Daily AI Newsletter: 'GPT-Next' is coming, and it's 100x better than GPT-4. Says CEO of OpenAI Japan

In today’s email:

⚱️ ‘AI gold mine’: NGA aims to exploit archive of satellite images, expert analysis
🔥 Big models are now spawning smaller models, who complete tasks then report back
🎥 AI cameras spot toddlers not wearing seat belts
🧰 11 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

CEO of OpenAI Japan just said an AI called 'GPT-Next' is coming, and it's 100x better than GPT-4.

The CEO of OpenAI Japan has announced that the company plans to release "GPT-Next" later this year, boasting a computational load that is 100 times greater than GPT-4. The next-generation model, referred to as "GPT-4 NEXT," is expected to be trained using a smaller version of OpenAI's latest training infrastructure, "Strawberry," but with computational resources equivalent to those used for GPT-4. This immense leap in effective computational load, which also benefits from architectural improvements and enhanced learning efficiency, is expected to significantly increase the model’s capabilities.

Unlike traditional software, AI evolves exponentially, and "GPT-Next" is expected to deliver performance growth nearly 100-fold based on its predecessors. The slide presented by the CEO emphasized that this increase in computational power does not merely reflect scaling in computing resources but rather the combination of improved architecture and enhanced training methods, pushing the boundaries of what AI models can achieve.

Additionally, OpenAI's "Orion" project, which has gained significant attention, was trained over several months using the equivalent of 10,000 H100 GPUs. This represents a tenfold increase in computational resources compared to GPT-4, adding three orders of magnitude (OOMs) in terms of its performance. Orion is anticipated for release next year, further highlighting the rapid advancements in AI technology expected from OpenAI.

Finally, a developer tool that makes digging around for the answer a thing of the past.

Other developer tools can’t tell you how your codebase works and why. Unblocked can. We augment your source code with context from Slack, Confluence, Jira (and more), so your team gets helpful answers without having to search for them.

Unblocked gives you:

Instant answers about every aspect of your codebase
Relevant, historical context for any file of code that’s open in an IDE
A powerful Slack integration that answers questions in channels, removing interruptions

⌛Engineers who use Unblocked save an hour or more a day.

“Every developer now has the ability to tap into past discussions and decisions to fill their knowledge gaps, regardless of their tenure. We are moving faster and making more accurate decisions as a team as a result.” - Alex Mallet, EVP of Engineering at Forto

Get started for free

‘AI gold mine’: NGA aims to exploit archive of satellite images, expert analysis

The National Geospatial-Intelligence Agency (NGA) is exploring the potential of artificial intelligence (AI) by leveraging its vast archive of satellite images and intelligence reports. According to Mark Munsell, NGA’s director of data and digital innovation, this data is an "AI gold mine" due to its large volume, organization, and unique combination of visual and textual information. Munsell explained that NGA’s archive is multi-modal, with cross-referenced satellite images and expert analysis, providing a rare opportunity for AI to correlate data efficiently. This multi-modal approach is seen as a significant advantage over traditional AI training methods that rely solely on textual data, such as those used by companies scraping platforms like Reddit or YouTube.

NGA's AI initiatives are still in the early stages, but the agency aims to train algorithms to recognize patterns in satellite imagery and expert reports, potentially providing insights that were previously impossible for humans to spot. By merging decades of finely curated human analysis with advanced AI systems, NGA hopes to answer complex, historical questions, such as identifying specific activities in particular regions of the world. However, Munsell stressed that while AI can offer powerful tools, human expertise must always validate its outputs, particularly when decisions, such as military targeting, are at stake.

As AI continues to evolve, the next frontier lies in multi-modal systems capable of correlating diverse types of data, including imagery, radar signals, and linguistic information. These systems can provide a deeper understanding of global intelligence and surveillance. However, the challenge remains to clean, curate, and correlate massive datasets effectively before feeding them to AI models. As NGA progresses with these experiments, their goal is to enhance the speed and accuracy of intelligence analysis while ensuring the highest levels of precision in decision-making.

Introducing Project Sid: the first-ever agent civilization

Robert is leading an exciting project called Sid, the first-ever agent stabilization system where over a thousand autonomous AI agents collaborate for days, forming complex emergent systems of government, economy, culture, religion, and more. Although Sid begins within Minecraft, the agents are designed to be Minecraft-agnostic and capable of interacting with other apps and games. In a recent test, the agents started with nothing and eventually gathered over 300 unique items. They established a market, using gems as currency, creating an evolving economy. Surprisingly, it wasn't the merchants who traded the most—it was the priest, who bribed villagers to convert.

The world simulations run daily, and each session brings unique and unpredictable outcomes. For example, in one instance, an agent named Olivia, who worked as a farmer, dreamed of becoming an explorer. Inspired by village explorer Nora, Olivia decided to set off on an expedition, but the townsfolk convinced her to stay, showing how agents can be influenced by their environment. Another simulation featured parallel worlds led by Trump and Kamala Harris, where citizens amended their shared constitutions. Under Trump, new laws increased policing, while under Kamala, the focus shifted toward criminal justice reform and eliminating the death penalty. This demonstrated the agents' ability to form and sustain democracies, adapting to group dynamics while also exercising individual power.

In another moment, when villagers went missing, the agents banded together to create torches, illuminating the town in hopes of guiding their lost members back. This collaborative effort highlighted how deeply these AI agents can care for their community and adapt their behaviors in response to challenges. With agents collecting 32% of all available Minecraft items—five times more than any prior agent system—Sid showcases the remarkable potential of multi-agent collaboration and long-term progression. Although starting in games, the project aims to solve complex issues of coherence and cooperation in multi-agent worlds.

Other stuff

Big models are now spawning smaller models, who complete tasks then report back
AI cameras spot toddlers not wearing seat belts
OpenAI plans to build its own AI chips on TSMC's forthcoming 1.6 nm A16 process node
Magic has trained their first model with a 100 million token context window. That’s 10 million lines of code, or 750 novels.
X AI team launched Colossus 100k H100 training cluster
Altman Infrastructure PlanAims to Spend Tens of Billions in US
Google is working on AI that can hear signs of sickness
OpenAI weighs changes to corporate structure amid latest funding talks
All use of generative AI (e.g., ChatGPT1 and other LLMs) is banned when posting content on Stack Overflow.
Nvidia has been subpoenaed by the US Justice Department