Lemon-flavored Data Centers

To state the obvious: a data center costs a lot of money.

When there’s a lot of money, there are contracts. Finish the foundation by this date, assemble the superstructure here. Get the zoning and land contracts ready. Acquire electricity, adequate data access, and other utilities needed to launch.

These all could be pretty straightforward. It might be difficult to negotiate and arrange the resources to accomplish them, but they’re one-off risks. Either the contract will be fulfilled or the penalties for non-performance will stop once the contract is completed or canceled.

a slice of lemon

Once the data center is launched, the contracts are ongoing. The sales teams guarantee a specific speed and reliability statistics for the managers to achieve. These might be evergreen contracts that are kept active while the owners offer those system advantages, If a data center consistently fails to meet those requirements, it can lead to business interruption and liability claims.

When I buy a new car, there might be a lemon law that applies. If it cannot be repaired adequately, the car is deemed a lemon and can be refunded. Would a data center that is consistently unable to meet contracted reliability levels be considered a lemon data center? Perhaps it might just be “lemon-flavored” since there’s no way to get a refund and return the massive air-conditioned box and its contents.

a pencil signing a contract

For most contractual obligations, a level of insurance can mitigate the risks of non-compliance. Since each stage of development has clear contractual obligations, there are insurers with an appetite for each stage of risk. Each policy, from start to finish, will have clear failure criteria and underwriters familiar with the issues can cover specific project phases. Well-placed insurers are a resource to guide the project to success and minimize their own exposure.

Many features of an active data center are also easy to assign risk for… Fire suppression systems, power failures, and cooling lapses that lead to inadequate temperature and humidity control. An insurance claim would be straightforward when something breaks like a cooling unit or power system.

Where the lemon-flavoring can come from is electronics and software configurations. Since a fault in a rack may not be identified precisely, insuring the capital behind such failures could be unattractive for underwriting. If there is a high risk of business interruption, the underwriters might change their expectations.

As a point of comparison, consider the backup company Backblaze. It collects detailed information about the reliability of their hard drives. They publish quarterly reliability data for around 350,000 drives. Their data allows outside researchers to evaluate drive model reliability. I noticed in one recent statistics chart that there can be a 10-fold range of reliability between different drive models.

One interesting message in their first quarter 2026 report was about some instances of a certain drive was having problems. Stephanie Doyle reported that the drives were having excessive failures. It was a complex situation that took an extended time to discover its solution. That reaction was not “we need to get rid of these drives”, but rather “we need to understand what was going wrong.” The problem ended up to be a complex interaction of several factors. Her conclusion was “our drive failure rates, and what we report on in general, is the function of people actively managing failure and risk.” The solution involved more than crunching numbers to make decisions. The problem was solved by careful analysis by the professionals maintaining the system.

a gaming PC

This speaks to the problem of getting optimal results for a computing system. Building a data center is more than putting together a warehouse of GPUs as if they were a collection of high-end gaming machines. Different components in the center will have different failure profiles. Perhaps one model of memory card has more faults than demanded. A GPU might fail in a particular manner that is characteristic of a build of the computer board. Perhaps certain manufacturer facilities and date ranges were deficient, leading to problems. A 10-fold variation in reliability can create a few outliers on a critical path for the maintenance schedule. Insurance claims for normal events will be difficult to negotiate.

As Backblaze has found, a data center requires careful analysis of each feature of the hardware and the software installed on it. It’s more than just counting the frequency of unscheduled downtime. The hardware is enough of an investment that preserving the value is essential. The software configurations can lead to complicated failure patterns to decipher. Detailed diagnostic records should help identify when the problem is just a $2 connector vs. when it is a $10k GPU when both lead to the same failure patterns.

A glass of iced tea with lemon

A lemon-flavored data center will have trouble meeting performance goals. It will take careful accumulation of data and human analysis to intervene. If the data is collected diligently (or compulsively?) an AI system might notice patterns in the failures. Skilled data center managers will need to develop tools to protect their investments.

With an advanced understanding of the varied configurations, some vexing failures could still have a solution. Having a robot able to quickly reboot a server will not be not enough. Diligent managers should be able to find a way to evolve their system into smooth iced tea rather than unpalatable warm lemonade that came with a pass code and a power bill.

No Boss, It’s not Superintelligence

If I had the ear of a hypothetical executive in a technical services company, here are some of the points I would make…

One of the things that would be prime in my discussion is that hiring an AI engine to replace employees is bad for morale. It will disrupt the relationships and the rapport that the team has built. The group I am in has a great synergy and we can fill in the holes of each other’s presentations. We are more together than we are separately and replacing one or two of us with a mandatory use of AI would make part of our enthusiasm leave with them. Going out for drinks after work would not be the same after many of us are gone.

I also have noticed that employees who use AI and then force another worker to correct the lies and “hallucinations” that the coworker produces, the worker doing those remedial edits will resent the other employee for producing poor quality product. It will cause a disconnect between our actual skills and a what an AI ends up foisting on us. I aspire to being the best engineer I can and when my work is evaluated rigidly and I become a clerical worker rather than a creative team member, I no longer see that aspiration as possible.

Yellow sign with a red lightning bolt warning of danger ahead

Another thing will happen when we need to hire another worker. Any real person we consider won’t be chosen for their experience in our business but rather whether they are an LLM whisperer who can get (usually) adequate work from the AI engine. That goes to another concern, the AI is not going to be aware of the proprietary skills that we’ve developed. It might give me something that is good from an outsider’s point of view but that is different that what we need.

Right now, using AI tools to replace employees is a fad that is hard to resist. Will we evaluate those tools as if they were a regular candidate? What will be our criteria for choosing one vendor vs. another? It may be able to write much faster than we can, but our engineers do more than write reports and put together slide shows. Some of its explanations may be plausible, but correcting plausible English text into a quality product can be harder than getting it right the first time. It might be able take an expert and help make them a virtuoso, but it won’t take an average worker and make them a superstar.

Although AI proponents talk about the impending creation of a superintelligence or “artificial general intelligence,” their boosterism obscures that it doesn’t exist and that deception gives us unrealistic expectations. Would we want a coworker who lies and that isn’t as intelligent as our combined team? After all, its intelligence is due to a convincing parlor trick; it can sound intelligent, use the right jargon and look good in the number of words it can create per day, but it won’t pass other measures of value. It can also have a reverse Pygmalion effect on the remnant: supervisors who treat their employees as less skillful than the AI will make them less capable workers. The former synergy would be transformed into a defect rather than being what made us better than our competition.

Another risk of the AI is that we should not be surprised if it reveals our trade secrets and how we solve problems. It does not have loyalty and can make as many harmful decisions as beneficial ones. The tendency will be that the easily produced content of the AI would not get a proper vetting. Editing the fully formed text becomes impractical and the defects would get passed on. The apparent productivity boost of the products obscures the actual work needed to stay in the front. It will harm the company’s reputation to release lies and chaos.

Although talk of superintelligence may make it seem that we’ve got to start using it before the competition does, such a fad will not serve the company well. The media may be enamored with the big talk by AI proponents, but they are just salesman like any other business and what they promise is coming soon is not what they offer.

Grokipedia v0.2 observations

I was poking around on Grokipedia v0.2 on Nov. 25, 2025 to contrast it with Wikipedia. I was not evaluating its political orientation. The media reports that its political attitude is a primary point of interest but not for me.

Here are some observations.

  • GP tends to have long sentences that could be refactored into improved prose. There are a lot of comma splices. The style can be stilted and unnatural. The articles seem to be consistently too long. They include redundant content and repeat the same information in different sections of an article. Each topic has two or more subtopics. This contrasts with Wikipedia pages, which often have a flat outline and make judicious use of sub-topics and, rarely, sub-sub-topics, Grokipedia articles are always subdivided into two levels. The text of each Wikipedia article has been individually organized in a logical fashion. Grokipedia’s organizing principle is usually hidden. In addition, the section headings are long which makes them hard to scan quickly as you review a page.
  • GP only includes pure URLs as its citations. Wikipedia includes the date that the reference was accessed along with other information such as title, source and publisher when its available. Wikipedia can include citations that are not available online or are paywalled (for example in books or bibliographic databases) when the information is not available freely. Wikipedia heavily uses archive.org’s Wayback archive to preserve access to references that are gone or altered. A pure URL doesn’t help me grade the quality of sources or decide whether it is relevant to my purpose for visiting the site. When I did go to Grokipedia citations, it was be difficult to find where the referenced information was located. It also is a valid question to ask whether the citation actually supports the statement being cited or is it being misinterpreted.
  • With only a direct URL for the citations, Grokipedia will be brittle as the internet churns and pages come and go. Theoretically, a Grokipedia page could regenerate with new material, but that fragility inhibits consistent content over time. When source material is deleted, Grokipedia can’t stay current.
  • GP has some formatting issues. One ugly example was the mathematical page Brauer Group that had red errors in the mathematical equations. When you mouse over a citation index ([15] for example), the domain name popup is centered and hidden behind the mouse. In addition, because the target is small, you can’t move the mouse to uncover the domain. My mouse is big, but Wikipedia aligns the popup when you mouse over an internal link to the right of the mouse and there is enough room to move the mouse out of the way. When you mouse over an actual citation in Wikipedia, the entire citation is presented directly above the index number. In the body of an article, GP’s headings and subheadings are too similar in size and font weight so that they blend together. Although there is a line between top level outline elements, it is such low contrast that it’s not actually visible. This defeats the goal of adding a visual cue that a new section has begun.
  • GP has sparse internal links. Where Wikipedia has many cross links to other Wikipedia articles, GP has dramatically fewer. For example, I didn’t see any crosslinks in the article about Tasha Yar from Star Trek: The Next Generation. In most articles there are many obvious places to put an internal link that aren’t used.
  • Although it contains an edit history for each page, that feature is not very useful. You can’t put edits in context. If the edit is complicated, you can’t interpret it. Of the several pages I visited, there was only one edit and it was still being reviewed. Allowing users to thumb up/down to an edit as Grokipedia does might be ok for Reddit, but an encyclopedia requires more detailed evaluation. It’s not clear what will happen when a page with edits regenerates.
  • The pages are white on black. Since many of the pages are very long, that makes them hard to follow along. As I scroll on the page, I lose track of where I am. The page map on the left has all of the text in the same font size and weight which makes it hard to identify the top-level outline entries. Since the lengths of the outline entries are long, they are usually folded onto two lines, making the page map even harder to scan.

Although, I’m sure Grokipedia v0.3 will have improvements, right now I wouldn’t use it as a replacement for Wikipedia. Grokipedia is written by an anonymous, evolving blob and there is no one to intervene. I can visualize the Wikipedia community because I know there are people behind the scenes. In addition, I can find Wikipedia projects that I can contribute to there and be a contributor as well as a user.

Wikipedia is organic and uses principles compatible with Douglas Engelbart’s Language, Artifacts and Methodology to augment human intellect. Why reinvent the wheel?

Fair use and ChatGPT

Magnifying Glass

A quick note about the copyright lawsuit centered on fair use and ChatGPT’s use of copyrighted texts. I think the lawsuits are looking at the wrong moment in time to demolish fair use arguments.

My understanding is that there are 4 factors for determining fair use. The Stanford libraries explain them as:

  • The goals that the material is being used for
  • The copyrighted item’s properties as a work
  • How much of it was copied and how substantial that part was
  • How will the use affect the value of the original source.

Unfortunately, these criteria are fuzzy and need a legal proceeding to come to a definite conclusion.

Most discussions of fair use in the OpenAI copyright lawsuits focus on the final product: the large language model developed from the documents OpenAI accessed. Looking at the ChatGPT product is the wrong moment in the process to identify infringement.

Before the large language model had been created, the text was downloaded to OpenAI’s servers. At this moment, the claim of fair use is most tenuous. The information has been copied in full and potentially can destroy the market for the copied text. It is a situation to deny fair use claims and to apply conventional copyright protection statues.

At this moment of text ingestion, the purpose of the use is to transform the information into an unpredictable level of fidelity. The work is textual, and its value is the information in contains. The majority, if not all, of the information has been taken and it could harm the potential market by making the information available without accessing the original service.

To make an argument for fair use more convincing, fair use evaluation should be made before the information has been taken. One can always say, after the fact, “this is fair use,” but unless the analysis has been made up front, OpenAI can throw out all the ideas they can think of and see if any stand up in court. It is similar to the situation where the police find a criminal using techniques they can’t bring up in court, for example by using a Stingray. The police can backtrack and wash the information they already know, hiding the improper methods from the defendant. OpenAI can backtrack and find any idea they can think of no matter how disingenuous.

Another issue of fair use is that the user of the information should be able to make a listing of the materials that have been accessed. If I’m going to take a painting and claim my use is fair use, I need to be able to present where the work came from. If OpenAI can’t identify what they have taken, they can deny a practice of violating copyrights by hiding the “low hanging fruit” of implausible fair use claims.

If they could show these secondary properties of a fair use access, it would bolster OpenAI’s claim of fair use. Did they make their analysis before accessing the data? Can they completely identify what information they used for fair use. Did they consider the different fair use situations for the different kinds of sources?

They took the text verbatim when they acquired the text. They intended to use it in a way that they can’t argue convincingly that it would not harm the existing market for the text. Those are strong violations of the fair use definitions, it’s just a different point of time when the problematic behavior happened.

News organizations such as the New York Times can argue that the infringement happened when the text was acquired. Whether or not the original text can be retrieved from the created system is irrelevant. The infringement happened long before the OpenAI was brought to defend itself.