Zero-Click Already is Changing Search

The implications of zero-click search (where a search does not end in a click to one of the results on the search engine results page) for search providers, advertisers and content providers are understandably huge. 

Oddly enough, the data so far might suggest that zero-click has had a neutral impact on Google search revenue, for example, whatever the possible future impact. Advertisers are working to shift their ad buys in ways that feature inclusion in the AI summarizes, for example.

Content providers arguably have been hardest hit, as the decrease in Web traffic affects their ability to monetize content using advertising placements.

In 2024, 65 percent of all global searches on Google are estimated to be zero-click, according to Briskon. Mobile searches tend to feature more than 75 percent zero clicks. Ahrefs analyzed 300,000 keywords and found that the presence of an “AI Overview” in the search results correlated with a 34.5 percent lower average clickthrough rate for the top-ranking page, compared to similar informational keywords without an AI Overview. 

Bain’s February 2025 research suggests that about 80 percent of consumers now rely on “zero-click” results in at least 40 percent of their searches, reducing organic web traffic by an estimated 15 percent to 25 percent. 

It is not hard to understand why the trend exists. Some 40 percent to 70 percent of generative AI model users use the platforms to conduct research and summarize information (68 percent), understand the latest news and weather (48 percent), and ask for shopping recommendations (42 percent).

   

source: Bain

Will AI Lead to Cognitive Costs "Near the Cost of Electricity?"

“The cost of intelligence should eventually converge to near the cost of electricity,” says OpenAI head Sam Altman. The economic implications could be quite significant. 

For example, the pricing of services that rely on intelligence (consulting, legal, creative) could approach the marginal cost of electricity, disrupting traditional business models. In other words, at least some cognitive tasks might become commoditized. 

Even if we might not be able to directly compare the “cost” of cognitive activity and equivalent operations conducted by an AI model, we might all agree that AI generally uses more energy and resources upfront than humans to achieve a similar single outcome, but then can scale to produce vastly more output with lower marginal cost. 

In other words, the AI advantage comes when we scale the activities. Looking at the matter in terms of water or electricity consumption, humans use relatively little energy and water in performing knowledge work, but throughput is limited and cost scales roughly linearly with quantity.

If one human produces one unit of work, then 10 units requires 10 humans. AI outperforms at scale. 

From an environmental perspective, a single human brain is “greener” than a single massive AI for one unit of task; however, to match an AI that can do one million tasks, you’d need an army of humans whose combined footprint (millions of computers, offices, and lives) might then rival or exceed the AI’s footprint.

There are other imponderables as well. At least some might speculate that we are entering an age where cognitive labor scales like software: infinite supply, zero distribution cost, and quality improving constantly. 

To be sure, the cost of employing cognitive workers is far more complicated than simple consumption of electricity and water. Still, at scale, AI impact on cognitive work does seemingly create economies of scale. 

By that logic, AI doesn’t just automate tasks; it commoditizes thinking. Of course, that looks only at cognitive input costs, not outputs. We probably are going to have to look at outcomes produced by using AI at scale, such as curing a particular disease or reducing production costs for some product. 

Still, it is shocking to ponder the economic implications of cognitive costs related to the cost of the electricity and water required to produce the models and inferences.

Waymo Features a Rider Cost "Premium" Compared to Uber or Lyft, By Design

It appears there is an early adopter pricing premium being paid for Wayno rides, compared to either Uber or Lyft, according to a study by Obi. 

To be sure, that is by design: Waymo entered the market with a premium pricing position. So going driverless doesn’t mean a cheaper ride; instead it is deliberately priced at a premium to either Uber or Lyft. 

source: Obi

Why AI Era of Computing is Different

If we can say computing has moved through distinct eras, each with distinct properties, it is not unreasonable to predict that artificial intelligence represents the next era. And though earlier generations are normally defined by hardware, that is less true of more-recent eras, where virtualization is prevalent and the focus is more on applications than hardware. 

But AI might shift matters further. 

Era	Key Feature	Key Technologies
Mainframe Era (1950s–1970s)	Centralized computing	IBM mainframes
Personal Computing Era (1980s–1990s)	Decentralization, personal access	PCs, MS-DOS, Windows
Internet Era (1990s–2000s)	Connectivity, information access	Web browsers, search engines
Mobile & Cloud Era (2000s–2020s)	Always-on, distributed services	Smartphones, AWS, Google Cloud

The AI era should feature software “learning” more than “programming.” Where traditional software follows explicit rules, AI models learn from data, discovering patterns without being explicitly programmed.

AI systems can generalize from experience and sometimes operate autonomously, as in the case of self-driving cars, recommendation systems or robotic process automation.

Voice assistants, chatbots, and multimodal systems mark a transition to more human-centric interfaces, moving beyond keyboards and GUIs.

AI can be considered a distinct era of computing, not because it introduces new tools, but because it changes the nature of computing itself, from explicitly coded systems to systems that evolve and learn.

Why Language Models Tackling Very-Complex Problems Often Provide Incorrect Answers

It perhaps already is clear that large language models using “reasoning” (chain of thought, for example) can provide much-better accuracy than non-reasoning models. 

Reasoning models consistently perform well on simple queries. Benchmarks such as MMLU (Massive Multitask Language Understanding) and ARC (Abstraction and Reasoning Corpus) show that models reach near-human or superhuman accuracy on tasks that do not require multi-step or abstract reasoning.

But reasoning models frequently exhibit "reasoning is not correctness" gaps as query complexity grows. Performance degrades sharply as step counts increase, as is required for complex tasks. Token usage grows three to five times for complex queries while accuracy drops 30 percent to 40 percent compared to simple tasks.

Reasoning models using techniques like Chain-of-Thought (CoT) can break down moderately complex queries into steps, improving accuracy and interpretability. However, this comes with increased latency and sometimes only modest gains in factuality or retrieval quality, especially when the domain requires specialized tool usage or external knowledge.

The biggest issues come with the most-complex tasks. Research shows that even state-of-the-art models experience a "reasoning does not equal experience" fallacy: while they can articulate step-by-step logic, they may lack the knowledge or procedural experience needed for domain-specific reasoning, for example. 

That happens because the reasoning models use logically coherent steps that contain critical factual errors. In scientific problem-solving, 40 percent of model-generated solutions pass syntactic checks but fail empirical validation, for example. 

Such issues are likely to be most concerning for some use cases, rather than others. For example, advanced mathematics problems such as proving novel theorems or solving high-dimensional optimization problems often require formal reasoning beyond pattern matching

Problems involving multiple interacting agents (economic simulations, game theory with many players) can overwhelm LLMs due to the exponential growth of possible outcomes.

In a complex negotiation scenario, an LLM might fail to account for second-order effects of one agent’s actions on others. 

Also, problems spanning multiple domains (designing a sustainable energy grid involving engineering, economics, and policy) require integrating diverse knowledge LLMs were not trained on. 

Of course, one might also counter that humans, working without LLMs, might very well also make mistakes when assessing complex problems, and also produce logically reasoned but still "incorrect" conclusions! 

But there are probably many complex queries that still will benefit, as most queries will not test the limits of advanced theorems, economic simulations, game theory, multiple domains and unexpected human behaviors. 

So for many use cases, even complexity might not be a practical issue for a reasoning LLM, even if they demonstrably become less proficient as problem complexity rises. And, of course, researchers are working on ways to ameliorate the issues. 

AI Text Only Suffers from Emotional Flatness In Some Use Cases

As with all generalizations, the claim that writing produced by artificial intelligence or generative AI suffers from a lack of emotion requires some elaboration. Not all writing tasks require or even allow much “emotional expression.”

Academic or essay writing; advertising and marketing content; history or instructional content might do just fine with a straightforward style. 

On the other hand, most of us might wonder how well present and future models will be able to handle fiction, where nuance, emotional depth, and subtlety or a unique voice do matter. 

AI might also not be so great for memoirs, reflective essays, or opinion pieces.

The reasons for the difference are pretty simple. Genres such as academic writing, advertising, and instructional content follow established structures and therefore are easier for AI to mimic.

Fiction and personal narratives require a level of creativity, empathy, and emotional understanding that AI systems currently struggle to replicate. AI can mimic certain tones and styles, but it often lacks the unique voice and perspective that human writers bring to their work.

The point is that AI content, which already is prevalent, will seem more appropriate in some genres, compared to others. One size, as they say, does not fit all. And as useful as AI might be for many humans, in many situations, writers are not going to stop writing because they could use AI for that purpose. 

No, writers write because they enjoy the craft of writing, just as musicians play music or artists paint. AI will not deter any of these creators from doing what they enjoy. My brother wouldn't get any enjoyment out of having AI paint a picture. My sister wouldn't prefer that an AI create and play music. I wouldn't be interested in using AI to write on my behalf. I write because I enjoy the process.

Still, to the extent that AI is a tool to automate or speed writing tasks for many, when it is simply a practical task, the inability to fully mimic human nuance will not be an issue. We don't expect nuance in our emails, ads, marketing copy, technical training manuals or instructional material, really. We never expect it for legal, academic or technical writing, either.

"Lack of emotion" is an issue mostly for creative or fiction writing or biographies; film and TV scripts and often musical lyrics.

How to Avoid or Reduce the Danger of Model Collapse

Recursive training on synthetic data, often referred to as "model collapse," is a significant challenge for artificial intelligence models. It is a point Apple researchers made recently. A paper entitled The Illusion of Thinking suggests reasoning language models underperform standard models for simple problems; show some advantage for medium-complexity tasks but collapse under the weight of high-complexity tasks. 

“We acknowledge that our work has limitations,” the authors note. “While our puzzle environments enable controlled experimentation with fine-grained control over problem complexity, they represent a narrow slice of reasoning tasks and may not capture the diversity of real-world or knowledge-intensive reasoning problems.” 

Some might note that the problem of complex problem processing is not new. In machine learning, recursive training generally refers to a process where a model is trained, and then its own outputs (or outputs from a previous version of itself) are used as part of the training data for subsequent iterations of the model. 

This creates a feedback loop where the model is essentially "learning from itself" or from other models that have similarly learned from generated data.

This leads to a degenerative process where errors compound, and the model's outputs become increasingly narrow, repetitive, and nonsensical.

As always, developers and architects have methods for reducing such distortions. 

The most crucial strategy is to retain a significant portion of original, human-generated, "real" data. By some estimates, even a small percentage of real data (10 percent) can significantly slow down or prevent model collapse.

Continuously introducing new, diverse real data into the training pipeline can counteract the degradation caused by synthetic data. 

Architects also can implement robust processes to verify the quality and accuracy of synthetic data before it's used for training.

Also, architects can focus on creating synthetic data to address specific model "blind spots" or underrepresented scenarios in the real data. 

Architects can also try to ensure that the synthetic data generated is diverse and representative of the desired data distribution, rather than simply mimicking the most common patterns. 

Human feedback also is helpful, and could involve humans evaluating the quality of generated data and providing guidance for the next generation.

Training can use regularization methods (L1/L2 regularization, dropout) during training to prevent overfitting and encourage the model to learn more robust representations. Reinforcement Learning with Human Feedback (RLHF) can be used to align the model's outputs with human preferences, effectively guiding the generation process towards more desirable and accurate results, even when using synthetic data.

The core principle to combat model collapse is to ensure that the model always has access to a reliable source of "ground truth" information, whether through direct inclusion of real data, or through careful curation and validation of synthetic data.

Will Content Companies Fair Better than Telcos with Content Business Separations?

With Warner Brothers Discovery slitting into two companies, one containing the linear video networks and the other the studies and streaming business, some of us might recall similar splits in other industries intended to separate low-growth, cash-flow-generating businesses from higher-growth activities. 

It might seem, in retrospect, an odd thing that telecom executives at AT&T and Rochester Telephone Co., for example, once considered “long distance calling” the growth business, where the local telecom operations were seen as the low-growth entities. 

As it turned out, long distance calling was not much of a business at all, once internet calling developed. Competition was an issue, of course, but the real profit killer was free and nearly-free IP-based calling, plus the rise of mobile calling that simply included domestic long distance calling as a feature. 

Period	Market Structure	Estimated Profit Margin
1985-1990	Post-AT&T Breakup Transition	60-80%
1990-1995	Early Competition	45-60%
1995-2000	Intense Competition	25-40%
2000-2005	Market Maturation	15-30%
2005-2010	Mobile Disruption	5-20%
2010-2015	VoIP/Internet Disruption	0-10%
2015-2020	Legacy Service	-5-5%
2020-2025	Vestigial Market	-10-0%

The point is that, as logical as it seems, such asset separations often do not work out as expected. 

The fundamental problem, some will argue, is that linear TV networks are experiencing accelerating decline rather than stable cash flow. 

source: PwC 

Year	Share of TV Households Without A Traditional TV Connection
2026*	75%
2023	60%
2022	53%
2021	47%
2020	41.70%
2019	36.10%
2018	30.60%
2017	26.20%
2016	22.60%
2015	20%
2014	18.80%

The theory behind separating growth and harvest assets assumes the declining business can generate predictable cash flows while the growth business receives focused investment. But some might note that linear TV's decline appears too steep for reliable cash generation.

The moves might still be “rational” from a management perspective, as it offloads assets in terminal decline (allowing a “harvesting” strategy) while unburdening the potential growth assets (reducing debt, boosting revenue growth rates, reducing overhead). 

Sure, people are going to propose all sorts of new business models or products that might be created. That sort of advice has been common since the 1990s for telecom service providers. And, fortunately, mobility service rose to become the driver of revenue and profits, with the rise of internet access helping replace lost voice revenues as well. 

It might be harder to envision what changes of product or revenue model could develop for linear video.