gpt2-chatbot and OpenAI search engine - Weekly News Roundup - Issue #465

Plus: Med-Gemini; Vidu - Chinese answer to OpenAI's Sora; the first race of Abu Dhabi Autonomous Racing League; deepfaking celebrities to teach math and physics; and more!

Hello and welcome to Weekly News Roundup Issue #465. This week, we will focus on gpt2-chatbot, a mysterious AI model that came out of nowhere, gained the attention of the AI community and then vanished. We will also discuss the rumoured upcoming release of OpenAI’s search engine.

In other news, Google DeepMind released a paper describing Med-Gemini, a promising AI model tailored for medical tasks. Chinese researchers, meanwhile, revealed Vidu, a text-to-video model and a response to OpenAI’s Sora. Over in robotics, DARPA tested autonomous racing tanks while proper autonomous racing cars raced in Abu Dhabi.

I hope you’ll enjoy this week’s issue!

---

In the last week, the AI community was perplexed by a mysterious model named gpt2-chatbot. The model has been available on LMSYS Chatbot Arena, an AI benchmark platform comparing and ranking AI models.

It challenged the best currently available models such as GPT-4 and Claude 3 Opus, showing remarkable capabilities and performance. Some have shown that gpt2-chatbot is very good at solving hard math questions, while others noted its responses are better than the responses from GPT-4.

The gpt2-chatbot caught the attention of the AI community, which now wanted to know more about the mysterious model. OpenAI quickly became the primary suspect. The name 'gpt2-chatbot' suggests a connection to OpenAI’s GPT series of AI models, although this alone is not strong evidence. However, more compelling links were found, such as the use of a GPT-4 tokenizer and system prompts mentioning ChatGPT and GPT-4. While this is not definitive proof, the consensus is that OpenAI was likely involved with the model. It might have been a covert test of an advanced version, possibly GPT-4.5 or even GPT-5. It could even be GPT-2 all along but fine-tuned with modern datasets. Sam Altman has a soft spot for GPT-2 after all.

gpt2-chatbot and its links to OpenAI add more fuel to the growing hype surrounding the imminent release of OpenAI’s next model, which Sam Altman has been teasing recently. In an interview with MIT Technology Review, when asked if he knew when the next version of GPT is going to be released, Altman just said “Yes”. Speaking at Stanford University, Altman said that “GPT-4 is the dumbest model any of you will ever have to use again by a lot”. At the same event at Stanford University, Altman also reiterated OpenAI’s commitment to iterative releases. As he said, he does not want to surprise and scare people with what AI can do. It is better to release new models gradually, step by step, than by releasing a massive leap forward. These remarks reinforce the possibility that we could see the release of GPT-4.5 soon before the full release of GPT-5 sometime at the end of this year.

Another story making waves this week was the rumour that OpenAI might be gearing up to release its own search engine. Over at X, Jimmy Apples shared a list of new subdomains created by OpenAI, which includes search.chatgpt.com. So far, there is nothing on that page. In a follow-up tweet, Apples says that OpenAI may hold an event on May 9th. If that were to happen and OpenAI were to announce a new search engine, it would likely occur just days before the Google I/O conference, potentially shifting all attention to OpenAI’s new product and stealing thunder from Google.

The rumoured search engine is not something new. Back in February, The Information leaked that OpenAI is working on a search engine to directly compete with Google. Interestingly, OpenAI has recently launched its redesigned webpage, which now looks a lot like a search page. Is OpenAI cheekily telegraphing something here?

In any case, May is shaping up to be an interesting month in tech. It will start with an Apple event on May 7th, where Apple is expected to release the new iPad lineup powered by the new M4 Apple Silicon chip with an upgraded Neural Engine to boost the performance of AI tasks. A week later, Google will hold the annual I/O conference where the tech giant from Mountain View will present their latest AI products and services. In the meantime, OpenAI might release a surprise product at any time.

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

---

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

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

Dutch startup to test hearing via brain-computer interface
MindAffect, a health tech startup based in the Netherlands, proposes a way of testing hearing and detecting any hearing impairments. By combining brain-computer interfaces (BCIs) and AI, their solution can detect any hearing impairments within minutes and only requires the patient to watch a video while the device monitors their brainwaves and compares them with what the patients are hearing and what they should be hearing.

🧠 Artificial Intelligence

Capabilities of Gemini Models in Medicine
Researchers from Google DeepMind have published a paper describing Med-Gemini, a version of Google's top large language model, Gemini, tailored for healthcare applications. According to the paper, Med-Gemini outperforms GPT-4 in medical tests and matches or exceeds the performance of human doctors in some scenarios. The model features several innovations, like using Gemini’s very long context window or checking how confident the model is in its answers. If the model is not confident in its answer, it can search the Web for more information. Med-Gemini is still in the research phase and will need “rigorous evaluation” before it can be deployed. Google DeepMind won’t be open-sourcing or releasing weights for Med-Gemini.

GitHub Copilot Workspace: Welcome to the Copilot-native developer environment
GitHub announced Copilot Workspace, a new coding workspace built around the GitHub Copilot coding assistant. Similar to other similar AI coders (like Devin) that have recently started to pop up, the user only has to describe in plain language what the code needs to do. The AI will then propose a plan on how to implement the requested change and, once the plan is approved, write the code. GitHub claims Copilot Workspace will make professional programmers more efficient while lowering barriers to entry for everyone else. If you are interested in a thorough review of the announcement, I recommend this 30-minute-long video (tl;dw - it does not look good for GitHub Copilot Workspace).

▶️ AI Is Turning into Something Totally New | Mustafa Suleyman (22:01)

In this TED Talk, Mustafa Suleyman, co-founder of DeepMind and Inflection AI, and now the head of recently formed Microsoft AI, asks the question: What are we actually building with AI? Suleyman calls to stop seeing AI as a tool but as something more. He calls to see AI as a new digital species. He argues that framing AI in this way will help us to better understand what we are creating and how to guide the development of AI to avoid catastrophic scenarios and usher in “the most productive decade in human history”, as he said.

▶️ How AI Is Unlocking the Secrets of Nature and the Universe | Demis Hassabis (25:33)

Demis Hassabis, the founder and CEO of Google DeepMind, shares in the conversation with the head of TED Chris Anderson the story of DeepMind and how the company went from making AI play Atari games to solving protein folding problem with AlphaFold. Hassabis believes AI can accelerate scientific discovery in ways that will benefit humanity and unlock the greatest mysteries surrounding our minds, bodies and the universe.

China Unveils Vidu: A Powerful Text-to-Video Generator

Researchers from Shengshu Technology and Tsinghua University unveiled Vidu, a text-to-video model and a competitor to OpenAI’s Sora. Vidu can generate 16-second clips at 1080p just from a text prompt. In contrast, Sora can produce clips up to 60 seconds long. Additionally, when comparing the output of Vidu with that of Sora, OpenAI’s model generates higher quality videos.

Why AI celebrities are teaching kids math and physics
This is an application of deepfakes that I approve and accept. Onlock Learning creates educational shorts featuring various deepfaked celebrities—from Elon Musk and Kim Kardashian to Logan Paul and Donald Trump—explaining basic concepts in math, physics, or engineering. Kabir Knupp, the creator of Onlock Learning, then posts these videos on TikTok, YouTube and Instagram to sneak some learning in between algorithmically served brain rot.

---

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

---

🤖 Robotics

DARPA RACER Speeds Into a Second Phase With Robotic Fleet Expansion and Another Experiment Success

DARPA’s Robotic Autonomy in Complex Environments with Resiliency (RACER) program, which aims to develop autonomous off-road vehicles for military use, has entered its second phase. One of the successfully tested vehicles is a 12-ton autonomous unmanned tank that not only navigates rough terrain easily with terrifying speed.

Microsoft taps Sanctuary AI for general-purpose robot research
Microsoft announced a collaboration with Sanctuary AI, a robotics company best known for its humanoid robot, Phoenix. Through this partnership, Microsoft aims to accelerate the development of general-purpose AI systems good enough to control the robot as it navigates in a real world full of humans. The insights and the data gathered through this process could then be used to help develop artificial general intelligence (AGI).

Ukraine Is Riddled With Land Mines. Drones and AI Can Help
Drones have proven to be a valuable weapon on the battlefields in Ukraine. However, the war will sooner or later end, leaving behind a large number of unexploded ordinance and land mines. In this article on IEEE Spectrum, Jasper Baur writes about how drones equipped with cameras and various sensors can be used to clear minefields or spot unexploded ordinance and help safely remove them before they can cause any harm.

In the first Autonomous Racing League race, the struggle was real

Recently, Yas Marina racing track hosted a rather unusual race. The cars looked like the ones used in Formula 1, but with one major difference—they had no drivers. In the inaugural race of the Abu Dhabi Autonomous Racing League (A2RL), nine teams competed to find whose AI was the best at driving a racing car on a track. The vehicles used were the same Super Formula cars that human drivers use, but they were equipped with onboard computers and cameras. The first race was rough, featuring many spins, crashes, and erratic behaviour, so human drivers can rest easy for now. However, the teams learned a lot and will improve their AI drivers. Perhaps next year, they could challenge Max Verstappen.

Will Human Soldiers Ever Trust Their Robot Comrades?
As more robots enter battlefields, soldiers form complex relationships with their robotic counterparts. This article explores these relationships, their psychological and ethical implications, and what the US military is doing to foster trust between human soldiers and robots, aiming to create effective fighting forces.

🧬 Biotechnology

Improving microbial phylogeny with citizen science within a mass-market video game
Inside Borderland 3, the game developers have hidden a mini-game called Borderlands Science, in which the players were tasked to solve simple mini-challenges. By solving these challenges, over 4 million players helped align DNA sequences to improve microbial phylogeny estimations (the process of determining the evolutionary relationships among different microbial species based on their genetic information). This paper describes the results of the experiment and highlights the potential of using gamification to harness the collective problem-solving capabilities of gamers to solve complex scientific challenges.

---

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

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!"