Cognitive Surrender

NYU professor, commentator and entrepreneur Scott Galloway promoted a blog from one of his staff warning that writing, attention and critical thinking are collapsing in American classrooms because of AI. Data released in early June showed that math and reading scores for 13-year-olds have steadily declined since 2012. One third of 12th graders don’t have basic reading skills and 80% of hiring managers believe today’s high school students are less prepared to enter the workforce compared to previous generations. Roughly 85% of high schoolers are now using AI (and the remaining 15% are probably lying). A college professor said, “Offloading the struggle to a chatbot does not ‘free students up for higher-order work.’ It deprives them of building the strength to do any substantial cognitive work at all.” Cognitive surrender “represents a deeper abdication of critical evaluation, where the user relinquishes cognitive control and adopts the AI’s judgment as their own… they stop deliberative thinking altogether.”

Scientific studies support this. In one experiment, two groups of college students completed a research project, with one group using an LLM, and one using traditional search. The group using the LLM experienced lower mental effort and produced weaker reasoning than peers using search engines. In another study, students used ChatGPT to prepare for a math test. They did well on practice sets, but when ChatGPT was taken away, they performed substantially worse. A Stanford University review of over 1,000 academic papers on AI in K-12 education came to a similar conclusion. The Stanford SCALE review showed only 20 studies showed strong causal inference and found no high-quality causal studies for U.S. K-12 schools on student-facing AI tools. For the studies showing causation, AI boosted students’ performance when they had access to the tool, but when taken away, those gains disappear, or in some cases, reverse. The Stanford survey concluded that AI use does “not necessarily result in deeper learning.”

Two case studies are Alpha Schools, an elite set of private schools, that garnered extensive press on their AI-driven adaptive learning process to replace traditional education. And there is Khan Academy, whose Khanmigo program, provides step by step guidance to students in a Socratic tutoring format.

Alpha claimed that its students consistently score in the top 1% on MAP (Measures of Academic Progress) tests. It promises to deliver a full day’s worth of instruction in two hours using AI tutors, with academic guides providing motivational and emotional support. The director of the Teaching Systems Lab at MIT, Justin Reich, pointed out that MAP is not usually used as a benchmark for students going to schools that cost an average of $40,000 per year. Other elite private schools may also score in the 1% on this metric. A WIRED study found that Alpha’s academic coaches were largely remote workers in the Philippines and Colombia and were not credentialed teachers. The AI tutors used adaptive learning algorithms, found in programs such as Atom to prepare children for independent schools testing. In essence, we don’t have good data on Alpha Schools performance despite the wide-spread praise the company received on technology friendly podcasts.

Khan Academy, the non-profit behind Khanmigo, is a nonprofit and is arguably more truth-seeking. An independent study compared Khanmigo against Google search in undergraduate physics and found that both groups showed significant learning gains but found there was no statistically significant differences in outcomes between students using Khanmigo and those using a search engine. Students praised Khanmigo and appreciated its step-by-step guidance but viewed it as a supplementary tool rather than a replacement for traditional instruction.

AI has the potential to be additive to education. Instead of just handing a paper in, a teacher could ask students to give oral defences of their paper, live annotation sessions, and iterative drafts tracked in real time. The other dynamic is teaching children how LLMs work. Why do they hallucinate? Why is outsourcing judgement dangerous?

The future employees who will thrive won’t be those who avoided AI, but who developed independent cognitive capacity to direct AI, catch its errors, and know when not to trust it.

Stealing Models or Helping the End User? Anthropic accused Alibaba’s Qwen lab of a distillation attack after operators linked to the Chinese tech giant ran 28.8 million exchanges with Claude through 25,000 fraudulent accounts between April and June. Anthropic said that this was the “biggest attempt so far by a Chinese company” to assess U.S. capabilities.

The numbers are large. They are not as large, however, as Anthropic would like you to think. The accounts conducted an average of 1,152 exchanges or 19 exchanges per account per day. In a full distillation attack you need billions of exchanges. That would be wholesale model replication. That is not what Alibaba is doing. Instead, they are doing systematic capability probing like how Stanford tried fine-tuning a Llama model on GPT-4 outputs (note that OpenAI accused Stanford of violating its terms of service and settled on a deal outside of court).

Anthropic’s edge is the reinforcement learning techniques it gets from its users utilizing its models and from the extensive data pipeline that U.S. labs have created. For instance, there are several companies hiring domain experts to essentially evaluate LLM responses. This gives the model the feedback it needs to improve. Anthropic refines the models and charges 60% gross margins. Anthropic is now valued at $965 billion after a $65 billion Series H round.

The incentives are interesting. Anthropic is lobbying against open-source competition saying its dangerous. They are asking the U.S. government to act against Alibaba. Meanwhile it tries to charge relatively high prices for tokens to generate high gross margins.

The U.S. government should condition any regulatory protection of closed AI models on public interest obligations, which could include interoperability requirements or open licensing of certain capabilities.

Startup Slop. Business Insider surveyed dozens of startup founders and found AI has become the primary author of startup code. However, the code is hard to maintain after launch. The original developer can’t read or reason about AI-generated code, they can’t debug it when it breaks in production, they can’t extend it cleanly, and they can’t onboard new engineers to fix the code efficiently. This then generates a massive clean-up tax, which then negates the productivity benefit. Creatr reported that the average rescue cost of vibe-coded apps that failed in production ran between $50,000 and $500,000, depending on how far the app had grown on top of its original shaky foundation.

Data from TechCrunch shows that companies are spending 44% of their tokens on bug fixes that their AI generated. Another survey from Pixelmojo and Autonoma AI shows that technical debt, or the project backlog, increases by 30%-41% after AI code adoption.

It’s important to know your code because the maintenance costs that accrue after launch can erase the speed advantage at the prototype stage.

Continuing Down the Yellow Brick Road. The PBOC and China’s General Administration of Customs jointly published a draft revision to change their gold import measures. The draft removes the provision requiring PBOC and customs authorities to jointly formulate rules for individuals carrying or mailing gold across the border, while keeping such movements subject to customs supervision. It seeks to improve convenience for businesses and the public to import gold. This follows the April 2026 reform which extended multi-use permit validity to nine months, removed per transaction limits and expanded approved ports.

China is attempting to implement what could be called a gold-as-a-service strategy, trying to court foreign central banks to store gold within China’s borders via the Shanghai Gold Exchange. This will also facilitate Chinese banks’ expanded gold accumulation.

We’ve covered how China seems to be trying to internationalize offshore yuan into a reserve currency by using gold collateral to build trust. This is an incremental data-point on the way to China developing a parallel monetary system.

End Note

A team at the Norwegian University of Science and Technology built a three-armed “Sashimi-Bot” that can cleanly slice and plate sashimi. One arm straightens the fish by gently putting it into position, a second holds the knife and slices, and a third picks up the finished slices with chopsticks and places them on the tray. In testing, the robot produced 34 slices and successfully plated 26 of 28 that reached the cutting board. A full cycle cut took 27.9 seconds.

Salmon loins are notorious for being slippery and hard to hold. They are limp and deformable with poorly understood elastoplastic properties. The robots learned to straighten fish through thousands of virtual simulations using deep reinforcement learning. They gave the cutting arm a sense of touch using machine learning from thousands of readings from a tactile sensor with an embedded camera mounted near the knife. They trained a separate model to detect when the blade contacts the cutting board so it can adjust its trajectory in real-time.

Could it disintermediate traditional sushi/sashimi preparation, particularly at the itamae (master sushi chef) level? A chef spends years of apprenticeship just learning to cook rice before touching fish. The knife work takes enormous skill from precise angles, pressure, slice thickness and understanding the grain and texture of different fish species.

The artistry, the reading of the fish, the presentation philosophy and the years of intuition built into master’s cuts are a long way from what the Norwegian team’s robot does. Chain sushi/sashimi places don’t hire itamae and may consider the robot as quite the positive. But omakase is not being replaced by AI robots anytime soon.

Omar Sayed