170: Консьюмеризация — AI, VPN и вообще всего

Summary

In episode 170 of IT Thoughts, Sergey Petrenko reviews GPT-5 and argues that its reception makes more sense when viewed as part of the broader consumerization of technology: tools once understood mainly by specialists are becoming everyday products for a mass audience.

GPT-5: Features and Hands-On Assessment

GPT-5 initially replaced the previous assortment of models with a more unified experience. A router can decide whether a request needs more reasoning, and users can ask it to think longer. After users objected to losing familiar models, OpenAI restored access to older options. Petrenko sees this as evidence that users can become attached to particular model versions.

The interface offers automatic, Thinking, and Pro-style options. In the API, Petrenko says reasoning effort can be set to low, medium, or high, though there is no direct equivalent to selecting Thinking or Pro.

Petrenko found GPT-5’s standard interface fast, while deeper reasoning—especially in Pro—could take minutes. He reports a 19-minute response for a task involving source comparison. He also found GPT-5 Nano surprisingly slow on simple tasks, such as generating commit messages, and temporarily returned to GPT-4.1 for those jobs.

He considers GPT-5 a meaningful improvement over GPT-4, including because multimodal capabilities are built in. He cites an API context window of up to 400,000 tokens, while noting that a large context window matters little to most users if they cannot see or use it directly.

Overall, Petrenko does not consider GPT-5 a major disappointment. He found it more stable than some previous models and useful for his work, while describing its progress over recent reasoning models as more incremental than revolutionary.

Coding and API Testing

Petrenko describes using GPT-5 as a coding assistant and reviewer. He supplied implementation requirements, repository contents, and documentation, then asked the model to check whether the code matched the intended design. He found it useful for reviewing and rewriting a small project.

One test also exposed a failure: GPT-5 rewrote code without producing the expected result. On inspection, Petrenko found that it had introduced a maximum output limit of 900 tokens, which he had not requested. He also notes that GPT-5 supports OpenAI’s Responses API rather than the older Completions API.

His takeaway is that the model can be capable and still make mistakes or take unrequested initiative.

Consumerization: Why Expectations Differ

Petrenko’s central point is that AI has rapidly become a household technology, like television or internet access. Most people use it through a simple chat interface without needing to understand model architecture, training cutoffs, context windows, or how to adjust settings.

That changes what product success means. For specialists, benchmarks, context size, and control over model behavior may matter; for everyday users, the defaults and the immediate answer matter more. As an analogy, Petrenko points to smartphones and television: once technology becomes commonplace, people often stop thinking of it as technology at all.

He also argues that AI assistants are shifting information-seeking away from keyword searches toward conversational questions, which may affect services such as Wikipedia and search engines.

Regulation, Age Checks, and VPNs

The episode connects mass-market technology to government restrictions. Petrenko discusses the UK Online Safety Act’s age-verification requirements, including concerns that verification could compromise the anonymity and safety of Wikipedia contributors. He notes reports of reduced traffic to some age-restricted sites alongside increased interest in VPNs.

His broader argument is that circumvention is not effortless for ordinary users. Technical users may know how to install and configure a VPN, but most people will not make that extra effort. He uses internet restrictions in Russia and China as examples of how people may shift to local alternatives—or simply use restricted services less.

For Petrenko, this is another consequence of consumerization: large-scale controls can affect everyday behavior even when technically informed users know ways around them.

Reviews, Guides, and Tutorials

• Hands-on review: GPT-5 compared with GPT-4 and other reasoning models, based on the host’s use in the chat interface, API, and coding work.
• Practical testing: A coding-review workflow using requirements, repository files, and documentation.
• No formal tutorial or step-by-step guide is presented.

Main Speaker and Sources

The main speaker is Sergey Petrenko, host of IT Thoughts. He refers to OpenAI and Sam Altman’s comments, Similarweb traffic figures, a reported analysis of publicly indexed GPT chats, and Wikipedia’s legal challenge to UK online-safety requirements.