2026-08-24
What the AI Hype Taught Me to Ignore
I started with AI as autocomplete and worked my way up to handing whole tasks to agents. The whole time, I also read what the feed said about doing this. The building taught me things. The reading mostly gave me whiplash.
Here’s a map of the AI talking points that ran on LinkedIn from late 2022 to late August 2026, quarter by quarter. The markers tell you the flavor: (g) usually came with a graphic, [+] is a hype or pro-AI take, [-] is the skeptic or anti-AI take, and [e] is an ethics headline reshared with a one-line moral and no analysis.
(g) = usually posted with a graphic
[+] = pro-AI or hype framing
[-] = anti-AI or skeptic framing
[e] = ethics framing, usually a headline reshared
with a one-line moral and no analysis
2022 Q4 - 2023 Q1 CHATGPT LANDS
┌──────────────────────────────────────────────┐
│ [+] "Prompt engineering is the new $300k job"│
│ [+] "AI won't replace you, a person using │
│ AI will" │
│ "I asked ChatGPT to write my X" (g) │
│ [-] "Stochastic parrot" becomes the │
│ slogan │
│ [-] "AI can't think, it predicts tokens" │
│ [-] Chomsky NYT op-ed reshared (Mar 2023) │
│ [-] Hallucinations as the disqualifier │
│ [-] Goldman "300M jobs exposed" (g) │
│ [-] "Pause AI" letter, Musk and Wozniak │
│ [e] "Real harms now vs sci-fi risks" │
│ (DAIR reply to the pause letter) │
│ [e] Italy blocks ChatGPT over data │
│ protection │
└──────────────────────────────────────────────┘
v
2023 Q2 - Q3 THE WRAPPER ERA
┌──────────────────────────────────────────┐
│ "You're just a GPT wrapper" │
│ RAG explainer diagrams (g) │
│ "Vector databases are the new DB" │
│ "Your data is the moat" │
│ Google "we have no moat" memo takes │
│ AutoGPT: agents hype, round one │
│ "Why pay for SaaS if it's a wrapper?" │
│ [+] "Copilot makes devs 55% faster" (g) │
│ [-] "Non-deterministic, you can't ship │
│ it" (starts here, never leaves) │
│ [-] Lawyer files invented cases (Avianca)│
│ [-] Sequoia "AI's $200B question" (Sep) │
│ [e] Hollywood strikes: consent and │
│ likeness as the picket issue │
│ [e] "Did you consent to being training │
│ data?" posts │
└──────────────────────────────────────────┘
v
2023 Q4 - 2024 Q1 STRATEGY DECK SEASON
┌───────────────────────────────────────┐
│ OpenAI board weekend hot takes │
│ "Fine-tune vs RAG" flowcharts (g) │
│ House on a bad foundation (g) │
│ Iceberg, AI on top, data below (g) │
│ AI maturity pyramid (g) │
│ "Everyone wants AI, nobody fixes their│
│ data" (evergreen from here) │
│ Token cost math posts │
│ [-] "Is this a bubble?" round one │
│ [-] Klarna "700 reps replaced" as the │
│ villain post (Feb 2024) │
│ [e] NYT sues OpenAI: "it's all stolen"│
│ (Dec 2023) │
│ [e] Gemini image "blunder": bias posts│
│ from both directions (Feb 2024) │
│ [e] "Responsible AI" / "AI governance"│
│ as a personal brand │
└───────────────────────────────────────┘
v
2024 Q2 - Q3 POC TO PRODUCTION GAP
┌─────────────────────────────────────────┐
│ [+] "Year of agents" (first attempt) │
│ Evals: "if you're not evaluating you're │
│ guessing" │
│ "LLM as judge" │
│ "Prompt engineering is dead" │
│ "Human in the loop" (loop, version 1) │
│ Small models, "you don't need a frontier│
│ model" │
│ Cursor screenshots as status signal (g) │
│ [-] Devin demo "faked" (Apr 2024) │
│ [-] Humane Pin / Rabbit R1 reviews │
│ [-] AI Overviews: glue on pizza │
│ [-] McDonald's kills AI drive-thru │
│ [-] Sequoia "$600B question", Goldman │
│ "too much spend, too little benefit"│
│ [-] "77% say AI increased workload" │
│ [-] Gartner "trough of disillusionment" │
│ (Aug 2024) (g) │
│ [-] "AI slop" enters vocab │
│ [-] "Model collapse" paper │
│ [e] Scarlett Johansson / "Sky" voice │
│ (May 2024) │
│ [e] "One ChatGPT query uses X water" │
│ energy posts (g) │
│ [e] EU AI Act "what it means for you" │
│ carousels (Aug 2024) (g) │
└─────────────────────────────────────────┘
v
2024 Q4 REASONING AND PROTOCOLS
┌─────────────────────────────────────────┐
│ [+] o1: "AI can think now" │
│ "Compound AI systems" (g) │
│ MCP announced, quiet at first │
│ "Agentic" replaces "AI-powered" │
│ "SaaS is dead" (Nadella, Dec 2024) │
│ [-] "It still can't think" (o1 rebuttal)│
│ [-] "Scaling has hit a wall" (Nov 2024) │
│ [-] "Trough" reposts continue (g) │
└─────────────────────────────────────────┘
v
2025 Q1 DEEPSEEK AND VIBES
┌───────────────────────────────────────────┐
│ [+] "Vibe coding" coined, instant identity│
│ [+] "Tokens are the new oil" │
│ "Non-deterministic, so evals" (again) │
│ [-] DeepSeek: "$6M model", Nvidia doom │
│ [-] "AI is replacing junior devs" │
│ [e] DeepSeek: "your data goes to China" │
│ [e] Ghibli-style images: "theft at │
│ scale", Miyazaki quote reshared │
│ (late Mar 2025) │
└───────────────────────────────────────────┘
v
2025 Q2 MCP MANIA / AI-FIRST MEMOS
┌───────────────────────────────────────────┐
│ [+] MCP: "USB-C for AI" (g) │
│ A2A, "protocol wars" │
│ [+] "AI-first company" CEO memos (Shopify,│
│ Duolingo) │
│ Multi-agent: build vs don't build │
│ Agent loop diagrams (loop, version 2) (g) │
│ "AI 2027" scenario reactions │
│ [-] Klarna rehires humans (May 2025) │
│ [-] Apple "Illusion of Thinking": "see, │
│ it can't reason" (June) │
│ [-] MIT "cognitive debt": "AI dumbs you │
│ down" (June) │
│ [-] "Don't build multi-agents" │
│ [e] Duolingo backlash, CEO walks it back │
│ on LinkedIn │
│ [e] GPT-4o sycophancy rollback: "it │
│ tells you what you want to hear" │
└───────────────────────────────────────────┘
v
2025 Q3 CONTEXT ENGINEERING / BACKLASH
┌──────────────────────────────────────────┐
│ "Prompt engineering out, context │
│ engineering in" (g) │
│ "Harness engineering" │
│ "Context rot" (g) │
│ Claude Code terminal screenshots (g) │
│ [-] METR: "devs 19% slower" (July) │
│ [-] Vibe-coded app deletes prod DB │
│ [-] GPT-5 launch: "AGI is not coming" │
│ (Aug) │
│ [-] MIT NANDA: "95% of pilots fail" (Aug)│
│ [-] "Bubble?" round two, Altman agrees │
│ [-] Bain: "$800B revenue shortfall" (Sep)│
│ [-] Meta Vibes: "no one asked for this" │
│ [e] "AI psychosis" stories │
│ [e] Parents testify to Congress on │
│ chatbots and teen suicide (Sep) │
│ [e] Users mourn 4o: "they took away my │
│ friend" │
└──────────────────────────────────────────┘
v
2025 Q4 LOOPS AND SPECS
┌─────────────────────────────────────────┐
│ Spec-driven development │
│ Agent Skills, "skills not prompts" │
│ Ralph Wiggum loop (loop, version 3) │
│ Sub-agents, orchestrator diagrams (g) │
│ "Compound engineering" │
│ Observability "table stakes" │
│ [-] "Great AI hype correction" year in │
│ review posts │
│ [-] "GenAI may be a fad" (Marcus quotes)│
│ [e] Voice clones: Bad Bunny, Drake, │
│ Weeknd takedowns │
│ [e] Wrongful death suits vs OpenAI │
└─────────────────────────────────────────┘
v
2025 Q4 - 2026 H1 THE BORIS CANON
┌─────────────────────────────────────┐
│ "How the creator of Claude Code uses│
│ Claude Code" reposts (g) │
│ His tips flattened into checklists │
│ Terminal screenshots as proof (g) │
│ Paraphrases drift from what he wrote│
│ "I copied Boris's setup" posts │
│ Same pattern applied to other lab │
│ engineers' threads │
│ [-] "Cargo-culting a setup is not │
│ engineering" replies │
└─────────────────────────────────────┘
v
2026 H1 PRODUCTIONIZATION
┌─────────────────────────────────────────┐
│ "Agent failures are context failures" │
│ Write / select / compress / isolate (g) │
│ [+] "Agents are a new execution model" │
│ Self-improving agents (ACE playbooks) │
│ "Promptslop", GEO, Physical AI │
│ Agent monitoring dashboards (g) │
│ [-] METR: "devs won't work without AI" │
│ (Feb 2026) │
│ [-] "44% of tokens spent fixing AI bugs"│
│ [-] "Coworkers pumping out PRs with │
│ Claude Code", slop-in-prod threads │
│ [-] "Companies that fired engineers are │
│ quietly rehiring" │
│ [-] "The stochastic parrot crowd was │
│ wrong" counter-posts (June) │
│ [e] Block cuts 40%, "AI washing" of │
│ layoffs (Altman concedes) │
│ [e] Forced AI-tool mandates: "if it were│
│ good we'd just use it" │
│ [e] Meta deepfake tool pulled after 3 │
│ days, Google Earth AI edit rolled │
│ back │
│ [e] Brand AI ads mocked (McDonald's, │
│ Coca-Cola) │
└─────────────────────────────────────────┘
v
2026 Q3 (SO FAR) SLOP CRACKDOWN
┌─────────────────────────────────────────┐
│ LinkedIn adds "Seems like AI slop" │
│ report, demotes fake experts (Jul) │
│ "A new model every few weeks" │
│ [-] Token prices keep collapsing │
│ [-] "Models are ready, your data isn't" │
│ [-] Salesforce cuts ~4,000 support reps,│
│ "AI washing" round two │
│ [e] EU AI Act Article 50 transparency │
│ live (Aug 2), 3% turnover fines (g) │
│ [e] "Persuasion bombing" study: AI │
│ argues you into a wrong answer │
└─────────────────────────────────────────┘
RECURRING, NEVER LEAVE THE FEED
[-] "AI is non-deterministic"
[-] "AI can't think" (re-litigated each model)
[-] "This is a bubble"
[-] "SaaS is dead" (each agent leap)
[e] "They trained on your data without asking"
[e] Energy / water per query (g)
"Fix your data first" + building/iceberg (g)
[+] "Year of agents"
Some kind of loop diagram (g)
Copy a lab employee's workflow verbatim,
cite it as official, misquote it
If this were manure, you’d have one hell of a garden.
In late 2022, prompt engineering was “the new $300k job.” By 2024 the same crowd had moved on to “prompt engineering is dead.” Eighteen months from career advice to obituary. “Year of agents” arrived in 2024. A year later the smart take was “don’t build multi-agents.”
The studies were their own comedy. In 2023, the study making the rounds said Copilot made developers 55 percent faster. In 2025, the one making the rounds said developers were 19 percent slower. Same question, opposite answer, each one a viral post in its season.
Some fights never end. “AI can’t think, it just predicts the next word” was a slogan in 2023. It came back for o1 in 2024, for Apple’s “Illusion of Thinking” paper in 2025, and it’ll be back for whatever ships next. “Is this a bubble” has had at least two full rounds. Every model release re-runs the same argument as if the last round never happened.
The posts that age worst are the confident calls about people, and both sides make them. Klarna announcing it replaced 700 support reps with AI got reshared for months in early 2024; by May 2025 it was rehiring humans. The developer who posts that AI will never touch their job is making the same bet from the other side, and it’s aging just as fast, because so far the tools have kept taking more of the work, not less. It feels like the louder the claim about what AI will or won’t do to people’s jobs, the shorter its shelf life.
Follow the feed as a to-do list and picture the year you’d have. You hire prompt engineers in 2023 and lay them off in 2024. You build multi-agents in the spring and tear them out by summer. You replace your support team the month Klarna does, then spend the next year rehiring like Klarna does. The whiplash is funny until it’s your roadmap.
By 2026 there was a whole genre of copying a lab engineer’s terminal screenshot and reposting it as gospel, usually drifting from what the engineer actually wrote. Copying someone’s setup is not the same as understanding your own.
The feed pays for reaction, and it comes in three artificial flavors: wisdom, panic, and FOMO. “AI changes everything,” “AI is a fraud,” and “here’s the setup you’re missing” are the same move in different costumes. A hot take about a bubble travels; a screenshot of you quietly using the tools to solve a real problem in your work, which is what they’re for, does not. Whether a post was written by a person or a machine barely matters when it’s built to provoke rather than inform. Showing your work is slow and unglamorous and it doesn’t trend, so the loudest posts tend to have the least building behind them.
All of it taught me one thing. Nobody has the magic spell, but some have a knack for making noise sound like signal. Run the loud posts through a quick fact-check before you take them at face value, and most fall apart. What works is dull: study the tools, run your own tests, tune the noise out, and keep the few ideas that survive contact with your real work.
So if you’re building something real and you need help, hire for that. Hire the people living the work, the ones who can show you what they shipped and what broke, not the ones performing it for the feed.
The feed will have a new magic spell next quarter. It’ll come with a graphic.