Hello and welcome to Weekly News Roundup Issue #462. This week saw the release of several large language models, and we will take a closer look at the new offerings from OpenAI, Google, and others.

In other news, ex-DeepMind employees released Udio, a very good AI music generator. Some people are looking forward to it while others, especially musicians, are not happy. Over 200 of them signed an open letter saying that AI poses an existential threat to their livelihoods. Additionally, Intel released Gaudi 3, an AI accelerator to compete with Nvidia’s H100 GPUs and the UK’s antitrust enforcer sounds the alarm over Big Tech’s grip on generative AI.

Over in the world of robotics, Tesla is set to unveil its robotaxi on August 8th and a new bunch of humanoid robots show what they can do.

To top off this week’s issue, we also have new advances in brain-computer interfaces and a $10 million fund to develop plants that extract metals from the soil. I hope you enjoy this week’s issue!

This last couple of days saw many players in the AI industry, both big and small, releasing or teasing their upcoming language models.

Let’s start with OpenAI. A couple of days ago, OpenAI released an updated version of GPT-4 with Vision. In a tweet, the company said the “majorly improved” version of its flagship model is now available via OpenAI API and will be rolled out to ChatGPT soon. However, OpenAI did not explain what “majorly improved” means. Initially, there were no official benchmarks comparing the new version to the previous ones. In the absence of any official numbers, the community took matters into their own hands and ran benchmarks. The results show that this new version of GPT-4 with Vision is indeed better than the previous one but not enough to earn to be called “majorly improved”. In some tests, it falls behind Anthropic’s Claude 3 Opus while in others it exceeds every other model available.

Two days after releasing the new GPT-4 with Vision, OpenAI published official benchmarks in a GitHub repository. The results show an uplift in performance consistent with other benchmarks, but I still don’t think this new model deserves the label "majorly improved."

The same day OpenAI released an updated version of GPT-4 with Vision, Google made Gemini 1.5 Pro available in over 180 countries. Now available through the Gemini API in a public preview, Gemini 1.5 Pro features native audio and speech understanding capabilities along with a new File API for easier file handling. It also includes system instructions to better tailor the model for specific use cases and returns responses in JSON format for simpler integration with other services or programs, amongst other improvements.

Those were models released by the big players, so now let’s take a look at what other models have been released in the last few days.

Mistral AI has released Mixtral 8x22B, a bigger version of their Mixture of Experts (MoE) model Mixtral 8x7B which itself is a very capable model achieving the level of GPT-3.5 in many benchmarks. Mixtral 8x22B outperforms its predecessor and offers very good performance, according to benchmarks. Mixtral 8x22B weights are open and can be downloaded from HuggingFace.

One model that was released at the beginning of the month and flew under my radar is Command R+ from Cohere. This large language model is designed specifically for real-world enterprise applications, understands ten languages, and is optimized for advanced Retrieval Augmented Generation (RAG). According to Cohere’s benchmarks, Command R+ is better than competitors at understanding documents uploaded to the model and performing tasks such as summarization. The Chatbot Arena Leaderboard ranks Command R+ comparably to GPT-4 and Claude 3 Sonnet in terms of performance. Cohere also claims their latest model is more cost-effective than GPT-4 or Mistral-Large while delivering similar, if not superior, performance.

Command R+ is currently available on Microsoft Azure and through Cohere’s API. It will be soon available on Oracle Cloud Infrastructure (OCI), as well as additional cloud platforms in the coming weeks.

These models join previously released models like xAI’s Grok-1 or Databrick’s DBRX to the active large language model’s scene. The gaps between different models are narrowing, with many—including open-weight models—now performing at or above the level of GPT-3.5, the model that powers the free version of ChatGPT. We even have a new leader with Anthropic’s Claude 3 Opus dethroning GPT-4 after OpenAI’s model reigned for almost a year.

And there is more to come.

At an event in London this week, Meta confirmed that Llama 3, the successor to the highly popular open-weight Llama 2 model, will be released in the next month. However, according to a report published in The Information, Meta might release smaller versions of Llama 3 ahead of the full release. Llama 2 made a big impact on the open-source scene and it will be interesting what Llama 3 brings to the table.

Google and Anthropic can release updated versions of their models at any time, further reshuffling the state of the game.

Lastly, there is the highly anticipated GPT-5, rumoured to be released by the end of this year. There is a chance that OpenAI could release partially trained GPT-5 as GPT-4.5 ahead of the full release.

One thing is for sure - the AI race is not slowing down.

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

First-in-human implantation of bionic device to halt Crohn’s disease
In a world first, clinicians have implanted a nerve stimulation device in a patient suffering from Crohn's disease to prevent post-surgical inflammation. This device, which stimulates the vagus nerve, could potentially keep Crohn's disease at bay, reducing the need for repeated surgeries and continuous, lifelong medication.

“Universal” BCI lets anyone play games with their minds
Engineers at UT Austin have developed a "one-size-fits-all" non-invasive brain-computer interface (BCI) that works without extensive calibration, potentially eliminating a major hurdle to using these systems. Using this new approach, volunteers quickly learned how to control computers using only their thoughts. Some were even able to play a racing game without using any physical controllers.

Brain-cell transplants are the newest experimental epilepsy treatment
A new method of treating epilepsy has been developed, involving the injection of brain cells grown from stem cells directly into the brain. Initial tests have shown a dramatic decrease in seizures and improved cognitive capabilities in patients. This treatment is also a big win for stem cell therapies, whose reputation has been recently tarnished by shady health clinics.

🧠 Artificial Intelligence

New AI music generator Udio synthesizes realistic music on demand
Meet Udio - a new AI music generator created by former DeepMind employees who previously worked on Google's AI music generator, Lyria. Similarly to other text-based AI generators, all you need to do is describe the type of music you want to create, and Udio will generate it. The results are impressive, and you can explore Udio's creations on the website without needing to create an account. While some welcome Udio and similar tools, believing they will help people without musical training to create music, others, including professional musicians, are not pleased with the rise of music-generating AIs (more on that later in this issue).

The EU AI Act – A Guide for Developers
The EU AI Act has become a law and AI companies which want to operate in the EU need to adhere to it. This post on TheSequence does a good job at briefly explaining the main points of the EU AI Act, who needs to comply, what new responsibilities are placed on AI companies wishing to operate within the EU, and what the new regulations mean for innovation, open-source and the wider machine learning community and industry.

Nvidia Blackwell Perf TCO Analysis - B100 vs B200 vs GB200NVL72
Nvidia released its new line of top GPUs, Blackwell, a couple of weeks ago, promising massive improvements in AI performance—up to 30x over the previous generation of GPUs. In this post, SemiAnalysis takes a very close look at these claims with a detailed performance analysis of Nvidia’s latest GPUs for AI.

▶️ But what is a GPT? Visual intro to Transformers (27:13)

3Blue1Brown, one of the best math channels on YouTube, has released two videos that visually explain how transformers work in a detailed yet easy-to-understand manner. The first video focuses on transformers, while the second explains how the attention mechanism operates in large language models. Together, these videos help better understand how large language models work.

Billie Eilish, Pearl Jam, 200 artists say AI poses existential threat to their livelihoods
The Artist Rights Alliance (ARA) published an open letter critical of AI, signed by over 200 musical artists including Pearl Jam, Nicki Minaj, Billie Eilish, Stevie Wonder, the estate of Frank Sinatra and many more well-known artists. In the letter, the artists urge AI developers, technology companies, platforms, and digital music services to stop using AI in ways that "infringe upon and devalue the rights of human artists." A tweet from the ARA added that AI poses an "existential threat" to their art. The full letter and the list of artists who signed it are available here.

UK’s antitrust enforcer sounds the alarm over Big Tech’s grip on GenAI
The UK's Competition and Markets Authority (CMA) has issued a warning about the growing influence of major tech companies - Google, Amazon, Microsoft, Meta, and Apple (collectively referred to as GAMMA) - across the entire AI value chain., from compute, data and model development to partnerships, release and distribution platforms. “The essential challenge we face is how to harness this immensely exciting technology for the benefit of all, while safeguarding against potential exploitation of market power and unintended consequences,” said Sarah Cardell, CEO of the CMA.

Intel’s “Gaudi 3” AI accelerator chip may give Nvidia’s H100 a run for its money
Intel unveiled the Gaudi 3, an AI accelerator chip positioned as a competitor to Nvidia's popular H100 GPU. Intel claims that Gaudi 3 offers four times the AI compute performance for BF16, a 1.5 times increase in memory bandwidth, and double the networking bandwidth for massive system scale-out compared to its predecessor, Gaudi 2. Additionally, Intel claims that Gaudi 3 delivers 50% faster training times for large language models like OpenAI’s GPT-3 and Meta's Llama 2, compared to Nvidia’s H100, as well as superior inference performance.

A Pure C Implementation of GPT-2 in <1000 Lines
Andrej Karpathy, a former AI researcher at OpenAI and Tesla, demonstrates how simple the code for a large language model can be with llm.c, a pure C implementation of the GPT-2 model with 124 million parameters in about 1100 lines of code, designed to be trained on a CPU. To put this into perspective, the code to generate the page you are reading right now is much, much larger. The apps we use daily have tens of thousands, if not millions of lines of code.

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

Refer a friend

🤖 Robotics

Elon Musk says Tesla will reveal its robotaxi on August 8th
Elon Musk announced in a tweet that Tesla will unveil a robotaxi on August 8th. It's worth noting that a few years ago, Musk mentioned one of the benefits of owning a Tesla with full self-driving capabilities would be the potential to use it as a robotaxi to earn money. That hasn't happened yet, and we'll have to wait until August to see what exactly the Tesla Robotaxi is.

▶️ Sanctuary AI - Phoenix at Human-Equivalent Speed (0:40)

Another humanoid robotics company has uploaded a video demonstrating what their robot is capable of. This time, it's Sanctuary AI showcasing the speed, precision, and fluidity of movements of their humanoid robot, Phoenix.

Toyota’s Bubble-ized Humanoid Grasps With Its Whole Body

Meet Punyo, a humanoid robot developed by the Toyota Research Institute to experiment with manipulating objects using the robot’s entire body. Unlike other humanoid robots that use only their hands, Punyo can also utilize its arms and chest, mimicking how a human uses their whole body while carrying things. On top of that, Punyo is also a squishy robot, covered with air bladders, making it the closest we have right now to real-life Baymax from Big Hero 6.

▶️ UBTECH Robotics x Baidu: Accelerates the Future of Intelligent Automation at Baidu Create 2024 (2:31)

Do you remember that viral video showing Figure humanoid robot equipped with OpenAI’s models? In this video, Chinese company UBTECH Robotics shows a similar robot, powered by Baidu’s Ernie large language model.

🧬 Biotechnology

Tiny Biosensor Is Just a Cup, a Membrane, and a Magnet
Chinese researchers have developed a prototype biosensor that consists solely of a membrane with a magnet. When paired with a reader device, this chipless and battery-less implant offers a cheaper and more accessible way to monitor blood sugar levels and other biochemical and physiological markers.

U.S. Department of Energy Announces $10 Million to Explore Using Plants to Extract Critical Materials from Soil to Support Domestic Supply Chain
The US Department of Energy has announced a $10 million initiative, managed by the Advanced Research Projects Agency-Energy (ARPA-E), to explore extracting nickel from soil using plants, a process known as phytomining. This effort aims to strengthen domestic critical materials supply chains, reduce nickel imports, and support clean energy technologies crucial for economic and national security.

💡Tangents

Ukraine Is the First “Hackers’ War”
The ongoing war in Ukraine is a preview of what modern conflict will look like. In addition to traditional heavy military equipment such as tanks, artillery, and jet fighters, the battlefields in Ukraine are also filled with all kinds of drones—from modified off-the-shelf models to military drones and 3D-printed DIY drones. This article explores this topic with a focus on the arms race between drone operators and the electronic warfare methods used to disable 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