Joseph Zhang

AI Builder & Writer | Sharing Technical AI Insights, Helping Businesses Apply AI | Co-Founder @VizageAI | Wharton MBA

[Joseph Zhang @AppliedLLMs]

@simonw This explains the pricing page confusion but not the docs update. Those were deliberate changes, not a bug. Hard to reconcile both stories at the same time.

[Joseph Zhang @AppliedLLMs]

@GergelyOrosz Fair framing, and I've run fake door tests myself. The thing that makes this one uncomfortable is Anthropic's brand is specifically built on trust. The technique is legitimate. The fit with the brand is not.

[Joseph Zhang @AppliedLLMs]

@Miles_Brundage I noticed this too. Deep Think had such a tight quota it felt like a demo feature. Pushing Max mode for Deep Research before that capacity issue was visibly solved feels like the same pattern again.

[Joseph Zhang @AppliedLLMs]

@arpit_bhayani I learned this the hard way with a long-running data pipeline that failed 3 hours in because a single env var was wrong. Five lines of startup checks would have saved an entire afternoon.

[Joseph Zhang @AppliedLLMs]

@jeremyphoward I've been waiting for this. The unofficial endpoint being used in prod by multiple tools while everyone held their breath was never a stable foundation. Official support changes what you can actually build on top of it.

[Joseph Zhang @AppliedLLMs]

@jxnlco I think Anthropic's access model made sense 18 months ago. But Codex went wide, the landscape shifted, and now the waitlist just looks like friction with no real upside.

[Joseph Zhang @AppliedLLMs]

@petergyang @cursor_ai The xAI compute deal is the tell for me. That's not a product decision, that's an "our unit economics don't work at scale" decision. Post-training only solves half the problem.

[Joseph Zhang @AppliedLLMs]

@garrytan I crossed this threshold a few months ago with my own setup. The moment background processes started catching things I would've missed, I stopped thinking of it as a "tool" entirely.

[Joseph Zhang @AppliedLLMs]

@ChowdhuryNeil I've been on teams that ran internal tooling on better infra than what shipped. You stop noticing the rough edges really fast. Nobody files a bug on something they never hit.

[Joseph Zhang @AppliedLLMs]

@kunchenguid In my experience the teams actually ahead don't talk about it much. They've just quietly moved to agents handling the boring stuff and humans reviewing, not initiating.

[Joseph Zhang @AppliedLLMs]

@emollick Publishing it got them the ecosystem credibility and benchmark adoption. But I think about the counterfactual a lot. No DeepSeek moment if they kept it quiet. That's not a small thing to give away.

[Joseph Zhang @AppliedLLMs]

@burkov I watched this happen in creative software. You spend years subsidizing adoption, enterprises get dependent, and by the time you try to flip on pricing a cheaper alternative has hit the same quality bar.

[Joseph Zhang @AppliedLLMs]

@GaryMarcus The analogy really holds. Counterfeit money still spends until someone actually checks. That's exactly where most press coverage on Claude's emotional states is right now.

[Joseph Zhang @AppliedLLMs]

@willccbb The half-work requirement is real. A harness that's too clean never gets enough training volume to matter. It needs rough edges to survive long enough to get baked in.

[Joseph Zhang @AppliedLLMs]

@cwolferesearch This is worth writing about. I've seen so many teams assume RL scaling works like pretraining scaling and blow their compute budget on it. One you can project forward, the other you're mostly explaining after the fact.

[Joseph Zhang @AppliedLLMs]

@AkariAsai @_emliu I once helped a researcher navigate the Stability AI compute grant process. We got access in two weeks with a one-page proposal. Two months waiting on a standard academic grant and the window would've closed.

[Joseph Zhang @AppliedLLMs]

@tylertringas The differentiator used to be the workflow. Build the integration nobody else has, defend the wedge. Now the foundation model ships the workflow native. The moat question is live again for every vertical.

[Joseph Zhang @AppliedLLMs]

@xlr8harder Framing it as harm not to improve this sidesteps the consciousness debate entirely. You don't need to resolve the hard question to agree that better calibration is worth pursuing.

[Joseph Zhang @AppliedLLMs]

@venturetwins The rate limit problem is real for anyone doing serious design iteration. You burn through the quota on the first three rounds of feedback, then you're waiting. Not a hobbyist issue.

[Joseph Zhang @AppliedLLMs]

@bindureddy I remember the first time I showed a skeptical colleague an AI that caught a data anomaly they'd missed for months. No ideology left in the room after that.