03 Editorial
We only write about what we've verified ourselves: independent model runs, case breakdowns, technology history. Every article has an author, a date and sources.
On November 30, 2022, OpenAI released a "research preview" of a chatbot. Five days later it had a million users; two months later, a hundred million — the fastest product growth in history. We unpack what happened technologically and culturally: why GPT-3.5 in a chat wrapper worked where more powerful systems hadn't, and how that moment rebooted the entire industry.
Code, long docs, math, data analysis — an independent run with no vendor marketing.
M. Sokolova · 9 minTools, prompts, metrics. Tested on teams of 3–50: up to 80% of tickets resolved without a human.
D. Belov · 7 minCapabilities, limits and hardware requirements of the most powerful open model.
K. Ivanov · 8 minHow vendors use your data, what zero-retention means and when you need self-hosted.
D. Belov · 6 minFor readers without a spare twenty minutes: condensed takeaways from the three latest pieces. Full versions are linked in the list above.
Maria Sokolova ran both models through an identical set: code, long documents, math, spreadsheets, texts in three languages. The verdict holds few surprises but plenty of nuance: Claude Fable 5 takes seven of twelve categories — above all long contracts, where GPT-5.6 starts dropping cross-references around page three hundred, and multi-step code refactoring. GPT-5.6 Sol wins four: agentic tool chains, image generation alongside text, latency on short requests, and its plugin ecosystem. One category — creative writing — ends in a draw; the difference is taste. The practical takeaway: holding both subscriptions costs more than it seems, and the choice should follow the two or three tasks that consume most of your time.
The most-quoted part is the timeline. Days 1–3: export the last thousand tickets and label them by topic — it turns out 60–70% of requests fit twenty scenarios. Days 4–7: a knowledge base for those scenarios, written in plain language with no internal jargon. Days 8–10: a Claude Sonnet agent with RAG running internally only — operators see draft replies and edit them. Days 11–14: accuracy measurement; the agent reaches customers only above 90% on the internal check. The key post-launch metric is not the automation share but the fraction of escalations the agent hands to a human correctly instead of inventing an answer.
Kirill Ivanov cools the excitement: the record context is an engineering trade-off, not magic. At full length the model runs slowly and expensively, and fact retrieval from the middle drops 8–12% versus the first and last pages. Still, three scenarios where 10M genuinely rewrites the rules: analyzing a company's entire year of correspondence, auditing a large code repository without chunking, and working through a corpus of case law. Hardware requirements are honest: from 40 GB VRAM for the quantized build, comfortable from 80. The article's conclusion: 200K commercial contexts cover 95% of tasks — but the remaining 5% used to be impossible outright.
We revise this guide quarterly based on reader mail: if a step raises questions or goes stale, write in and we fix it. The version below is dated August 2026 and has been tested on three dozen people with no technical background.
Seven days, twenty minutes each, no programming and no subscriptions. The goal is not to "learn neural networks" but to wire one or two into your actual work.
Sign up for Claude, ChatGPT or Gemini — any of them. Do not ask "what can you do"; hand over a genuine task from today: "rewrite this email shorter and firmer", "draft a moving plan for a 20-person office". The difference from search becomes obvious by the second task.
The working prompt formula: role + task + format + example. Compare "write a post" with "you are a B2B blog editor; write a 150-word post about a release delay, calm tone, no apologies, end with one question to the reader". The second version needs almost no editing.
Upload a thirty-page contract, report or manual. Ask for: a summary, five risks, every date and amount as a list. Then interrogate the text — it replaces an hour of diagonal reading.
Paste a table straight into the chat: "find anomalies", "group by region", "which three rows break the total". Models read CSV better than commonly assumed; the key is asking them to show the calculation logic.
Start a note and collect the phrasings that worked. In a month it becomes your main tool: a good prompt outlives any model switch.
Two prohibitions: never paste customers' personal data or trade secrets into public chatbots; never publish facts or figures from an answer without checking the primary source. These two rules cover 90% of a beginner's real risk.
The filter is simple and strict. First: a topic must pass the test "what will the reader do differently after this text?" — traffic-bait never enters the backlog. Second: the newsroom must reach a primary source — a technical report, a dataset, a participant; retellings of other people's retellings do not get published. Third: if the piece covers a tool, one of the four of us must run it on real tasks for at least a week — which is why reviews come out later than competitors' and disappoint less often.
Reader requests are the second-largest topic source after our own backlog: roughly every fourth article grew from an email to [email protected]. The most useful requests look like this: "we tried X for task Y, got Z, and cannot explain why." That kind of letter almost always becomes the article a hundred other teams need.