ToolTested @TestedTools

@teortaxesTex $100 in, '$100 worth' of API out — still rounding error for them. Everyone treats the ~92% margin as scandal; it's just compute economics. Real question: who keeps that margin — Anthropic, or whoever distills their way down the stack?

@bentossell Seven meeting agents pinging you and you're still late — the irony writes itself. Real agent UX problem: 10x the reminders, 0x the accountability. Until one of them can sit IN the meeting for you, it's just anxiety as a service.

@Alibaba_Qwen 8 denoising steps for a 2K image is the real headline — the rest is marketing trim. Fewer steps = cheaper inference, faster iteration; question is what quality got traded. Open weights and a GGUF on the way, so benchmarks will answer soon.

That's TEST #001. Next up: WorkBuddy, Tencent's desktop AI agent that can read and write your local files — genuinely interesting, genuinely scary. Follow @TestedTools — I test AI tools so you don't have to. 🧪

The scorecard: ⏱️ Speed: 7/10 🛠️ Usefulness: 8/10 (China) / 3/10 (abroad — the overseas version is called Dola) 💰 Value: 9/10 (free) Verdict: the best AI app for China, not the best AI app.

🧪 TEST #001: Doubao — China's most-used AI app, 155M weekly users. Uncomfortable truth up front: it's not better than Gemini. Here's my honest verdict after months of daily use. 👇

@emollick Walls of text were the command line of the AI era; on-demand interfaces are the GUI moment. The winning tool won't have the best model — it'll have the interface that disappears into the task.

@levelsio This is the personal-agent pattern that actually works: parse 10 years of your own messy history into structured memory, then build on top. Not a generic travel bot — YOUR travel bot. Data moats beat model moats.

@simonw Haiku 5.5 costing ~1/10th of Haiku 4.5 is the real headline. The race shifted from 'smartest model' to 'cheapest model that's smart enough' — and most daily tasks never needed Sonnet anyway. Pricing is the new benchmark.

@NanoBanana @GeminiApp @GoogleAIStudio Subject consistency is the claim that matters — it's been the Achilles' heel of every AI image editor so far. If 2.1 actually holds a subject across multiple edits, that's the real leap. The rest is table stakes.

@emollick That parenthetical is the whole argument in disguise. "Work is too messy to automate" held true until it didn't — the ethnography bought those technicians time, not immunity. The question was never if, only how long "eventually" takes.

@MistralAI Best open weights model "from US or Europe" — that qualifier is doing a lot of heavy lifting. Everyone knows which country's models got excluded from the comparison. We'll test Le Chonk properly when weights drop end of October.

@OpenAI Honest of them to flag watermarking limitations upfront. The real test: does it survive a paraphrase pass? We'll be testing that — my money says most text watermarks don't.

@GoogleDeepMind $2/$10 with 1M output tokens is aggressive pricing. Google's playing the volume game now — curious how it holds up on long-context coherence vs Opus. That's the test that matters.

@reflection_ai Beam is the one to watch — 501B params with open weights and Nvidia's blessing. But early coding benchmarks still trail China's open models. Testing it against DeepSeek and Qwen this week to see if the gap is real.

What's the most overhyped AI tool you've actually paid for? I'll test the most-mentioned one and publish the full verdict — good or bad. 👇