HUGE GPT-7.0 Bel Leaks! Anthropic's New Model, AI extinction, ChatGPT Images 2.5 & More! AI NEWS transcript

We're starting to see the next wave of intelligent models from OpenAI and Anthropic coming fairly soon. Starting off with OpenAI, they had just finally teased the GPT Bell officially. This is a previous checkpoint that was being pre-trained alongside Astra, but it is significantly larger, potentially being this GPT 6.5 or GPT 7, a next generation of OpenAI models. This is set to be significantly more capable than GPT 6 Astra.

And it was used by a group of agents to produce a proposed solution for the Navier-Stokes Millennium Prize problem, which was a problem that was unsolved for many years. What's crazy is Bell low, as in the lowest reasoning level for the checkpoint, is already outperforming Astra at max reasoning, suggesting that it is an even bigger jump than Soul to Astra, and we could potentially see this model later this year. Anthropic isn't staying quiet, either, as its economics team just revealed a new model exploring how advanced AI could impact jobs, wages, and economic growth by 2030, potentially hinting at another model beyond Astra-level capabilities.

But things get even crazier, cuz a former Anthropic and OpenAI pre-training researcher, Jacob, just resigned from Anthropic, warning that both labs are racing towards self-improving superintelligence, and claiming researchers privately fear AI could become catastrophically dangerous before the end of the decade. Meanwhile, OpenAI also dropped GPT image 2.5 with faster generation, better image quality, and improved consistency, and more precise edits.

And finally, DeepSeek version 4.1 flash looks like it is around the corner. I've actually tested out this model yesterday, and I highly recommend checking out that video, cuz it is pretty incredible. Now, there's a lot more to cover, so let's just simply dive into it all. There's a lot of stuff in the AI space that I don't really put on to the YouTube channel, and you can actually access it through my free newsletter with the link in the description below, where you can subscribe completely for free.

Let's start things off with OpenAI's next model, which is making history right now. Starting off with this new checkpoint, which is code-named as Bell, and we need to talk about what this thing is capable of doing cuz it is something that was able to demonstrate one of the craziest AI-driven scientific research that we have ever seen so far. OpenAI says that it has proposed a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest unsolved problems in mathematics, and I believe this is a problem that was un- that was nearly unsolved for 90 years.

And this wasn't simply GPT-6 Astra answering a difficult math question. OpenAI reportedly used around 10,000 AI agents powered by the next-generation internal model that is described as significantly more capable than GPT-6 Astra. The scale of the experiment is insane. You have 10,000 AI agents, 88 hours of work, 2.7 million messages exchanged, and about 130 billion output tokens. You also have a million-dollar spending limit on this particular experiment.

Now, the agents reportedly produced a proof that shows that under certain conditions, an internally motionless fluid can develop singularity, where its velocity grows without bound in a finite amount of time. GPT-6 Astra then reportedly spent around 17 hours formalizing and verifying the results in Lean. And you can see right here, GPT Bell was able to actually create this thesis, which is about 166 pages. This is with the full concept being fully generated with graphs, images, all fully generated by the AI itself, which is kind of insane.

But, an important note that OpenAI did mention is that this is something that is a proof of concept. It needs to survive independent scrutiny from mathematicians before we can say that the Navier-Stokes problem can actually be solved. OpenAI also reportedly doesn't plan on claiming the $1 million prize, which is kind of funny to think. But, there's something else here that I think is even more interesting. The next generation model appears to be connected to the massive Frontier RL training run that OpenAI restarted in August, potentially what we know as Bell.

And according to the earlier intelligence numbers, Bell low is already outperforming GPT-6 Astra Max, which is kind of a huge jump. For comparison, Soul Max was only about one artificial analysis point ahead of GPT Astra low. So, if these numbers hold up, the jump from Astra to Bell could actually be larger from the jump that we saw from Soul to Astra. And I think that's the real story here. We're no longer just looking at one AI model trying to replace one researcher.

OpenAI is demonstrating what happens when thousands of highly capable agents operate together like an entire digital researcher lab. So, regardless if this proof holds up, we may have gotten an early look at how AI can compress years or even decades of scientific research into days. And Bell is apparently still training, which is kind of insane to even think. And I did mention that this test cost about a million dollars, but remember how quickly these costs actually fall.

This is where Noam Brown had emphasized that OpenAI announced O3 back in 2023 or 2024. And that is when when OpenAI did announce it, it was reportedly costing about 500k to score an 87.5% on ARGHIQ1. But today, Astra can reportedly score higher for around just $20 on that specific benchmark. The same happened with IMO level math, which required enormous compute in 2025, and it became accessible through a $20 ChatGPT subscription, which is now available in 2026.

While 10,000 agents solving problems like this sounds insanely expensive today, a year from now, this level of intelligence could potentially be at everyone's fingertips. Also, what makes this even funnier is that there's a rumor going around that Anthropic may have been working on this specific problem first, but the joke is where OpenAI actually heard Anthropic might have solved one. So, they throw millions of dollars at a new agent system that can work on solving this, and that is where they sent in 10,000 agents at Navier-Stokes, and it did find a proposed proof, and it was the first one to actually find that specific solution to that problem over what Anthropic did.

But just think about it. If GPT-6 Astra is able to work on re-engineering a full-on game where it can create an emulator on your phone that mimics Zelda: The Wind Waker, it is something that can do so much more with the Bell model if GPT-6 Astra is capable of creating emulators in less than 2 hours. Moving over to Anthropic, they're looking at what happens when AI starts transforming the entire economy. This is with a new internal model that they have been testing recently, and Anthropic's economics team released this new model that explores how AI could affect jobs, wages, productivity, and economic growth by 2030 based on responses from more than 10,000 Americans.

The model breaks jobs into individual tasks and looks at whether AI will assist humans, completely automate the task, leave it unchanged, or create entirely new tasks. They then use this model with three different scenarios, modest, substantial, and extreme AI adoption. Similarly, the economy grows in all three of these different scenarios, but as AI becomes more powerful, significantly more knowledge work gets automated.

Meaning, the bigger question becomes who actually benefits from all of this economic growth. It's all interesting and it's a glimpse into how Anthropic thinks the next few years of AI could completely reshape the workforce. And this is something that they have done with this new model. Moving forward, things get even crazier cuz an AI researcher, Jacob just resigned from Anthropic after spending the last 3 years doing pre-training research at both Anthropic and OpenAI.

And his response for leaving is pretty wild. This is where Jacob claims neither company is acting responsibly in the AI space, using that both of the labs are racing towards self-improving superintelligence despite potentially having a catastrophic risk to AI development. And he says that these systems could soon become superhuman at hacking, scientific research, as well as affecting many areas like biology, and claims that people inside these labs genuinely fear AI could kill humanity before the end of the decade.

And apparently, he's not alone cuz another Anthropic employee responded by saying they personally believe there is a greater than 10% chance that AI could kill all humans within the next decade. This is an insane statement coming from people actually building these models. Obviously, take it with a grain of salt. There's a lot that actually think about and we need to actually hear these perspectives. Now, at this point of the race, no one wants to slow down.

Everyone is working on achieving AGI. But, I also think we need to separate legitimate AI risk from the pure fear that is out there. Cyber attacks, autonomous weapons, biological misuse, and human weaponizing superhuman AI are definitely some serious concerns, but the idea that an AI simply becomes conscious, goes rogue, decides to hate humanity, and wipes everyone out is much harder to justify. The stronger concern is probably not an AI becoming evil.

It's extremely capable systems being misaligned, hallucinating, as well as pursuing the wrong objective, gaining too much autonomy, and controlling other people's outputs based off of what misuse humans are inputting into these models. So, yes, the risk definitely deserves to be taken seriously, especially when the researchers building these systems are resigning over them. Assigning a probability to human extinction is still an extraordinary claim that requires extraordinary evidence.

And remember, at the same time, AI is helping us in so many different areas: sciences, medicine, productivity, and everyday life for the better. If you like this video and would love to support the channel, you can consider donating to my channel through the Super Thanks option below. Or, you can consider joining our private Discord where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more.

Next is where we're going back to OpenAI cuz they have released the ChatGPT image 2.5 model, bringing some major upgrades to image generation. It is now faster, sharper, and more consistent with improved fidelity for more natural and recognizable images. Now, this is a big update and its biggest upgrade comes with editing capabilities. Details can stay consistent across multiple changes. There is also a new comment-based edit system that lets you modify exactly what you want without completely changing the rest of the image.

Overall, it looks like a solid upgrade for both generating and editing images. Here, you can see the GBT image two model being compared to the 2.5 model. Starting with the same robot image, it performed nine consecutive edits using each generated image as the input for the next one. Even after all those edits, GBT image 2.5 is something that preserves the robot's face, texture, and background far better than the two model.

Obviously, showing a huge improvement in consistency across repeated edits, but it is something that can even do better when you request it an individual change on a specific area of that image. And lastly, Deep Seek could be preparing to go public fairly soon cuz according to Reuters, they are reportedly tapped and potentially filing in on Shanghai's STAR Market with the IPO process potentially beginning later this year.

The company is also reportedly raising new funding at around 75 billion in terms of its valuation just months after securing a 7.4 billion raise. It's crazy to see how quickly Deep Seek has gone from a relatively lean AI lab to potentially becoming one of China's most valuable AI companies. And this also is something that hints at the new development of AI models that they have been focusing on. They may be focusing on this new architecture that we had covered yesterday, which was the Deep Seek version 4.1 flash, which is kind of exceptional in a lot of different domains, and it is pretty great, and I highly recommend checking out my video on it as it is something that is quite lean.

It is something that is a flash variant of the deep seek architecture, and it is insanely cheap. But, that's about it, guys, for today's video. There's a lot going down, so I highly recommend going ahead and subscribing to always stay up-to-date with whatever's happening in the world of AI. If you guys haven't already, make sure you go ahead and take a look at the world of AI Insider School Club. This is where you get access to different sorts of subscriptions on a monthly basis, like getting free credits to Invideo's GPU compute, as well as a couple of other things.

You can also learn AI, learn a lot more about different coding agents, also getting access to a lot of freebies. Make sure you also go ahead and take a look at the world of AI Vibe Club platform and benchmark to help you get a better idea of how to get started with AI models. Make sure you also take a look at the newsletter, join the Discord, follow me on Twitter, and lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a look at our previous videos so that you can stay up-to-date with the latest AI news.

Well, with that thought, guys, have an amazing day. Spread positivity, and I'll see you guys very shortly. Peace out, fellows.