Daily TEA – A Subway Company Solved AI’s Business Model

AMD cracks the CUDA moat, a Hong Kong railway is the AI-lab blueprint, agents spend real money on-chain, and why you should pick your model last

Hello, dear TEA-mates! Here is what you need to know today.

1. 🔴 AMD Just Ran a Frontier Model at 80% of NVIDIA’s Speed for Half the Price

Inference startup Wafer served GLM-5.2 on AMD’s MI355X (about 2.75x cheaper per GPU than NVIDIA’s B300) and hit 2626 tokens/sec/node at 2.4 requests/sec on a 20k-in / 1k-out, 60% cache-hit workload, roughly 80% of the throughput they measured on a B200 at over 2x lower cost. They quantized the bf16 model to MXFP4 with AMD Quark (lossless vs the official FP8 on GSM8K, GPQA-Diamond, and tau2), ran it on sglang, and unblocked speculative decode with two small fixes: re-listing a bf16 shared expert under the module prefix sglang actually reads, and adding one #ifdef USE_ROCM guard to a CUDA-only kernel. No custom kernels were written this round. Wafer’s conclusion: SOTA on AMD is now a matter of software support, not silicon, and the CUDA moat is eroding in real time. (Read More)

🫖 TEA For Thought: “More like this will happen.”

2. 🎯 The Case for Learning Something Completely New This Week

Marginalia’s Viktor Lofgren makes a practical argument for picking up a new skill (pixel art, touch typing, 3D modelling, a language, woodworking) at about an hour a day. His core insight is about how learning actually feels: your first sessions will genuinely suck, you will perform worse toward the end of each session as your brain and muscles tire, and you will see little or no improvement during practice itself. The gains happen overnight, because practice is when you gather data and sleep is when the brain processes it, so the skill that felt impossible yesterday is tangibly easier the next day. He recommends 30 to 45 minutes of daily deliberate practice, stopping when you start making a lot of mistakes (so you do not ingrain them), and focusing on basics rather than mainlining advanced Reddit threads. The deeper payoff: long-term projects are how you build a real sense of control over your circumstances, since almost nothing meaningful changes within a day, but a lot can change across months and years. (Read More)

🫖 TEA For Thought: “Some inspirations for Monday after the long weekend!”

3. 🤖 AI Agents Are Now Spending Real Money On-Chain, in Tiny Bursts

Base spotlighted the emerging on-chain agent economy running on x402, Coinbase’s open protocol that uses the HTTP 402 “Payment Required” status code to let an AI agent pay for a resource in stablecoins (mostly USDC) with no login, card, or subscription. When an agent hits a paywalled endpoint, the server returns a 402, the agent signs a stablecoin transaction, attaches the proof, and retries, with the whole cycle settling on-chain in seconds. The activity is real but granular: from May 2025 to April 2026, autonomous agents processed more than 176 million transactions worth roughly $73 million, averaging about 31 cents each, meaning agents are not making big-ticket buys but constantly paying for micro-bursts of value like API access, data feeds, cloud resources, and AI inference. By Q1 2026, over 104,000 autonomous agents had registered across 15+ directories, and AWS, Coinbase, and Stripe have all shipped agent-payment infrastructure, with Base and Solana as the most-used settlement networks. (Read More)

🫖 TEA For Thought: “What are AI agents buying nowadays on chain?”

4. 🧠 Three Big-Question Essays: Ending Pandemics, Fixing Broken Countries, and the Subway That Solved AI’s Business Model

Dwarkesh Patel announced the winners of his AI essay contest, which drew 600 submissions. First place, Jassi Pannu (Johns Hopkins biosecurity professor), argues the OpenAI Foundation should run a state-scale operation to end airborne disease transmission, spending $40 to $60 billion over 10 years on passive physical infrastructure like far-UVC lamps, which she says could cut seasonal flu mortality 60% and unlock over $1 trillion in annual GDP. Second place, Ege Erdil (Mechanize co-founder), argues countries outside the AI supply chain should mostly resist populist panic and enact timeless pro-growth policy: strong property rights, low capital taxes, and open regulation, betting that incumbents like the US and China will make self-inflicted blunders. Third place, Michael Li (Harvard Kennedy School), draws the sharpest analogy: Hong Kong’s MTR railway never recovers construction costs through fares, but it is one of the only self-financing transit systems on earth because it owns the property that appreciates around its stations. His reframe for AI labs: the API is the rail and will never be profitable enough, so the survivors will own the appreciating assets around it (deployment rights, RL reward data, forward-deployed integration). (Read More)

🫖 TEA For Thought: “Great read!”

5. 🧭 Stop Picking Your AI Model First. Build the Router.

Tomasz Tunguz argues most teams building agents pick the model first and the architecture second, which is backwards: the model choice should be the last decision, not the first. What matters is the router, a small piece of code that decides which tier of model handles each request. Get it right and 70 to 80% of traffic runs on local models that cost nothing per call, or on async batch models that cut AI spend by 90%+. He splits the routing problem into three distinct jobs: a skill classifier (a language problem: what is the task?), a router (a scheduling problem: which tier runs it, based on complexity, context size, and historical success, not the raw prompt), and a model selector (pick the cheapest model in that tier that clears a confidence bar). The unlock is queueing: a draft reply, a repo summary, a diligence memo, or a nightly evaluator run do not need to return in a second, and async batch inference runs about two orders of magnitude cheaper than real-time. He cites Coinbase’s Brian Armstrong, who kept AI spend flat while token usage grew exponentially, not with spend alerts but with better defaults, routing, and caching. (Read More)

🫖 TEA For Thought: “This is a great piece, especially when you have to think about lowering costs. Look at how much Meta itself has to pay for all the expensive frontier models, because after all you don’t need the most frontier model to answer your emails. And you don’t have to have answers immediately for all requests. They could be done overnight, and the next morning you just read and verify. This is something people can do manually too: schedule work during the day, models work at night, in the morning you check, and there you go!”

🛠️ Skill of the Day

The Skill Ramp: turns “I want to learn X” into a realistic first-30-days plan that respects how learning actually works.

You are a patient coach who has taught total beginners to reach a usable
intermediate level. I want to learn: [SKILL, e.g. touch typing / watercolor /
Python / conversational Spanish]. My honest constraints: about [30] minutes a
day, [4] days a week, and my starting level is [complete beginner / rusty].

Design my first 30 days. Rules:
1. Pick ONE free, non-salesy starting resource and one backup. Say why.
2. Give me a week-by-week focus. Weeks 1 to 2 = only the true fundamentals,
   no advanced tricks.
3. For each session, give a concrete 30-minute structure (warm-up, the one
   thing I am drilling, a stop cue for when I am too tired and starting to
   ingrain mistakes).
4. Tell me exactly what "normal and fine" struggle looks like versus a real
   warning sign I am practicing wrong, so I do not quit on a bad day.
5. End with 3 tiny milestones that prove I am actually improving by day 30.

Keep it blunt and encouraging. Do not overload me with information.

Paste into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own skill and your real weekly time.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:
Twitter/X