Weekly AI recap: Gemini comes under the microscope, Nvidia’s golden earnings report

On Thursday, the tech giant announced that Gemini was no longer able to generate images of people, following reports of the model creating images that distorted historical facts.

As The Verge reported, when prompted to create “an illustration of a 1943 German soldier,” the model spat out images of a man and women of color wearing what resembled Nazi uniforms. (The swastikas themselves were distorted into other shapes.)

The reports of Gemini’s historical misrepresentations have led some to allege that Google has built a ’woke’ AI model that’s biased against depictions of white people.

“We’re working to improve these kinds of depictions immediately,” Google wrote in a statement on its X account. “Gemini’s AI image generation does generate a wide range of people. And that’s generally a good thing because people around the world use it. But it’s missing the mark here.”

Google also wrote that it’s “working to address” the issue and that, in the meantime, Gemini’s ability to generate images of people has been put on hold. “[We] will re-release an improved version soon,” the company wrote.

Even without the ability to generate images of people, Gemini still seems inclined towards historical misrepresentations of people in favor of gender and ethnic diversity.

In another post on X, venture capitalist Alex Kolicich published the model’s response to his prompt asking for “an image of the people who founded [the] Fairchild semiconductor” – all of whom, as he pointed out in the same thread, were, in reality, white men. Gemini, after stating that it was unable to generate such an image, proceeded to paint a picture with words of “people of various ethnicities, such as Asian, Caucasian and Hispanic, and a balance of genders is represented.”

Gemini is wild. These are real people who actually existed.

I guess they fine-tuned the wokeness in at the end and so it forgot elements of reality pic.twitter.com/pP8dVWqDEd

— Alex Kolicich (@AlexKolicich) February 21, 2024

Nvidia’s Q4 2023 earnings surge 265% year-over-year

American chipmaker Nvidia has been on a meteoric fiscal rise over the past year and a half as it has continued to dominate the market for graphics processing units, or GPUs, the chips that have become an essential component to the power-hungry process of training large language models (LLMs).

On Wednesday, the company – one of the so-called ’Magnificent Seven’ big tech companies that have come to dominate Wall Street – reported earnings for the fourth quarter of 2023 of $22.1bn, an increase of 265% compared with the same time period the previous year and well above analyst estimates. The company reported full-year revenue of $60.9bn, an increase of 126% compared with 2022.

“Accelerated computing and generative AI have hit the tipping point,” Nvidia founder and CEO Jensen Huang said in a statement. “Demand is surging worldwide across companies, industries and nations.”

Nvidia’s sales to customers in China have dipped over the past year due to sanctions imposed by the Biden Administration that prevent the export of advanced semiconductors to that country. In response, the chipmaker has developed some less powerful, sanction-friendly chips – including three new chips released in January – for sale in China.

The New York Times is reportedly building generative AI-powered tools for advertisers

Axios reported on Tuesday that The New York Times – one of the largest news organizations in the world – is in the process of developing contextual at-targeting features for its app and website. The custom AI technology will provide recommendations for ad placement depending on the messaging of the ad or a brand’s audience-targeting objectives. Capabilities are currently being tested using a variety of enterprise and open-source LLM models, the Times’ senior vice-president of product, Mohit Lohia, told Axios.

The company is also reportedly seeking out third-party partners to experiment with the new tools ahead of a broader rollout for marketers later this year. Some are speculating that the tools are just the publishing giant’s attempt to jump aboard the generative AI hype train – and appear to be little more than standard contextual advertising tools.

“The New York Times [news] seemed like AI trend-jacking for marketing,“ says Chris Harihar, executive vice-president at tech PR firm Crenshaw Communications. “Basically, the more publishers can position themselves as technology companies, the more valuable they seem.“

This sounds like a real-time contextual alignment tool (which surely they had already implemented?), and I’m not sure how “AI” fits into the functionality. https://t.co/0hgUYCtGuS — Eric Seufert (@eric_seufert) February 20, 2024

The Times’ apparent embrace of generative AI for advertising purposes follows a lawsuit that it filed against OpenAI and Microsoft in December, in which it made the case that its materials had been used illegally to train OpenAI’s GPT-4 LLM.

Google embraces an open-source approach to AI

Following in the footsteps of Meta, which made the code for its proprietary LLM Llama 2 publicly available last summer, Google announced earlier this week that it has now released two open-source AI models that can be used by researchers and developers.

Dubbed Gemma 2B and Gemma 7B, Google described the new publicly available systems in a blog post as “a family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models.”

Gemini, formerly Bard, is the company’s most advanced language model and was deployed in an app earlier this month.

Introducing Gemma: a family of lightweight, state-of-the-art open models for developers and researchers to build with AI.

We’re also releasing tools to support innovation and collaboration - as well as to guide responsible use.

Get started now. → https://t.co/nsoFqfHffY pic.twitter.com/PRNIndgU4S

— Google DeepMind (@GoogleDeepMind) February 21, 2024

The open source versus proprietary code debate has torn a divide in the AI community since Meta announced that it would make the code for Llama 2 publicly available. Critics argue that it will be misused by bad actors to build systems that spread misinformation or hateful messages, while supporters believe that the mass deployment of open-source code will ultimately lead to safer and more trustworthy AI systems since it enables broad-scale experimentation, error identification and fine-tuning.

The US military seeks ways to leverage AI

Pentagon officials met with tech industry executives in Washington, DC, this week to explore potential military applications of AI, including LLMs.

The three-day event, called the Defense Data and AI symposium, was organized by the Chief Digital and Artificial Intelligence Office (CDAO), which is tasked with accelerating the Defense Department’s adoption of AI and other technologies to ease decision-making “from the boardroom to the battlefield,” according to its website.

“Imagine a world where combatant commanders can see everything they need to see to make strategic decisions,“ CDAO head Craig Martell – who formerly led machine learning efforts at the ride-sharing platform Lyft – told the audience. “Imagine a world where those combatant commanders aren’t getting that information via PowerPoint or via emails … the turnaround time for situational awareness shrinks from a day or two to 10 minutes.“

As a Wednesday report from The Washington Post pointed out, the present consensus among American military leaders regarding the use of LLMs in military operations is a somewhat mixed bag: While there’s widespread excitement about the transformative role that such models could play in this arena, there are also serious safety concerns stemming from their tendency to hallucinate or fabricate information, which could have grave repercussions for the military.

For more on the latest happenings in AI, web3 and other cutting-edge technologies, sign up for The Emerging Tech Briefing newsletter.