6/25: Leviathan Wakes
The government blocks GPT-5.6, Apple raises prices, autonomous weapons are coming, Alibaba steals Claude, elites launch RAISE US
It’s Thursday, monitors, and we had one of our best days yet on stream. As always, be sure to catch us on X and YouTube, and join our Discord to chat with our hosts live.
Today’s Experts
Pim de Witte (General Intuition)
Andrew Trask (Google DeepMind, OpenMined)
Tarek Alauri (Stuut)
Rick Song (Persona)
Neil Movva (Sail Research)
Meb Faber (Cambria Funds)
Jamin Ball (Altimeter Capital)
Timour Kosters (Edge City)
Making Sense of the World
Leviathan Wakes
In May, AI policy researcher and friend of the show Dean W. Ball, now head of Strategic Futures at OpenAI, wrote an article titled “Before Leviathan Wakes”. Borrowing Thomas Hobbes’ 1600s conception of the state as the Biblical Leviathan — a vast, uncontrollable being of immense power1 — he argues that if the state truly realizes the potential of the models being created right now, it will exert a level of control never seen before with any digital technology.
The Leviathan has stirred before, such as when it designated Anthropic a supply-chain risk or imposed a voluntary framework for testing models before release, but today it did something unprecedented: it asked OpenAI not to release GPT-5.6 to the public, and requested to approve access to every preview customer individually. After months of encroachment, the precedent has now been set. The labs no longer decide when and how models are released; the government does.
I can already hear the accelerationists making a counter-argument: “the labs asked for this”, they say, “by demanding regulation. This is what they wanted.” But this isn’t regulation like that of the FAA or the RAISE Act. In the systems suggested by OpenAI and Anthropic, the government works within a framework of properly scoped rules to determine if a given AI model poses an unacceptable risk in any of a few clearly defined areas, and has the limited authority to block or delay its release until the developer can demonstrate that the risks of the model have been mitigated. Instead, we’ve sleepwalked into a system where the government can use the authority of a “voluntary” testing regime to indefinitely and arbitrarily block the release of any frontier model (except, ironically, open-source Chinese ones) for any reason, or no reason, at all.
This is also no victory for AI safety or for Pause AI advocates. For years now, rather than building superintelligence in secret, companies have been gradually working towards it through iterative deployment — broadly and frequently releasing minor increases in capability. This helps people adapt to the technology over time and avoid future shock, helps the labs learn from real-world feedback that they couldn’t simulate internally, and makes models safer because there are strong economic incentives not to put misaligned models in your B2B SaaS product.
If iterative deployment gets replaced by siloed AI research right before labs kick off recursive self-improvement, we should expect the result to be less safe rather than more. The labs will have limited oversight and little real-world data to inform the process, and if one AGI goes rogue, there will be fewer aligned AGIs to help humanity defend against it. “The Project” to have one government-anointed group of researchers build superintelligence in a bunker while the rest of the world is stuck on weak 2026-level models may end up increasing the likelihood of misaligned AI takeover, the very thing it was designed to prevent.
For now, the immediate effects of this policy seem limited. There’s still enough commercial pressure and enough laissez-faire sentiment in the government that it’s likely we’ll get GPT-5.6 (and Fable) relatively soon. But we are no longer in the early game of AI policy. The training wheels are off; the stakes are now real. Decisions being made right now in Washington determine in the short term whether you continue to access the benefits of widely available machine intelligence, and in the longer term who controls the creation of superintelligence itself. The Leviathan has woken.
More Stories
Apple raises prices of Macs and iPads due to skyrocketing memory and storage costs, which have risen so much so fast that CEO Tim Cook says he’s “never seen anything like it”. Prices across both product lineups, as well as HomePods, Apple TV, and Vision Pro were raised ~15-35% across the board2. The iPhone, Apple Watch, and AirPods are intact for now.
Also, Microsoft is raising the price of the Xbox. Starting in August, the Xbox Series X will cost $800 and the cheapest Xbox Series S will cost $500 (ironically, the cost of the premium Series X when it launched in 2020). Microsoft said they’ve seen console storage and memory prices increase by more than 2.5x and expect another doubling by fall 2027.
The Pentagon will increase the scope of AI’s military decisions. An unpublished document says that the speed of future warfare “may require the joint force to adopt completely autonomous systems”. This has always been more or less inevitable barring international treaty, as in the limit, autonomous systems can be much deadlier and more effective than systems with humans in the loop. Still, it’s interesting to see it explicitly mentioned. Notably, one of Anthropic’s “red lines” that caused the Pentagon to declare it a supply chain risk was “no fully autonomous weapons with current tech”.
Nonprofit RAISE US is launched to prepare workers for the AI economy. Its co-chairs are Democrat Gina Raimondo, Biden’s Commerce Secretary and former Governor of Rhode Island, and Republican Eric Holcomb, former Governor of Indiana. Founding partners include Anthropic, the OpenAI Foundation, Amazon, Microsoft, Eli Lilly, Coefficient Giving, and the Schwarzman, Ford, Rockefeller, Arnold and Walton foundations. Worth keeping an eye on as an elite organization, like the World Economic Forum, that will help create AGI-era labor policy.
Anthropic accuses Alibaba of stealing Claude. According to a letter sent by Anthropic to multiple US Senators, Alibaba created tens of thousands of fake accounts and collected tens of millions of Claude’s responses in order to distill Claude’s capabilities into its own Qwen models. Anthropic called for tighter export controls on chips and sanctions on foreign labs that engage in distillation attacks.
DeepSeek plans to double its headcount in order to better compete with AGI labs in China and globally. DeepSeek recently released their V4 model series and is raising at a $50 billion valuation, but Z.ai now has a better frontier model (GLM-5.2) and a higher valuation.
It’s amusing how closely Hobbes’ metaphor of the Leviathan resembles the AI community’s metaphor of the shoggoth despite being written nearly four hundred years earlier. Both conceptualize the subject (either the state or a language model) as a vast, alien, uncontrollable being of immense power. Both are formed out of a composite of humanity — either literal humans or the corpus of their writings. Both have a human-accessible “face” — the sovereign or the RLHF-ed chatbot persona. And of course, both represent one of the most important phenomena of the modern world.
Some of the price increases:
Apple Vision Pro: $3,499 → $3,699 (+6%)
MacBook Neo: $599 → $699 (+17%)
MacBook Air: $1,099 → $1,299 (+18%)
MacBook Pro M5: $1,699 → $1,999 (+18%)
iPad Pro: $999 → $1,199 (+20%)
HomePod mini: $99 → $129 (+30%)
Apple TV: $129 → $199 (+54%) (!)
A fully maxed out MacBook Pro — 16-inch, nano-texture display, M5 Max, 18-core CPU, 40-core GPU, 128GB unified memory, 8GB SSD — now costs $10,149 before tax.