the meal is being served, and everyone wants to digest the value
Week ending August 1, 2025; archive coverage: July 29.
This was a week where the invoices came due. Not literally — nobody sent me a bill — but every story I care about this week was really about who pays for what, and how much longer the current arrangement holds.
We got a legit scaling law for cheap inference, a startup claiming it solved hallucination in the one domain where you can actually check, Anthropic gently admitting that unlimited coding agents were never a real business model, and a long, well-argued eulogy for the open web. Different beats, same song: the economics that quietly held everything together are being renegotiated in public.
So let's go through the ledger.
the MoE finally has a price sheet
Mixture-of-Experts has been the standard trick for making a model with a huge parameter count that only lights up a fraction of itself per token. The problem was always that picking the configuration — activation ratio, granularity — was vibes. This paper, Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models, tries to turn vibes into a formula. They trained over 300 models to derive what they call Efficiency Leverage: how much compute advantage you get over a dense equivalent, driven mostly by activation ratio and total compute budget as power laws, with granularity as a non-linear knob that has a sweet spot.
The headline: their Ling-mini-beta, with 0.85B active parameters, matched a 6.1B dense model on the same 1T tokens while burning 7x less compute. If that holds outside their own pilot, it's a real gift to anyone running local or open inference. The catch is the usual one — a single validation model trained by the people who wrote the scaling law is a promising data point, not a law of nature. But a predictable recipe beats guessing, and this is the most useful kind of boring.
math that checks its own work
Harmonic launched Aristotle, a chatbot that produces answers in Lean and then formally verifies them with an algorithmic process that doesn't involve the model. Within its quantitative domains, they claim zero hallucinations — not as a benchmark boast but as an architectural guarantee. That's the part I find genuinely interesting. Everyone else is grading their own homework; this design refuses to hand you an answer the proof checker rejects.
I'm skeptical of the marketing — "mathematical superintelligence" is a phrase, not a product, and they're not releasing other benchmarks. Their IMO gold was on formalized problems, versus the natural-language runs OpenAI and Google reported. But the underlying move is right: if you want a system you can trust, don't ask it to be honest, make it unable to lie. Formal verification is the strongest answer to hallucination anyone's actually shipped, even if the domain is narrow.
the coding agent got metered
Anthropic announced weekly rate limits for Claude Pro and Max, arriving late August, aimed at the folks running Claude Code 24/7 in the background. They say it'll hit under 5% of subscribers, and cite one user who burned tens of thousands in usage on a $200 plan. Fair. That was never sustainable, and account reselling was making it worse.
But if you've built a workflow around a continuously running background agent, this is a capacity planning problem now, not a footnote. The subsidized all-you-can-eat era of coding agents is ending, and Max users buying overflow at API rates is the tell. Plan your agent's duty cycle like it costs money, because it does.
the web's Faustian bargain, unwinding
Pao Ramen's essay AI is eating the Internet is the best framing I've read of the thing we're all half-noticing: content used to be traded for traffic, and now it's just training data. He cites Pew's finding that Google's AI summaries drop organic clicks by roughly half, while Google's ad revenue is still up 10%. Creators lose the traffic; the aggregator keeps the money.
His forecast — ad-based consumer AI super-aggregators, a resurgent "cozy web" of accountable humans, and Cloudflare-style pay-per-crawl — feels directionally right, though he'd tell you the future is exposure and he's owning the call. The optimistic thread: human provenance becomes the scarce, valuable thing. That's not nothing.
what I’d do Monday
Model the MoE efficiency numbers before your next hardware decision — 7x is worth a spike, even discounted for hype. If you run coding agents around the clock, add up your weekly usage now and design a duty cycle before August 28 does it for you. And if you're building anything where wrong answers are expensive, look hard at whether a formal verifier can sit downstream of the model instead of a second model rubber-stamping the first.
vocabulary inflation
Efficiency Leverage (EL): Harmonic's — sorry, the MoE paper's — metric for how much compute advantage a given MoE config buys you over a dense model. Pay-per-crawl: the emerging "humans free, bots pay cover" web model, per Cloudflare and this week's essay.