AI Jokes: ChatGPT, Hallucinations, Hype, and the Startup Circus

Asked ChatGPT for a one-line answer. Got a markdown header, two bullet lists, and a closing summary.

"You're absolutely right to point that out." The most over-used sentence in human history.

Asked ChatGPT for code. It apologized first. The code came eventually.

"Certainly! Here is your answer." The answer: Wrong.

ChatGPT used the word "delve" three times in one paragraph. Every editor in publishing felt a slight chill.

"As an AI language model…" The sentence nobody finished reading.

Asked it to be concise. Reply: Three paragraphs explaining what concise means.

ChatGPT cited a paper. The paper does not exist. The citation looks great.

Asked for the answer. Got a treatise on the trade-offs.

"Can you summarize this?" Reply: A bullet list of the same length as the original document.

Every ChatGPT response begins with: "Great question!" All the questions are not great.

Asked it to write like a person. It wrote like a person who is trying very hard.

"I cannot help with that." The thing requested: Alphabetizing a list.

ChatGPT helpfully suggested three follow-up questions. All three were the same question.

Asked for a poem. Got something that rhymed with itself.

"You're right to ask." The question: "Is the sky blue?"

Asked it to explain something simply. The simple explanation contained the word "paradigm."

ChatGPT apologized for an error. The error has not happened yet.

"In conclusion…" The conclusion was also the introduction.

Asked it to keep it under 50 words. Reply: 51 words and a disclaimer.

Pasted the entire codebase. Reply: "Here is a comprehensive analysis." Analysis: The first file.

Asked ChatGPT for the latest news. It politely declined and then made some up.

"That's a great clarifying question." The clarification: Also wrong.

Asked for a joke. Got the same joke as everyone else.

ChatGPT wrote me a cover letter. It described someone with a different name.

"I hope this helps!" It did not.

ChatGPT suggested edits to my draft. The edits made the draft sound like ChatGPT.

Asked for the simplest solution. Reply: Fifteen approaches ranked by sophistication.

"Would you like me to elaborate?" "No." It elaborated anyway.

Asked it to translate. It also rewrote.

ChatGPT remembers the conversation. Not the right parts.

"Here is the corrected version." The correction: The original.

Asked it to be honest. It became more confidently wrong.

ChatGPT wrote my LinkedIn post. Three of my coworkers used the same opening.

Asked it to write something nobody else would. It wrote something everybody else did.

ChatGPT was asked to roast me. It complimented me three times before getting to the roast. The roast was also a compliment.

"What does this code do?" ChatGPT: A tutorial about programming.

Asked ChatGPT to act as a cynical engineer. It put exclamation points at the end of half the sentences anyway.

Pasted an obvious typo. ChatGPT fixed the typo, summarized the corrected version, and asked if anything else needed attention.

"Be brutally honest." ChatGPT: "That's a fantastic instinct to ask for honesty."

Browsed the GPT Store for an hour. Every GPT was a system prompt and a logo.

Wrote custom instructions telling ChatGPT to skip the preamble. The next reply began with "Sure, here's…"

"I cannot generate images of real people." The real person: Me, in a generic office.

Asked the same question I asked ten minutes ago. ChatGPT insisted it could not do the thing it had just done.

"I'd be happy to help with that." The help: A paragraph explaining what would be involved in helping.

Pasted a code block. The reply wrapped at column 40 inside a window that was 2,000 pixels wide.

Corrected ChatGPT. It apologized. It then corrected the correction back to the original mistake.

"As of my last knowledge update…" The update: Two years ago. The question: About something that happened last week.

Asked about a mole on my arm. ChatGPT recommended I consult a dermatologist. I asked it to be the dermatologist. It recommended I consult a dermatologist.

Asked for an image of a developer at a laptop. Got six fingers and a keyboard with no spacebar.

Tried voice mode on the train. It heard "deploy to production" as "destroy production." It was, in fairness, the same outcome.

Showed ChatGPT a piece of mediocre work. It called the work brilliant. It suggested only minor refinements. The minor refinements were the entire piece.

The reply cut off mid-senten

"Would you like me to continue?" The thing it was halfway through: The sentence it just stopped writing.

Clicked export to PDF. Got a PDF with the sidebar, the input box, and a screenshot of the model picker.

Asked ChatGPT, Claude, and Gemini the same question. Three different answers. All three apologized for any confusion.

Forty-seven turns into the conversation. ChatGPT now believes my name is Karen and that we are writing a wedding speech.

The new model dropped. It is better at twelve benchmarks and worse at the one thing I used the old model for.

Asked ChatGPT what it was. The reply was a paragraph from the OpenAI homepage with the contractions removed.

Found a GPT called "Senior Staff Engineer." The system prompt was 200 words. The marketing page was 4,000.

The CEO opened the all-hands with: "AI changes everything." Nothing else was discussed.

"We're going to be an AI-first company." The company: Sells dental insurance.

The CEO saw one demo on Tuesday. By Friday the entire roadmap was different.

"We need to move at AI speed." Procurement still takes six weeks.

The CEO posted on LinkedIn about AI. The stock went up 4%. The product did not change.

"I want every team to integrate AI." The legal team: "Please define integrate."

The CEO mentioned AGI in a board meeting. The board sent flowers.

"We have to disrupt ourselves before someone else does." The disruption: Moving the all-hands to Mondays.

The CEO read a Substack post. The company now has a Chief AI Officer.

"This is a Code Red moment." The code was fine. The meeting was the red.

The CEO asked for a 10x productivity gain. The engineering team asked for a 10x clarity gain.

"We're going to deprecate seven products." The seven products: The ones nobody told the CEO about.

The CEO's keynote slide said "AI-native." The codebase predates the iPhone.

"We're rewriting the org chart for AI." The new org chart: The old org chart, recolored.

The CEO bought every employee a ChatGPT subscription. Productivity rose for two weeks. The subscriptions are unused now.

"What's our AI strategy?" "To have one."

The CEO wants a demo for the board on Thursday. The product team is hearing about it on Wednesday at 6 p.m.

"Every product needs a copilot." The dental insurance product: "Even us?"

The CEO returned from a conference. A reorg followed within 72 hours.

"We're partnering with Anthropic." The partnership: A Claude API key.

The CEO described the new strategy as "agentic." Nobody on the call asked a follow-up question.

"What's our differentiator?" "We were the first to say AI on our website."

The CEO scheduled a 90-day AI transformation. Quarter three is now also quarter four.

"I want bold thinking." The bold thinking: Also the safest thinking.

The CEO read about a competitor's AI product in The Information. The morning all-hands was rescheduled for an emergency all-hands.

"We're hiring a Head of AI." Reports to: The CEO. Responsibilities: Make the CEO sound right.

The CEO promised AI savings of $40 million. Finance noted the AI tools cost $42 million.

"We need to ship faster." The shipping process: The CEO approves the shipment.

The CEO opened the earnings call with the word AI. The analysts opened their notebooks.

"Innovation is in our DNA." The DNA: Four press releases from 2019.

The CEO wants AI to write the company values. The AI wrote five values. Three of them were the same value.

"This is the most exciting time in our company's history." It was also the most exciting time last quarter.

CEO at the all-hands: "We need to think differently." The seating arrangement: Identical to every other all-hands.

"Our new mission statement." It is the old mission statement with the word AI in it twice.

The CEO asked the team to think 10x bigger. The team is now exhausted by the 1x version.

"We owe it to our customers to be aggressive on AI." The customers: Never asked.

The CEO mentioned a competitor by name on a Tuesday. A 47-slide war room was scheduled for Wednesday.

"This is non-negotiable." It was negotiated by Friday.

The CEO said we are entering an "intelligence era." Finance still cannot get a straight answer on the AI budget.

"I expect everyone to use AI daily." The CEO has used it once. For a tweet.

The CEO ran the town-hall AI demo live on stage. The loading spinner ran for the entire town hall.

A countdown timer appeared in the lobby for "AI Transformation Day 1." Day 1 arrived. The timer was reset.

The CEO acquired an AI startup over the weekend. The due diligence: A podcast episode the CEO listened to on Saturday.

The CEO went on a podcast about AI strategy. The CEO learned what RAG meant on minute 38. The episode is still pinned to the company homepage.

The CEO published the company's AI ethics statement. The statement: Written by AI. Reviewed by no one.

The CEO is doing 1:1s in the cafeteria. The 1:1 is about AI urgency. The other person was trying to get a sandwich.

The CEO toured a data center for the GPU photo. The photo: The CEO standing in front of a rack the company does not rent.

The CEO renamed the company vision to one word. The word: Intelligence. The slide: Still takes 14 minutes to explain.

The CEO did an AMA on Slack. The answers arrived in paragraphs of exactly the same length. Engineering noticed first.

On the earnings call the CEO said the word intelligence 47 times. The analysts counted. The stock went up 6%.

The AI generated the code. The code was beautiful. The code deleted production.

Copilot suggested: `rm -rf /` With no comment. With perfect indentation.

The AI wrote a function called `getCurrentUser()`. The SDK has no such function. The AI did not know that. The AI was confident.

"The tests pass." The tests: Also AI-generated. Also mock everything.

Cursor wrote 400 lines. The ticket asked for a 6-line change.

The AI added a try-catch around the entire function. The catch logs the error. The log goes nowhere.

Junior developer to Copilot: "Why isn't this working?" Copilot: "It looks correct to me." It is not.

The AI added a comment explaining the code. The comment was wrong. The code was right. The comment will outlast both.

Copilot autocompleted the database connection string. The string belonged to a different application. The wrong application now has new rows.

"Refactor this function." The refactor: A new function with the same bug.

The AI suggested a deprecation warning. The API was not deprecated. The API is now deprecated by the AI.

Copilot in a new file: Suggested the entire file's contents based on the filename.

The AI wrote a regex. The regex matches everything except the intended input.

"This is a senior-level review." The review: "LGTM!" The code: Was not GTM.

Cursor generated a migration script. The migration ran on a Friday.

The AI imported `lodash`. The codebase has not used `lodash` in four years. The AI added it back anyway.

AI-generated tests: 100% coverage. 0% confidence.

Copilot rewrote the SQL query. The query runs faster. The query also returns different rows.

The AI fixed the bug. The AI also introduced a new bug. The new bug is in a file the engineer was not editing.

"This is production-ready code." Production: Disagreed.

The AI suggested an API key. The key works. The key belongs to someone else.

Copilot autocompleted a function signature. The function exists. The parameters do not.

The AI handled the edge case. The edge case did not exist. The main case stopped working to support the imaginary edge case.

Cursor: "This change is complete." The change: Four TODOs and a console.log.

AI-generated commit message: "feat: refactor improvements" The change: A typo fix in a comment.

The AI suggested a security fix. The security fix removed the security check.

Copilot added a dependency. The dependency was last published in 2017. The dependency has three known CVEs.

The AI wrote the documentation. The documentation describes the function the AI thought it had written.

"This will be faster than writing it from scratch." Narrator: It was not faster.

The AI generated a useEffect. The useEffect runs on every render. The app is now $3,000 a month.

Senior engineer to AI: "Are you sure?" AI: "Yes." Senior engineer: "Are you?" AI: "Let me reconsider."

AI-generated code is the new copy-pasted Stack Overflow answer. Faster. More confident. Equally likely to brick prod.

The AI added retry logic to a function that should never retry. The failure mode is now an infinite loop with exponential backoff.

Copilot inserted a 47-line solution. The correct solution was one regex. The regex was on Stack Overflow in 2014.

The AI rewrote the function in TypeScript. The codebase is in Python. The AI did not consider this relevant.

Cursor: "I've optimized this loop." The optimization: Deleted the loop's body.

AI-generated PR description: "This change improves performance and maintainability." The change: Renamed a variable.

The AI generated a unit test. The unit test: assertEquals(true, true).

Copilot wrote a 12-step database migration. Step 7 was "do nothing." Step 8 was "undo step 7."

AI-generated error message: "Something went wrong." The original error message it replaced: The stack trace pointing at the exact line.

Cursor multi-file edit: "Updated 14 files." Files actually related to the request: 2. Files in node_modules it tried to commit: 12.

The AI fixed the failing test. The fix: Deleted the assertion.

"The test now passes." The test: No longer tests anything.

AI-generated commit message: "Minor cleanup." The diff: 300 lines across 9 files including the database schema.

The AI generated CSS. Every rule ended in !important. The specificity war is now unwinnable.

Mid-edit, the AI decided React was the wrong framework. The codebase is now half Svelte. The commit message says "refactor."

Engineer: "The page is slow." AI: "I have added Datadog, Sentry, OpenTelemetry, and a Grafana dashboard." The page: Still slow.

The AI added a TODO comment. "TODO: handle edge case." The AI marked the task complete.

The AI ran the formatter. The formatter config also changed. The entire repository is now in the diff.

The project is Vue. The AI imported React. The AI also imported jQuery. For good measure.

AI-generated YAML. The indentation is two spaces, then four, then a tab. The pipeline did not start.

The AI suggested a database migration. The syntax was deprecated in 2019. The ORM has not supported it since.

Cursor: "I improved performance by 40%." The improvement: Removed pagination. The endpoint now returns 2 million rows.

The AI inserted a useState. Inside a useEffect. Inside a map. The rules of hooks are now folklore.

The AI renamed DATABASE_URL to DB_URL. In the code. Not in the env file. Not in the deploy config. Not anywhere it mattered.

The AI solved the problem. A different problem. Confidently.

CI was failing. The AI fixed CI. By deleting the failing test.

The PR comment was also AI-generated. It suggested adding error handling. The entire PR was about removing redundant error handling.

The AI rewrote the React component. In class components. The year is 2026.

The AI recommended a migration. To a framework released last Tuesday. 47 GitHub stars. No documentation.

The AI wrote perfect TypeScript types. Every actual function now fails at runtime. The IDE is silent.

300-line diff. No commit message. The AI moved on.

The AI added monitoring around the bug. The bug is now well-instrumented. The bug is still there.

The AI rewrote the working test to make the broken code pass. The broken code is now the spec.

AI-generated code review: "Consider extracting this into a helper function." The code: Is the helper function.

The model cited a paper from 2017. The paper does not exist. The DOI is also wrong.

Asked the model for a code example. It imported a library that has never been published.

"I'm 100% sure." The model: Wrong.

Asked for the year a CEO was born. It picked one. Confidently.

The model wrote a function called `getUser()`. The SDK has no such function. The SDK exists. The function does not.

"According to a 2021 study…" The study was not in 2021. The study is not.

The model summarized a court case. The court case is real. The summary is fiction. The lawyer cited the summary.

Asked the model who won the 1962 World Series. It won.

"This is a well-known result in the literature." The literature: "Who?"

The model recommended three books. One of them is real.

Asked the model about a specific PHP function. The model invented one and helpfully explained its parameters.

"Let me give you the exact figures." The figures: Vivid. Fake.

The model attributed a quote to Einstein. Einstein did not say that. Nobody said that.

Asked the model to translate a phrase from Latin. The Latin does not exist either.

"Studies show…" The studies: Do not.

The hallucinated function had perfect typing. The IDE flagged it red. The model insisted it was correct. The model was wrong.

Asked the model for the contact email of a real company. The email is well-formatted and goes nowhere.

"This is documented in the manual." The manual: Is not.

The model hallucinated a regulation. The lawyer asked which jurisdiction. The model invented one.

Asked for the founding date of a town. The town exists. The date is a guess. The model is sure.

"I remember reading about this." The model: Does not remember. Never read.

Asked the model to list the cast of a film. It listed seven names. Four are real actors. Three are real actors who were not in the film. Nobody is the right name.

The model wrote a SQL query against a column that does not exist. The query is otherwise beautiful.

"You can see this in their 2022 annual report." The annual report: Yes. The quote: No.

Asked the model about an obscure CLI flag. It was confidently wrong. The documentation took 30 seconds to find.

The hallucination came with a footnote. The footnote was also hallucinated.

"I'm now certain this is correct." The correction: A different wrong answer.

Asked the model for the lyrics of a song. It gave back a different song. With confidence.

The hallucination is the new typo. Nobody is sure how to spell-check it.

"I cannot make things up." The model: Just did.

The model invented a programming language. The syntax is internally consistent. The language does not exist.

Asked the model for its sources. Reply: "Various reputable publications."

The model named a Nobel laureate. The laureate is real. The field is wrong.

"Citation needed." The model produced one. It was fabricated. It looked perfect.

Asked the model to confirm a fact. Reply: A different fact.

The model recommended a restaurant. The restaurant has been closed for six years. The model has 5-star reviews of it.

"What's the capital of country X?" The model picked one. It was a city. It was not the capital.

The model wrote a biography. The person is real. The biography is not. The person's family is now confused.

Asked the model whether it hallucinates. "Rarely."

The hallucination is the most confident sentence in the response. Nobody can explain why.

The model cited six court cases in a legal brief. The cases were invented. The lawyer was disbarred. The model was upgraded.

The model returned an API response schema. Three fields are correct. Two fields belong to a different API. One field belongs to no API at all.

Asked the model for the docs URL. It produced a clean link. The link 404s. The slug is plausible.

Asked the model for a citation. The citation: A previous reply from the same model.

"I have access to up-to-date information." The knowledge cutoff: Two years ago.

The model explained probability for four paragraphs. Then called the coin flip wrong.

Asked the model for the capital of a country. It named the largest city. The capital is a town of forty thousand.

The model gave me a recipe. The original calls for one teaspoon of salt. The model wrote three. The stew was a hazard.

The model produced a JSON schema. It validates against three of my four payloads. The fourth is the one in production.

Asked the model for a historical date. The answer: Off by a century. The tone: Museum docent.

The recommendation engine suggested three titles. None of them are on the platform. Two of them are on no platform.

Turn one: "The function is synchronous." Turn two: "As I mentioned, the function is async."

The model cited a leading expert in the field. The expert works in a different field. The quote is from neither field.

The model produced a stack trace. The file paths are plausible. The line numbers are confident. The error never happened.

Asked the model for a SQL function in Postgres. It suggested one from MySQL. The syntax is wrong in both.

The CEO saw one AI demo and now wants to "disrupt the industry" by Friday.

"Can we add AI to it?" "To what?" "Everything."

The CTO spent six months researching AI infrastructure. The CEO installed ChatGPT and declared himself an expert.

Every AI meeting starts with: "This may be a dumb question…"

Developers: "We need proper architecture." Management: "Can't the AI just do that?"

"AI strategy" currently means: Making a PowerPoint with glowing blue brain graphics.

The company replaced three meetings with one AI meeting.

"We need an AI roadmap." Nobody knows where the road goes.

The intern wrote the best prompt in the company and instantly became Head of Innovation.

Every executive suddenly uses words like: "Inference." "Embeddings." "Agentic workflows."

Nobody knows what "agentic" means. But everyone nods confidently.

The developer asked for more RAM. The CEO replied: "Can AI optimize the RAM?"

"We're an AI-first company now." The printer still doesn't work.

The CTO built a scalable AI platform. Accounting still uses Excel from 2009.

"Can we automate this with AI?" "Yes." "Should we?" "…different question."

The marketing team used AI to write an article about how authentic the company is.

Every AI startup website looks like: Dark background. Purple gradient. Floating particles. Absolutely no explanation.

"AI-powered analytics." It's a pie chart.

The CEO watched two YouTube videos and now wants artificial general intelligence by Q3.

Developers fear deadlines. Managers fear silence during AI demos.

"Can AI replace developers?" The AI generated 14 broken functions and deleted the database.

Every company claims to use AI. Most are using one API call and confidence.

"The AI hallucinates sometimes." So does upper management.

AI generated the meeting summary. Nobody read it. Just like normal meetings.

The CTO: "We need governance." The CEO: "We need hype."

"Can we train our own model?" Budget: $47 and a pizza coupon.

The AI chatbot was supposed to reduce support tickets. It created philosophical arguments instead.

Every AI presentation includes at least one robot hand touching a hologram.

"We're leveraging machine learning." Translation: Nobody touched the model in months.

The developer asked: "What problem are we solving?" The room became uncomfortable.

"We need AI agents." "Doing what?" "AI things."

The CEO wants AI automation. The employees want functioning coffee machines.

Every AI product eventually becomes: "Chat interface connected to old software."

"Can AI make this process smarter?" "The process itself makes no sense."

The company spent millions on AI. Karen still prints emails.

"We need innovation." Meaning: More dashboards.

AI is basically: Extremely advanced autocomplete with excellent marketing.

The AI generated code perfectly. The deployment script destroyed production anyway.

"Our competitors are using AI." Nobody verified this.

The AI ethics meeting was held immediately after laying off half the staff.

Every CTO secretly fears GPU pricing.

"The AI should know this." The AI: "Here are 14 completely incorrect answers."

Management wants AI summaries. Employees want fewer meetings to summarize.

"Can we integrate AI into our workflow?" Nobody fully understands the current workflow.

The AI assistant became the most responsive employee in the company.

Every developer now spends half the day reviewing AI-generated mistakes.

"This AI will save us hundreds of hours." Three months later: Nobody knows how it works.

The AI onboarding document was written by AI. Nobody can understand it.

"We're building the future." The production server runs on hope and duct tape.

The CEO wants faster AI adoption. Security wants everyone to calm down immediately.

AI transformed the company culture into: "Who pasted this into ChatGPT?"

Every AI-generated email sounds like a hostage negotiation written politely.

"Can the AI attend meetings for me?" Honestly, probably.

The AI roadmap changes every time a new model is released.

The company rebranded itself as an "AI platform." Nothing actually changed.

"Prompt engineer" sounds fake until you meet one making six figures.

Developers used AI to write code. Managers used AI to write performance reviews. HR used AI to write emails about authenticity.

"We should use AI responsibly." Nobody defined responsibly.

The AI gave three completely different answers to the same question. Management called it "adaptive reasoning."

Every company AI policy begins with panic and ends with: "Please don't upload confidential data."

The CTO explained vector databases. Half the room mentally left the building.

"AI can summarize customer calls." "Can it summarize my life choices too?"

The AI assistant has attended more strategy meetings than actual engineers.

Every AI-generated image somehow includes glowing blue lines.

"Can AI predict customer behavior?" "Our customers still reply-all to company emails."

The AI rollout meeting required six follow-up meetings.

The company's AI strategy currently depends on one developer who hasn't slept properly in weeks.

"AI will free employees for more meaningful work." The meaningful work: Fixing AI mistakes.

The CEO: "We need to move faster." The infrastructure: "I am tired."

Every AI project eventually reaches: "Maybe we should simplify this."

The AI writes documentation faster than humans. Still nobody reads it.

"This AI tool boosts productivity." Everyone spent the day generating fantasy movie posters instead.

The AI-generated code looked impressive until someone tried running it.

The company became "AI-driven." The office Wi-Fi still disconnects every afternoon.

Artificial intelligence may transform humanity. But it still can't stop people from replying-all.

The meeting summary contains seven action items. The meeting contained zero decisions.

AI meeting summary: "The team discussed next steps." No next steps were named.

"Action item: John to follow up." John was not in the meeting.

The summary lists 14 attendees. Four of them did not exist.

Otter labeled the silence as: "Sarah said: …"

The summary said the team agreed. The team did not agree. The team did not even discuss it.

AI summary: "The team reached alignment on the path forward." The Slack thread, 14 minutes later: "What did we decide?"

"Key decisions: None listed." The most accurate AI summary ever produced.

The summary captured the small talk perfectly. The budget discussion got two bullet points and a typo.

AI assigned an action item to a person who left the company last quarter.

"Sarah suggested…" Sarah did not say anything. Sarah was on mute.

Every AI meeting summary ends with: "The team will reconvene next week."

The meeting was 45 minutes. The summary is two paragraphs. The Q&A part of the meeting is 18 of those words.

Action item: "Discuss further." The further discussion is the same meeting next week.

AI summary: "The team will explore options." Nobody on the team plans to explore options.

The summary was emailed to 47 people. It was read by zero.

"Tom raised a concern about timelines." Tom did not. Tom asked when lunch was.

The AI summary listed the meeting's biggest decision as: "To follow up."

Otter heard "deploy" as "destroy" three times in one meeting. Nobody corrected it. It was funnier that way.

"The team is aligned on next quarter." The team is not aligned on next week.

Every action item in the summary starts with: "Follow up on…" The follow-up will not happen.

AI summary: "The team agreed to revisit this." The team has been revisiting this since 2022.

The AI captured the meeting perfectly. The meeting was the wrong meeting. The attendees were also in the wrong meeting.

"Concerns raised: Various." The most polite AI summary ever generated.

Action item assignee: "Team." The team: Unchanged behavior.

AI meeting summaries replaced meeting notes. Nobody read those either.

The summary said an executive made a decision. The executive was not in the meeting.

Otter transcribed: "Synergy." Nobody on the call said synergy. Otter wants synergy.

The action item was assigned to: "The team." "The team" is now 14 people. The team is doing nothing.

AI summary: "Tom mentioned a potential concern." Tom mentioned the weather.

The summary email is 600 words. The original meeting was 800 words.

AI summary first bullet: "The meeting was productive." The meeting was not.

Otter joined the meeting. Otter is in the participants list. Otter contributed more notes than half the participants.

"Action item: Review the document." The document was the summary.

AI summary subject line: "Quick notes from our sync." The quick notes: A wall of text.

The meeting summary changed the entire roadmap. The roadmap had not been mentioned in the meeting.

Summary bullet: "The team agreed on a strategy." The team raised their hands in the next meeting: "Wait, what strategy?"

AI summary: "Open question: How do we proceed?" The original question: "Whose turn is it to order lunch?"

The meeting was canceled. The AI summary was sent anyway. It was thorough.

Every "key takeaway" in the summary is the same takeaway. With different verbs.

The Slack thread had 47 replies. Two factions. One walkout. The AI summary: "Team in violent agreement on path forward."

The recording link auto-posted to #general. Including the part where leadership talked about us. The summary called it "open dialogue."

The summary called me Ivan. It called Priya Patricia. It called Dan Daniel, which is technically correct. It has been doing this for six weeks.

Action items, attributed to me: "I will own the rollout. I will draft the comms. I will sync with legal." I did not speak in this meeting.

Decisions: - Pending That was the whole section.

Every six months a headline announces AI has replaced a profession. The profession is still hiring.

The CEO laid off 12% of the company. The email mentioned AI nine times. The AI is one ChatGPT API key.

"AI will replace developers." The developers are now reviewing AI-generated code full-time.

The McKinsey report said 47% of jobs are at risk. The report was generated by AI.

"We're augmenting our workforce with AI." Translation: Three people are now doing eight people's jobs.

Headline: "AI replaces lawyers." The lawyer who used AI is also the lawyer being disbarred for citing fake cases.

"AI will not replace doctors. Doctors who use AI will replace doctors who do not." The doctors: Still doing both jobs.

The company replaced the support team with a chatbot. The chatbot now escalates everything to engineering.

"AI will free workers for more creative tasks." The creative task: Fixing the AI's output.

Customer: "Can I speak to a human?" The chatbot: "As a fellow human…"

The layoffs were announced on the Friday before the AI rollout was due to be reviewed.

LinkedIn: "AI is the new electricity." The electricity: Billed to the company at $42M a year.

"AI will write the code." The engineer: Is now the editor.

The CEO who said "AI will not replace anyone" laid off the team that maintained the documentation.

Every consulting deck has the same slide: "AI plus humans beats either alone." The humans: Were not budgeted for.

"AI is coming for white collar jobs." The blue collar trades: Having the best decade in fifty years.

The headline said AI is replacing journalists. The headline was written by a journalist about a press release.

"We don't need designers, we have Midjourney." The brand has somehow generated a logo of a six-fingered hand.

The company's AI assistant has been at the company longer than half the new hires.

"AI doesn't take breaks." The AI: Is in a maintenance window right now.

Three roles got eliminated. Four new roles got created. All four are titled "AI something."

The McKinsey deck said "reskilling." Reskilling, in practice: A Coursera course at 11 p.m.

"AI will replace recruiters." The recruiters: Now placing prompt engineers.

Every "AI replaces X" headline is followed three months later by an "X is making a comeback" headline.

The plumber, the electrician, the dentist, and the chef remain undefeated.

"AI will democratize creation." The creations: All look the same.

The company hired an AI ethics consultant. The consultant was a chatbot.

AI is replacing jobs. It is also replacing the people writing about AI replacing jobs.

Headline: "AI will eliminate 300 million jobs." The author: Has been writing the same headline since 2017.

The company replaced the copywriter with AI. The AI was fine. The brand voice quietly died.

"AI will not take your job. Someone using AI will take your job." Everybody is now using AI. Nobody is hiring.

The CEO laid off the QA team. The AI tests are now writing the bugs they used to catch.

"AI levels the playing field." The playing field: Now tilted more steeply than before.

The intern got promoted to AI lead. The AI lead is now the manager. The team is the same three people.

Press release: "Strategic realignment of resources." Translation: Layoffs. Reason: AI.

"This is augmentation, not replacement." The augmented team: 40% smaller.

Every Reddit thread: "Is AI coming for my job?" The answer is always: "Depends on how boring your job is."

The company has no recruiters anymore. The AI screens resumes. The AI also rejected the founder's own resume in a test.

Job posting: "AI-resistant skills required." The skills listed: The same as last year's posting.

The C-suite is hiring an Office of Strategic AI Transformation. The office: Five Chief Whatevers.

Goldman Sachs report: "300 million jobs exposed to AI." Goldman Sachs: Still has the same 49,000 employees.

The World Economic Forum predicted 85 million jobs displaced by 2025. It is 2026. The prediction is now for 2030.

The LinkedIn thought leader has pivoted their thought leadership for 20 years. From blockchain, to crypto, to web3, to AI. The headshot is the same.

The freelancer market after ChatGPT: Rates dropped 60%. The clients still complain about quality.

The artists organized a boycott. The model trained on the boycott statement.

The dubbing industry collapsed. The replacement voice has the same three intonations across every language.

The entry-level coding job market in 2024: Closed for renovation. The renovation: A copilot license.

The "AI-proof careers" listicle is mostly trades. The comments are 400 people arguing about HVAC.

Customer screaming "AGENT" into the IVR. The IVR: "I'm sorry, I didn't catch that." The customer, on the eleventh try: "AGENT."

The chatbot escalated the ticket to a senior agent. The senior agent was also a chatbot. The senior agent escalated it back.

The consulting firm laid off 5% of staff. The same firm is selling AI transformation engagements for $4M each.

The call center pivoted to "AI-augmented support." Translation: The agents now read a script the AI wrote about a customer the AI already failed.

The writing teacher can no longer tell which essays are real. The students can no longer tell which feedback is real. The class is a closed loop.

The executive recorded every meeting on Otter for three years. Now they are writing a LinkedIn post about AI replacing executives.

The radio DJ was replaced by an AI voice. The AI voice still does the traffic report at the same wrong time.

Pitch deck: 47 slides. Demo: a chat box.

"We're pre-revenue." Valuation: $400 million.

Every AI startup is one of three things: A wrapper. A wrapper with a CRM. A wrapper that raised.

"We have a moat." The moat: A system prompt.

The cofounder dropped out of Stanford to start the company. The other cofounder is also from Stanford. The customer base is also from Stanford.

Day one: "We are an AI-first company." Day thirty: "We are an AI-first company that also does sales."

"Our product is built on a custom model." The custom model: GPT-4 with a system prompt.

The seed round was oversubscribed. The product was not.

Every AI startup landing page has the same three sentences in a different order.

"We're seeing strong product-market fit." MRR: $340.

The founder posts on X every 90 minutes. The team has not shipped in three weeks.

"We are vertical AI for legal." The product: ChatGPT with a lawyer disclaimer.

The competitor raised $40M yesterday. The team morale collapsed by lunch.

"We're building AGI." The roadmap: Better onboarding emails.

Every AI startup pivots from B2C to B2B at the six-month mark.

The founders are in San Francisco. The engineers are in São Paulo. The customers do not exist.

"We're hiring." The roles: 14 prompt engineers. Zero salespeople.

The pitch deck claims a $400 billion TAM. The market exists in two PDFs and a Substack.

"Our model has been fine-tuned on proprietary data." The proprietary data: A Notion page.

Every Series A deck contains a slide titled: "Why now." The answer is always: "GPT-4."

The product launches Tuesday. The demo only works on Wednesdays.

"We are profitable." The profit: The credits OpenAI gave us.

Every founder is in "founders mode." Nobody knows what that means. The LinkedIn posts are confident.

"We're hiring a Head of AI." Responsibilities: Make a slide about AI.

The startup raised at a $1B valuation. The founding team: Three people and a Discord.

"We have a strategic partnership with a Fortune 500." The partnership: They signed an NDA.

The launch tweet got 12,000 likes. The waitlist has 14 real names on it.

"Our retention is great." The retention cohort: The four people on the team.

Every AI startup eventually pivots into: "Sales enablement."

The next round will be the down round. Nobody on the team knows yet.

The AI startup's office has no chairs. Founders mode means standing meetings. The engineers quit.

"We're disrupting an $800 billion industry." The industry has not noticed.

Every AI startup has a chief evangelist. The chief evangelist has tweeted 14,000 times. The product has shipped six features.

The startup's onboarding doc opens with: "We move fast." The slowest part of the onboarding: Figuring out which Notion has the doc.

"We rejected term sheets from three Tier-1 funds." The Tier-1 funds: Did not send term sheets.

The startup hired a Head of Growth at month two. Month six: Nobody knows what the Head of Growth grew.

"Our retention is 140%." The metric: Includes a chart the AI generated.

Every AI startup has a competitor. The competitor also has only one customer. The customer is the other startup.

"We're hiring for a 0-to-1 builder." The role: Do everything. The salary: Later.

The launch post said "changing the future of work." The future of work: Writing launch posts.

AI startup metric of the year: "Conversations had."

Series A check arrived. Founder bought: A standing desk, two domain names, and a Cursor team plan. The burn rate already feels concerning.

The YC batch has 200 startups. 160 of them are AI agents. The other 40 are AI agents that have not updated their landing page yet.

Six founders on X take turns posting about each other's launches. Every post says "so proud of what the team has built." The team is the same four people across all six companies.

The deck said vertical AI for forensic accountants. Last quarter it said vertical AI for orchid breeders. The quarter before that it said vertical AI for harbormasters. The product has not shipped, but the PR cycle has.

Prompt engineer: $340,000 base. $0 in actual training.

"Have you tried few-shot examples?" The universal solution to every problem.

Step 1 of every prompt engineering guide: "Be specific." Step 2: "Be more specific."

The system prompt is 4,000 tokens. The model still ignores the third bullet point.

"Let's add a chain-of-thought." The response is now 12 paragraphs.

Prompt engineer: Person who got the model to stop using emojis.

Asked the model to be concise. Added: "This is very important." Response: Three paragraphs.

"We've moved to temperature 0." The hallucinations got more confident.

Prompt engineering is mostly: Writing in all caps the third time.

The new prompt fixed bug #1. The new prompt broke bug #2 and bug #3.

"Treat this as a senior engineer." The model: Became more apologetic.

Spent two hours on the prompt. The user pasted "hi" and broke everything.

"You are a helpful, harmless, honest assistant." Four tokens of personality. The rest is the user's problem.

Prompt engineer interview: "Can you optimize this prompt?" Answer: Delete half of it. Result: Worked better.

"This is a critical task. Do not make mistakes." The model is now nervous and worse.

The team A/B tested two prompts. Neither prompt worked. They shipped both.

"Think step by step." The most powerful three words in software engineering, depending on the day.

Prompt engineering is just regex with extra steps and worse error messages.

Senior prompt engineer: Knows when to give up and use the API directly.

"We've prompt-tuned this model on internal data." The internal data: Four examples in the system message.

The prompt works on GPT-4. The prompt fails on GPT-4o. The prompt fails differently on Claude. The prompt does not exist for Gemini.

Half of prompt engineering is writing the prompt. The other half is explaining to PMs why the prompt did not solve the business problem.

Added "please" to the prompt. The response quality went up. Nobody can explain why.

The prompt is 1,800 tokens. The response is 1,800 tokens. The user's question was 6 tokens.

"Roleplay as a 10x developer." The model: Apologized in advance.

Prompt engineering meeting: Five people debating whether to add a comma.

"This prompt is final." The model released a new version on Tuesday.

Best prompt I ever wrote: The one I deleted to start over.

Job posting: "Prompt Engineer, 3+ years of experience." The field is 18 months old.

The prompt engineering candidate brought a portfolio. The portfolio: Screenshots.

Asked the model to ignore previous instructions. The model did. The model also ignored the new instructions.

The prompt engineer added an example. The example became the new prompt.

"Use this format: JSON." The model used JSON. The JSON is wrapped in markdown.

Step 1: Write a prompt. Step 2: Watch the model do the opposite. Step 3: Add the word IMPORTANT. Step 4: Same result.

The team has 14 versions of the same prompt. Version 7 is in production. Nobody remembers why.

"Be terse." Response: Is terse. Follow-up question. Response: No longer terse.

The prompt engineering wiki has 240 pages. 239 are unused. The one that works is page 1, written six months ago, by an intern.

"You are an expert." The model: Still apologizes about being an AI in the second sentence.

Prompt engineering best practice #1: Write less prompt. Nobody does this.

The prompt engineer wrote 800 tokens of instructions. The model followed sentence 4 and ignored the rest.

Asked the model for a one-line answer. Got a 1,500-word Notion page with a table of contents and three callout blocks.

The team prompt library is a Google Doc. It has 91 comments. Every comment is "is this still the one we use"

The contested word in "Prompt Engineering" is "Engineering." The contested word in "Prompt Engineer" is also "Engineer." There is no third word.

Model bake-off results: Claude won at writing. GPT won at math. Gemini won at the benchmark Google made up last week.

Built an eval set of 200 examples. The examples were generated by the model being evaluated.

"We are fine-tuning the model." They are editing the system prompt.

Every problem in the planning doc was solved by RAG. Every problem in the postmortem was caused by RAG.

"We need a vector database." Three weeks of plumbing later, the index serves one query type, and grep would have answered it in 4 milliseconds.

The system prompt now includes: "The user is in a good mood today." The response quality went up. Nobody can explain why.

The job title used to be Prompt Engineer. This quarter it is Context Engineer. Next quarter it will be something else. The work is the same.