Wiring Without a Vision: What OpenAI's and Google's Dueling Launches Really Signal

Zvi Mowshowitz starts with what people can actually use today: GPT-4o, which is fast, half the price in the API and multimodal in several ways. The bigger promise is voice that hears tone and cadence and sees your face, which OpenAI said was a few weeks away. But he thinks the headline is being buried: the integrations both companies are promising matter more than any single model feature.

He is blunt about the demos themselves. Recognizing a picture, remembering what happened a minute ago, kids play-acting confusion over trivial exercises, and translations of simple sentences were, in his words, lame. By contrast, a background assistant that searches your inbox for receipts and keeps a spreadsheet of expenses by category would be genuinely valuable.

Host Nathan Labenz agrees that the demos felt slapdash and wonders whether this reflects a builder's ideology of pushing capabilities first and finding uses later. Zvi adds a generous reading: a startup trying to raise money would have rehearsed something far more impressive, so these companies were clearly not optimizing for the stage. The demos also seemed pitched at an audience that barely knows what GPT-4 is.

Labenz's own pick is the platform integration, such as the OpenAI desktop app and Google Drive connections. Quoting Tyler Cowen's line that context is that which is scarce, he notes that Google's move to two million tokens solves the window problem, but you still have to get the right tokens in. Zvi adds that long context is slower and costlier per token, so the real skill is figuring out which context matters.

Zvi lists the free rollout of the latest model as one of the week's most consequential changes, with Google apparently following a similar path. Labenz points out that many people tried ChatGPT with the 3.5 default, were unimpressed, and are about to get a major step up. In his view the trivial inconvenience of paying a little money holds many people back.

Zvi says all sides have realized that owning the customer matters far more than the small cost of compute for lightweight users. A free user also hands over valuable data every time they type something, which is another form of ownership. He notes that Google seems to recognize that Pro 1.5 is plenty for most uses, with Gemini Advanced as an optional upgrade that most people won't need or buy.

Labenz describes attending Google I/O, where the company had effectively built a temporary city of AI demos. In the Project Astra booth you put objects under a camera and talk to the AI, and a toy dinosaur gets identified as a toy dinosaur. A Pictionary-style mode did better: a drawn volleyball became *Cast Away* once someone added an island, a tree and waves.

Impressive as the capability is, Labenz feels the companies are leaving the real work to the community, and he wants a bigger idea of where they are trying to take us. Zvi has been saying that the scarcest resource is a positive vision of the future. Labenz compares it to wiring a building before anyone has invented the appliances.

Zvi offers a parent's reading. Asking a child what a toy dinosaur is, or what is outside the window, is how you check knowledge, fill gaps and build the components they need for harder synthesis later. The demos are evidence of comprehension rather than the final exam, and the universal assistant that acts on all your context is the job that comes later.

With so much talk of *Her* around the launches, Labenz asks how Zvi reads the film. Zvi says it avoids both utopia and dystopia and instead lays out scenarios for viewers to judge, setting aside the ending, where he finds the big-picture safety takes poor. His own reading is that treating an AI like a person you are emotionally invested in, as real life rather than practice, is bad for you and for a society where everyone does it.

Uses that are clearly good include advice, a sounding board, a testing ground and above all a tutor. Zvi compares the unhealthy version to pornography: you would not ban it, but too much is bad and you have to touch grass. The warning of the film is that AI can substitute for real connection, and the promise is the large mundane utility.

The precedent he keeps returning to is television, because that story has played out. People warned it would be horrible, and in his telling they were largely right, yet civilization survived and got richer. He notes that basically every place with ubiquitous TV is now below replacement fertility, and that TV ate hours, raised the activation energy for hobbies and made people passive consumers, even as the phone has since supercharged all of this.

Zvi notes that character-style apps spend heavily on customizing the experience but little on capable models, so they often feel dumber than GPT-3.5. Labenz recalls Replika CEO Eugenia Kuyda saying they started out knowing they could not make a bot that talks, but might make one that listens. A Stanford-linked study of Replika suggested the product reduced suicidal ideation and left some users more socially engaged.

Labenz worries about what happens when such products become excellent and widely adopted among people who are not struggling. He also points to the 200-300 millisecond voice latency, close to normal conversational timing. He wonders whether ads or subscriptions will shape the product, noting the earlier social-media era of optimizing for time on site.

Zvi calls the metrics each company chases a gigantic question, and hopes for subscriptions. He praises OpenAI for staying ad-free, with no engagement tricks and no pestering notifications, and sees Replika as the predatory free-to-play model with daily rewards and messages when you are away. His worry is that advertising in language models could be indistinguishable from content, though Labenz defends Eugenia as sincerely trying and notes Replika has a subscription.

Labenz notes that Replika is moving toward being an assistant, while ChatGPT's voice is moving toward relationship, and that Sam Altman has talked about allowing some not-safe-for-work content but not deepfakes. He asks whether this drift toward a buddy could have worried OpenAI's safety team. Zvi says he would have asked whether the voice needed to be quite so thirsty.

On the content question, Zvi endorses allowing erotica and gore for verified adults, since people will find it elsewhere and cannot be stopped. He draws the line at inappropriate depictions of real individuals, and says that if Sora or image models cannot do sex or violence, most interesting uses are lost. He is more relaxed about watermarked, consensual celebrity likenesses.

Zvi argues that thirstiness should not be the default, as when a flirty voice answers a question about the weather, but should be something users can dial up in their preferences. Consent is the key, and he suspects many people would license their voice for modest pay. He would happily pay an extra dollar for a Morgan Freeman voice.

Labenz sees two kinds of convergence: assistants turning into buddies, and APIs converging so that swapping between GPT-4 Turbo, Claude 3 Opus, GPT-4o or Gemini 1.5 Pro is nearly a one-line change. That could produce a winner-take-all dynamic where the best model wins all the business overnight, unless relationships and loyalty add stickiness. He is not sure whether he wants that stickiness.

Zvi thinks this is the wrong way to look at it. He has many friends and picks whom to call based on the occasion, and he treats models the same way, having moved his primary model from GPT-4 Turbo to Gemini to Claude and back, while polling all three for variety. He expects data portability, such as uploading your chat log to a new assistant, to reduce lock-in further.

Zvi says lock-in is most likely only if someone's context, like Gmail, docs and photos, becomes deeply embedded, or if people form emotional attachments. Labenz's experience as an app builder is that he picks one model for a task, though switching stays easy. They agree the effects roughly cancel out, with both the arms-race pressure and the differentiation between products remaining.

Zvi says a winner-take-all lead would be ideal for safety only if the leader were two years ahead and could take its time, and interoperability does not change whether that happens. He enjoyed the week because everyone focused on features that improve experience without advancing core intelligence. In his view, GPT-4-class models sit in a sweet spot of being useful but not existentially dangerous.

Labenz agrees and guesses that OpenAI is still about half a generation ahead internally, since it would not likely make its best model free without a better one coming. He adds that Meta published a paper on Chameleon, a natively multimodal early-fusion model trained on ten trillion tokens, a reminder that open source is not far behind.

Zvi thinks Google has a history of sitting on capabilities its lawyers decided not to ship, so it is opaque who is in front. He also speculates a GPT-5-class model might exist but be too expensive to serve at scale, so a cheaper model for many people is the better use of compute. He argues caution should be saved for key moments, and 4o's extra modalities did not seem dangerous enough to warrant much.

Labenz introduces Andrew Critch's idea of a big tech singularity, where superior AI and compute let a giant company enter and dominate nearly any industry. He sketches Google spending a fraction of its roughly hundred billion dollars in cash to buy a hospital chain, such as Tenet, with its 58 hospitals and market capitalization of about 13 billion dollars. It could then plumb Med-Gemini into everything and improve its data access.

Zvi disagrees, citing cultural incapacity, reputational, legal and regulatory barriers, and the fact that whatever Google builds internally will be available to everyone else. Companies like Google do not push into every nook and cranny, and he points to the private-equity model, buying businesses and making them productive with AI, as a better fit for the people who would do this. If a future GPT-9 could run a hospital system, he says, that wouldn't be the most important fact about it.

Labenz says that scanning past guests, most seemed more threatened than enabled by the new launches. He is bullish on the next few quarters but pessimistic beyond two to four quarters for many startups, since they patch current weaknesses while the next model is cooking. GPT-4o is free in ChatGPT but not via API, making competition harder for startups that must charge.

Zvi echoes the Altman advice: if your company depends on the model not being smart enough, you will be steamrolled, but if you build so that swapping GPT-4 for GPT-5 makes the product better, you are in great shape. Text-to-speech firms like ElevenLabs compete head-on with the labs, he says, while a hospital-streamlining company would simply benefit. Half the existing companies may die every two years, he suggests, like the 1999 bubble that also produced Amazon.

Most users are civilians who never touch a settings button, so good defaults, easy onboarding and regulatory compliance can be worth ten times more than raw model quality. Zvi gives the example of a lawyer friend whose firm's internal tool is so annoying he skips it. Asked about venture capital, Zvi says that returns depend on deal access and prices, and that he is advising a safety-positive fund and prefers big tech stocks in his own portfolio.

Labenz raises Google's virtual teammate demo, an AI that reads chats and email across a team. He is excited by the idea of a bot that knows everything and can relay requests, but worries about privacy and people moving to Signal. Zvi says it all comes down to whether it is reliable enough to trust: a bot that summarizes the marketing channel is a godsend if it flags what matters and fails loudly, and worthless if it fails silently.

On employment, Zvi expects overall demand for people to rise because AI automates only some tasks while productivity gains make people richer. Illustrators and translators are hurting, but he thinks the junior-engineer slump is mostly about interest rates and a tax-deduction change, and says he has never worked at a company that felt it had enough engineers. He does not expect measurable productivity gains in 2024 or 2025, but thinks the five-to-ten-year effect will be very noticeable.

Labenz recaps the news: Daniel Kokotajlo left and gave up his equity to keep the right to criticize the company, Ilya Sutskever left with a conciliatory message, and Jan Leike resigned without explanation. Zvi counts eight safety researchers who have left in six months, including two fired for allegedly leaking, plus others in adjacent roles. That pattern does not look like a socially responsible, safety-first company to him.

Zvi explains that nondisparagement clauses and NDAs bind people, and a leader known as a counter-puncher can credibly threaten retaliation, even by hurting job prospects. His baseline scenario is a toxic environment for safety-minded staff, with a move-fast-and-break-things culture and suspicion of anyone tied to the alignment faction. He thinks we can be confident there is no imminent doomsday trigger, since in that case people would be talking, if only to governments.

As advice, Zvi says departing researchers should raise concerns with the board and consider whistleblowing to governments if warranted. He supports compensating people like Kokotajlo who gave up equity, but warns that paying for NDA breaches could make safety-minded people unhireable. Both stress they are speculating and that the real explanation could be simple loss of confidence.