What AlphaGo and a $100 Pressure Washer Reveal About Judging Performance

Two recent articles make a similar editorial point from entirely different fields: impressive results are easy to misinterpret when we focus on appearances, reputation, or headline specifications.

One examines whether modern large language models genuinely reason. The other compares an inexpensive electric pressure washer with a considerably more expensive machine. In both cases, the useful question is not whether the cheaper or newer system looks impressive. It is what mechanism produced the result, how was it tested, and where does the evidence stop?

AlphaGo’s “creative” move was not machine intuition

In March 2016, AlphaGo defeated professional Go player Lee Sedol by four games to one. The match became famous partly because of move 37 in the second game: AlphaGo placed a stone on the fifth line, a move that appeared so unusual that some commentators initially wondered whether it was an error.

Lee later described the move as evidence of creativity. That interpretation is understandable, but the source material argues that it was incomplete. Move 37 was not simply a flash of intuition emerging from a system that guessed the most likely human move. It resulted from a more elaborate process involving reasoning and evaluation.

The distinction matters because Go is not like chess in the way a computer evaluates positions. Deep Blue’s 1997 victory over Garry Kasparov relied on hard-coded rules and an enormous search through chess positions. Go has a vastly larger range of plausible moves, and the value of a stone may depend on distant groups, territory, and sequences extending many moves into the future.

A system attempting to calculate every possible continuation would quickly encounter an unmanageable number of positions. AlphaGo therefore needed more than brute-force calculation. It had to estimate the overall state of the board and identify promising moves without exhaustively examining every possibility.

Two different kinds of capability

The AlphaGo system described in the source material contained two components:

• A policy network, trained to predict moves that a strong human player might make.
• A separate process capable of evaluating positions and selecting a move through a deeper form of search and assessment.

The policy network treated move 37 as unremarkable rather than as an obvious breakthrough. The move became remarkable because the broader system was able to assess the board beyond the immediate expectations of human play.

That is why calling AlphaGo’s decision “intuition” can obscure the mechanism behind it. The move looked intuitive to observers because it was difficult to explain in familiar human terms. But the result came from a system designed to compare possible actions and estimate their consequences in a complex environment.

Why the AlphaGo comparison does not automatically explain today’s LLMs

Large language models can produce fluent explanations, solve some problems, revise answers, and show patterns that resemble deliberation. Those behaviors can make it tempting to treat them as reasoning systems in the same sense as AlphaGo.

The source’s central argument is that this conclusion goes too far. Today’s language models may generate reasoning-like text without possessing the kind of underlying machinery that allowed AlphaGo to evaluate a Go position and discover an unconventional but effective move.

That does not mean every output from an LLM is useless or that language models cannot perform multi-step tasks. It means that surface behavior is not enough to establish a reliable reasoning process.

A model can write a persuasive explanation while making a basic error. It can produce a chain of steps that appears orderly without ensuring that each step is valid. It may also arrive at a correct answer for reasons that are difficult to inspect or reproduce.

The practical consequence is straightforward: users should separate:

• Fluent explanation from verified reasoning
• A correct answer from a dependable method
• Novel wording from a genuinely novel insight
• Confidence from evidence

The argument does not prove that no LLM has any reasoning capability. Rather, it challenges the assumption that language fluency itself demonstrates the deeper reasoning powers needed for trustworthy scientific or medical discovery.

What trustworthy reasoning would require

For AI systems to produce dependable results in fields such as science and medicine, a convincing answer is not enough. A useful system would need to support more reliable evaluation of alternatives, consequences, and uncertainty.

The AlphaGo example suggests several properties worth looking for:

1. A representation of the problem’s underlying state, rather than only a prediction of likely language.
2. A way to compare possible actions or explanations.
3. Evaluation that extends beyond the next immediate step.
4. A mechanism for recognizing when an unusual option is stronger than a familiar one.
5. Testing against outcomes, not merely against how plausible an answer sounds.

These are design goals, not proof that a particular current model possesses them. The source material presents AlphaGo as an example of why reasoning should be understood as a capability with a mechanism behind it—not as a label applied after a system produces an impressive result.

The $100 pressure washer that challenged the premium model

The second evaluation concerns a much more practical question: whether spending several hundred dollars on an electric pressure washer necessarily buys a better cleaning experience.

The comparison tested the LawnMaster LT506-2000A, a budget model priced at roughly $100 to $120 depending on its pressure rating, against an AR Blue Clean 383SS purchased for about $250 several years earlier. The specific AR model was no longer readily available, while comparable models were described as costing substantially more.

Both machines were tested on the same surfaces, using the same water source and comparable spray settings. The reviewer found that the LawnMaster performed extremely well, particularly given the price difference, and judged it sufficient for most homeowners.

That conclusion is useful because it is based on observed performance rather than brand position or purchase price alone. It does not establish that the LawnMaster is superior in every respect. It shows that, in the reported test, the less expensive machine delivered enough cleaning performance for many ordinary household tasks.

Specification comparison

Feature	LawnMaster LT506-2000A	AR Blue Clean 383SS
Advertised maximum pressure	Up to 2,300 PSI	Up to 1,900 PSI
Rated pressure listed in the manual	1,650 PSI	Not provided in the source material
Water flow	1.4 GPM	1.3 GPM
Motor	13-amp, 120-volt	13-amp
Approximate weight	22.6 pounds	32 pounds
Reported price context	About $100–$120	About $250 when purchased; similar models described as $300–$400
Test conditions	Same surfaces, water source, and comparable spray settings	Same surfaces, water source, and comparable spray settings

The figures suggest a modest advantage for the LawnMaster on the listed maximum pressure and flow ratings, as well as a lower weight. But specifications alone do not determine cleaning performance. Nozzle design, pressure regulation, durability, hose and accessory quality, and long-term reliability can all affect the ownership experience.

The test result therefore carries more weight than the price tag, but it should not be stretched beyond what was actually examined.

Why the pressure-washer result needs qualification

The comparison has several useful controls: both machines were used on the same surfaces, with the same water source and comparable spray settings. That makes the observed difference more meaningful than a casual impression formed from using each machine in unrelated circumstances.

There are still limitations:

• The AR Blue Clean model was older and may not represent the performance of newer premium machines.
• The specific model was no longer available, so the comparison does not directly evaluate every current alternative.
• The source material describes a single hands-on comparison rather than a broad, independently replicated test across many machines.
• The result focuses on cleaning performance, not necessarily durability, maintenance, noise, warranty support, or accessory quality.
• Maximum PSI figures should not be treated as a complete measure of practical cleaning ability.

For a homeowner, the reasonable takeaway is not “the cheapest pressure washer is always best.” It is that a budget model may provide enough performance for common tasks, and paying more should be justified by a specific need rather than by the assumption that a higher price guarantees better cleaning.

The shared lesson: test the mechanism, not the impression

AlphaGo and the LawnMaster pressure washer have little in common mechanically. One selected a move in a strategic board game; the other removed dirt from household surfaces. Yet both examples illustrate the same discipline of evaluation.

AlphaGo’s unusual Go move looked creative because it differed from established human expectations. The explanation lay in the system’s ability to evaluate the board, not necessarily in a human-like spark of inspiration. The LawnMaster’s performance looked surprising because its price suggested a compromise, but the test indicated that the machine could still deliver adequate results for many users.

The examples support a distinction between an impressive output and a dependable mechanism. An LLM’s polished response is not proof of a reliable reasoning process, just as a pressure washer’s maximum PSI is not proof of superior real-world cleaning.

A result is also not the same as a universal verdict. The AlphaGo example supports a distinction between prediction and deeper evaluation, but it does not by itself settle the entire debate about how modern LLMs work. Likewise, the pressure-washer comparison supports the view that the LawnMaster can be a strong value for many homeowners. It does not show that it will outperform every more expensive model, last longer, or satisfy every demanding use case.

Practical takeaways for readers

When evaluating an AI system, ask:

• Does it merely produce a plausible answer, or can its reasoning be checked?
• Does it evaluate alternatives and consequences?
• How does it handle uncertainty and correction?
• Has its performance been tested on the kind of problem you actually care about?
• Is an unusual answer genuinely supported, or simply expressed with confidence?

When choosing an electric pressure washer, ask:

• What surfaces and tasks will it handle?
• Are the PSI and GPM figures clearly identified as maximum or rated values?
• Does an independent hands-on test use comparable conditions?
• Is the machine’s lower price offset by weaker accessories, limited durability, or fewer features?
• Do you need premium performance, or only enough power for routine household cleaning?

Bottom line

AlphaGo’s famous move was not evidence that machines had suddenly acquired human-style intuition. It demonstrated the value of a system capable of evaluating complex possibilities beyond the most obvious human choice. That distinction remains relevant when judging large language models, whose fluent outputs can resemble reasoning without necessarily providing the same dependable machinery.

The pressure-washer comparison reaches a different but related conclusion: a lower-priced product can perform impressively when tested fairly against a more expensive alternative. Yet the result should be read as a measured buying recommendation, not a universal ranking.

The common lesson is simple: look past the impressive output, the premium price, or the headline specification. Ask how the result was produced, what was tested, and what the evidence can actually support.

Sources

• Technology Review, “Don’t be fooled—LLMs don’t reason,” published October 2, 2026: https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/
• CNET, “This $100 Power Washer Whupped My More Expensive Model. I Was as Shocked as Anyone,” published October 3, 2026.