Written by Lincoln

momentarily useful

Linus Torvalds’ oft-repeated statement that AI is “90% marketing and 10% reality” points out the obvious hype machine burning through the software industry today. Hyper-inflated claims, buoyed by placebo & addictive effects, have made communication impossibly challenging between programmers and executives. It was a relationship often fraught to begin with, and figuring out how to re-bridge that gap has been at the back of my mind these last two years.

I’ve come up with a simple way to encapsulate the benefits of LLMs we refer to as AI: momentarily useful.

To explain, let me first present my TWO CENTS (it’s an acronym, see) on good use cases for LLMs. Then, I’ll contrast that with what they are disastrously bad at doing, and why “momentarily useful” explains the gap.

Good uses:

  • Test Cases — What else should I add to my testing plan?
  • What-Ifs — Thought experiments that are interesting, but not worth the effort to play out.
  • Omissions — What didn’t I think about in this scenario?
  • Comparisons — What are the differences, similarities, or tradeoffs between options?
  • Examples — Please just give me an sample implementation so it clicks for me.
  • Names — What was this specific thing called again?
  • Terms — What words do folks use to talk about this topic?
  • Suggestions — I’m feeling a bit stuck and just need something to react to.

The common theme here is that all these use cases are trivially verifiable in a single step and center your own critical thought process. The AI is momentarily useful, and then you move on. They all lean into it being a language model and what it does well: Report on word relationships through a prose-based UI.

Now, disastrous uses:

  • Generating — Any prose, code, or images it creates should be discarded at the end because it’s mid-tier trash. If you don’t understand why it’s mid-tier trash, you should consult someone who does; publishing it is telling on yourself. 1 2 3 4
  • Automating — We already have many great tools for this. A non-deterministic one is particularly heinous at it because their inevitable mistakes compound breathtakingly fast. Even when that’s acceptable, traditional ML is simply cheaper.5
  • Reasoning — Systems thinking, critical thinking, and understanding simply do not exist in them. They cannot collaborate, investigate, nor innovate. Worse, they purposefully produce artifacts that give a dangerous illusion of all that.6 7 8 9

You may have noticed that salesmen created marketing terms for exactly those weak points: “genAI”, “agentic”, and “reasoning model”. What a crazy, random happenstance! It’s almost like they’re using a classic playbook to obfuscate their products’ damning shortcomings to make it seem unbeatable and inevitable.10

If these systems lack accountability, deterministic outcomes, reasoning, understanding, expertise, or creativity, then what hope do they have of operating independently? Pretty much zero. Which is why I call them momentarily useful. A human can sit down and use it for narrow use cases, then set it aside.

If it’s a tool like any other, that makes the decision to use it a local one: Is this tool a good choice for me here today? Only you have enough context to know that. Top-down mandates of tool use are universally expensive propositions that require considerable upside to justify, but we’re taking it to a new extreme with AI by explicitly ignoring the costs and claiming immeasurable upside.

As a general rule, I believe any company that invests heavily in insufficiently tested technology will cost itself more than it gains.11 12 For a small to medium-sized company, those costs13 14 can quickly become lethal.

I’ve seen this pattern repeat itself again and again in our industry15, and the cause is mundane: We haven’t collectively learned where to put the guardrails yet AND we haven’t quantified the upside yet, so you’re volunteering to find both on your own. That’s incredibly expensive, because the inevitable “off-roading” experiences will constantly disrupt your team’s momentum and distract from its mission. It centers the technology over the mission.

The folks selling you AI would like nothing more.16

Further reading:

  1. “Maybe We Need Some More Examples:” Individual and Team Drivers of Developer GenAI Tool Use (Jul 2025) “Our findings imply that widespread organizational expectations for rapid productivity gains without sufficient investment in learning support creates a “Productivity Pressure Paradox,” undermining the very productivity benefits that motivate adoption.” ↩︎
  2. I Went All-In on AI. The MIT Study Is Right (Oct 2025) “Twenty-five years of software engineering experience, and I’d managed to degrade my skills to the point where I felt helpless looking at code I’d directed an AI to write. I’d become a passenger in my own product development.” ↩︎
  3. Faulty Towers, vibe sickness, and the vibe bobsled (Jul 2026) “The thing is that all of this tooling is useful for some things, but the term “genAI” points at exactly what it’s worst at: generating things.” ↩︎
  4. Pen tests show AI security flaws far more severe than legacy software bugs (May 2026) “New attack surfaces, larger blast radii, and unclear remediation ownership compound the risks.” ↩︎
  5. Out-Of-Distribution Checks – How Human Intelligence Stabilises Agentic Workflows (Sep 2026) “The bottom line is that reliable fully-autonomous agentic workflows are extremely improbable, and all the credible data taps that sign.” ↩︎
  6. Epistemological Fault Lines Between Human and Artificial Intelligence (Dec 2025) “Large language models (LLMs) are widely described as artificial intelligence, yet their epistemic profile diverges sharply from human cognition. Here we show that the apparent alignment between human and machine outputs conceals a deeper structural mismatch in how judgments are produced.” ↩︎
  7. AI Mania Is Eviscerating Global Decision-Making (Jul 2026) “The world’s organizations have been captured by people in the throes of frothing excitement, and saner people who now live in a state of constant commingled fear and frustration.” ↩︎
  8. BEST-OF-N JAILBREAKING (Dec 2024) “We introduce Best-of-N (BoN) Jailbreaking, a simple black-box algorithm that jailbreaks frontier AI systems across modalities.” ↩︎
  9. Cognitive outsourcing (3/3): Analogue Asymmetry (Sep 2026) “Maintaining cognitive diversity is not a values position. It is a functional requirement for the kind of computation that collective intelligence depends on.” ↩︎
  10. Anthropic May Know A Lot About Dice, But Do They Understand The Game? (Aug 2026) “Anthropic have demonstrated quite unambiguously – check out the Claude status page and the Claude Code issues page for the receipts – that they don’t really understand software engineering.” ↩︎
  11. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (Jul 2025) “Surprisingly, we find that when developers use AI tools, they take 19% longer than without—AI makes them slower.” ↩︎
  12. The Anthropic Glasswing Receipts Are Starting to Trickle In (Sep 2026) “So five months into Project Glasswing, only 9.8% of the findings have made it to a project maintainer […] The data shows that Claude’s assessment of the vulnerability severity appears to be overestimated.” ↩︎
  13. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (Jun 2025) “EEG analysis presented robust evidence that LLM, Search Engine and Brain-only groups had significantly different neural connectivity patterns, reflecting divergent cognitive strategies. Brain connectivity systematically scaled down with the amount of external support: the Brain‑only group exhibited the strongest, widest‑ranging networks, Search Engine group showed intermediate engagement, and LLM assistance elicited the weakest overall coupling.” ↩︎
  14. Staying or Leaving? How Job Satisfaction, Embeddedness and Antecedents Predict Turnover Intentions of Software Professionals (Nov 2025) “Job satisfaction and embeddedness were significantly negatively associated with software professionals’ turnover intentions, while work-life balance showed no direct effect. The strongest antecedents for job satisfaction were work-life balance and job quality, while organizational justice was the strongest predictor of job embeddedness.” ↩︎
  15. What Goes Around Comes Around… And Around… (Jun 2024) “Two decades ago, one of us co-authored a paper commenting on the previous 40 years of data modeling research and development. That paper demonstrated that the relational model (RM) and SQL are the prevailing choice for database management systems (DBMSs), despite efforts to replace either them. Instead, SQL absorbed the best ideas from these alternative approaches. We revisit this issue and argue that this same evolution has continued since 2005.” ↩︎
  16. The Second Derivative: Why No One Understands the AI Boom (Jul 2026) “The market misremembers 2008. That same blind spot sits at the center of the AI boom.” ↩︎