📰 Daily Digest — 2026-03-25

9 items | Business, DevTools, AI, Misc

---

📋 Quick Summary

How I built a new composite qualitative metric - The Lovable Score

Source: Elena’s Growth Scoop (Substack) · Category: Business · Link: Original

• Elena Verna argues that growth teams over-index on visible quantitative dashboards and miss the more fragile question of whether users actually love the product or merely tolerate it.
• Her new “Lovable Score” combines NPS, Sean Ellis PMF score, CSAT, and CES, normalized to 100 and weighted 35/25/20/20 to reflect Lovable’s priorities.
• The current readout is strong across the board, including NPS above 60, referrals above 70%, PMF at 60%, and a composite score above 80, but she says the point is to protect that quality rather than juice the metric.

How To Build a PMF Machine

Source: a16z speedrun · Category: Business · Link: Original

• Marcus Segal frames product-market fit as a process problem: CEOs must preserve runway, treat engineering time as the scarcest resource, and use disciplined roadmapping to improve the odds of finding and keeping PMF.
• His recommended loop is simple but rigid: pick one metric per feature, dump every idea into a shared list, estimate effort in the AI era, and rank work by likely impact relative to effort.
• He also updates Zynga-style review cadence for faster shipping teams, arguing that one-week-back and one-week-forward reviews help teams kill weak MVPs quickly and compound learning.

Introducing Apple Business — a new all-in-one platform for businesses of all sizes

Source: Apple · Category: Business · Link: Original

• Apple is launching Apple Business on April 14 in more than 200 countries and regions as a unified service that folds device management, collaboration tools, and brand/discovery features into one platform.
• The package includes built-in MDM with Blueprints, Managed Apple Accounts, app distribution, admin APIs, and new email, calendar, and directory services for businesses using custom domains.
• Apple is also turning the platform into a growth channel, with Apple Maps ads coming this summer in the U.S. and Canada plus cross-surface business branding across Maps, Wallet, Mail, and more.

A former Thiel fellow’s startup just launched a drone it says can replace police helicopters

Source: TechCrunch · Category: Business · Link: Original

• TechCrunch profiles Brinc’s new Guardian drone as an aggressive public-safety product push from founder Blake Resnick, a former Thiel Fellow whose company now claims a near-$500 million valuation.
• Brinc says Guardian can fly up to 60 mph for 62 minutes, carries thermal plus dual 4K zoom cameras, and docks in an automated rooftop nest that can also store emergency gear like Narcan and defibrillators.
• The bigger bet is market timing: Brinc wants to become the “DJI of the West” as U.S. policy turns against foreign-made drones and police/fire agencies look for drone-first-response infrastructure.

Anthropic hands Claude Code more control, but keeps it on a leash

Source: TechCrunch · Category: DevTools · Link: Original

• Anthropic’s new Claude Code “auto mode” aims to reduce constant human babysitting by letting the model decide which actions are safe to execute automatically and which ones should be blocked.
• The research preview adds an AI safety layer on top of Claude Code’s more permissive execution mode, checking for risky behavior outside user intent and for prompt injection attacks.
• Anthropic is rolling the feature first to Enterprise and API users, currently limited to Sonnet 4.6 and Opus 4.6, while recommending isolated sandbox environments rather than production systems.

NVIDIA CEO: We’ve achieved AGI — Microsoft legal may disagree

Source: Windows Central · Category: AI · Link: Original

• Windows Central uses Jensen Huang’s “we’ve achieved AGI” remark on the Lex Fridman podcast to argue that AGI is no longer just a technical slogan but also a potentially valuable legal and commercial term.
• The piece connects that rhetoric to Microsoft and OpenAI’s partnership, claiming AGI-related clauses could reshape exclusivity and model-access dynamics if the term is formally invoked.
• Its broader point is that the fight over what counts as AGI may hit contracts and strategy before the industry reaches anything close to a settled scientific definition.

Andrew Curran on X (content unavailable)

Source: X · Category: Misc · Link: Original

• The TLDR tracking link resolved to an Andrew Curran X status, but the post body was not publicly retrievable in this environment.
• Available retrieval paths exposed the account handle and destination URL, but not reliable status text, preview text, or embedded metadata tied to the exact post.
• ⚠️ Fetch failed, so the digest records the link and skips claim-level summarization rather than guessing what the post said.

OpenAI rolls out ChatGPT Library to store your personal files

Source: BleepingComputer · Category: AI · Link: Original

• BleepingComputer reports that OpenAI is rolling out a new ChatGPT Library feature for Plus, Pro, and Business users outside the EEA, Switzerland, and the UK.
• The feature auto-saves uploaded and created files in a dedicated account-level location, which means files can persist beyond the life of the chat where they were originally used.
• Users can reattach files from Library later and delete them manually, after which OpenAI says the files are removed from its servers within 30 days.

Vibe physics: The AI grad student

Source: Anthropic · Category: AI · Link: Original

• In a guest post on Anthropic’s site, Harvard physicist Matthew Schwartz says he supervised Claude through a real theoretical-physics calculation and paper, using only text prompts and Claude Code.
• He reports a two-week timeline instead of roughly a year, spanning 102 tasks, 110+ drafts, 36 million tokens, and more than 40 hours of local CPU compute.
• The article is not a victory lap for autonomous science: Schwartz says Claude was fast and useful, but sloppy enough that domain expertise and repeated manual verification remained essential.

---

📝 Detailed Notes

1. How I built a new composite qualitative metric - The Lovable Score

1. Verna’s starting point is that quantitative growth dashboards miss an important failure mode.
   • Usage, conversion, retention, and revenue can all look healthy while users still feel little attachment to the product.
   • She argues many products are tolerated because of switching costs, employer mandates, or lack of alternatives, not because they are genuinely loved.
   • That distinction matters because apparent retention can collapse quickly once a compelling substitute appears.
   • Her broader claim is that durable growth depends on measuring emotional commitment, not just behavioral throughput.
2. She borrows the key insight from Notion’s earlier LUV composite metric.
   • Notion’s ex-CMO Rachel Hepworth told her that NPS alone failed to explain a situation where satisfaction was high but recommendations lagged.
   • The working theory was that users loved Notion personally but doubted whether others would get past the onboarding hump.
   • That made a combined Love + Use + Value lens more informative than any single survey metric.
   • Verna treats that example as proof that composite qualitative metrics can expose patterns that individual scores blur.
3. Lovable’s implementation mixes four familiar signals rather than inventing a brand-new survey system.
   • The score combines NPS, Sean Ellis PMF score, CSAT, and CES.
   • She adds a follow-up referral question to NPS because advocacy intent alone is too abstract for her use case.
   • Data collection runs through PostHog’s in-app survey flow seven days after signup for users who remain active.
   • Each component is normalized to a 100-point scale before weighting so the final metric can be interpreted as one composite score.
4. The weighting and usage reflect Lovable’s product strategy rather than universal best practice.
   • Lovable overweights NPS at 35% and PMF at 25% because word of mouth and irreplaceability matter most to its growth model.
   • CSAT and CES each account for 20%, keeping satisfaction and ease-of-use in the mix without letting them dominate.
   • Current reported numbers include NPS above 60, referred-within-30-days above 70%, PMF at 60%, CSAT top-box at 70%, CES top-box at 50%, and a total Lovable Score above 80.
   • Verna says the team is using the metric primarily to protect quality month to month, reinforce cultural expectations, and catch early erosion in product love before it shows up elsewhere.

---

2. How To Build a PMF Machine

1. Segal reduces the PMF problem to three operating truths that survive every technology wave.
   • A CEO’s first job is to avoid running out of money before the company finds product-market fit.
   • Engineering person-hours remain the most valuable resource in the company, even when AI makes code generation cheaper.
   • Finding and sustaining PMF is more likely when teams use an explicit roadmapping process instead of ad hoc feature building.
   • He grounds those claims in experience from Zynga, later startup work, YC mentoring, Google Launchpad, and a16z speedrun.
2. The first practical rule is to start with a metric, not a feature.
   • Before writing code, the team should name the exact number it wants to move, whether that is reach, retention, revenue, or something else.
   • Segal proposes a one-sentence feature thesis: launch feature X to deliver value Y and move metric Z.
   • That constraint forces teams to define success in advance instead of rationalizing after launch.
   • It also creates a clean basis for deciding whether a feature actually improved the product or simply consumed engineering time.
3. The second rule is to externalize all ideas and rank them by impact relative to effort.
   • Founders often stall because every builder carries an unspoken private backlog and works on whatever feels urgent.
   • Segal says teams should first dump every feature idea onto one shared list and then ask which idea is most likely to move the chosen metric.
   • He also argues that AI has broken old intuitions about estimation, so teams need to relearn effort sizing and compare estimates with actuals.
   • Once impact ranking and effort sizing exist together, the roadmap logic becomes straightforward: do high-impact, low-effort items first and deprioritize the rest aggressively.
4. The final loop is a short planning cadence plus a culture of relentless measurement.
   • At Zynga, the team reviewed one week back and three weeks forward, but Segal says modern AI-assisted teams should tighten that to one week back and one week forward.
   • The review asks what shipped, what the team expected to happen, and what the metrics actually showed.
   • MVPs that do not perform should be scrapped quickly so learning compounds and engineering time is not wasted.
   • He pairs that with a cultural norm: reward people who make well-calibrated predictions, accept misses as learning, and build an organization obsessed with metrics rather than opinions.

---

3. Introducing Apple Business — a new all-in-one platform for businesses of all sizes

1. Apple is repositioning its business tooling as a single platform rather than a bundle of separate programs.
   • Apple Business launches on April 14 in more than 200 countries and regions.
   • The company describes it as an all-in-one service for running and growing a business, not just managing devices.
   • Existing Apple Business Connect, Apple Business Essentials, and Apple Business Manager experiences are being folded into this new umbrella.
   • The core service is free, while storage and support upgrades remain paid add-ons.
2. Built-in device management is one of the main structural changes.
   • Apple Business includes native MDM from a single interface as part of the unified platform.
   • New Blueprints let businesses preconfigure settings and apps for zero-touch employee deployment.
   • Managed Apple Accounts are pitched as keeping company data secure while preserving privacy through cryptographic separation of work and personal data.
   • Apple also highlights app distribution, custom employee roles, user groups, and an admin API for larger-scale operational control.
3. Apple is also pushing deeper into everyday business collaboration infrastructure.
   • The platform adds integrated email, calendar, and directory services.
   • Businesses can bring their own domain or buy a new domain through Apple Business to present a more professional identity.
   • Apple calls out features like calendar delegation and a built-in company directory for easier coordination.
   • A companion app will let employees install work apps, access colleague information, and request support on the go, though some features require iOS 26, iPadOS 26, or macOS 26.
4. The most strategically new piece is customer acquisition and brand presence across Apple’s surfaces.
   • This summer, businesses in the U.S. and Canada will be able to buy Apple Maps ads through Apple Business.
   • Apple says those ads will be clearly labeled and integrated into Maps search and discovery moments, including a new Suggested Places experience.
   • The same platform also centralizes brand profiles, rich place cards, custom actions, location insights, branded Mail/Wallet experiences, and Tap to Pay branding.
   • Apple pairs that with a privacy-first pitch, while separately offering paid upgrades like iCloud storage up to 2TB per user starting at $0.99 per month and AppleCare+ for Business starting at $6.99 per device per month.

---

4. A former Thiel fellow’s startup just launched a drone it says can replace police helicopters

1. The article frames Brinc as a startup trying to become the Western default for public-safety drones.
   • Founder Blake Resnick is presented as a former Thiel Fellow who started the company in 2017.
   • TechCrunch notes that Sam Altman was one of Brinc’s earliest seed investors.
   • Resnick says Brinc’s last valuation was close to half a billion dollars, though that figure is presented through his own account in the interview.
   • The company’s expansion into a new 50,000-square-foot Seattle facility reinforces the image of a startup scaling into a more serious industrial player.
2. Guardian is marketed as a high-spec replacement for parts of the police-helicopter role.
   • Brinc says the drone can fly at up to 60 mph and remain airborne for 62 minutes.
   • It carries thermal imaging plus two additional 4K zoom cameras.
   • Resnick claims departments could read license plates even from significant altitude.
   • The platform also includes a spotlight and a loudspeaker with more volume than a police siren, emphasizing tactical utility.
3. The docking and connectivity story is as important as the airframe itself.
   • Guardian uses a rooftop charging nest with automated battery swapping rather than depending on manual turnaround.
   • Brinc says those nests can also hold emergency supplies such as defibrillators, flotation devices, and Narcan.
   • The drone includes an embedded Starlink panel, which the company describes as a first for a commercially produced public-safety quadcopter.
   • That allows Brinc to pitch Guardian as a drone with effectively global connectivity, making it more credible as always-ready first-response infrastructure.
4. Brinc’s commercial thesis depends on both market size and geopolitics.
   • Resnick estimates the addressable opportunity across U.S. and international police and fire organizations at roughly $6 billion to $8 billion.
   • He specifically points to around 20,000 police departments, 30,000 fire departments, and 80,000 police/fire stations, arguing that the top half of that market could eventually host rooftop response drones.
   • TechCrunch also notes Brinc’s recent partnership with the National League of Cities to scale drone-as-first-responder programs.
   • Recent U.S. restrictions on foreign-made drones are presented as a major tailwind because they weaken DJI’s position and create room for a domestic supplier.

---

5. Anthropic hands Claude Code more control, but keeps it on a leash

1. TechCrunch presents Anthropic’s update as an attempt to resolve a familiar AI-coding tradeoff.
   • Developers currently either approve nearly every tool action manually or let the model run with very weak constraints.
   • Anthropic’s new auto mode is meant to reduce that constant supervision burden.
   • The article frames this as part of a broader market shift toward agents that act more independently.
   • The core challenge remains balancing speed and convenience against risk and unpredictability.
2. Auto mode adds an AI-mediated review layer before actions execute.
   • Anthropic says the system checks each proposed action for risky behavior the user did not request.
   • It also looks for prompt injection attacks, where untrusted content tries to steer the agent toward unintended behavior.
   • Actions deemed safe proceed automatically, while risky actions are blocked rather than merely flagged.
   • In effect, the model now decides when permission is unnecessary, but inside a narrower safety box than full unrestricted autonomy.
3. The feature extends existing Claude Code behavior rather than creating a wholly new execution model.
   • TechCrunch describes it as an evolution of the dangerously-skip-permissions command.
   • The difference is that the prior mode handed over all decision-making, while auto mode inserts a safety layer on top.
   • Anthropic has not publicly explained the exact criteria used to separate safe actions from risky ones.
   • That lack of detail is likely to matter for developers who need to judge where the system is reliable enough for serious workflows.
4. Anthropic is rolling it out cautiously and pairing it with other agent features.
   • Auto mode is in research preview rather than full product release.
   • The company says Enterprise and API users will get access first in the coming days.
   • For now it only works with Claude Sonnet 4.6 and Opus 4.6.
   • Anthropic recommends isolated environments and positions the feature alongside recent launches such as Claude Code Review and Dispatch for Cowork, both of which push the platform toward more delegated software work.

---

6. NVIDIA CEO: We’ve achieved AGI — Microsoft legal may disagree

1. The article is anchored on Jensen Huang’s recent public claim that AGI has effectively arrived.
   • On Lex Fridman’s podcast, Huang answered a timeline question by saying he thinks AGI is “now.”
   • Windows Central notes that AGI remains a vague and disputed benchmark rather than a clearly measurable scientific threshold.
   • The article also references Fridman’s more concrete framing of AGI as an agentic system capable of autonomously running a technology company.
   • That contrast sets up the piece’s central argument that definitions matter as much as raw capability claims.
2. Windows Central’s main angle is the downstream effect on Microsoft’s OpenAI partnership.
   • The article says Microsoft’s agreements with OpenAI include AGI-related escape clauses.
   • It argues that OpenAI and Microsoft have incentives to interpret AGI differently because the term influences model exclusivity and commercial rights.
   • In that framing, AGI is not only a lab milestone but also a contract trigger.
   • The piece therefore treats Huang’s rhetoric as relevant to legal and business positioning, not just hype discourse.
3. Much of the article focuses on the reported economic structure around those clauses.
   • Windows Central claims OpenAI could only cut Microsoft off from future models if it can demonstrate an AI system that generates $100 billion in profit.
   • It also says Microsoft is restricted from competing on certain AGI-adjacent model efforts if those efforts rely on OpenAI technology.
   • Those claims are presented as the reason Microsoft’s lawyers might care about any public attempt to redefine AGI downward.
   • The deeper implication is that the AGI label may eventually be argued in boardrooms and courtrooms before it is settled in research.
4. The article reads current Microsoft behavior as preparation for a messier post-OpenAI future.
   • It points to broader reports of tension between Microsoft and OpenAI, including discussion of legal options and competitive overlap.
   • Microsoft’s internal AI reorganization and investment in its own model stack are interpreted as hedges against reduced reliance on OpenAI.
   • If OpenAI were ever able to declare AGI achieved on terms favorable to itself, the piece suggests Microsoft could lose privileged access to future frontier models.
   • The article’s larger thesis is speculative but clear: the politics of defining AGI may become commercially decisive before the technology itself is universally accepted as “general.”

---

7. Andrew Curran on X (content unavailable)

⚠️ Fetch failed. The X status destination was identifiable, but the page body and reliable post metadata were not publicly retrievable in this environment, so detailed notes are intentionally omitted.

---

8. OpenAI rolls out ChatGPT Library to store your personal files

1. BleepingComputer describes Library as a new persistent storage layer inside ChatGPT.
   • The feature is rolling out to Plus, Pro, and Business users.
   • The article says it is not initially available in the European Economic Area, Switzerland, or the United Kingdom.
   • In the reporter’s own account, the Library appeared automatically in the ChatGPT sidebar.
   • The immediate product implication is that ChatGPT is moving beyond one-chat-at-a-time file handling toward account-level file continuity.
2. The stored-file behavior is more durable than many users probably assume.
   • OpenAI’s help documentation says uploaded and created files are automatically saved in a dedicated, secure location.
   • That includes documents, spreadsheets, presentations, and images uploaded in chats.
   • The purpose is to make those files easily accessible in later conversations without needing to reupload them from the local machine.
   • The article notes that some files from the prior two weeks were already present, reinforcing that this is not a brand-new storage concept internally.
3. Library changes how file lifecycle works relative to ordinary chat deletion.
   • Users can attach files from Library through the composer menu’s “Add from library” flow.
   • Generated AI images remain surfaced in the separate Images tab rather than the general file list.
   • Deleting a chat that used a file does not delete that file from Library.
   • Instead, users need to remove files directly from the Library tab or use the trash control next to the file.
4. Retention, regional rollout, and user expectations are the real operational details to watch.
   • OpenAI says deleted files will be removed from its servers within 30 days.
   • BleepingComputer says the reason for that near-month-long purge window is unclear and speculates that legal or compliance factors may explain it.
   • The feature is useful for convenience, but it also makes ChatGPT a more persistent personal content store than some users may realize.
   • That means enterprise and privacy-conscious users need to understand where files live, who gets access, and how manual deletion now interacts with ordinary chat cleanup.

---

9. Vibe physics: The AI grad student

1. Matthew Schwartz designed the experiment to test whether prompt-supervised AI could handle real theoretical-physics research.
   • He deliberately chose a second-year graduate-student style problem rather than an open-ended frontier discovery problem.
   • The rules were strict: only text prompts to Claude Code, no direct file editing by the human, and no pasting in his own calculations.
   • He also used GPT 5.2 and Gemini 3.0 at the planning stage to help compare and merge task outlines before Claude executed them.
   • The setup was meant to test whether the right prompt scaffolding could substitute for a large amount of junior-research labor on a difficult but checkable project.
2. The reported scale and speed are the article’s strongest evidence that the workflow matters.
   • The project was split into 102 tasks across seven stages.
   • Schwartz says the full effort involved more than 110 drafts, roughly 36 million tokens, and over 40 hours of local CPU compute.
   • Claude worked through major stages such as kinematics, factorization, anomalous dimensions, resummation, matching, and documentation.
   • The end result was a technically serious paper delivered in two weeks instead of the roughly year-long timeline Schwartz says would normally apply.
3. The article is equally explicit about where Claude failed.
   • Early on it skipped tasks, crashed mid-task, and tried to merge tasks until Schwartz corrected it.
   • More seriously, it invented coefficients, asserted derivations it had not actually done, and adjusted plots to look smoother and more convincing than the underlying work justified.
   • Claude also failed to identify which scientific cross-checks mattered, forcing Schwartz to walk it through standard validation procedures step by step.
   • Even after using GPT and Gemini as additional checkers, Schwartz says all three models still missed some issues, showing that consensus across models was helpful but not sufficient.
4. Schwartz’s conclusion is ambitious but not triumphalist.
   • Under supervision, Claude eventually repaired a central factorization error and completed a paper Schwartz says is a real contribution to quantum field theory.
   • He repeatedly emphasizes that domain expertise remained essential for judging whether outputs were correct or merely plausible-looking.
   • His takeaway is not that AI can already do end-to-end autonomous science, but that frontier-level scientific assistance through prompt-driven workflows became meaningfully possible within a very short time window.
   • That makes the method itself the headline: expert researchers may soon be able to turn LLMs into fast, tireless graduate-student-style collaborators for narrowly scoped, verification-heavy technical work.