The Moonspeaker:
Where Some Ideas Are Stranger Than Others...
LLMs Are Not Search Engines (2026-06-29)
I know, I know – strictly speaking, since search engine algorithms are generally not publicly code, they may well be based on at least similar mathematical algorithms to those used in large language models. So having cheerfully made up a clickbait-style title, it is as well to get a bit more specific as to what I mean. A great deal of coding time, electricity, and water (to start with) has been poured into trying to make LLMs "more powerful" since the big splash of ChatGPT's first public release. The difficulty with dubiously named "machine learning" and subsequent programs flogged to the general public is that they represent a vast effort to create a tool for which there is no general public problem to solve, let alone a cultivated "want" to satisfy. Branded LLMs are not the equivalent of the word processing or spreadsheet programs that made computers approachably useful to people who were not already computer programmers. But so much money has already gone into making these LLMs, including venture capitalist funding, that the desperation to make them somehow, some way into products people will buy ongoing subscriptions for is all but palpable. Based on my observations, they seem to be undertaking a two-pronged strategy. First, deliberately code the LLM interface to imitate a person at least enough to convince real people who want to be convinced, while constantly labelling program outputs with emotive terms for human behaviours and experiences. Hence the penchant for referring to LLMs as "hallucinating," "lying," or "making things up" all of which attribute non-existent intentionality to the LLM. Second, doing all they can to reduce the usability and accessibility of the web. Their original mainstay here was to deploy computing resources to running their associated web crawling bots, and not being too careful about it so that they regularly crash websites of all sizes and types by abusively downloading their contents, wiping out their bandwidth, or absorbing all connections so they are unusable. Now they turn a blind eye to the proliferation of "AI slop," which now includes entire pseudo-websites of LLM-generated imitations of articles and pictures. If the web is ever more difficult to actually use because sites are down or crashed, and kingpin search engine google is in a state of ruin, then the LLM-sellers can try to position their models as replacements for search engines and actually browsing the web.
UPDATE 2025-08-02 - For another perspective on what LLMs are and what they can do from people doing some intriguing technical work, see the Asahi linux documentation on Generative AI. Don't let the "technical" descriptor put you off, the page is well written in plain language.
I don't think people are actually taken in by this iteration of a crook causing a problem and then selling a "solution" to it that lines their pockets. There are many people who have played around with say, chatGPT or google's johnny-come-lately response, finding these programs to be something of a novelty and going on their way, others seeking to use them for specific tasks, and the least lucky ones ordered to use "AI" at work to "improve their efficiency." There is a potent minority of people who seem to be utterly convinced that LLMs like microsoft's copilot or chatGPT are near-perfect at such tasks as summarizing documents, writing papers and so on. Among them seem to be many managers and employers engaged in cynically following the trend of "increasing efficiency by adopting AI" in expectation of forcing a significant portion of their staff to train the LLMs into making their jobs redundant. Another portion of this influential minority seem convinced these LLMs are a sort of oracle which they can ask all manner of questions, including for personal and medical advice. This latter view is certainly encouraged by attempts to make the LLMs sound "human." Then again, people viewing LLMs as oracles are probably very much in the minority, but their at best alarming stories make great media copy.
Unlike search engine algorithms, the basics of how LLMs work is public knowledge, and there is even an entire series of course lectures by revered mathematics professor Gilbert Strang about the deeper technical elements via MIT's opencourseware project. Practically speaking, LLMs are based on applications of linear algebra, mathematical operations on matrices of numbers. To take an extremely simplified illustration here, take two linear equations such as 2x + 3y - z = 0 and x - 7y + 4z = 2. They can be represented as a grid of numbers, with the numbers from the left side in a "matrix" and the numbers on the right side in a corresponding "vector," as shown below.
2 3 -1 | 0
1 -7 4 | 2
It is astonishing what can be done by representing data in this way. One of the most commonly encountered applications is in signal processing, where manipulating systems of equations in matrix forms does things such as removing reverberation effects from seismic data or processing tracks in a music studio. A more day-to-day use is in basic digital photo processing. Processing language in a similar fashion may seem like a bit of a stretch, though it isn't necessarily. In fact, we have an everyday, very simple equivalent to how LLMs generate sentences and larger blobs of text in the form of the game often referred to as "mad libs." And it does make sense that a linear algebra-based program can be written and developed to imitate certain types of language, and even to do machine language translation. Those "certain types" have a strongly patterned nature. The programmers developing LLMs strive to write code that can accurately identify language patterns that can then be used to produce an output based on a prompt to get the ball rolling. I have read descriptions of this being a matter of the LLM predicting the most likely word one at a time, along the lines of starting with a properly selected noun from the prompt (assuming the language is english as I am writing in here), predicting the most probable verb, and so on. It seems likely there are some ready-made patterns already available, such as one for a five-paragraph essay or a book report. There is no denying that LLMs can generate some pretty impressive, sometimes useful, and sometimes startlingly human-like results. However, this approach is not at all similar to the underpinnings of search algorithms.
Suppose LLMs are "trained" on, that is set to apply clever mathematics to find patterns in, however many billions of known searches and search results that may be stored and available to acquire for analysis. Assuming for now the patterns identified are always real and not spurious or accidents, what the LLMs will be able to sort out first are the typical shapes search queries take, and the shapes of the sort of results people at least actually clicked on. This sounds like it could also be quite similar to the algorithms well-suited to generating and updating an index of interconnections and relevance ratings for searching. But LLMs like chatGPT are not indexers, they are building up "answers" based on a type of pattern based on mathematical abstraction and application of probabilities. They are not designed to simply return a result, but to generate something that seems plausible, at least to the person who typed in the query. This can have pernicious results. Take for example reference citations, for which there are several formal styles, all highly structured. It has turned out that LLMs are quite good at generating plausible-looking but utterly fictitious citations, even if their training data included accurate citations that would be appropriate responses to a person's query.
A search engine in the original, and yes I think proper sense, reports the results of a search. It does not in effect generate a result from bits and pieces put together in plausible order from a mass of unsorted data. I doubt LLMs are carrying out even a modified version of the most basic version of "Pagerank" from the early days of google as search engine. If they were, then the results would not be generated based on calculating a probable sequence of tokens, they would return pre-existing results. In the end it seems the real logic behind trying to position LLMs as search engines is not because that is what they are or what they could be, but because their boosters are desperate for "data." If only they could get enough data, supposedly that would make the LLMs "more powerful," by which they seem to mean "more convincing." However, it is hard not to see them as seeking ever more data in hopes of having a nice fat lot of it to sell off to so-called "security services" when the "AI" bubble bursts.
Previous Thought Pieces
|
