Openai

OpenAI is the company that lit the match. ChatGPT shipped on November 30, 2022, and the world hasn't looked back. Three years later they're the most valuable AI company on the planet, the most controversial, and arguably the most overstretched. Worth knowing what's hype and what's real.

The actual track record

Founded 2015 by Sam Altman, Elon Musk, Ilya Sutskever and others. Nonprofit to start, then a strange hybrid "capped-profit" subsidiary so they could take Microsoft's $13 billion. They shipped GPT-1, GPT-2, GPT-3, GPT-3.5, GPT-4, GPT-4o, GPT-5 — each one bigger and broader than the last. DALL-E for images. Sora for video. Whisper for audio transcription. The API has eaten roughly 70 percent of the developer-tools AI market.

If you want to actually use the technology, the consumer entry is ChatGPT Plus at $20/month — worth it if you write or code for a living, not worth it if you're a casual user. For developers, the API key is free to create; you only pay for what you use. A good primer book like Andrew Ng's AI for Everyone is the cheapest 4 hours you can spend before forming opinions.

What it does well

Generative text at a level that wasn't possible three years ago. Translation. Summarisation. Code generation good enough that anyone writing Python by hand for routine work is wasting their time. Image generation that took years off the timeline for graphic design.

The honest user experience: ChatGPT is the best general-purpose assistant most people will ever use. For specific technical work, Anthropic's Claude is often better (and we use it heavily for content). For image work, Midjourney and now Stable Diffusion XL beat DALL-E on style. For coding, GitHub Copilot (which runs on OpenAI under the hood) is the workflow that has actually changed software development.

If you want to build with this stuff, Chip Huyen's Designing Machine Learning Systems is the systems book to read. Hands-on Large Language Models for the practical end.

What it doesn't do well

Truth. Models hallucinate confidently and constantly. The fix isn't "newer model" — the fix is verifying every fact it produces if it matters. Anyone who copy-pastes ChatGPT output into a brief, a memo, or a code review without checking is going to get burned. This hasn't gotten meaningfully better since GPT-4.

Reasoning that requires multi-step inference. The new "o" reasoning models help, but they're slower and more expensive, and they still trip on problems a smart 14-year-old would solve. Don't trust them with anything where being wrong has cost.

Long-context reliability. The models can technically read a 100k-token document but they don't always weight the information correctly. For serious research, you still need to chunk and verify.

The Sam Altman saga and what it tells you

Altman was fired in November 2023 by his own board, then rehired five days later after the staff threatened mass resignation. The episode revealed two things: the company is built around one person more than is healthy, and the original safety-oriented governance structure is functionally broken. Most of the original safety leadership has since left or been pushed out — Ilya Sutskever, Jan Leike, Dario and Daniela Amodei (who founded Anthropic), Mira Murati.

The pattern matters because OpenAI's safety story used to be the company's whole pitch. Now it's a commercial AI company in safety drag. That's not necessarily bad — but it's worth being clear-eyed about.

For the politics, Mustafa Suleyman's The Coming Wave and Cade Metz's Genius Makers are the two most-cited reads.

Where it goes

The next 18 months are about agents — AI systems that take actions in the world, not just generate text. OpenAI is racing Anthropic and Google there, and nobody has won yet. The first company to ship a reliable agent that books your travel, writes your code reviews, and manages your inbox without supervision will own the next era. OpenAI is the favourite, but they're not the lock they were two years ago.

If you're using the tools, keep using them. Verify outputs. Don't trust marketing. And if you build with the API, write tests — the model behavior changes between versions and what worked yesterday may not work tomorrow.

The hype is real and the limitations are real. Both are true. Anyone selling you only one half is selling you something.