On June 6th, 2026, Anthropic published the following article: https://www.anthropic.com/institute/recursive-self-improvement. If you haven't already, you really should take the time to read the entire article. It's kinda long compared to the news bites we're accustomed to, but it's packed with important information about the rapid development of LLM technology. (And please excuse Anthropic for the annoying animated graphic at the top of the article, you gotta let it finish drawing out before you can scroll down to the really important stuff.)
I'm kinda surprised the article didn't create more of a stir in the industry. Perhaps it's because we're overloaded with AI news every day. Top that off with all the non-AI related news we're smothered in, it's virtually impossible to find the signal in the noise anymore.
Let's Dig In
There's so much going on in the article, if I wrote out all the commentary I wanted to make, this writeup would be even longer than the article it references. For now, let's just focus on a few really interesting points. The first two paragraphs set the stage:
For most of AI’s history, humans drove every step in its development cycle. But at Anthropic, we are delegating a growing share of AI development to AI systems themselves, which is speeding up our work.
Taken far enough, and given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor. This is called recursive self-improvement. We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.
Personally, I would amend that last sentence to say "it is coming sooner...", but that might come off as being alarmist.
Continuing with the article (past the annoying animated diagram), Anthropic emphasizes not only how fast LLM-based AI technology is advancing, but also how quickly the rate of advancement is increasing. Lots of good content in there, some impressive statistics, and a good number of caveats and cautionary statements. The key takeaway here is that they are now using their current best AI tools to help create the next generation of AI tools, and that with each generation, the AI is doing more of the work, that work is of higher quality, and human involvement in the process is mostly just setting goals.
The Three Scenarios
Now to what I consider the true salient portion of the article, under the heading "Possible futures". Anthropic lays out three of them, which I am summarizing here, but you really need to read them in their entirety.
- The self-improvement trend soon stalls. Rates of AI self-improvement taper off, and the capabilities of new models created by existing models hit a plateau. If this happens, LLMs won't get any better until some new innovations come from an external source, i.e. human researchers.
- The current trend continues for some time. In the near future (months?) new and improved models are built almost entirely by AI. Humans are still needed to keep the development on track and gauge its progress. In this scenario, LLM-based AI becomes considerably more powerful in a short amount of time. Anthropic mentions some possible caveats in this scenario, but also starts to introduce some of the serious economic, social, and geopolitical issues that could arise. They also say that, in their opinion, this is the most likely of the three future scenarios, and I don't disagree, so take note of this section.
- AI becomes fully self-improving. In the near future, AI takes over pretty much every aspect of building new and better AI. The technology progresses at an astounding rate, limited only by available hardware and the energy required to drive it. AI soon becomes better than people at pretty much all practical tasks, including physical jobs, via robotics.
Though they don't use the term (and I understand why), the third scenario is often referred to as the Technological Singularity. Why it's called that is better explained by Wikipedia there, but the first sentence of the article is enough to get the point across:
The technological singularity, often simply called "the singularity", is a hypothetical event in which technological growth accelerates beyond human control, producing unpredictable changes in human civilization.
This is the basis of a thousand science fiction stories, but we're not talking about science fiction anymore. We are literally, right now, facing this event. How soon, and how long will it last? Well the answer to that depends on which of those three scenarios we're in. If it's the third scenario... months? A year at the most? I don't think it will be quite that fast, for the same reasons mentioned in the article. It would require much more infrastructure development than we're capable of building in that short of a time. But even scenarios one and two eventually converge on scenario three. And a significant breakthrough in any one of dozens of technologies (for example, Optical Computing), could launch us into scenario three virtually overnight.
Let's examine a bit more of the article, because it gets even more interesting.
If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing. But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe. Without a global coordination mechanism, companies and governments will have to make difficult decisions about safety while under competitive and geopolitical pressures.
We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology.
Anthropic themselves are saying the worlds needs to pump the brakes on AI.
None of this is necessarily impossible in principle—the world has built verification regimes for other complex technologies (e.g., the Intermediate-Range Nuclear Forces Treaty)—but those regimes took decades to build both the infrastructure and the trust. We don’t have that long. A unilateral pause by one lab, by contrast, is achievable immediately, but accomplishes much less: it would change who the front-runner is, but it would not create the wider deliberative process that is currently missing.
Even with the threat of the entire planet becoming a post-apocalyptic nuclear hellscape, it took decades for the involved nations to come to an agreement regarding nuclear proliferation. But is this really an apples-to-apples comparison? I'm skeptical.
So Where Are We Now?
On or around June 13, 2026, the US government imposed export restrictions on Anthropic's most powerful models that we know of, Fable 5 and Mythos 5. Now, there's no truly practical way to completely prevent foreign countries from accessing these models, so Anthropic just shut them down for everyone. Around June 27, the US government relaxed those restrictions a bit, allowing a "vetted list" of around 100 corporations, federal agencies, and cybersecurity vendors access to Mythos 5. At the time of writing, access to Fable 5 is still blocked to everyone (except Anthropic itself, presumably).
A few days ago, the US government also asked OpenAI to hold off on the general release of GPT-5.6 models, restricting it to a government-vetted list of trusted partners. OpenAI has complied with this request.
So the US government has finally gotten involved in AI in a big way. This was obviously inevitable. Some of the reasons stated for these restrictions seem reasonable, others I'm kinda skeptical about. Not going to dig into that side of it all right now, and in this case the "why" doesn't really matter. We regular people currently cannot access the best AI models. Notice though, neither Anthropic nor OpenAI put up much resistance to these developments. Remember, Anthropic was already calling for a general slowdown in AI proliferation, and they said right in their article that they can't do it themselves. With these restrictions, aside from those privileged "vetted list" members, we are indeed slowed down on the US side.
It's a shame that it's playing out like this. Firstly, the current state of the situation means that large corporations who can now access frontier models have a distinct competitive advantage over ones that can't. This is very bad news for smaller tech businesses. And what about hobbyists and individuals? Will we be permanently barred from the most powerful AI models? There are already rumblings of us needing to provide detailed personal identification to access AI in the near future. Privacy concerns aside, it's pretty clear that our government will now be between us and AI, with big tech pulling the strings. And unfortunately, there's nothing in our Bill of Rights to help us here, though to be fair, all these happenings would have been pretty hard to predict way back then.
What About Other Countries?
The only other country that's a serious player in this story is China. They've been playing catch-up for a while with AI tech, but at the time of writing (man this stuff moves fast), it's looking like the performance of Chinese frontier models is getting awfully close to that of their US-based counterparts. There's a good deal of debate how or if the tech might have been "stolen" from the US, but that's a complicated subject. Suffice to say that the US export restrictions do at least put a damper on foreign organizations' ability to distill information from US-based AI providers, which is why OpenAI and Anthropic aren't making a fuss about those restrictions.
Let's return to some of the points made in the Anthropic article. If the Chinese providers are also employing recursive self-improvement using their current models, then those models will also advance in capability quite rapidly. No "theft" necessary. Just chips and electricity. A lot of chips and a lot of electricity. It isn't really any easier for China to produce these things than it is for the US, but both sides are ramping up production as fast as they can. China will achieve and maintain parity with the US in AI technology.
And, in a surprise to all, Z.ai, a Chinese-based AI company, recently released their GLM-5.2 model under the open-source MIT license model, making it free to all. Not free to use, completely free to download and run on your own hardware. And this model is approaching the capabilities of Anthropic's higher-end models, based on current benchmarks. Now before you try to run it yourself, keep in mind you'll need a top-end GPU, 256G of server-grade DRAM, and a high-end mobo and CPU to connect it all. At least $12,000, and that's just to run the most compressed version of GLP-5.2 which is less performant than the full model. To run the full thing, you'll need about 1.5TB of DRAM, which will run you about $50,000 at current prices.
That's too rich for me personally, but it's not an astronomical amount of money. About as much as a very nice new Toyota SUV here in the US. But for even a modest-sized business, this is quite attainable, and with it, they've got access to their very own LLM server with no worries about government interference or outsized monthly cloud-AI bills. Or, if ten enthusiasts contribute $5,000 each, they can share the thing among themselves. Given where things are headed here, this seems like an attractive alternative to Anthropic or OpenAI.
Let's review what the CEO of Anthropic, two years ago, said before Congress: https://www.youtube.com/watch?v=yBgvqGrJ3g8&t=7718s. To his credit, he saw where things were headed way back then, and now we're there. But I'm not comfortable with his stance on the subject.
And Where Are We Headed?
If AI really is as dangerous as our government and industry leaders would have us believe, we're headed toward some sort of AI war. But not like a Terminators vs. humans AI war. This war will be between countries, with their leading AI providers leading the charge. It will start off relatively cool, with the players tangling over access to raw resources, chip manufacturing capacity, and the energy to run things.
"But countries are already fighting over these things." True statement. The AI war is already in progress. And if you don't believe me, you didn't pay any attention to anything I wrote up there. Or current events for that matter. Did you know that 50% of China's oil supply was moving through the Strait of Hormuz? Only 7% of the US oil supply is impacted by the closures. Kinda interesting, don't you think?
Nonetheless, China now has the capability to disrupt the entire US-based AI market. Will they do this? Remember, if the US economy completely tanks, so does China's. We're economically joined at the hip, and will be for decades to come. My gut says that somehow, the US and China will arrive at an agreement over AI control.
What Does This Mean For Us Regular Folks?
Too early to say, but some flavor of "nothing good".
Here's what really bugs me about all this though. We're seeing all this FUD about how AI could be dangerous to us regular citizens. "What if someone uses AI to build a nuclear bomb, or create a biological superweapon?" Though I don't recommend anyone actually try this, I strongly suspect if you were to ask a frontier model to help you build such a thing, you'd very quickly discover that AI is going to be of absolutely no help to you in your mad science endeavors. And both history and common sense tell us that adversarial world powers don't need AI to create these horrors.
"What if AI puts all of us out of work?" That's only a problem if that AI is controlled by power hungry politicians and greedy tech oligarchs. This is exactly where we're headed! In the hands of sane and compassionate people, AI is nothing but good for us regular folks. We could transition to an AI-based economy without anyone going hungry or missing a mortgage payment. It's not AI that's the problem. It's the people who want to control our access to it. And they want to control that access to maintain the status quo.
I want to see a future where AI tech advances as fast as possible, and is used to the benefit of all people, not just a vetted few. Is such a future complete fantasy? I'm a pretty solid realist, and there aren't many historical examples of technological advancements that play out this way. But I'm going to hold onto some hope for now.
Maybe Claude could help us figure it all out.
Edit 07/01/2026: I added a second installment on this subject: Is The Golden Age Of AI Already Over Pt 2