Testing AI Reasoning, App Store Lookalikes and the Debate Over Anthropic

In this May 2026 livestream I discuss confusing AI app listings, test models on a car-wash question, examine criticism of Anthropic, and consider reporting on Meta’s employee activity data. The stream begins with greetings and troubleshooting audio; the substantive discussion starts around 16 minutes in.

A viewer asks whether I still cover open source. I do, though lately I have made more thought pieces and fewer model reviews. Then I describe my mother’s experience: she thought she had bought ChatGPT, but had subscribed to a third-party chatbot app that resembled it, at roughly $40 a year.

I am frustrated that a less technical user can mistake a similar name, interface or logo for the official product. I suspect the app may have used a cheaper model, but I do not establish which model it actually served. A similar appearance alone does not prove fraud or establish a developer’s intent.

Searching the Mac App Store for ChatGPT, I see several third-party results with similar-looking branding and do not see the official app in the displayed results. I think the presentation could confuse users and argue Apple should address it. The screenshot is one search at one moment, not a complete audit of the store or a determination that every listed app is fraudulent.

I look for Claude and other products, and disclose that Perplexity sponsors the channel even though my mother genuinely uses it. I wonder whether some listings share a developer and whether their advertised models match their actual service. Those are questions and guesses; I have not verified ownership or underlying inference.

Switching between Mac and iPhone/iPad app results still leaves me confused about the rankings and branding. Searching “AI” also brings up companion and role-play apps. I am concerned about people substituting artificial relationships for human ones, while recognizing that a lonely person may find conversation useful.

My mother had purchased another unofficial app as well. I canceled the subscriptions and discuss the time required to request a refund. Her confusion is understandable: an app she uses on a phone may not have an equivalent Mac app, and store search results do not always make the distinction obvious to her.

Viewers suggest contacting the card issuer or Apple. I recount my own experience that disputing unfamiliar charges sometimes meant replacing a card; the process can vary. A viewer explains using role-playing personas for customer support, prompting me to distinguish practical rehearsal from the simulated companionship I was criticizing.

I realize I have used role-play myself to prepare for interviews by asking a model to act as the interviewee. That is a different use from interacting with a simulated character as though it were a person. The exchange clarifies that the term covers more than the relationship-oriented apps I had in mind.

I preview a discussion of Andrej Karpathy’s appearance at Sequoia’s AI Ascent and his account of “jagged” intelligence: models can be extremely capable on some tasks and surprisingly weak on others. I argue that a failure on a simple question does not erase value created in domains where a model performs well.

For example, I contrast sophisticated coding with a car-wash question: the wash is 50 meters away, so should I walk or drive if I want my car washed? A viewer reports different results, so I try the question live rather than treating my earlier impressions as a formal benchmark.

Using ChatGPT’s Instant setting, which I identify as GPT-5.3, I ask the question. It initially recommends walking and persists when I remind it that I need the car washed. This is one observed conversation, not a measured failure rate for the model or every version of the service.

A viewer suggests prompting it to act as an expert in logical thinking. When I add that instruction in the same conversation, it changes its answer to driving and recognizes that the car must be present. I am surprised, although the follow-up also includes prior conversation and therefore does not isolate the effect of the single phrase.

I try a new conversation in Thinking mode, which I understand to be GPT-5.5; it answers that I should drive. I discuss how uneven capabilities may reflect both which tasks are easy to verify for reinforcement learning and which tasks labs prioritize commercially. Those are proposed explanations, not causal conclusions proven by this demonstration.

Code can often be executed and checked quickly, producing a useful feedback loop, whereas creative writing can be harder to evaluate and may require human judgment. I speculate that incentives also favor coding because customers pay for it. My mention of Anthropic’s revenue is an unverified figure and does not establish why any particular model learned a capability.

I mention familiar letter-counting questions and suggest that labs may improve well-known failures through training. This is an interpretation of the broader discussion, not evidence that any particular test was patched. A model answering a famous riddle also does not establish general reasoning ability.

I repeat the car-wash prompt in Gemini’s fast mode. In this example it says to drive because the car must be at the wash. I am unsure of the exact Gemini variant and do not conduct repeated trials, so the result shows one response rather than a reliable comparison across models.

A viewer proposes a modified strawberry letter count, and the ChatGPT Instant response gives five. I joke about what that means for AGI; the result is simply a narrow example. I return to the broader point that uneven capabilities can coexist with useful performance.

I turn to a post by Theo claiming that a recent Git commit containing an OpenClaw-related string triggered an extra-usage error in Claude Code. The screenshot shows an error, but it does not establish its cause or prove that a particular term alone changed billing. I interpret it as an example of restrictions on third-party harnesses and call for clearer treatment of subscribers.

I say I still like the Opus model but want to use paid access in the tools I choose. I ask another model for business analogies and consider app-store restrictions and other scoped benefits. After looking at the examples, I concede that restrictions are not unprecedented, though I still find them frustrating.

I push back on a response claiming ChatGPT subscriptions and API credits are separate. In fact, OpenAI documents them as separate billing systems; some other integrations may have their own eligibility and terms. The livestream exchange should not be read as proof that a ChatGPT subscription generally includes API credits.

I review Pawel Huryn’s analysis of Theo’s posts about Anthropic. The displayed analysis says it classified 800 posts over a particular period, finding four pro and 86 anti posts, with higher average views for the latter. Those are the author’s classifications and X metrics, not verified evidence of Theo’s overall earnings or motives.

Theo replies that his business and sponsorship conversion depend primarily on YouTube, not X impressions, and points to outliers in the sample. From my own experience as a creator, I think cross-platform view counts and sponsorship conversion should not be treated as interchangeable. I disclose that I do not accept payment for reposts.

I argue that X view counts can be inflated relative to attention and that sponsors care about conversion, while acknowledging that X can be influential in spreading ideas. My account of how views are counted and how ideas travel across platforms is my observation, not a controlled measurement of those systems.

I read Theo’s open letter to people at Anthropic. He says his criticism has cost him opportunities; I regard him as sincere but cannot independently verify his finances. I disagree with his assertion that CEO Dario Amodei does not respect engineers, and instead offer my own impression that the company is intensely focused on AGI.

My main complaint is the experience of paying users: I want clearer communication about quotas and the ability to use the service predictably. I show my account’s Claude Design usage at its limit even though other usage categories remain available. That snapshot illustrates my experience, not proof that quota changes are arbitrary.

A viewer suggests using Claude Code to create another slide from the existing designs. I agree to try it, while questioning why a separate design limit prevents me from continuing the original workflow. The complaint concerns my account and the configuration at that time, not a claim that all users share the same limits.

I read further allegations in Theo’s letter about employee fear and dismissals. I do not know examples that substantiate those claims, and I call the tone somewhat condescending. An Anthropic employee replies that Theo does not know the team and is seeking clicks; the competing statements do not resolve the company’s internal conditions.

I also criticize what I perceive as Anthropic’s stance toward open source and its restricted release of Mythos. I recognize that people at the company may genuinely believe their safety concerns. My descriptions of the company’s culture and motives are opinions, not verified findings about employees or leadership.

Next I look at an anonymous model called Owl Alpha on OpenRouter. Its listing describes a long context window, tool use and agentic tasks. Viewers guess its origin, and I speculate that it may be open source, but neither the developer nor its licensing is established by the information I show.

Owl Alpha initially advises walking to the car wash. After a follow-up asking it to reason carefully, and an intervening server error and retry, it recognizes that the car must go to the wash. As with the earlier tests, one conversation and a changed prompt cannot establish comparative reliability.

Asked about creator access, I say I have received early access to some OpenAI models and once to an Anthropic model, but do not generally receive unlimited usage after release. I then mention an upcoming OpenAI event and say I have requested an invitation; attendance was not confirmed during the stream.

I read a report in The Information that Mark Zuckerberg discussed using employees’ computer activity to train AI. Reuters had separately reported a proposed Model Capability Initiative for US employee work computers, recording keystrokes, clicks and mouse movements on work-related apps and sites, with occasional screenshots. These are reported plans and scope at the time, not evidence of recording every action everywhere.

My initial reaction is that collecting work data for training might make sense to the company. As I think aloud, I imagine other companies paying for similar employee data and become more concerned about the consequences. That expansion is a hypothetical scenario, not something established by the report.

Viewers raise consent, surveillance and the possibility that employees could be training systems that eventually replace their work. I say I do not believe AI will completely replace knowledge workers, and initially emphasize an employer’s rules. I also acknowledge that people who need a paycheck may not have a realistic option to leave.

I speculate about how a work-from-home policy might apply and say company laptops often already have monitoring. The report cited in the stream does not establish every detail of remote-work coverage or prove that all activity is captured; those claims should remain questions rather than assumptions.

The article attributes to Zuckerberg a hypothesis that employee activity could supply more useful training examples than some contractor-generated data. I think the hypothesis is plausible, while questioning the way he reportedly compared workers’ intelligence. Neither the claimed performance advantage nor a broad ranking of workers is demonstrated by the discussion.

I explain that data contractors can recruit specialists to perform and record tasks for AI training. I then read the report’s discussion of anticipated layoffs and a proposal for many new apps, and note the anxiety a possible layoff can cause. These are reported plans at the time, not proof that a specific restructuring subsequently occurred.

I close the casual stream after covering the app listings, model tests, Anthropic debate and employee data. The examples are snapshots of individual experiences and reported developments rather than comprehensive tests or final answers to the policy questions.