Skip to content
OpenAI · Apple · ChatGPTEpisode 007 · 26 May 2024 · 45:05

Why ChatGPT Is ‘Getting Dumber,’ Apple Is Moving Slowly on AI, and Our Data Is the Most Fragile Part of the System

What to watch for

1Compare “When a new version of ChatGPT responds faster, speaks, and understands images, it can feel as though progress moves in only one direction” with “Competition disguised as cooperation between Apple, Google, and Microsoft”: they provide different criteria for judging the same issue.
2Test the conclusion from “New concept of basic unconditional income and AI” in your own use case—what actually changes in the process and what remains a promise.
3Before choosing a product or approach, record the constraint identified in “10-year data loss”.
4Define the owner of the outcome and the quality metric for the situation described in “Artificial intelligence in iPhones: What will Apple do?”.
Signals to track afterwards
Watch for actions by Apple and Google that confirm or challenge the episode’s central claims.
Compare new launches and policy changes with “New concept of basic unconditional income and AI”: have access, quality, price, or constraints changed?
Check whether the scenario in “Artificial intelligence in iPhones: What will Apple do?” becomes repeatable practice rather than a one-off demonstration.
Most useful for
Product teamsEntrepreneursAI usersExecutives and managersEveryday usersInvestors

Key takeaways

00:00How the issue moves from news to product: when a new version of ChatGPT responds faster,

The “When a new version of ChatGPT responds faster, speaks, and understands images, it can feel as” issue should be assessed with one constraint in mind: progress does not move only forward: users notice the model sometimes answering worse, behavior changing without explanation, old functions disappearing, and familiar data becoming inaccessible — a powerful technology does not automatically become a strong product.

01:16The practical meaning of the issue: chatGPT and the destruction of archives: how OpenAI worsens

The decision in “ChatGPT and the destruction of archives: how OpenAI worsens conditions for new startups” depends on one criterion: OpenAI settled with the Financial Times and Reddit, while over the book archive it wriggled out — the training was long ago, the staff has left, the archive is deleted; a gulf opens for startups trying to catch up: collecting datasets pirate-style, as OpenAI once did, is no longer an option for them.

04:13Where the promise meets reality: openAI and the courts

The “OpenAI and the courts” scene leads to a working conclusion: the terms of OpenAI's deals with publishers are opaque — unlike Google's more transparent creator-payout rules — so it is unclear whether the company is setting an industry standard or simply paying to be left alone.

07:36What determines the outcome: new ChatGPT-4o: what opportunities are available to a

In the context of “New ChatGPT-4o: what opportunities are available to an ordinary user,” this criterion applies: GPT-4o showed that the interaction model matters as much as raw capability — voice, images, and natural dialogue bring the model closer to an everyday assistant, and the test is how many steps it actually removes from routine tasks.

14:20Why an announcement is not enough: artificial intelligence in iPhones: What will Apple do

For the “Artificial intelligence in iPhones: What will Apple do?” scene, the decisive point is this: Apple cannot simply add another chat window to the iPhone — it has to integrate AI so that it works with personal context, preserves the familiar interface, and does not hand control of the user to a competitor.

17:22Any partnership between Apple and OpenAI or Google remains a form of competition

The decision in “Competition disguised as cooperation between Apple, Google, and Microsoft” depends on one criterion: OpenAI is tied to Microsoft, Google competes with Apple directly, and Apple itself is used to controlling the critical parts of its products: a deal buys fast access to a strong model, but long-term dependence on someone else's system inside the smartphone — its key interface — is something the company is unlikely to accept.

20:08The boundary between value and constraint: review of key players in the AI market

The decision in “Review of key players in the AI market” depends on one criterion: OpenAI keeps timing announcements a day before Google's — Sora before Gemini, GPT-4o before the new assistant — investing in the feeling of being ahead; meanwhile nobody audits the quality of its models, which sometimes degrades — unusual for software that used to only improve.

23:36Who owns the outcome: the unprofitability of Perplexity and Uber

The working conclusion from “The unprofitability of Perplexity and Uber” is that running a large model for every search query is expensive, and unworkable unit economics get flooded with investor money, as Uber once did; Perplexity has $10–20M revenue against a $3B valuation, and the risk is higher than Uber's: the big companies never built their own Uber, but they do build search.

31:56Against this backdrop, the treatment of data looks especially strange

The decision in “New concept of basic unconditional income and AI” depends on one criterion: Sam Altman proposes basic income not as money but as a per-person quota of compute — and that is a question of power: select people and projects already have personalized models without resource limits, so the key issue is not the payout but who gets what level of access.

39:52The question ‘Why is ChatGPT getting dumber?’ is therefore broader than the quality of a single answer

In the context of “10-year data loss,” this criterion applies: archives, request histories, and personal materials can sit for years until one error or a company decision wipes out access: information has become cheap to create, but its preservation is not guaranteed, and users often do not know where their digital memory lives or how to take it with them.

What this episode is about

GPT-4o made AI noticeably easier to use, Apple is preparing to integrate models into the iPhone, and users are increasingly noticing instability and lost data. Rapid progress does not remove a basic requirement: a product must be reliable, understandable, and must not force people to live inside someone else's opaque system.

When a new version of ChatGPT responds faster, speaks, and understands images, it can feel as though progress moves in only one direction. Yet users are noticing something else at the same time: the model sometimes answers worse, its behavior changes without explanation, old functions disappear, and familiar data can become inaccessible. A powerful technology does not automatically become a strong product.

GPT-4o showed that the interaction model matters as much as raw capability. Voice, images, and more natural dialogue bring the model closer to an everyday assistant.

That is precisely why Apple cannot simply add another chat window to the iPhone. It has to integrate AI into the device so that it works with personal context, preserves the familiar interface, and does not hand control of the user to a competitor.

Any partnership between Apple and OpenAI or Google remains a form of competition. OpenAI is tied to Microsoft, Google competes directly with Apple, and Apple itself is accustomed to controlling the critical parts of its products. A deal can provide fast access to a powerful model, but in the long run Apple is unlikely to want dependence on someone else's system inside its most important interface: the smartphone.

Against this backdrop, the treatment of data looks especially strange. Archives, request histories, and personal materials can be stored for years until a single error or company decision removes access.

Information has become cheap to create, but its preservation has not become guaranteed. Users often do not even know where their digital memory is stored or how to take it with them.

The question ‘Why is ChatGPT getting dumber?’ is therefore broader than the quality of a single answer. We depend on systems that change constantly without explaining their internal logic. A proper AI product should do more than impress us with a new model. It should preserve data, behave predictably after updates, and allow people to leave the ecosystem without losing their own history.

A sound AI product must do more than impress with a new model: it must preserve data, behave predictably through updates, and let people leave without losing their own history.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 59 segments: 47 identified, 2 mixed, 7 probable, and 3 unresolved.

Loading…