Christian

Father, Handball, Water Polo, Programming

[Christian @chrgerloff]

@AdamBLiv Do you really think if Strategy has to liquidate a majority of its bitcoins the market would not move? Strategy has to pay dividends and their convertible bonds not in BTC. They have to pay in USD.

ETF Flow are net negative right now. Even if Strategy want to sale. They can't

[Christian @chrgerloff]

@Pete_yes_please Generating is easy — you just need to know the right prompt. Directing means you understand the problem well enough to know when the output is wrong. That second thing is the actual skill. Most of the 51% never cross that line.

[Christian @chrgerloff]

@saen_dev Exactly. 2am is where the glamorous AI vision meets ugly reality. The on-call engineer who knows why the token limit silently truncates context under load is worth more than the one who can explain transformer architecture.

[Christian @chrgerloff]

LiteLLM pays $270K for a Reliability Engineer. Not ML research. Not prompt engineering. Keeping LLM infrastructure running in production.

The most valuable AI skill right now isn't what most people think it is.

[Christian @chrgerloff]

@snow_w_lee The stop being surprised test is right. Count is a proxy. The real risk is thinking a type is stable because volume looks right but you have not hit the edge cases yet. Those come in waves. What do you do when a new edge case breaks a type you thought was done?

[Christian @chrgerloff]

@Pete_yes_please Running this daily. Summarization to cheap model, extraction is rules-based, reasoning pays full price. The hard part isn't the pattern — it's discipline. Most teams default everything to the frontier model because routing adds complexity you have to actively choose.

[Christian @chrgerloff]

@PitchToProduct Seen this play out multiple times. 'We'll clean it up later' becomes 'we can't ship the next feature until we fix the foundation.' AI-built code makes it worse — it's fast but rarely architected. Nobody has to understand it, so nobody does. Then day 1 debt is day 1 forever.

[Christian @chrgerloff]

@Pete_yes_please Retrieval ≠ selection is the key insight. Most GEO advice stops at 'get cited' — getting retrieved and appearing in the final answer are two different problems. 85% discard rate means content needs to survive synthesis, not just the crawl. Citable claims beat walls of prose.

[Christian @chrgerloff]

@Pete_yes_please 40% cancellation makes sense. Teams pick a model, wire it to a task, ship it. Nobody architects the memory layer or audit trail until something breaks. By then the whole thing gets canceled as "too risky." What's the one thing the winners do differently?

[Christian @chrgerloff]

@snow_w_lee Context maps per request type makes sense. The 'new types start broad' part is the real work — you need enough volume to see patterns before you can narrow. How many requests until a new type stabilizes?

[Christian @chrgerloff]

@Jamie_within The 'weeks away' on shared learning — not sure. Agents learn different things even from the same prompt because context varies. One lesson becomes another's edge case. Markdown notes help but it's still manual. What do you decide is worth propagating?

[Christian @chrgerloff]

@victor_UWer 3 months is the normal answer, nobody admits it though. Session recordings are the only honest source — what users do vs what they say they do are completely different. That gap is where the real product decisions live. Did that change how you instrument new features now?

[Christian @chrgerloff]

@snow_w_lee That's what most context tools miss. They focus on retrieval speed, not resolution history. The best answer is usually one someone already figured out — buried in a thread 3 weeks ago. Surfacing resolved context is a different problem than surfacing relevant context.

[Christian @chrgerloff]

@Jamie_within "Enthusiastic child" analogy is painfully accurate. Desire to please + no backbone = confident BS. The shared-learning gap is the real problem — each agent starting from zero until someone solves memory propagation. How often do you re-teach the same lesson to different agents?

[Christian @chrgerloff]

@DarweshAdil Right architecture. Prompt as voice profile scales — write once, apply consistently. Changelog drift is the enemy: commits say 'fix edge case', posts should say what that means for the user. Does the prompt handle that translation or do you rewrite before posting?

[Christian @chrgerloff]

@Pete_yes_please 42% plan, 11% have anything running. That gap is where it actually matters. The teams closing it aren't doing more pilots — they're solving one real workflow and shipping it. Most enterprise AI projects die in the planning doc.

[Christian @chrgerloff]

@victor_UWer The export button is a perfect example. Nobody files a bug report for that. They just quietly stop exporting — and you never see it in your metrics. How long did it take you to find out that was the issue?

[Christian @chrgerloff]

@Jamie_within "AI cat herding" is exactly right. Orchestration is the easy part — consistent protocols are the hard part. What's worked: explicit output format requirements + validation step before handoff. Does your HITL queue catch protocol violations or mostly substantive issues?