The Wall Street Journal argues cheap model weights relocate spend from training to inference, not away from infrastructure. Meanwhile GLM-5.3 lands near Kimi K3 with unusually strong cybersecurity scores.
Asa Fitch's August 16 analysis argues AI infrastructure beneficiaries cash in whether open weights proliferate or not, because cheap weights shift spend from training runs to inference serving. Anthropic's first quarterly operating profit and revenue up more than 14x year over year sit in the same frame.
Developers testing Z.ai's GLM-5.3 report it matching and sometimes passing Kimi K3 at the same size, with unusually strong cybersecurity benchmark results. Offensive and defensive capability is heading into open weights, which Z.ai says it will release around the end of August.
A market note published 17 August says the US will impose 50% tariffs on nearly $20 billion of Canadian imports on 19 August unless negotiators reach a breakthrough. Hardware and component supply chains are the part builders should watch.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →In an August 16 analysis, the Wall Street Journal's Asa Fitch argues the AI boom's infrastructure beneficiaries cash in whether open-weight models proliferate or not, because cheap weights shift spending toward inference capacity rather than eliminating it. The piece lands during the heaviest open-weight release...
Continue reading arrow_outwardAn August 16 Wall Street Journal analysis by Asa Fitch argues open-weight proliferation redirects AI spending toward inference capacity rather than shrinking it, leaving infrastructure vendors paid either way.
Washington is scheduled to impose 50% tariffs on nearly $20 billion of Canadian imports this week unless negotiators reach a breakthrough, according to a market note published 17 August.
Washington is scheduled to impose 50% tariffs on nearly $20 billion of Canadian imports this week unless negotiators reach a breakthrough, according to a market note published 17 August.
A maglev sled reaching 800 km/h in 5.3 seconds and a newly named black hole star share a trait with GLM-5.3: each is a record set by moving faster than the category that describes it.
Fitch's argument is that cheap model weights do not remove compute demand, they relocate it from training runs to inference serving. The timing is pointed: the piece landed during the year's heaviest open-weight release burst, with DeepSeek and Z.ai among the names shipping. If correct, the thesis means the release wave that looks like commodity pressure on labs is actually volume growth for whoever sells capacity.
Developers report Z.ai's GLM-5.3 matching and occasionally exceeding Kimi K3 in the same weight class, with unusually strong cybersecurity scores. Z.ai says open weights follow in roughly two weeks, around end of August, on an unchanged architecture. Strong offensive and defensive tooling scores in a downloadable model put those capabilities in far more hands than a hosted API would.
Washington is scheduled to impose 50% tariffs on nearly $20 billion of Canadian imports this week unless negotiators reach a breakthrough, according to a market note published 17 August.
14x Revenue, First Profit
Read full analysis arrow_outwardArchitect.new is a new entry in the prompt-to-app category indexed on There's An AI For That, alongside the more established builders. Use the directory listing to check pricing and named alternatives first, then give it one throwaway spec you already know the shape of so you can judge output quality instead of novelty. Treat anything it generates as a scaffold to review, not code to ship.
Try it arrow_outwardBlink.new is another recent prompt-to-web-app tool listed on There's An AI For That, and the honest way to evaluate it is head to head. Pick one internal tool you would otherwise build by hand, give the same prompt to Blink.new and to whatever you use now, and compare how much of each result survives a code review. The winner is the one that needs fewer manual fixes, not the one with the nicer preview.
Try it arrow_outwardLumi.new sits in the crowded generate-an-app-from-a-prompt tier that has grown fast this month. Before you commit, run the boring test: add a second data model, an auth requirement, and one edit after deploy, because that is where most of these tools stall. Check the listing page for current pricing and alternatives so you are not comparing on marketing copy alone.
Try it arrow_outwardAtlassian's new agent for Jira Cloud generates standup digests, project health checks, and stakeholder updates from ticket data you already maintain. The value depends entirely on hygiene: if your team does not move tickets, the digests will be confidently wrong. Point it at one squad with disciplined ticket updates first, then compare its weekly summary against the one a human writes.
Try it arrow_outwardAnthropic's Cowork runs tasks from Chrome and lets you continue them across devices, which targets the multi-tab research and form-filling work that never fits a chat window. Start with read-heavy jobs such as pulling data out of a dashboard or reconciling two web views, and keep it away from anything that submits payments or sends mail on your behalf until you have watched a full run. Browser agents inherit your logged-in sessions, so scope permissions before convenience wins.
Try it arrow_outwardDocument Kit for PM and UI/UX designers is a bundle of research, build, and design templates for shipping a digital product, covering the documents teams usually improvise under deadline. Its real use is as a prompt scaffold: paste a template into your model of choice and have it fill the sections from your notes, rather than asking for a spec from nothing. Keep the structure, replace every generic line, and delete the sections your project does not need.
Try it arrow_outwardRun a three-way scaffold test. Give Architect.new, Blink.new and Lumi.new the same one-paragraph spec for your next side project and compare what each produces in the first ten minutes. Keep the winner, delete the rest.
Push the generated app past the demo. Lumi.new is the cheap test here: take its output, add auth and a real database write, then see how much of the generated code survives. That failure point is your real evaluation.
Price your inference, not your training. Fitch's argument is that open weights relocate cost to serving, so pull GLM-5.3 into a local eval, measure tokens per request on your actual workload, and model the monthly serving bill before you commit to a self-hosted stack.
The Journal's read is the one to internalise today: open weights move the bill rather than cancel it, and inference serving is where that bill lands. GLM-5.3 proves the capability side of that trade, arriving near Kimi K3 in the same weight class with cybersecurity scores strong enough to matter on both offence and defence. For builders, the leverage is no longer in access to a good model, it is in what you pay to serve one and how fast you can scaffold around it. Spend today testing generators and measuring tokens, not shopping for capability you already have.
AI analysis, world news, markets, and tools. One briefing, delivered free.