Week 33 · 2025-08-11 → 2025-08-17 · 14 newsletters

The Layer Above Execution

design-and-talent-bifurcation · ai-as-situational-leverage · vertical-fintech-and-data-rails · creative-process-and-attention

A mid-August week, sixteen emails across seven days, the kind of inbox where everyone is either at the beach or writing about who is not. The through-line that emerged was thinner than usual but real: the layer of work that AI does not touch is also the layer the market suddenly cannot find enough people for, in design and in engineering both. A small cluster on vertical fintech and AI data infrastructure ran underneath that, and a handful of writers on craft and attention closed it out.

Design and the Talent Bifurcation: The Execution Layer Is Hollowing Out

The week's strongest piece was Carly Ayres at Good Graf on the bifurcated market for designers. The setup is the whiplash on Design Twitter: half the feed is gallows humor about disappearing jobs, the other half is founders posting roles and asking how to land a great designer. Julie Zhuo's line that the market for startup design talent "has never been more competitive" sparked hundreds of replies and 250k views. Carly's answer to which it is, peak demand or mass extinction, is that both are true at once. The past year has been a bloodbath at the execution layer (Autodesk, Intel, Google, Salesforce, agencies, studios), full-time roles swapped for contract work, basic visuals outsourced to AI or overseas shops. Junior designers have lost the production rungs where craft used to get built. The part AI does not touch (framing the right problems, translating complexity into experiences people trust) is the part companies are now bidding for. Preston Attebery's line, that once everyone can make an app we will remember the hard part about apps is not making the app, is the version of the argument that travels.

Anthony Kline at The General Partnership ran the same argument in a different vocabulary on the engineering side. Six months after writing about software's deflationary era (fewer people doing more, founders in builder mode, lean teams with tall mandates), the unexpected counter-pattern is that the most capitalized companies in the world cannot hire. Meta is poaching from OpenAI with founder-level equity. Engineering management is making a return at Series A and B because CEOs are rediscovering that you cannot manage scores of direct reports and still build, ship, sell, hire, and raise. The wave moves systemically, from technical talent to seasoned operators to sales leaders to marketers, and the tell that the war has reached every corner of the org chart is when the best-capitalized companies start prioritizing recruiter headcount. The ZIRP-era playbooks (careers pages, junior sourcers, top-of-funnel volume) no longer deliver.

The take: both writers are describing the same phenomenon from different sides. The execution layer in design and the junior-sourcing playbook in engineering are getting compressed at the same moment that the layer above (judgment, framing, leverage) is in shortage. The arbitrage right now is for anyone willing to do the harder part of the work, in either discipline. The middle of the ladder, the production specialist, is where the squeeze is.

AI as Situational Leverage: The Productivity Story Gets Honest

Addy Osmani at Elevate ran the week's most useful AI piece, a long read on the actual state of AI-assisted software engineering productivity. The framing: AI functions as a situational force multiplier, providing modest, uneven boosts that augment rather than transform engineering productivity. The Stack Overflow 2025 numbers are the ones to remember. 84% of devs are using or planning to use AI tools, up from 76% a year prior. Favorable views dropped from over 70% in 2023 to ~60% in 2025. 46% of developers say they do not trust the accuracy of AI output. Studies show 20-30% productivity improvements, far from the 10x claims. The killer stat is that 66% cite AI's "almost correct" solutions as their biggest time sink due to debugging. AI excels at greenfield projects and struggles with complex legacy codebases. Adoption soars, trust plummets.

Sahar Mor at AI Tidbits ran the companion craft piece on DeepWiki, the Cognition tool that turns any GitHub repo into a navigable wiki by swapping github.com for deepwiki.com in the URL. The framing he opens with is the one Osmani's numbers point at: we are generating more code than ever, so the challenge is no longer producing code but understanding it. Fast mode answers instantly from the code graph, Deep Research spends extra cycles for higher-confidence multi-hop answers, and every answer ships with clickable line-level citations. The pairing of these two pieces is the read of the week for any working engineer: take the productivity story seriously, then pick the tools that actually attack the bottleneck (comprehension) rather than the ones that just produce more output.

The take: the AI productivity conversation has moved from "look what it can do" to "what is actually getting measured, and where." Osmani's piece is the most honest accounting of the gap between marketing and lived experience I have read this year. The operators who win the next year will not be the ones who use AI for everything; they will be the ones who know which tasks AI is actually a multiplier on and which tasks it is a tax on.

Vertical Fintech and AI Data Rails: The Quiet Infrastructure Bets

The week's two investor letters from Nikhil Basu Trivedi at next big thing hit the same beat from two sides. The first was Footwork's investment in Confido, the OS for CPG brands to manage inventory and financials with retailers and distributors. Justin Hunter saw the pain working for Dan Gilbert's family office: patchwork systems, broken trade workflows, hours wasted reconciling deductions. The thesis Footwork is laying out is that AI is finally letting COOs, CFOs, and CROs run vertical financial operations as software, with measurable impact on both costs and revenue. The second was Protege, the data exchange for AI training, now working with over 100 data providers across healthcare, media, audio and speech, and motion capture. Most major foundational model companies are customers. GMV is up more than 20X in 2025 from 2024. The thesis is that the four substrates of the AI revolution (energy, compute, data, models) all need transaction layers, and data was the one without one until Protege.

Matt Brown ran the broader frame in a short and quotable post on embedded fintech. The joke that airlines are credit card companies that happen to own planes is closer to truth than most realize: loyalty program fees generate the bulk of US airline profits. Toyota, John Deere, D.R. Horton, AT&T all run lending or insurance arms on top of their core products. The modern version is platforms like Toast, Shopify, ServiceTitan, Yardi: restaurant owners no longer go to Wells Fargo for card payments, landlords no longer wait in line at Chase for loans. The pattern is when a non-financial company offers financial products that are highly contextual, curated, and customized for the company's existing customers.

The take: the connective tissue across Confido, Protege, and Matt Brown's piece is that the cleanest fintech bets right now are the ones embedded in workflow software where the customer is already showing up to do the underlying job. The horizontal financial product is harder to win; the vertical one rides on top of demand that already exists. That has been true for a while in retail and restaurants. The Confido bet is that it is now true in CPG-to-retailer financial operations too.

Creative Process and the Defense of Diversion

The week's quieter cluster came from three writers on craft, attention, and the cost of efficient self-management. Ami Vora opened with the line her manager once asked her: "Ami, you always think of ten reasons why something won't work. Can you think of ten reasons it will work this time?" The wake-up call was that her skill at spotting risks had become a reflex, a self-reinforcing cycle where she would call out a risk, things would go well, so she would call out even more, until she stopped experimenting and lost the data to check whether her instincts were still right. The practice she lands on (deliberately asking what would happen if the moonshot worked, what is the quality worth, when to veto rather than ask leading questions) is leadership hygiene rather than insight, but the framing of strengths casting shadows is the cleanest version of the argument.

Piera Luisa Gelardi at Noomalooma ran the week's most surprising piece, a defense of delightful diversions built around the childhood story of her family trying to invent a Ben & Jerry's flavor (root beer float) and getting so absorbed in the process that they forgot to submit an entry. Years later she spotted "Root Beer Float My Boat" on the Ben & Jerry's menu as a limited-edition flavor. The argument she is making, against the pure utility frame on attention, is that the diversions that go nowhere are often the ones that matter most. It is a piece designed for a August week and lands accordingly.

Michael Goldstein writing at AIR ran the creator-economy version with Clay Campaigns. The Steve Jobs line that customers do not know what they want until you show them needs an update for an algorithm-ruled marketplace: customers do not know what they want until other customers show them. Brands zig-zagging their values, the rise of creator intimacy, culture speeding up: rigid slow-developed campaigns cannot keep pace. The platform reality is that the algorithm is the market, and smart brands spin the wheel with a wide range of bets rather than betting everything on a single 30-second spot.

The take: these three pieces are not a coherent theme so much as a coherent mood. The week was light on news and the writers who landed best were the ones who used the slowness to make a case for something soft (curiosity, diversion, iteration over rigidity). That is the right read of an August week.

Three Takeaways from the Week

The execution layer in knowledge work is getting compressed from two sides at once, and Carly Ayres and Anthony Kline are describing the same arbitrage from different industries. The play, if you are a designer or an engineer, is to spend the next year building the part of the craft that AI does not touch and the org chart cannot easily replace. The play, if you are hiring, is to stop running the ZIRP-era top-of-funnel motion and start hunting for judgment.

The AI productivity story has gotten honest enough to act on. The 20-30% number is the one to plan around, not the 10x. The 66% "almost correct as biggest time sink" stat is the one that should change which tasks you actually delegate to an agent. Pair that with comprehension tools rather than production tools and the equation flips.

If you only revisit three pieces from the week, I would suggest Carly Ayres on the bifurcated design market for the cleanest read on where craft is going, Addy Osmani on the reality of AI-assisted engineering productivity for the most honest accounting of what AI actually does at the desk, and Piera Luisa Gelardi's "In Defense of Delightful Diversions" for the post most worth reading slowly in a week that finally lets you.