[Jun 22–28] | Taste Beats Building · AI Middle Squeeze · Trust as Growth

Taste is becoming the scarce operating skill

Mark Pincus, Andrew Ambrosino and Scott Wu all circled the same shift from different angles: building is getting cheaper, but choosing is getting harder. Pincus separates the underlying instinct from the first product idea, then runs ideas through Proven / Better / New. Ambrosino says the implementation is no longer the expensive part; taste is. Wu argues AI coding is already useful, but the hard problem is context-rich software engineering, not generating code in a clean demo.

That changes the slogan of the AI era. “Everyone is a builder” sounds empowering, but it is incomplete. If output becomes abundant, the scarce skill is not production; it is selection. The danger is that teams mistake a polished prototype for a real product, or confuse novelty with an obvious improvement users can feel. AI makes it easier to create ten plausible versions of an idea and harder to know which one deserves the company’s attention.

The practical move is to put judgment before velocity. Write the instinct memo before the feature memo: what human behaviour or market wound are we betting on, separate from this exact implementation? Then ask what is proven, what is better, and what is actually new. Label artefacts by stage so exploratory demos do not acquire false authority because they look finished.

For product leaders, this is also a hiring and promotion change. The best people will not just be those who can ship more with Codex or Claude Code. They will be the ones who can delete complexity, pick the right medium, spot when a B+ idea is blocking the A idea, and preserve craft as roles blur. Taste is no longer a soft skill; it is operating leverage under abundance.

Open source is squeezing the AI middle

The AI economics thread was unusually concrete this week. The DeepSeek discussion framed open source as an industrial-policy story, not a hobbyist one: a reported $7.4bn Series A at roughly a $50bn valuation, fewer than ten investors, and Chinese state voting control. Nikesh Arora expects long-term token prices to fall to about one-tenth of today’s levels, while also noting compute is costing 2–4x more than two years ago. Brian Armstrong claims open-source inference can be 99% cheaper and only a few months behind frontier models.

The obvious read is that frontier labs are under pressure. The sharper read is that the exposed layer is the middle: model vendors without frontier pull, wrappers with thin differentiation, and services businesses whose margin depends on expensive humans or brittle API economics. The number-three closed-source model is an uncomfortable place to stand if the leaders own the narrative and open source attacks price.

Operators should treat model choice as a supply-chain decision, not a religious one. Build routing across closed-source, open-source and fallback tiers before procurement forces the issue. Put an ROI gate on AI spend now, especially where usage is growing faster than proof of labour saved. If your product is “good enough model plus workflow wrapper”, have a clear answer to why open source plus a cheaper wrapper will not compress you.

The same logic applies to services and implementation businesses. If revenue is priced by seat, hour, or body count, assume customers will ask why AI has not lowered the bill. The defensible work moves towards judgment, accountability, proprietary context and outcomes. Everything else starts to look like margin waiting to be repriced.

Enterprise AI is a decision system, not a copilot layer

Nikesh Arora’s enterprise AI critique was blunt: most companies are still layering AI onto old workflows and getting 10–20% efficiency, rather than redesigning who decides, who reviews, and what no longer needs a human. Armstrong described Coinbase running about 1,200 full-time-equivalent AI agents internally, with pods shrinking from ten people to two to four, sometimes one human plus ten agents. Wu’s view of AI coding points the same way: the system has to understand logs, repos, prior decisions and workflow context, not just produce text.

That means the real product may not be the copilot UI. It may be the governance layer: routing, permissions, memory, audit trails, approval thresholds and the record of what happened last time. Enterprises do not need more chat boxes if they are willing to hand over more decision rights. The uncomfortable part is that handing over decision rights raises the bar on false positives, traceability and organisational trust.

The operator question is not “where can we add AI?” but “where is decision density high?” Look for workflows where humans move information, enforce consistency, make routine judgments, or forget brittle handoffs: contracts, invoicing, CRM updates, collections, reconciliation, support triage, release coordination. Start with one repetitive workflow where errors are costly and handoffs are weak, then design the approval and audit layer before the flashy interface.

This also changes org design. Smaller pods only work if knowledge is captured in a company brain that agents can use: customer feedback, incident data, product decisions, code patterns and prior trade-offs. Without that memory, AI usage stays as individual hacks. With it, the company starts to compound its own context.

Distribution is no longer downstream of product

FOMO, Pincus and Armstrong all made the same point in different markets: distribution is not a department you add after the product exists. FOMO’s social graph is the glue of the trading experience, with users following traders, copying ideas and sharing positions. The company also used a 140-angel round and in-house management of 30–40 creators as distribution infrastructure. Pincus’ Zynga lesson was that consumer products need to live where people already are. Armstrong says traditional media is effectively dead for many under-50 customers, so companies need podcasts, blogs and X as owned channels.

The contrarian version is that “build a better product and market it” is often backwards. In consumer and prosumer markets, the product is the thing users touch plus the loop that brings them back plus the channel that lets the company speak without permission. Super-app breadth only works when the bundle has one intentional thesis; otherwise it becomes a feature pile in search of a habit.

Founders should audit whether growth is embedded in the product or bolted on as paid acquisition. What public or shareable artefact does the product create? What repeated surface does it occupy? Where does social proof live? If a creator programme is material, run it like an operating system: creator-level CAC, LTV, conversion quality, creative iteration speed and audience fit, not vague “brand” reporting.

The tactical consequence is focus. One strong loop beats three weak channels. If the product depends on daily attention, design the reminder, the status mechanic, the social object or the owned narrative at the same time as the core feature. Distribution is now part of product architecture.

Trust is becoming a growth asset

Armstrong, Vlad Barbalat and Wu each described trust as something operational, not ornamental. Armstrong’s argument for tokenized equities is that one-to-one backing matters more than the wrapper; stablecoins worked because people trusted the reserves. He also frames the Clarity Act and Genius Act as product-enabling policy after years in which he says 80% of crypto trading volume moved offshore. Barbalat’s line is that transparency enables autonomy. Wu’s Windsurf rescue showed that, in a distressed market, a fast clean answer can preserve value before gossip fills the gap.

The less obvious implication is that regulation, disclosure and speed can become distribution. In fragile or regulated categories, users are not just buying UX; they are buying legibility. What is backed? Who owns the asset? What happens if something breaks? Who is accountable? The company that answers plainly can move faster because stakeholders give it more room.

This cuts against the startup instinct to stay vague until later. If you touch money, ownership, regulated assets, enterprise risk or customer continuity, trust is not back-office compliance. It is part of the product surface and the sales motion. Coinbase going public, seeking licences and putting Armstrong’s name visibly behind the company are not just institutional details; they are GTM choices.

Operators should make trust concrete. Publish what is backed and by whom. State the assumptions behind multi-year targets if you want investor or board latitude. In a crisis, move quickly enough to reassure customers, staff and the market while the asset is still intact. Transparency is not the opposite of speed; in the right market, it is what makes speed possible.

Permanent capital only helps if it stays sharp

Barbalat’s Liberty Mutual Investments conversation was a reminder that capital structure is not a personality trait. Managing about $120bn, including roughly $70–75bn of insurance reserves, gives Liberty the ability to make decisions that are right rather than expedient. But his stronger point was that duration is only useful if paired with investment hygiene, transparency and a broad toolkit.

The interesting framework is “exposures first, products second.” Liberty asks what risk it wants, then chooses the best instrument: direct, co-invest, club deal, GP relationship, LP allocation, structured credit or another route. That translates well beyond investing. Product and GTM teams often start with the instrument they know — sales-led, PLG, partnership, acquisition, pricing change — before naming the exposure they actually want.

The contrarian warning is that permanent can become lazy. Long duration can buy flexibility, but it can also hide weak thinking and slow response times. Barbalat’s comparison between four-year and 30-year credit in an AI-shifted world is useful: the same company can look safe over one horizon and fragile over another if technological relevance becomes harder to forecast.

For founders, the move is to separate capital availability from business-model fit. Do not go public just because you can; do it if public-market cadence fits the operating model. Do not raise long-duration capital and then let it become a comfort blanket. Write the exposure map first: what risks, partners, customers and time horizons do we actually want, and what structure gives us flexibility without dulling the edge?

The distressed asset may be worth more than the departure story suggests

Scott Wu’s account of the Windsurf deal was a clean case study in speed and asset specificity. Cognition heard about the situation on Friday, reached out cold that evening, agreed verbally on Saturday, papered Sunday and signed Monday morning. Wu pushed back on the idea that Windsurf was only a husk after talent departures, pointing to the remaining product, customer book, code, data, proprietary IP and team.

That is a useful corrective to how markets read talent shocks. The public story often becomes “the stars left, therefore the company is dead.” Sometimes that is true. But in software, especially enterprise software, value can remain in usage, workflow integration, data, customer trust and code paths that still solve a painful problem. The asset may be impaired without being worthless.

The operator lesson is to diligence retained value, not vibes. In any distressed deal or team shock, list the assets that remain and how quickly they decay: customers, contracts, data rights, product quality, support obligations, talent, brand, roadmap, IP. Then ask whether speed can stop the decay. In some situations, the first deliverable is not synergy or integration; it is reassurance.

There is also a talent covenant underneath this. Founder mythology says you go down with the ship. Wu’s disappointment at that covenant weakening is notable, but the practical point is narrower: when trust breaks, the clock starts. Leaders need a prepared playbook for fast clarity, because silence is not neutral. It lets the market write the ending for you.