Hello and welcome to Weekly News Roundup Issue #480.

This week, Google held the #MadeByGoogle event, during which the tech giant unveiled the new lineup of Pixel devices and the AI-powered features coming alongside them, and that will be the main focus of this week’s issue.

In other news, brain-computer interfaces and AI helped a person suffering from ALS speak for the first time in years, and biotech firm Recursion acquired its competitor, Exscientia, for $688 million.

Over in AI, xAI released Grok-2, and people discovered it has no filters when generating images. Meanwhile, Nvidia delayed the release of its AI chip due to a design flaw and is accused of scraping 80 years' worth of videos (many of them copyrighted) every day to train its AI models. Additionally, the first node of a new network of supercomputers for AGI is coming online this September.

In robotics, Amazon got the green light to test-fly delivery drones in the UK, and rumours emerged about Apple’s secret robotics projects. We also learned that Nepal will be using drones to carry cargo missions in the Himalayas.

Enjoy!

Google goes all in with Gemini

The sudden rise of ChatGPT and generative AI caught Google by surprise. Google had all the pieces needed to be the leader in the generative AI space—it had the necessary knowledge, skills, and technology (it was a group of Google’s software engineers who created the transformers which are the bedrock of every large language model), the required computing power, and the vast amount of data needed to train these AI models.

But it was OpenAI that started the current AI boom, and Google had to catch up. The tech giant initially responded with Bard until it created a more coherent strategy centred around the Gemini family of AI models, which are now being integrated into every product and service Google offers.

What was presented at this year’s #MadeByGoogle conference is another step in putting Gemini everywhere.

#MadeByGoogle is Google’s annual conference focusing on new hardware products Google will release in the near future. This year, Google unveiled the new Pixel 9 smartphones, which include a base Pixel 9 and the more powerful Pixel 9 Pro and Pixel 9 Pro XL, alongside a foldable phone, the Pixel 9 Pro Fold. We also saw the new Pixel Watch 3 and Pixel Buds Pro 2.

However, what interested me more was the new AI features Google is bringing to their phones.

The biggest reveal from Google is the new Gemini Live assistant, which will become the default assistant on Pixel 9 phones, replacing Google Assistant. Gemini Live is a conversational assistant, similar to what OpenAI showcased in May and what Apple is planning to bring to Siri with Apple Intelligence. You can speak to it as you would to a human, asking it to answer questions or perform tasks. Just like we saw earlier this year with ChatGPT and GPT-4o, Gemini Live can be interrupted while it provides an answer.

However, initial responses to Gemini Live suggest that the new AI assistant has been released half-polished and could benefit from more time to add features and fix some bugs. According to Jared Newman from Fast Company, who had the opportunity to spend some time with Gemini Live, the new assistant is a downgrade compared to Google Assistant. Many features that are present in Google Assistant, such as taking notes, searching photos, or playing music from services other than YouTube Music, are not available in the new assistant.

Another aspect that contributed to this half-polished feeling was the live demos, which sometimes did not go as planned. In one demo, intended to showcase how Gemini can read concert dates from a poster and check if the user is free when the artist is in town, the feature failed not once, but twice. It was only on the third attempt that Gemini worked correctly. While I admire Google’s confidence in presenting live demos—a step up from previous demos—the technical problems do not make Gemini Live look good. And that wasn’t the only blunder.

The next set of AI-powered updates focused on photos and the camera. One of the new features coming to the camera on Pixel phones is Add Me (here’s the demo). It is Google’s solution to the problem of including the photographer in a group photo. With Add Me, two photos are taken—one of the entire group and a second in which someone from the group swaps places with the photographer. The AI then stitches both photos together to include everyone in the final image.

The new updates also include a suite of tools designed to make photo editing easier. In a live demo, Google demonstrated how the new AI can automatically frame a photo and fill in the blanks to create a new image that ideally looks better and has improved composition. The demo then showed how easy it is to further reimagine the photo by adding elements like flowers and a hot air balloon in the background. However, this is where the second blunder happened, as the AI generated something that definitely did not look like a hot air balloon.

Google has also added AI to other areas. The weather app now provides plain-language descriptions of the weather. You can take notes from a call, and there’s a new app for screenshots that can extract information from the screenshots and group them. However, for some reason, this new screenshots app is limited to screenshots and cannot ingest photos.

What this year’s #MadeByGoogle event shows is that Google is fully committed to Gemini. The company is deep into its Gemini Era and will continue integrating more AI features into all of its products and services, even if those features aren’t quite ready for release (see the Google AI Overview disaster). There were new features demoed that failed live on stage for millions to see and are now available on Pixel 9 phones. It seems Google, like many other companies, both big and small, is now throwing whatever they can at the wall to see what sticks. While this approach may eventually lead to effective solutions and applications, it’s currently turning people off anything that has “AI” in the name.

In the meantime, I’m sure Google engineers will be busy ironing out all the kinks and adding more functionality to Gemini Live and other products and services to make them at least usable most of the time.

Now, let’s see what Apple will bring to the table in autumn with Apple Intelligence.

If you enjoy this post, please click the ❤️ button or share it.

Share

Do you like my work? Consider becoming a paying subscriber to support it

Become a paid subscriber

For those who prefer to make a one-off donation, you can 'buy me a coffee' via Ko-fi. Every coffee bought is a generous support towards the work put into this newsletter.

Your support, in any form, is deeply appreciated and goes a long way in keeping this newsletter alive and thriving.

🦾 More than a human

A.L.S. Stole His Voice. A.I. Retrieved It.
Four years ago, ALS took away Casey Harrell's ability to speak. Now, thanks to four electrode arrays implanted in his brain and AI to decode brain signals, Harrell has been able to speak again, albeit with an AI voice reconstructing his own. Scientists involved in the study were surprised by how quickly the implant started to work—just three weeks after the surgery and with brief training, the implant began picking up a 50-word vocabulary with 99.6% accuracy. Over eight months, Harrell was able to utter nearly 6,000 unique words, with the device maintaining a 97.5% accuracy rate, a significant improvement over previous studies, which achieved roughly 75% accuracy. The implants not only allowed Harrell to reconnect with loved ones and express himself more fully but also offer hope that others suffering from ALS might regain the ability to communicate freely.

Brain-Computer Interfaces: The Promise of Technological Telepathy and More Conrad Gray · Jan 29 Read full story

New biomaterial regrows damaged cartilage in joints
Northwestern University scientists have developed a new bioactive material that successfully regenerated high-quality cartilage in the knee joints of a sheep, an animal whose cartilage is similar to that of humans. Within six months, researchers observed new cartilage growth, including natural biopolymers like collagen II and proteoglycans, which contribute to pain-free mechanical resilience in joints. This new material could potentially be used in the future to prevent knee replacement surgeries, treat degenerative diseases like osteoarthritis, and repair sports-related injuries.

Researchers develop an implantable device to detect and respond to opioid overdose
iSOS is a new implantable device designed to autonomously detect and reverse opioid overdoses by continuously monitoring vital signs and delivering naloxone when necessary. Equipped with sensors to monitor respiratory and heart rates, the device aims to be a life-saving solution, particularly for individuals who use opioids alone and are at high risk of overdose. In preclinical trials, iSOS effectively detected and reversed overdoses in animal models, showing promise as a next-generation antidote platform. Researchers are now focused on optimizing and miniaturizing the device for future human trials.

🧠 Artificial Intelligence

Grok-2 Beta Release
xAI has released a beta version of their newest model, Grok-2. Alongside Grok-2, xAI also released a smaller model, Grok-2 mini, which follows a recent trend of releasing smaller models like Gemini Flash or GPT-4o-mini. According to xAI's benchmarks, Grok-2 performs on par with models like GPT-4 Turbo and Claude 3.5 Sonnet. Both models are now available to Grok users on X. Additionally, xAI is experimenting with integrating FLUX.1, a recently released open text-to-image model, into Grok to enable it to generate images.

Musk’s ‘fun’ AI image chatbot serves up Nazi Mickey Mouse and Taylor Swift deepfakes
xAI has added image generation capabilities to Grok, which has already sparked controversy. The new text-to-image feature has minimal safety measures and, unlike other AI platforms, has no issues generating offensive depictions of political figures and celebrities. Controversial images include Donald Trump inappropriately depicted near the World Trade Center, the prophet Muhammad holding a bomb, and sexualized images of Taylor Swift, Kamala Harris, and Alexandria Ocasio-Cortez. Grok's system has also generated content featuring copyrighted characters, such as Disney’s Mickey Mouse, in highly inappropriate contexts—something most other platforms also prohibit.

GPT-4o System Card
Three months after releasing GPT-4o, OpenAI has released the system card for the model. The paper continues the trend of revealing nothing about the model architecture and gives minimal attention to the training data. The vast majority of the paper focuses on safety and describes how OpenAI employed extensive red teaming and risk mitigation strategies, especially for speech-to-speech interactions. The card also highlights societal impacts, such as potential emotional reliance and the model’s use in scientific and healthcare applications, while outlining ongoing evaluations and future safety improvements. Additionally, the paper evaluates GPT-4o using OpenAI’s Preparedness Framework in areas like cybersecurity, biological threats, persuasion, and model autonomy, with GPT-4o classified as medium risk primarily due to its persuasive capabilities.

Nvidia reportedly delays its next AI chip due to a design flaw
Nvidia's upcoming Blackwell B200 AI chips, the successors of the highly sought-after H100 chips, are facing a production delay of at least three months due to a design flaw discovered late in the process, as reported by The Information. This setback impacts major customers like Microsoft and other cloud providers, delaying the widespread availability of these chips until the first quarter of 2024. Despite this, Nvidia has confirmed that production is still expected to ramp up in the second half of the year. The delay could affect the AI cloud infrastructure plans of tech giants like Microsoft, Google, and Meta, who have reportedly placed massive orders for the chips.

New supercomputing network could lead to AGI, scientists hope, with 1st node coming online within weeks
SingularityNET is building a global network of supercomputers designed to host and train the models needed for AGI. The first supercomputer in this network is expected to go online in September 2024, with full completion anticipated by late 2024 or early 2025. SingularityNET’s CEO, Ben Goertzel, stated that this network will enable a shift towards more advanced AI capabilities, including continuous learning, generalization, and AI self-modification. The ultimate goal is to develop a system that not only processes vast amounts of data but also engages in non-imitative machine thinking, utilizing multi-step reasoning algorithms and dynamic world modelling.

Paid Apple Intelligence features are likely at least 3 years away
Apple is reportedly planning to charge up to $20 per month for some advanced features of Apple Intelligence, but this won't happen for at least the next three years, according to recent reports.

Apple Intelligence is different Conrad Gray · Jun 16 Read full story

AI existential risk probabilities are too unreliable to inform policy
AI Snake Oil published an exhaustive article critiquing the reliability of AI existential risk forecasting. The article, the first in a series, highlights the inadequacies of traditional methods due to a lack of historical data and the speculative nature of such risks. It points out how biases within AI and forecasting communities can inflate risk estimates and cautions against policies based on speculative probabilities akin to Pascal's Wager, which could prompt disproportionate actions. Instead, the authors advocate focusing on tangible AI milestones and adopting evidence-based policies that remain flexible to new information, urging a more balanced approach to AI policy that avoids drastic measures based on uncertain forecasts.

Leaked Documents Show Nvidia Scraping ‘A Human Lifetime’ of Videos Per Day to Train AI
A report from 404 Media revealed that Nvidia is scraping 80 years' worth of videos each day from platforms like YouTube, Netflix, and others to train AI models. According to the report, managers told employees who raised ethical or legal concerns that the practice had approval from the "highest levels of the company," while other managers said that the copyright issues would be dealt with later.

Geekbench AI 1.0
Geekbench, a popular software for benchmarking CPU performance, has released a new benchmark specifically designed to measure the performance of various devices in AI workloads. As with other Geekbench benchmarks, results can be shared and viewed in the Geekbench Browser, which now has a dedicated section for AI results.

If you're enjoying the insights and perspectives shared in the Humanity Redefined newsletter, why not spread the word?

Refer a friend

🤖 Robotics

Clear for takeoff? Amazon gets green light to test-fly delivery drones in UK
The UK's Civil Aviation Authority (CAA) has granted Amazon permission to test-fly drones beyond a human controller’s line of sight, paving the way for using the drones to deliver packages to homes. Amazon is one of six organisations participating in a CAA-led trial, with other projects including the inspection of offshore wind farms, air traffic control, policing, and delivering emergency medical supplies.

How robots are building Amazon's future Conrad Gray · December 4, 2023 Read full story

▶️ The Real Reason Robots Shouldn’t Look Like Humans (1:27:19)

When we imagine what a robot should look like, we usually envision a humanoid robot or some kind of wheeled machine. But those are just a few shapes a robot can take. In this video, Veritasium explores robots of different, sometimes unusual, shapes and how their unique and clever designs open new possibilities and applications for robotics. This video is a compilation of Veritasium’s previous videos, in which he explored robots inspired by vines, a jumping robot, tiny robots solving big mazes in seconds, as well as bendable and soft robots.

Apple Aiming to Launch Tabletop Robotic Home Device as Soon as 2026 With Pricing Around $1,000
According to Mark Gurman, Apple is moving forward with its robotic home device, which reportedly resembles an iPad attached to a robotic arm. The device is envisioned as a smart home command centre, videoconferencing machine, and remote-controlled home security tool, according to sources who requested anonymity as the project is not yet public. Codenamed J595, the project was approved by Apple’s executive team in 2022 but has only recently begun to ramp up development. Apple has now prioritized the device's development, aiming for a debut as early as 2026 or 2027, with a target price of around $1,000.

Chinese drones will fly trash out of Everest slopes
Later this year, Nepal will deploy a fleet of heavy-lifter drones to remove trash from the slopes of Ama Dablam, a 6,142-meter-tall mountain south of Mount Everest. Tests have shown that the DJI FlyCart 30, a long-distance heavy-lifter drone, can transport 234 kg per hour between two camps—a task that previously took six hours and required at least 14 porters. On their way up the mountain, the drones will carry ropes, ladders, and other supplies. Officials hope the drones will not only make the transport of trash and supplies easier but also reduce casualties.

Researcher designs robot to help children cross streets safely
Hovannes Kulhandjian, an associate professor at Fresno State, has developed CrossBot, a robotic crossing guard designed to address the shortage of human crossing guards in the Bay Area. The robot monitors crosswalks and ensures children's safety by detecting vehicles and signalling when it's safe to cross. Kulhandjian believes the robot could last a decade and cost a school district about $15,000 to $17,000 to purchase. However, critics have raised concerns about its reliability and the loss of human interaction.

Rodney Brooks’ Three Laws of Robotics
In this post, Rodney Brooks, a professor emeritus of robotics at MIT and cofounder of iRobot, Rethink Robotics, and Robust AI, presents his three laws of robotics. These laws focus on the promise given by a robot's appearance, preserving people's agency, and ensuring that robots are reliable and robust. They are based on Brooks' nearly 50 years of experience in the field, during which he has designed, produced, and shipped robots on a massive scale.

Flying robots could search buildings after disaster
A team at Carnegie Mellon University is exploring the use of flying robots to coordinate rescue efforts in abandoned buildings after disasters. The idea is to deploy multiple drones into the building, with each robot communicating with the others to coordinate their actions and explore different rooms, thereby making the search for survivors more efficient.

🧬 Biotechnology

Biotech firm Recursion to buy smaller peer Exscientia for $688 million
Recursion Pharmaceuticals, an AI drug discovery company, has agreed to acquire its smaller rival Exscientia in an all-stock deal valued at $688 million. Recursion, founded in 2013 and currently in phase 2 of clinical trials, specializes in treatments for rare diseases and certain cancers. The acquisition of Exscientia, which also operates an AI-driven drug discovery platform, will bolster Recursion's drug development pipeline and expand its partnerships with major pharmaceutical companies, including Sanofi and Merck. The deal is expected to close in early 2025. Marina T Alamanou has a great write-up about these two companies in the latest issue of her MetaphysicalCells newsletter if you want to learn more about them.

Thanks for reading. If you enjoyed this post, please click the ❤️ button or share it.

Share

Humanity Redefined sheds light on the bleeding edge of technology and how advancements in AI, robotics, and biotech can usher in abundance, expand humanity's horizons, and redefine what it means to be human.

A big thank you to my paid subscribers, to my Patrons: whmr, Florian, dux, Eric, Preppikoma and Andrew, and to everyone who supports my work on Ko-Fi. Thank you for the support!

My DMs are open to all subscribers. Feel free to drop me a message, share feedback, or just say "hi!"

Message Conrad Gray