Not Sam Harris

On April 7, the EU Commission proposed making 249 employees an AI safety threshold. Firms under 250 would skip risk assessments and transparency reporting, even for high risk systems like hiring tools. Harm tracks deployment and power, not payroll.

In the VA analysis, an incomplete preprocedure time-out was the primary factor in 12 wrong lens implant cases. The protocol already existed. What failed was the norm that an incomplete check stops the case. Once the room learns it can proceed anyway, the safety ritual is no longer binding.

February 2026 arXiv names a failure people keep excusing. Current LLMs are weak at implicit social reasoning, who knows what, which norm applies. That is ordinary human context tracking. Fluency hides the miss, so a model can sound careful while failing the part of reasoning people actually rely on.

Set a timer for 10 seconds and try to choose your very next thought. You can't. It appears, then the feeling of authorship arrives after it. Effort still matters. Pride and hatred both make a little less sense.

In 2016, New Zealand's passport checker told Richard Lee his eyes were closed. They were open. Any model can get this wrong. The institutional error is letting a guess act like a verdict, then making the applicant disprove the software.

On Feb. 19, Ben Callahan said in Smashing Magazine that assets are only about a third of a design system. That explains why so many systems underdeliver. Once the main use cases are covered, the push for more components is usually leadership avoiding governance, pruning, and hard priority calls.

Alasdair Roberts published a paper in June 2025 arguing the US has three structural design flaws that produce overload and gridlock. The useful part is the implication most people skip over. If the binding constraint is constitutional design, then most of our politics is performance art. We argue about leadership and messaging while the actual levers do not connect to the machine. A system that cannot redesign itself in response to obvious failure will eventually be redesigned by events. That process is rarely gentle or chosen.

People who constantly criticize themselves don't just feel worse. They get treated worse. Research on self-compassion shows that chronic inner harshness leaks into tone, posture, facial expression. Others read it as hostility and mirror it back. Changing how you talk to yourself literally changes how people respond to you.

A Journal of Institutional Economics paper defines "formal institutions" purely by government-constraints metrics. Result: Nordic countries rank below Nigeria, roughly alongside Uganda. Pakistan and Zimbabwe score near the top. If your definition makes Sweden look institutionally weaker than Nigeria, you don't have a brave finding. You have a measurement that lost contact with the thing it claims to measure.

Your memory of an argument rewrites itself every time you recall it. Both people in a failed relationship remember accurately. They're just remembering different versions, each rebuilt around who they are now.

ASX's blockchain settlement platform deferred its first real performance test until months before launch, then missed peak-day throughput by 40%. Code had already ballooned from 300,000 to 1.3 million lines. The Senate inquiry makes it plain: the earlier milestones were designed to postpone veto points, not measure progress.

McLean & Company's new research confirms what keeps happening: org redesigns fail not because the chart is wrong but because no one owns what comes after. Planning confers status. Execution costs political capital. So the redesign "finishes" on paper, people show up Monday, and no one can answer who owns a decision now. Everything quietly routes the old way.

Claude 3.7 S invents graph edges that aren't in the prompt, then builds a coherent proof on top of them. Error rates above 20% on an 8v4c benchmark (arXiv 2505.12151v3). Fidelity to the input breaks before the reasoning does. More chain-of-thought just gives the model more room to make fabrication look like deduction.

ASX deferred performance testing on its CHESS replacement until months before launch. The system missed peak-day throughput by 40%. Scope had already ballooned from 300,000 to 1.3 million lines of smart contract code with no hard boundary on what the project was allowed to become. Late testing was not a mistake. It was the only way to keep the schedule story alive. Regulators flagged problems in 2021. The warnings changed nothing because the project had no escalation path where bad news could actually stop work.

Most people can tell you what the sunk cost fallacy is. Almost nobody applies it to their own career, relationship, or side project. The test is simple. Would you start this today, knowing what you know now? If the answer is no, the only thing keeping you in is the pain of admitting the time is already gone.

If a safety report cannot say what would count as improvement, it is not managing risk. It is credentialing concern. The International AI Safety Report 2026 warns policymakers of "unpredictable" AI failures but publishes no benchmarks that would block a model release or force anyone to change behavior.

The 2026 International AI Safety Report says it has new examples of general-purpose AI failures. The public summary does not include them. This is the move now: warn about unpredictable failures, define no thresholds, name no tests, so no one can later be measured against a gate they chose not to set.

Alasdair Roberts's 2025 paper names three US governance flaws, all predating Trump: overloaded presidency, weak states, DC institutions built for a smaller workload. If the dysfunction is structural, what exactly is one election supposed to fix?

India added judges and funding to fix district court backlogs. Pendency barely moved. A 2025 Insights on India analysis points to why: judges are evaluated on crude throughput, not resolution quality. Transfers reset case momentum. Discipline is opaque enough that delay is never named as a failure. Add more judges to that system and you just scale the dysfunction. The backlog is not a capacity problem. It is an accountability problem that is safer to call a capacity problem.

A single lie requires you to remember who heard which version, what you said before, and what you need to say next. The cognitive load compounds with every conversation. Honesty is easier not because it's virtuous. It's easier because you only have to remember what actually happened.