Cheerleading for the Google: A Soft Counter Howl
May 9, 2022
I have noted several posts which champion Google’s approach to smart software. I find it difficult to think about the cheerleading for Google’s “quantum supremacy” approach to its systems and methods. Dissent, disagree with the Jeff Deans of Google, or point out known flaws such as less useful results from a simple query — and what happens? The Google terminates people. The most recent example concerns a full fledged member of the Google High School science club. Dr. Satrajit Chatterjee suggested the emperor was wearing PR clothes. Yep, one can see things when the Big Dogs of Google parade around at conferences. For some “color” about Dr. Chatterjee’s misstep, check out this New York Times’ write up: “Another Firing Among Google’s A.I. Brain Trust, and More Discord.” (Paywall in place, but don’t complain to me.)
I read “Google AI Sparks a Revolution in Machine Learning.” Oh, really. I thought the Google’s machine learning was crafted from such methods as those presented by Dr. Christopher Ré, the Snorkel outfit, and the labors of engineers who recycle the original work of DeepMind. The novelty may be the PR, not the engineering.
The write up exclaims:
In all the hype around PaLM, people have not spent enough time understanding Google’s Pathways. But, when you do look into it, you will see that it is nothing short of a revolution. I’m not exaggerating.
See. “People” are not making an effort to understand the wonderfulness of Google’s method for reducing the cost of training machine learning models and then tuning those puppies with synthetic data. Why? “Real” data is increasingly difficult to get, even for the Google. Efficiency and cost reduction are the drivers. The PR push is designed to be a turbo charger.
Let’s take one small example of how the hype does not match what Google delivers. “Enterprise search” is a bound phrase. The idea is that if the bound phrase appears in a document, that document belongs in the search result set.
I get a sometimes daily and sometimes weekly summary of “important” and “relevant” documents sent to me via email. I use a competitive system as well, but the details of how that compares is not my concern. I want to focus on a result set of three:
enterprise search
NEWS
Global Enterprise Search Software Market â?? Recent Industry Trends and Projected Industry …
Global Enterprise Search Software Market â?? Recent Industry Trends and Projected Industry Growth 2021 â?? 2026. Admin Published: 8 minutes ago …
SYSTEM DEVELOPMENT SPECIALIST (ENTERPRISE SEARCH ENGINEER) – City of Toronto Jobs – City of Toronto Jobs
Desperate search for survivors in Cuba hotel blast; 27 dead – Beaumont Enterprise
They checked the morgue, hospitals and if unsuccessful, they returned to the partially collapsed Hotel Saratoga, where rescuers used dogs to hunt for …
Flag as irrelevant
Enterprise applies The Trust Project standards to increase news transparency
These indicators are a collaborative, journalism-generated standard that help both readers and search engines assess the authority and integrity …
Notice that the bound phrase “enterprise search” is interpreted by Google’s smart software in these ways:
The first citation is to a junk-type market research report which purports to provide a look at the future of the enterprise search market. Keep in mind that this is a market which is dominated by open source options and a handful of vendors chasing niches; for example, Coveo and customer support and Fabasoft Mindbreeze the Microsoft market. Other vendors are just desperate to make sales and try to sell to another outfit, do and IPO, or get more financing. Enterprise search is a tough sector, and it is now almost a commodity.
The second citation is to a job posting in Toronto. What? No job posts in Berlin. I saw one the other day for an enterprise search engineer with NLP expertise. Plus there are “contact us” pleas from numerous vendors in the search-and-retrieve game just focusing on the law enforcement, intelligence, and business intelligence sectors.
The third citation is about the natural gas explosion in a hotel in Cuba. The separate words “search” and “enterprise” appear in the citation. The problem is that Google’s smart software ignored the bound phrase and did key word matching unaware of the location of words (title and body of article) and the order of the words. Enterprise before search, right? Not for Google’s smart software.
The fourth citation is interesting. Same problem. No bound phrase but the order of the disconnected words is close. Of the four citations, the results are incomplete because I get alerts on the subject from another outfit. But the Google results are 50 percent accurate.
That’s what I mean by Google’s methods generating results that a close enough for horseshoes: 50 percent accuracy. That’s a high water mark when a Google user relies on one of Android’s forthcoming medical outputs. Do you want to stand in front of a Google self driving car and see if that system is 50 percent accurate? I sure don’t.
Observations:
- Google wants its methods to become the one true way to implement machine learning and, hence, smart software. Disagree and you get fired. Just ask Dr. Chatterjee or Dr. Gebru.
- Some of Google’s methods can be used to deliver high-value outputs. A good example is Heron Systems’ smart software which virtually killed Animal in about 50 seconds. Just bits, gentle reader, no bullets killing a US Department of Defense Top Gun.
- The PR is disconnected from what the Google system is doing: In my view, Google wants to cut costs, eliminate insofar as possible the subject matter experts who build training sets and updates, and find methods that justify displaying endless Grammarly and Liberty Mutual ads to me when I watch a YouTube video about differential equations or Russian bloggers explaining that life is okay in St. Petersburg. (Sure it is.)
I have a suggestion. PR fluff needs to be labeled. Otherwise, you may be in the path of a self driving Waymo and your 50/50 chance may not work out as you assume. Advertisers, so far, remain unaware of what’s shakin’ with their expensive bacon.
Stephen E Arnold, May 9, 2022
Comments
One Response to “Cheerleading for the Google: A Soft Counter Howl”
-
- Subscribe to Beyond Search
Feature archive
News archive
-
[…] alert is about half right or half wrong when it flags an item I am supposed to need. See this “Cheerleading” […]