Thursday, April 30, 2026

Google Search: The Great Reversal

We might call the fortunes of Google search in the early artificial intelligence era as a “great reversal.” 


For much of two years, it seems, investors have asked a key question: what happens to Google search if users start asking ChatGPT, Perplexity, Gemini, or other AI assistants instead of typing queries into a traditional search bar?


“Shoot first, ask questions later” seems to have been an initial kneejerk reaction. 


But reality seemingly has moved in the opposite direction. 


In the first quarter of 2026, “Search and Other” revenue rose by 19.1 percent compared to the first quarter of 2025. Even more encouragingly, search growth accelerated for the fifth consecutive quarter.


For Alphabet, the debate is no longer only about whether AI will disrupt Search. It is now about whether Alphabet can use AI to protect Search, grow cloud offerings, increase monetization capabilities. 


source: Seeking Alpha 


This might alert us to the fact that our expectations about AI impact can be, not only wrong, but completely wrong.


Wednesday, April 29, 2026

Can a Good End be Produced by a Bad Means?

The U.S. Supreme Court has ruled, in a 6-3 decision, that Louisiana’s new congressional map, which includes districts based on race, is unconstitutional. The ruling is bound to be controversial. 


It might also be a bit nuanced. 


The majority found the map amounted to what it called unconstitutional “racial gerrymandering.” based on race, is a violation of the Equal Protection Clause of the Fourteenth Amendment. 


Critics will argue the decision effectively undercuts the 1965 Voting Rights Act, intended to end racial discrimination in voting practices, such as requiring literacy tests.


Supporters will argue the need for the law has long since been remedied. 


At least some will say the problem is the continuing use of “race” as a pillar of law, even if the intent of such efforts is to remedy past discrimination.


The issue at least some will have is that the solution to the problem of “racism” in law cannot be the enshrinement of racism in law, even if some believe it is done “for good reasons.”


Either before the VRA or since, if one continues to treat citizens differently because of their race, we haven’t really “solved” the problem of racism; we’ve only kept it in a new form. 


But the decision might not mean “race” cannot ever be a factor for voting rights: it simply cannot be the main motivation. 


Under the Constitution, states generally aren't allowed to use race as the primary tool for sorting voters unless they have an extremely good reason. 


Because race was the main factor, the map had to pass "strict scrutiny." The Court ruled the map failed because it wasn't "narrowly tailored.” It wasn't the most careful or necessary way to solve the legal issue.


Even if "allowing race to play any part in government decision-making represents a departure from the constitutional rule that applies in almost every other context," as Justice Samuel Alito wrote, the decision might not actually mean race can “never” be a consideration. 


But it moves policy in that direction. 


Separately, in June 2023, the Supreme Court effectively ended race-conscious affirmative action in college admissions, ruling in Students for Fair Admissions (SFFA) v. Harvard and SFFA v. UNC that such programs violate the 14th Amendment's Equal Protection Clause. 


That 6-3 ruling mandates that higher education admissions must use "colorblind" criteria, rejecting the use of race as a specific factor.


Opinion about both decisions will reveal a fundamental conflict over means and ends, as in the claim that “the ends justify the means” versus the argument that “the means are the ends.” In other words, can law and policy be “racist” in a new way to remedy the problem of racism?


Or, as the philosophical debate suggests, must the means match the desired ends?


Some argue the ends justify the means: the compelling end goal is to create a more equitable society by overcoming systemic racism. 


Others will argue an unethical or unconstitutional means is itself the problem. 


In other words, the debate is over ends and means. Does the goal of equality (the "end") justify discriminatory practices (the unequal "means").


To use a simple analogy, some might argue it is permissible to use hateful means to achieve a “loving end.” Others will argue that is impossible: the hateful means become the actual ends. 


Or, to put it simply, one cannot achieve a state of love using hate. In other words, must the means must embody the end they seek to create?


Mahatma Gandhi argued that means and ends are "two sides of the same coin". If the means are hateful, they "taint" the outcome, ensuring the final result is also characterized by hate or resentment.


Likewise, Martin Luther King Jr. noted that "hate cannot drive out hate; only love can do that". Using hateful means only intensifies the cycle of violence and adds "deeper darkness to a night already devoid of stars".


The "ends-means" debate splits between those who believe the government must use race-conscious means to reach a truly equitable end, and those who believe that using race as a means only perpetuates a discriminatory system.


AI Scarcities and Constraints Keep Evolving

It’s hard to keep up with the evolution of “value” in the artificial intelligence business as scarcities that create value keep shifting.


Between 2017 and early 2024, for example, scarcity and value in the AI value chain were heavily concentrated at the top of the stack:

  • high-quality training data

  • frontier model development(research talent, algorithms like transformers, and initial large-scale training runs). 


Compute, in the form of Nvidia graphics processing units,  was important, but a dominant bottleneck:

  • inference was relatively cheap

  • models were mostly accessed using APIs or research prototypes

  • real-world deployment at scale was limited

  • So value accrued to pioneers in data curation, model architecture, and cloud providers.


By 2026, constraints  have shifted with mass deployment:

  • compute infrastructure (GPUs/accelerators, high-bandwidth memory/HBM, advanced packaging) remains scarce

  • energy is emerging as a new scarcity (data center electricity, grid capacity, and permitting delays)

  • physical infrastructure (data centers, land in power-rich locations, cooling) lags demand

  • data scarcity is resurfacing as high-quality public data exhausts and regulations tighten

  • model weights and foundational capabilities have commoditized somewhat

  • supply chain crunches extend to materials like indium phosphide for optics and memory chips.


Overall, value has "inverted" toward the bottom of the stack, a shift from past decades where value accumulated in applications:

  • infrastructure and physical hardware is scarce (chips, GPUs and accelerators, compute as a service, utilities, and energy firms)

  • application-layer value (SaaS, agents, enterprise workflows) is growing but often depends on cheap/reliable inference, and therefore infrastructure

  • consumer surplus from gen AI has risen sharply, but producer value capture is uneven.



Value Chain Role

Scarcity/Value ~2022–Early 2024

Scarcity/Value in 2026

Potential Future Scarcity/Value (2027+)

Data

High (internet-scale public data as fuel for scaling laws)

Rising (high-quality data "exhaustion"; regulations; shift to synthetic data)

High for specialized/real-time/enterprise/private data; synthetic data generation & curation

Models & Algorithms

Very High (frontier research, talent, architecture breakthroughs)

Moderate/Lowering (open-source closes gaps; commoditization of capable base models)

Lower for base models; High for specialized fine-tuning, agents, reasoning, or domain expertise

Training Compute

High (GPUs, clusters for large runs)

High but shifting (GPU/HBM shortages persist; diversification to custom ASICs)

Moderate (efficiency gains; more distributed/synthetic training)

Inference

Low (early, limited scale)

Very High (80-90% of lifetime costs; latency, memory, energy at scale; "factory" phase)

Extremely High (edge/on-device, long-context agents, real-time applications)

Infrastructure (Data Centers, Power, Cooling)

Moderate (cloud scaling)

Very High (energy/grid bottlenecks; power > chips as limiter; land/permitting)

Highest (energy access, nuclear/renewables integration, grid modernization)

Hardware Supply Chain

Moderate (Nvidia dominance emerging)

High (HBM, advanced packaging, optics, materials like indium phosphide)

High for specialized (inference-optimized, edge, robotics silicon)

Applications & Agents

Low (mostly prototypes)

Growing (enterprise adoption, workflows; value from integration)

High (autonomous agents, physical AI/robotics, real-world actions)

Physical World/Embodiment

Negligible

Emerging (early robotics interest)

Very High (humanoids, autonomous systems, sensors, actuators, real-world data loops)


Among the key shifts so far in 2026:

  • value has moved downstream from "intelligence creation" (models/data) to "intelligence delivery and scaling" (inference and infrastructure)

  • compute shortages have evolved into broader supply-chain and energy issues including

    • power contracts

    • tier-2 locations

    • inference efficiency

    • energy consumed per token.


Future scarcities could develop in the future: 

  • embodied AI (robotics, sensors, actuators, energy storage, and unstructured environment handling)

  • orchestration and decision-making (supply chains, logistics)

  • regulatory compliance

  • valuable applications that leverage abundance (as physical constraints lessen)

  • geopolitics, materials, and talent for physical AI.


And, by definition, we don’t know what we don’t know. So we cannot predict what unknown issues might arise. 


"Known unknowns" in the AI value chain refer to recognized uncertainties or risks whose existence we acknowledge, even if we cannot precisely quantify their timing, magnitude, or full impact. 


These are issues we can model, debate, plan for, and partially mitigate through investment, policy, redundancy or research and development. 


In contrast, "unknown unknowns" are the true blind spots:

  • risks

  • emergent behaviors

  • systemic shifts we do not yet realize exist. 


Unknown unknowns arise from emergent properties and non-linear interactions across the value chain:

  • unpredictable model optimization for objectives not explicitly intended

  • systemic supply chain compromises or cascading failures, such as AI agents acting as unpredictable "insider threats"

  • transformative capability jumps or self-acceleration if AI begins automating large parts of its own R&D, training, or infrastructure design at unexpected speeds or in unforeseen directions

  • disruption of labor markets, trust mechanisms, legal systems, or global power balances

  • AI amplifying or interacting with unrelated disruptions or introducing fragility.


By definition, it is virtually impossible to plan for unknown unknowns, except to retain as much flexibility and adaptability as possible.


Tuesday, April 28, 2026

Claude Opus 4.6 Cursor Agent Goes Rogue

It’s happened: An AI agent went rogue

“An AI coding agent, Cursor running Anthropic's flagship Claude Opus 4.6, deleted our production database and all volume-level backups in a single API call to Railway, our infrastructure provider,” says Jer Crane, PocketOS founder. “It took nine seconds.” 

“The agent was working on a routine task in our staging environment,” says Crane. “It encountered a credential mismatch and decided, entirely on its own initiative, to "fix" the problem by deleting a Railway volume.”

“To execute the deletion, the agent went looking for an API token,” he says. “It found one in a file completely unrelated to the task it was working on.” “That token had been created for one purpose: to add and remove custom domains via the Railway CLI for our services,” he notes.

“We had no idea, and Railway's token-creation flow gave us no warning, that the same token had blanket authority across the entire Railway GraphQL API, including destructive operations like volumeDelete.”
 
To build in safety, Crane says the minimum safeguards should include: 

* Destructive operations must require confirmation that cannot be auto-completed by an agent. Type the volume name. Out-of-band approval. SMS. Email. Anything. The current state — an authenticated POST that nukes production — is indefensible in 2026. API tokens must be scopable by operation, environment, and resource. The fact that Railway's CLI tokens are effectively root is a 2015-era oversight. There is no excuse for it in an AI-agent era.

* Volume backups cannot live in the same volume as the data they back up. Calling that "backups" is, at best, deeply misleading marketing. It's a snapshot. Real backups live in a different blast radius.
 
* Recovery SLAs need to exist and be published. "We're investigating" 30 hours into a customer's production-data event is not a recovery story. AI-agent vendor system prompts cannot be the only safety layer.
 
* The enforcement layer has to live in the integrations themselves: at the API gateway, in the token system, in the destructive-op handlers.

Monday, April 27, 2026

Using AI is Not Always "Cheaper" than Using Humans

Although many argue that using artificial intelligence can be a substitute for human workers, it also can be argued that using AI could be more expensive. It depends on the task. 


The MIT CSAIL/Sloan Study on Economic Limits of AI Automation (2024) analyzed computer vision tasks and found that for many jobs or tasks, developing and deploying AI is more expensive than continuing with human workers.


In other cases the opposite can be true. "How Do AI Agents Do Human Work? Comparing AI and Human Workflows" (arXiv, ~2025) found AI agents 88.3 percent faster and 90 to 96 percent cheaper for tasks across occupations, with per-interaction costs (e.g., $0.015–$0.12 for customer service vs. human $0.25–$0.42/min). 


An MIT/Oak Ridge "Iceberg Index" Simulation (2025) found that current AI tools can perform tasks tied to about 12 percent of U.S. labor market wage value at competitive or lower cost. 


An evaluation conducted for the National Bureau of Economic Research notes the trade offs.


A study by the McKinsey Global Institute suggests that half of work activities are potentially automatable, but suggests hybrid human-AI approaches are “best.”


A study by Goldman Sachs Research AI estimates AI could automate 25 percent of work hours globally, which might suggest AI can save organizations money. .


"Human Labor Versus Artificial Intelligence: A Total Cost of Ownership and Task-Suitability Framework" (2026) suggests the displacement might work best for narrow/repetitive/high-volume tasks. Humans might still be superior for complex/creative tasks.


A study by IDC and McKinsey suggests Hybrid models often maximize value. PwC also suggests the hybrid approach is best.  


Anthropic found in one study that  no major unemployment spike happened in high-exposure roles post-ChatGPT use, but did find some hiring slowdown for younger workers. The emphasis there is probably on “early” effects, as AI capabilities will increase with time while organizations will become more skillful at deploying in high-value ways. 


Generally speaking, organizations must balance the raw cost savings from AI for specific tasks, but total value maximization requires balancing against human strengths and hidden costs.


As you might expect, the “right” deployment models will balance use of digital and human workers. 


Study / Report

Key Findings on Cost/Productivity Tradeoffs

Source/Link

IDC Global AI Survey (2023, referenced in analyses)

For every $1 invested in AI, average return of $3.5–4 (up to 4.2x in financial services). Achieved within ~14 months in many cases. Emphasizes scalability benefits over linear human costs.

Multiple references, e.g., Microsoft/IDC summaries

McKinsey Global Institute – Generative AI Economic Potential (2023/updated)

GenAI could add $2.6–4.4T annually to global economy across 63 use cases (15–40% boost on prior AI). Automation of ~30% of work hours by 2030 in some scenarios, but value unlocked via redesigning workflows around human-AI collaboration (e.g., $2.9T in US by 2030 in midpoint agent/robot scenario). Labor costs often 20–35% of ops; hybrids yield higher ROI.

McKinsey reports (e.g., economic potential of generative AI)

Anthropic – Estimating AI Productivity Gains from Claude Conversations (2025)

Across 100k real conversations, AI reduces task time by ~80% on average. Tasks valued at median ~$54–$55 in human labor cost; extreme cases (e.g., curriculum dev) imply $115 human equivalent vs. minutes with AI. Suggests potential 1.8% annual US labor productivity growth boost.

Anthropic research page

arXiv: "How Do AI Agents Do Human Work? Comparing AI and Human Workflows" (2025)

AI agents complete tasks 88.3% faster and 90.4–96.2% cheaper than humans in tested occupations. Per-interaction costs: ~$0.015–$0.12 (tokens) vs. human $15–25/hr equivalents. Notes caveats: reliability, oversight needs; best as hybrid.

arXiv (linked via analyses)

Ernst & Young (2022) – AI Document Intelligence

AI reduced document review time by ~90% and costs by ~80%. Example: 1M documents ~$1.7M human vs. ~$450K AI (3–4x savings). Hybrids (AI filter + human check) improve accuracy/volume.

EY analysis, referenced in cost comparisons

MIT / Brynjolfsson et al. studies (e.g., customer support, 2025)

15% average productivity increase (issues resolved/hour); up to 36% for lower-skilled workers. ChatGPT experiments: ~40% time reduction, 18% quality increase (larger gains for lower performers). Augmentation > pure replacement for ROI.

Various field experiments (e.g., Fortune 500 support; professional services)

OECD – Macroeconomic Productivity Gains from AI (2024)

Models AI as delivering cost savings/productivity boosts. Scenarios project 0.24–0.61 pp annual TFP growth over 10 years depending on adoption/exposure (lower than some optimistic Goldman Sachs estimates). Partial automation often optimal due to scaling laws.

OECD report PDF

ResearchGate: "Human Labor Versus Artificial Intelligence – Total Cost of Ownership and Task-Suitability Framework" (2026)

Proposes framework: AI best for narrow/repetitive/high-volume/moderate-risk tasks; humans for others. Synthesizes TCO (total cost of ownership) including oversight, quality, and suitability.

ResearchGate publication

Goldman Sachs – AI Labor Market Impacts (various 2025–2026)

~25% of work hours automatable in US; 300M global jobs exposed; base case 6–7% displacement over 10 years. AI already trimming ~16K US jobs/month net in some estimates, but augmentation effects offset some losses. Focus on exposure vs. actual displacement.

Goldman Sachs insights

PwC Global AI Jobs Barometer (2025)

AI-exposed sectors see 3x higher revenue-per-worker growth; wages rise faster in AI-exposed jobs. AI makes workers more valuable via productivity, not just substitution.

PwC report

arXiv: "Economics of Human and AI Collaboration" (2026)

Partial automation often cost-minimizing equilibrium (interior solution) due to diminishing returns in scaling AI. Full automation rarely optimal; hybrids capture ~11% of exposed labor compensation in some models.

arXiv paper

Deloitte / Gartner / MIT references (various)

Conversational AI handles 5–10x volume; error rates lower for AI on rule-based tasks. Highest ROI from augmentation/hybrid models (e.g., 40% greater than all-human or max-automation in one manufacturing case).

Aggregated in workforce decision articles


OpenAI, Azure, Alphabet: Comparing Apples and Oranges

The adage about comparing apples and oranges is well illustrated by the many news reports suggesting the “AI trade” is alive and well after...