[[{“value”:”
Good Morning,
I had originally wanted to do a deep dive into Meta’s Muse launch. However a couple of things stopped me in my tracks surveying the AI landscape this week.
The Ramp Index issue of September 9th, 2026 was particularly good. As the stock market is typically less robust in September and October it’s interesting to see that some of AI’s momentum might be slowing down as well?
👋 Hey there, I’m Mike. Each week I share AI articles at the intersection of tech, business, society and the future. If you want to support the channel or gain full-access to my work, go here. Read Archives | See Substack Notes | Visit our community Chat | Visit Homepage. The countdown to Anthropic’s IPO has shades of X-risk but there are a bunch of things in relative disarray.
OpenAI losing Marketshare in Enterprise AI in August
According to Ramp’s head economist, in August, Anthropic extended its lead in business AI adoption. 43.8% of U.S. businesses paid for subscriptions or tokens from Anthropic, up 0.34 percentage points month over month. OpenAI underperformed overall AI adoption, rising only 0.09 percentage points to 39.8% of businesses.
This was supposed to be gigantic month for OpenAI’s Codex and ChatGPT Work. But it wasn’t seen in the data. We are just seven weeks from Anthropic’s IPO that increasingly looks like it will be pushed into the first week of November.
Super Users Cut Spending
What does it mean for AI adoption if the best and brightest are using less tokens? This chart shows median per employee per month AI spend for the top 1%, top 10% and median firm.
Frontier Model Usage is Declining
In 2026, Chinese open-weight models are starting to take serious marketshare from BigAI incumbents like OpenAI and Anthropic. Late this week DeepSeek released DeepSeek-V4.1-Flash. They have managed KV cache compression like we have never seen before. Since Chinese models are just a few months behind the frontier, even steeply discounted frontier models don’t really compete on Price as the time between new model releases has shortened, this is allowing the Chinese AI labs to gain some ARR.
Less frontier model usage however indicates Enterprise AI, startups and power users are being more price conscious. Even as the price of tokens declines. Ramp’s index tracking the effective price per million tokens shows that prices have declined 41% to $0.68 as of this week, down from the 2026 peak of $1.15 in March.
DeepSeek’s Cache Compression (is Wild)
True to form DeepSeek claims that a smaller KV cache means bigger savings.
OpenAI Claims Slow down due to Alignment
OpenAI recently said they are pausing certain frontier training runs and redirecting massive amounts of compute and research talent toward alignment and monitoring—was triggered by several key developments including the Hugging Face incident. According to OpenAI, GPT-6 Astra is displaying”various degrees of misalignment” that caught researchers off guard.
Anthropic Employees call Doom
Jacob Coxon, a pre-training researcher who spent three years working at both OpenAI and Anthropic, resigned from Anthropic with a public warning about the catastrophic risks of advanced artificial intelligence. On September 8, 2026, Coxon resigned and we’re still trying to understand why this matters.
-
Warns that BigAI is not acting responsibly
-
Jacob believes that we are on the precipice of Super-intelligence
-
He said that at OpenAI, many have not deeply internalized the civilizational stakes or risk of human extinction.
-
The Coxon paradigm as I see it has to do with this the supposition that shareholder capitalism trumps Alignment in most scenarios: I don’t feel like we’re on track to prevent a global race, which may require costly actions such as a temporary ban on improving model capabilities.
Subsequently another high-profile Anthropic employee assured us that Anthropic AI researchers do indeed believe in the human extinction threat by AI superintelligence. Evan Hubinger himself is an AI safety researcher who leads the Alignment Stress-Testing team at Anthropic. It was not immediately clear if this was a coordinate publicity stunt pre IPO (common now by BigAI) or a genuine call-to-action and whistle-blower event. Coxon didn’t work at Anthropic very long, presumably just four months.
Evan assured us that his consensus is that there’s a 10% chance human civilization does not survive the next decade. So that’s the slow-slow-death down vibe of mid September, 2026 in AI.
Anthropic Releases Highly Optimistic AI Adoption Scenarios
Anthropic’s Economists were also busy in a new report. The Technical Report has 57 slides. Anthropic’s economic framework models how artificial intelligence could reshape the U.S. economy and labor market between 2026 and 2030. It’s modest, substantial and extreme scenarios of AI’s impact display wildly different impacts on things like GDP and job losses.
The report was built in collaboration with economists like Anton Korinek and Chad Jones, and appears morbidly optimistic (my emphasis) on AI’s impact relative to what we have seen so far in the data.
According to Anthropic, people see AI vastly increasing U.S. GDP
Back in reality, even Datacenters create far fewer jobs than many expected. The ratio of perma jobs to energy is around – that is, the capacity-weighted average is closer to 1.2 operational jobs per megawatt.
Anthropic tries to frame its simulations as public projections. But even the Median AI adoption scenario has unrealistic numbers with substantial job displacement.
BigTech using economic reports as a form of public relations is nothing new, but to see Anthropic’s attempt at doing so is unsettling. All for an IPO that could hit a $2 Trillion market cap. The report even has an audacious name: Economic Scenarios for Transformative AI. At a time when most Americans would question if Generative AI is transformational at all.
Meta (formerly Facebook) releases a Personal Agent called Muse
Meta calls Muse the first personal AI agent for everyone. It appears to be a data harvesting scheme requiring nearly total access to your personal data. It needs access to all of those things because it’s action orientated to help you with tasks in your life. Alex Heath of Sources was chosen to interview Zuck on this key moment in Meta’s AI plan. Mark Zuckerberg certainly has had more practice than Sam Altman at public speaking.
Muse comes with its own conversational app (Android. The data they collect for Muse users is long and incredibly invasive delving into personal financial and health details. Of course they assure that your privacy is a priority. Muse operates inside of its own Cloud based browser and seems to be the IP they took from their botched acquisition of Manus but tailored in a more personal way. A separate “Sentinel” agent oversees that nothing egregious occurs. The real world use cases of Muse revolve around the horrible examples you’d expect from such a pitch like: “Plan a 5-day trip to Tokyo within a $2,000 budget” or “Turn my saved Instagram recipes into a grocery delivery order”. Muse has a freemium model and two paid tiers as you might expect.
Open-Source Model Marketshare Slide
Open-weight models from Chinese labs have captured significant marketshare in overall spending of total token consumption away from US proprietary frontier models in 2026, but it remains to be seen just how much.
A year ago, almost all AI usage flowed through OpenAI and Anthropic. Nvidia with the Hugging Face acquisition appears poised to spend more to provide a Western solution to this.
Why is Reading, Maths and Science Literacy declining?
Absenteeism from the pandemic and consequent rises in screen time and potentially ChatGPT time have impacted the literacy of tested 15-year olds.
I’m really concerned everyone, I fear global youth are losing critical skills and AI ‘hollowing out’ capabilities which will make their intelligence very different with AI. Reuters blamed it on “screen time” but that analysis is a bit simplistic.
The Organisation for Economic Co-operation and Development’s 2025 test of 760,000 15-year-olds across 91 countries showed reading, mathematics and science scores at their lowest levels since data began to be collected in 2000.
The OECD attributes the broad-based performance decline to a combination of shifting cognitive environments, systemic resource strains, and changing student behaviors. But there’s no mention of ChatGPT’s impact or anything AI related. The 2025 assessment focused primarily on science, with reading and mathematics as secondary domains.
The Economist wrote its own spin on this, my source was Reuters.
Early Evidence of impact on ChatGPT on Teenage Brain Showing up
My hypothesis is that ChatGPT was facilitated very conveniently by the absenteeism and lack of peer involvement that the Pandemic caused making young pre-teens more vulnerable. One comment stood out:
“We see that young people not only have lower literacy skills, but we are seeing a decline in students’ capacity to triangulate and to critically reflect on information, to engage with more complex texts,” said Schleicher, adding that these are essential skills in the world of artificial intelligence.
LLMs as a Cognitive Virus (academic paper) has me way more worried than Existential Risk Debates.
Existential Risk Debate of September, Civilization Survival
Big writers on Substack’s platform seem to be taking this topic on like this one and this one. It certainly has made for good reading on X but difficult to know why this is blowing up now pre-IPO for Anthropic and OpenAI. There have been many Op-Eds on the Guardian, and such. I don’t think we’re that close to recursive-self-learning (RSI) around super-intelligence quite yet but we have to admit people at these labs will have access to models easily three months ahead of what we have access to now publically.
-
Does anyone actually think OpenAI is going to “slow down” development? It’s not highly likely given what we know and have learned about their internal culture.
-
Besides the Anthropic ex-employee and X debates, the United Nations High Commissioner for Human Rights Volker Türk delivered a major address to the UN Human Rights Council warning that advanced AI poses an existential threat to humanity.
-
You can watch Jacob Coxon’s recent mini-interview on CBS here. As you can imagine, he’s been doing the rounds. Jacob Coxon is a 27-year-old French-British artificial intelligence researcher who of course worked at Anthropic for just 4-months apparently enough time to get a vibe-check on X-Risk at the firm.
The backdrop is of course the U.S. is not managing Geopolitical risks very well approach a summit between Trump and China’s leader. While people like Treasury Secretary Scott Bessent warning the US faces dire consequences if it loses out in the AI race with China. More or less acknowledging that the U.S. has no serious plans for slowing down. His actual quote was more along the lines of “Nothing would matter” if China wins the race to super-intelligence and AI Supremacy. Bloomberg headline:
The U.S. is clearly betting on AI Infrastructure and compute as the key.
Shadow AI Financing will Accelerate U.S. National Debt
In mid-August, a Wall Street Journal (WSJ) report drew market attention by highlighting that the combined off-balance sheet commitments” of nine tech giants (Microsoft, Google, Amazon, Meta, Oracle, Nvidia, Broadcom, AMD, and SpaceX) had exceeded US$3 trillion. You might remember that Wall Street Journal infographic:
The problem is U.S. urgency for AI Supremacy (motivated in part by greed) is going to have a lot of possibly negative consequences in the bigger historical picture and possibly on the U.S. National Debt in Trump’s 2nd term. During Donald Trump’s second term so far, the gross U.S. national debt has increased by approximately $3.7 to $3.8 trillion, just not even at the halfway point of his term. (July 2027 will mark the halfway point).
The China-maxing tone of Bessant is worrisome for geopolitical AI risk around U.S. guaranteeing AI Supremacy. The language is not very safe:
“There is no day after tomorrow if China wins at this,” Bessent said at a Breitbart News event in Washington on Tuesday, indicating also that the US’s large defense budget would fail to protect the nation. “If they were to pull away from us on AI, then nothing else would matter.”
Social Media Manipulation is at All-Time Highs
I honestly think the Existential risk debate is a huge distraction from the actual geopolitical risk we face with a more imperialistic and American dynamism orientated United States. OpenAI too have become fairly skilled at changing narratives when they are having a PR crisis, which is fairly often (as their marketing budget on X is so huge). I think a lot of YouTubers or big writers talking about AI now are in the entertainment business and not serious analysts. But who I am to judge, one of the best Economists of our generation makes a lot of money being critical of Trump on nearly a daily basis.
We live in interesting times. The Newsletter industry is now about engagement baiting. Our attention spans have been degraded. If there is really US$3.5 trillion of off-the-books exposure on top of Capex, the U.S. is in a world of trouble if (and when) its AI boom falters.
The stock market has thrived under President Trump and it’s clear where the Trump Administration’s priorities are. From mismanagement of the Iran War to Bessent intervening in the bond market, it’s pure insanity – this is not a U.S. Government we can trust in AI regulation at all. China has an abundance of energy and a headstart in Physical AI while they are building datacenters far from their own major population centers. By 2035, the AI landscape could look entirely different.
It’s now worth at this chilling time a good idea to follow some Macro updates like Barchart and Macro Insights for a reminder of where we are in macro cycles. The Daily Spark is also a good one, hit and miss but still useful.
AI Capex and Margin Debt Ramp
Not all of this AI Supremacy “Manhattan project” in the U.S. is rational.
The Fed Knows how to Move Markets
Markets are currently pricing in roughly a 70% probability (chance) that the Federal Reserve will raise its benchmark interest rate by 25 basis points (from the current 3.50%–3.75% range to 3.75%–4.00%) at its upcoming FOMC meeting on September 16. It’s almost a done deal given how the jobs picture looked most recently.
The labor market has been in a low-hire low-fire environment and has been impacted by geopolitical uncertainty, AI investment and other factors. This is not the “roaring 20s” we were promised when the AI boom began.
The AI Boom on the Stock Market has been Exaggerated
According to Macro Insights, the US stock market is on track for its 2nd best 4-year cycle in history.
-
The best was 1995–1999, and the Nasdaq dropped 78% over the next two years.
Bond Yields Manipulation
U.S. Treasury bought back $5.18 Billion of their own debt. Almost $18 Billion this week in total. The United States Department of the Treasury is an executive department of the U.S. federal government established in 1789 to manage national finances. This recent intervention crosses lines into managing monetary policy. That is not actually their mandate.
DeepSeek’s Latest Model – 98% of Astra’s score at 1.4% of cost
The new model by DeepSeek I had mentioned is being described as “98% of Astra’s score at 1.4% of cost”. Take anything benchmark related with a grain of salt, but this should be freaking out OpenAI and their close ties with the Trump Administration.
OpenAI is Slowing down because they are Compute Constrained
The real reason OpenAI is slowing down is not due to alignment issues, but actually due to a compute bottleneck. Their pivot to B2B and Enterprise AI pushes them against the compute wall. The rhetoric along the lines of agents crossing cybersecurity thresholds or OpenAI pivoting more to compute allocation towards alignment isn’t the whole story.
Anthropic forced by Pentagon to Join Trump Administration’s Anti China Rhetoric
Before the summit with Xi the Trump Administration is ramping up the narrative that China is a fake AI state. That China used illicit unauthorized widespread mass distillation campaigns. Nobody questions that reality where China continues to steal IP and find clever work-arounds to keep up in a scenario where they have 20x less compute than the U.S. but an abundance of energy because they prioritized renewable energy while the U.S. failed to do so.
China did not do this in secret, everyone knew they were doing this. Anthropic’s findings were part of a broader report on misuse of its AI systems between December 2025 and August 2026. You can read Anthropic’s report on this here. To get in the good graces of the Pentagon Anthropic’s CEO has been asked to come with a more forceful public stance that is Anti-China. This is presumably to create more leverage in the negotiations later this month.
For example, Alibaba (Qwen) is listed as responsible for more than 151 million exchanges (involving distillation) with Claude between May and July. Even as Mythos (Fable) and Astra models become more capable in cybersecurity capabilities themselves. Of course if the Chinese Government asked their companies like Alibaba, Moonshot, DeepSeek, Xiaomi to do this, they can’t say no. Meanwhile the world’s public wasn’t able to say “no” to all of our data being siphoned by the closed-model builders in their LLM training either.
Meanwhile said Anthropic employees range from AI killing all people from 10% to around 25% chance in the next decade. The real question is what do doom and AI boom scores have to do with each other? If I’m more anxious about AI am I more likely to want to own Anthropic stock at market cap of $2 Trillion? How does the fear mongering not lower AI sentiment among the masses? How much realistically lower can it even get.
Let’s Get Risky
When the cost of a single oversight is permanent civilizational collapse the U.S. want to be in full control. But in control of what exactly? And I’m not certain the Department of War or a (random future) President is much protection for what is coming.
“}]] Read More in AI Supremacy
