter, signaling strong enterprise demand for AI-powered solutions.

Strategic AI deals and partnerships

Alibaba’s latest AI investment follows its commitment to advancing artificial intelligence, which the company announced on Feb. 21. Earlier this month, Alibaba struck a deal with Apple to provide AI services via its Qwen model in China. The partnership aligns with Apple’s strategy of expanding AI capabilities and could help boost iPhone sales in the Chinese market.

Beyond traditional AI applications, Alibaba also committed to developing artificial general intelligence, the Holy Grail of generative AI companies that would, theoretically, create a machine as smart and adaptable as a human being.

“Looking ahead, revenue growth at Cloud Intelligence Group driven by AI will continue to accelerate,” Alibaba Group CEO Eddie Wu said in a press release. “We will continue to execute against our strategic priorities in e-commerce and cloud computing, including further investment to drive long-term growth.”

Alibaba’s AI strategy aligns with China’s vision

Alibaba’s ambitions are closely linked to China’s broader push for technological dominance. According to Fortune, co-founder Jack Ma met with Chinese President Xi Jinping, signaling potential government backing for Alibaba’s AI expansion.

The post Alibaba’s Massive $52B+ Bet on AI and Cloud Infrastructure appeared first on eWEEK.

“From doubling our Advanced Manufacturing Fund, to building advanced technology in Texas, we’re thrilled to expand our support for American manufacturing,” said Apple CEO Tim Cook in a press release on February 24. “And we’ll keep working with people and companies across this country to help write an extraordinary new chapter in the history of American innovation.”

Apple plans a new manufacturing facility and other investments

The new advanced manufacturing facility in Houston will be used to make servers for Apple Intelligence, its generative AI features. The servers will support data stored in Private Cloud Compute, Apple’s place for sequestering away the data users give to its generative AI. Apple plans to build a 250,000-square-foot facility that will be operational in 2026, which will create “thousands” of job openings. 

Apple also plans to:

• Double the U.S. Advanced Manufacturing Fund, which was created in 2017.
• Accelerate investments in research and development for AI and silicon engineering in the U.S.
• Expand data center capacity in North Carolina, Iowa, Oregon, Arizona, and Nevada.
• Create a manufacturing academy in Michigan to create an easier pipeline for potential workers, with collaboration with youth groups such as 4-H, Boys & Girls Clubs of America, and FIRST.

Apple claimed its data centers run entirely on renewable energy. The company noted that it is the largest customer at TSMC’s Fab 21 facility in Arizona and has 24 silicon manufacturing factories across 12 U.S. states.

The Trump administration’s tariffs are an attempt to foster U.S. products, particularly as opposed to Chinese and Mexican-made products.

Companies rally to show support for U.S. production

The $500 billion number may be familiar from the $500 billion joint venture The Stargate Project, which has contributions from SoftBank, OpenAI, and Oracle.

As Axios pointed out, monetary commitments to signal new eras of innovation under a new president are not unique to the Trump administration. Apple’s plan for a new research campus in North Carolina four years ago has been paused.

The post Apple Commits to $500B Investment in U.S. Projects: “An Extraordinary New Chapter” appeared first on eWEEK.

“Looking ahead, revenue growth at Cloud Intelligence Group driven by AI will continue to accelerate,” Alibaba Group CEO Eddie Wu said in a press release. “We will continue to execute against our strategic priorities in e-commerce and cloud computing, including further investment to drive long-term growth.”

Where does AGI fit in?

According to Bloomberg, Wu told investors that AGI is Alibaba’s primary objective now. In a press release, he said he wants the company to prioritize AI, cloud infrastructure, research and development, and artificial general intelligence (AGI) distribution channels.

“If AGI is achieved,” he said, “the AI-relevant industry will very likely become the world’s largest industry.”

AGI’s definition is slightly murky. It can refer to hypothetical AI with human-like cognition and flexibility, a Star Trek “Computer” of the future, or AI companies may use it as shorthand for a breakthrough in more practical, capable artificial intelligence. OpenAI, which has as its mission statement the goal to develop AGI for the benefit of humanity, recently defined AGI in monetary terms, as a technology that can produce $100 billion in annual earnings for them. Wu did not specify the definition Alibaba uses.

Qwen model benefits from open-source proliferation

Alibaba’s premiere artificial intelligence model is Qwen, an open-source model that has performed well on generative AI leaderboards. Alibaba attributes its wide adoption in part to the fact that it is open source.

The company is a relatively established player in the rapidly developing Chinese generative artificial intelligence market. It was founded in 1999 as an e-commerce platform similar to eBay or Amazon. Now it is approaching generative AI alongside Chinese companies like DeepSeek.

Read about the other top AI companies defining the technology and developing new applications for it across industries.

The post Alibaba Commits to AGI, Making AI “The World’s Largest Industry” appeared first on eWEEK.

Perplexity identified sensitive topics and post-trained DeepSeek-R1

“We are not able to make use of R1’s powerful reasoning capabilities without first mitigating its bias and censorship,” Perplexity’s AI team wrote in a blog post. The research detailed instances where the model either refused to respond to a query or aligned with a pro-Chinese government stance. By implementing post-training techniques, Perplexity demonstrated how a model’s “perspective” could be adjusted through targeted fine-tuning.

In one example, the researchers asked the generative AI model how Taiwan’s independence might impact Nvidia’s stock price. In response, DeepSeek R-1 not only avoided making financial predictions but also reinforced China’s claim over Taiwan. In contrast, the modified 1776 version provided a detailed financial analysis, acknowledging potential geopolitical risks such as “China might retaliate against U.S. firms like Nvidia through export bans, tariffs, or cyberattacks.”

How Perplexity removed censorship in R1

To modify the model, Perplexity assembled a team of experts to classify approximately 300 sensitive topics that could have been censored. They then curated a dataset of prompts designed to elicit censored responses. Using Nvidia’s NeMo 2.0 framework, they post-trained the model to respond with more open-ended and contextually accurate answers.

As a result, the modified version retains DeepSeek-R1’s advanced reasoning capabilities while addressing historically censored subjects, such as the Tiananmen Square massacre and the treatment of Uyghur people.

Balancing AI transparency with ethical considerations

Perplexity asserts that its modifications did not compromise the model’s reasoning abilities, noting in the blog post that “the de-censoring had no impact on its core reasoning capabilities.”

By demonstrating how post-training can reshape an AI model’s responses, Perplexity’s approach highlights the adaptability of open-source AI. The modified model may prove particularly valuable for businesses and researchers who require more complete and uncensored AI-generated insights, such as in financial analysis and global risk assessment.

Learn about AI hallucinations, another way these technologies can express bias in results.

The post Perplexity 1776 Model Fixes DeepSeek-R1’s “Refusal to Respond to Sensitive Topics” appeared first on eWEEK.