OpenAI Taught a Model to Think Longer—but Made Choosing an AI Even Harder

The boundary of the “ToTheMoon is a podcast and a hello from Silicon Valley!” case is defined by this point: o1-preview exposes intermediate reasoning and handles tasks where one fast answer is not enough, but the quality gain brings a new chaos: users must choose among several models, and the black box gets deeper.

In the context of “Google Wallet — an interesting novelty from Google,” this criterion applies: Google Wallet now lets you add a passport to present the document electronically — first in the U.S., tied to the TSA system where your ID is already scanned at security: the document finally moves into the phone.

In the context of “We are not Apple fanboys!,” this criterion applies: the hosts take the commenters' charge of Apple bias with a caveat: their phones really are all Apple — for style, security, or habit — but they follow Android too: Google's Gemini-linked Wallet ID feature was the first thing they covered.

The “New model from OpenAI! How does o1-preview work” issue should be assessed with one constraint in mind: the novelty changes the mode of operation itself — before answering, the system spends time reasoning and shows it "thought" for several seconds; for hard math and code that is a step forward, but users now must know when to pick fast 4o, when o1, and when mini.

The “The technical side of o1-preview. Why this model is unlike all the others” issue should be assessed with one constraint in mind: with o1 the market moves toward verticalization: one model suits everyday conversation, another code, a third expensive analysis; the interface grows more complex, and picking the wrong mode means overpaying or getting a weak result.

The working conclusion from “How o1-preview was trained” is that reinforcement learning always needs an auxiliary model checking the first model's results — and that built-in verification reduces hallucinations; the hosts expect a tool within a year that picks between 4o, o1, and mini automatically, because no one will do it by hand.

The boundary of the “B2C or B2B? Where AI solutions will go” case is defined by this point: Meta and other open-model builders face their own dilemma — developers fine-tune the mass-market Llama for themselves, but a reasoning model demands entirely different economics and infrastructure: open and closed markets diverge no longer by license but by product type.

The boundary of the “Will ChatGPT cost $2,000 a month? For whom?” case is defined by this point: the talk of plans costing hundreds or even thousands of dollars is a bet on professional scenarios: the price is justified only where the model's long reasoning replaces expensive human work rather than merely speeding up chat.

The discussion of “ChatGPT competitors: why the discussion still centers on OpenAI” yields a practical test: Anthropic is strong as a developer platform, and Claude 3.5 Sonnet already beats ChatGPT on some real tasks, but OpenAI has the chance to build an Apple-grade brand: expensive, mass-market, and clear to people who have no wish to study architecture.

The boundary of the “Black box in the black box” case is defined by this point: we see a polished description of the chain of thought but do not know how closely it mirrors the real internal process: one black box has learned to explain another — and trusting it has become harder, not easier.