Replace em dashes in latest post for a less AI-generated feel

content/articles/2026-06-15-closed-ai-vendor-lockin.md

@@ -7,13 +7,13 @@ coverImage:
   url: "https://images.unsplash.com/photo-1564313677573-481bd55237ad?q=80&w=1287&h=600&auto=format&fit=crop"
 ---
 
-Three days. That's how long it took for a flagship closed-source AI model to go from public release to global blackout. A leading frontier lab launched what it called its most capable model ever. By the end of the same week, a regulatory directive had pulled the plug for *every* customer on the planet — paying enterprises, internal staff, everyone. Sessions errored out. New queries were silently rerouted to older, less capable fallback models. The model wasn't deprecated. It was simply unreachable, by order of someone who doesn't know your roadmap.
+Three days. That's how long it took for a flagship closed-source AI model to go from public release to global blackout. A leading frontier lab launched what it called its most capable model ever. By the end of the same week, a regulatory directive had pulled the plug for *every* customer on the planet, paying enterprises, internal staff, everyone. Sessions errored out. New queries were silently rerouted to older, less capable fallback models. The model wasn't deprecated. It was simply unreachable, by order of someone who doesn't know your roadmap.
 
 If your stomach dropped reading that, you already understand the thesis.
 
 ## The New Shape of Vendor Lock-In
 
-For two decades, vendor lock-in was a *cost* problem. Switching databases, cloud providers, or CRM platforms was painful and expensive — but technically possible on a sane timeline. "Vendor lock-in" meant migration friction.
+For two decades, vendor lock-in was a *cost* problem. Switching databases, cloud providers, or CRM platforms was painful and expensive, but technically possible on a sane timeline. "Vendor lock-in" meant migration friction.
 
 Closed-source AI introduced a different shape: **availability lock-in**. The same week, a major cloud provider restricted its own employees from using a peer lab's newest model over data-retention concerns. Earlier this year, that lab was publicly designated a "supply chain risk" by a defense agency and cut off from an entire government vertical overnight. These aren't pricing changes. They're access changes, and your team has zero say in them.
 
@@ -25,7 +25,7 @@ As I argued in [Your Codebase is an Asset](/articles/2026-02-24-codebase-is-asse
 
 The economics are shifting under anyone who built production workflows on closed APIs. The newest frontier-tier model launched at roughly **double the per-token price** of the previous flagship, and burned through subscription quota at about twice the rate. Independent testers drained a top-tier subscription in under nine minutes. The interactive tier's flat rate was always an implicit subsidy; on the new model, the subsidy no longer covers a working session.
 
-And then today — on the day this post goes live — that same vendor cut the subsidy further. Agent workflows, headless scripts, CI integrations, and third-party tools authenticated through subscription credentials no longer draw from the same usage pool. They consume a small per-user credit — $20 to $200 a month depending on plan — billed at full API rates. Once that credit runs out, automated requests simply stop. No automatic overflow, no rollover, no team pooling. For teams running production automation on subscription credentials, this is the day a CI pipeline started failing silently at 2 a.m.
+And then today, on the day this post goes live, that same vendor cut the subsidy further. Agent workflows, headless scripts, CI integrations, and third-party tools authenticated through subscription credentials no longer draw from the same usage pool. They consume a small per-user credit ($20 to $200 a month depending on plan), billed at full API rates. Once that credit runs out, automated requests simply stop. No automatic overflow, no rollover, no team pooling. For teams running production automation on subscription credentials, this is the day a CI pipeline started failing silently at 2 a.m.
 
 It's the third such billing change in five months. The "cheap subscription" era was a customer-acquisition tactic, not a pricing model. The math caught up, the subsidy is being clawed back, and the workflows you built on top became the bill.
 
@@ -35,7 +35,7 @@ The line items above are bad enough. The things that don't show up on the invoic
 
 **Data retention as a feature gate.** The same top-tier model ships with mandatory 30-day data retention, which effectively excludes any European team operating under GDPR data-minimization, any healthcare or legal workflow with zero-retention contracts, and any organization that has promised its customers that prompts are not stored. A leading open-source CEO put it bluntly: this policy "delegated a lot of European companies to the permanent underclass." The same week, a major hyperscaler restricted internal use of the model over exactly these data concerns. If they don't trust it for their own staff, that's a signal.
 
-**The alignment tax.** The model card itself documents that for certain categories of prompt — anything related to building frontier AI systems — the model silently downgrades its own responses through prompt modification and steering vectors, *without telling the user*. You don't have a hallucination problem. You have a transparency problem, baked into the platform.
+**The alignment tax.** The model card itself documents that the model silently downgrades its own responses through prompt modification and steering vectors, for certain categories of prompt (anything related to building frontier AI systems), *without telling the user*. You don't have a hallucination problem. You have a transparency problem, baked into the platform.
 
 These are not bugs. They are policy. And policy can change, unilaterally, with an email.
 
@@ -43,7 +43,7 @@ These are not bugs. They are policy. And policy can change, unilaterally, with a
 
 - **Routing layer first, model second.** Your production code should never `import` a vendor. Abstract the model behind a thin interface so swapping providers is a config change, not a sprint.
 - **Keep an open-weights fallback.** Frontier-class open models (M3, and the local-inference ecosystem that matured this year) are competitive enough to serve as same-day failover. They are immune to export controls, retention mandates, and pricing whiplash because they run on your hardware.
-- **Tier by risk, not by hype.** Reserve the closed frontier model for tasks where its marginal capability genuinely matters — a shorter list than vendor marketing suggests. A 12% quality lift on a CSV parser is not worth a single point of vendor dependency.
+- **Tier by risk, not by hype.** Reserve the closed frontier model for tasks where its marginal capability genuinely matters, a shorter list than vendor marketing suggests. A 12% quality lift on a CSV parser is not worth a single point of vendor dependency.
 - **Re-read your contracts.** Yesterday's "flat rate" is today's metered credit. Plan for the next reprice now, not the morning it ships.
 
 The teams that survive the next wave of AI consolidation won't be the ones with the best prompt libraries for a single provider. They'll be the ones who treated the model as a *substitution* rather than a *platform*.
@@ -54,4 +54,4 @@ Closed-source AI isn't going away, and frontier models will keep delivering real
 
 > **Don't let your tools own your code. Don't let one vendor own your runtime either.**
 
-What does your team's AI exit plan look like? And if the honest answer is "we don't have one" — what's the smallest step you could take this week to change that?
+What does your team's AI exit plan look like? And if the honest answer is "we don't have one", what's the smallest step you could take this week to change that?