Ignoring the Big Thing: Google and Its PR Hunger
December 18, 2023
This essay is the work of a dumb dinobaby. No smart software required.
I read “FunSearch: Making New Discoveries in Mathematical Sciences Using Large Language Models.” The main idea is that Google’s smart software is — once again — going where no mortal man has gone before. The write up states:
Today, in a paper published in Nature, we introduce FunSearch, a method to search for new solutions in mathematics and computer science. FunSearch works by pairing a pre-trained LLM, whose goal is to provide creative solutions in the form of computer code, with an automated “evaluator”, which guards against hallucinations and incorrect ideas. By iterating back-and-forth between these two components, initial solutions “evolve” into new knowledge. The system searches for “functions” written in computer code; hence the name FunSearch.
I like the idea of getting the write up in Nature, a respected journal. I like even better the idea of Google-splaining how a large language model can do mathy things. I absolutely love the idea of “new.”
“What’s with the pointed stick? I needed a wheel,” says the disappointed user of an advanced technology in days of yore. Thanks, MSFT Copilot. Good enough, which is a standard of excellence in smart software in my opinion.
Here’s a wonderful observation summing up Google’s latest development in smart software:
FunSearch is like one of those rocket cars that people make once in a while to break land speed records. Extremely expensive, extremely impractical and terminally over-specialized to do one thing, and do that thing only. And, ultimately, a bit of a show. YeGoblynQueenne via YCombinator.
My question is, “Is Google dusting a code brute force method with marketing sprinkles?” I assume that the approach can be enhanced with more tuning of the evaluator. I am not silly enough to ask if Google will explain the settings, threshold knobs, and probability levers operating behind the scenes.
Google’s prose makes the achievement clear:
This work represents the first time a new discovery has been made for challenging open problems in science or mathematics using LLMs. FunSearch discovered new solutions for the cap set problem, a longstanding open problem in mathematics. In addition, to demonstrate the practical usefulness of FunSearch, we used it to discover more effective algorithms for the “bin-packing” problem, which has ubiquitous applications such as making data centers more efficient.
The search for more effective algorithms is a never-ending quest. Who bothers to learn how to get a printer to spit out “Hello, World”? Today I am pleased if my printer outputs a Gmail message. And bin-packing is now solved. Good.
As I read the blog post, I found the focus on large language models interesting. But that evaluator strikes me as something of considerable interest. When smart software discovers something new, who or what allows the evaluator to “know” that something “new” is emerging. That evaluator must be something to prevent hallucination (a fancy term for making stuff up) and blocking the innovation process. I won’t raise any Philosophy 101 questions, but I will say, “Google has the keys to the universe” with sprinkles too.
There’s a picture too. But where’s the evaluator. Simplification is one thing, but skipping over the system and method that prevents smart software hallucinations (falsehoods, mistakes, and craziness) is quite another.
Google is not a company to shy from innovation from its human wizards. If one thinks about the thrust of the blog post, will these Googlers be needed. Google’s innovativeness has drifted toward me-too behavior and being clever with advertising.
The blog post concludes:
FunSearch demonstrates that if we safeguard against LLMs’ hallucinations, the power of these models can be harnessed not only to produce new mathematical discoveries, but also to reveal potentially impactful solutions to important real-world problems.
I agree. But the “how” hangs above the marketing. But when a company has quantum supremacy, the grimness of the recent court loss, and assorted legal hassles — what is this magical evaluator?
I find Google’s deal to use facial recognition to assist the UK in enforcing what appears to be “stop porn” regulations more in line with what Google’s smart software can do. The “new” math? Eh, maybe. But analyzing every person trying to access a porn site and having the technical infrastructure to perform cross correlation. Now that’s something that will be of interest to governments and commercial customers.
The bin thing and a short cut for a python script. Interesting but it lacks the practical “big bucks now” potential of the facial recognition play. That, as far as I know, was not written up and ponied around to prestigious journals. To me, that was news, not the FUN as a cute reminder of a “function” search.
Stephen E Arnold, December 18, 2023
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
