Inbox Signal

Teams ship about 3x more code and still review it the same way: output is up, measured value and trust are not

Linear's report (199,000 users, June 2024 – August 2026) has agent-using teams going from 21 to 65 PRs a week against 8 to 10 for others, yet nobody saved time; Linear's conclusion is that deciding what to build, not execution, is now the limit. METR had four maintainers (scikit-learn, Sphinx, pytest) review 296 SWE-bench-passing AI PRs: merge rates ran 24 points below the grader's scores once adjusted for how strict the maintainers are, and the same maintainers merged only 68% of the original human patches.

Sonar says AI writes 42% of committed code, 96% of developers don't trust it, and only 48% always check it. DORA's ROI view: 35-40% gains on greenfield work, 10% or less on legacy code. Several figures are second-hand summaries of other reports, and the cause given for the 8-hour GitHub outage is Milan's own summary. Free, full text.

Anthropic left Git because agent-driven merges outgrew it, and rebuilt code review around risk tiers rather than reading every diff

Merges averaged about 2,500 a day in May and passed 30,000 a day by September, so a seven-person team (per Meng Li) built a replacement, Scion, and moved the monorepo over in five months. Reported results (TLDR Dev and Rohit): 200x less CPU and memory than the old Git-mirroring setup and 25% of CI CPU freed (TLDR Dev). Rohit notes the post itself says Git remains right for most repos.

The Claude Code team's post says continuous review, automated verification, multi-layer checks and a tiered risk system for approvals now sit in front of any human reviewer. Meng Li's teaser adds that about 90% of Claude Code's code is written by Claude, merged PRs more than doubled from June to September and the share needing rollback fell by more than 40%. His four rules for AI-written PRs: the opener owns it, add a 'How was this verified?' field, ask different review questions, and only the code owner merges. The three practices he would copy are not named in the preview.

Uber's AI developer infrastructure: a model gateway, tiered spend controls, and code review as the new bottleneck

Distinguished Engineer Uday Kiran Medisetty says Uber had 2 services in 2015, thousands by 2019 and about the same now, though the business is roughly 10x larger. Its Model Gateway (2023) keeps tooling model- and harness-agnostic, with AI Guard, per-user spend tracking and audit logs, and engineers can switch between Claude Code, Codex and OpenCode. Spend is allocated in tiers with manager sign-off to exceed it; token costs fell about 50% in recent months.

All PRs are still human-reviewed, but review is now the bottleneck, so AI review, screenshots/videos and risk scores come first. CI jobs more than doubled in 12 months, forcing an internal 'code red', and PRs are now tagged by type rather than tracked as 'PR velocity'. Podcast summary; a useful template for client governance conversations.

'Code is a commodity now' — and AI labs, hyperscalers and consultancies are building deployment companies to sell what is left

The efp.asia essay says agencies should shift from selling hours to domain expertise, product judgment and outcome accountability, with smaller generalist teams; it admits the productivity evidence is mixed and verification is costly in mature codebases. TLDR Dev's blurb is all the email gives on it.

'The rise of DeployCos' (TLDR IT) describes labs, hyperscalers and consultancies forming companies that embed forward-deployed engineers in enterprises to close the 'diffusion gap': workflow integration, systems access and change management. Neither blurb has numbers and both are opinion, but together they describe your direct competition and the opening for a consultancy.

Anthropic's own workforce study: agents now outwork people inside the company, and junior technical hiring has collapsed

By August agents worked about 6 hours for every human hour in employee-directed sessions and 19x in research orgs, and commits per engineer were 5.1x their December 2025 level (The AI Corner). The junior share of technical hires fell from 27% to 9% since early 2025; TLDR AI and The AI Corner agree on that figure. Agents sit idle about 80% of an interactive session waiting for a human, so human attention is the bottleneck.

Caveats: it is Anthropic's own data (HR records, telemetry and 17 interviews), and the AI Corner author discounts the 5.1x commits figure against an outside study of 500,000+ developers where commits rose about 240% but releases only about 30%. The email also prints a weekly reading load that exceeds the hours in a week, so that figure is left out. The workbook and prompts are paywalled.

Open-weight models now carry 71.5% of gateway tokens but 23% of spend, and Meta's Muse Spark Open is the newest contender

Vercel's AI Gateway index has open-weight models at 71.5% of September tokens and 23% of spend; the spend share more than quadrupled since June and almost one in three teams changed their most-used lab in September, an index record. TLDR Dev adds that price per token is down 47% since December.

Meta's Muse Spark Open (816B total / 42B active, 1M context) scores 39 on Artificial Analysis's index per AINews: first among US open-weight models but behind GLM-5.3 (45) and Kimi K3 (44), and weakest on agentic evals. Sources disagree on hardware: AINews says two 512GB M5 Ultra Mac Studios, Meng Li says two or four.

Agent teams cost 1.8-5.1x more than one agent with no statistically significant gain, except in one configuration

Vals AI's Vibe Code Bench compared a single agent with agent teams on GPT-6 Sol and Claude Opus 5.5 at medium and max reasoning effort; teams cost 1.8-5.1x more, and only Sol at medium effort showed a significant improvement. Vals suggests higher reasoning effort may be the better lever.

Latent.Space's workshop on Anthropic's dynamic workflows (Claude writes the program that orchestrates sub-agents) agrees directionally: more detail but 'much more' cost than a single agent, and the model does not know its budget while running, though a session spending limit can be set. No benchmark numbers in the workshop notes.

Anthropic's Q3 revenue is reported at $16.8B, but IPO valuation talk is meeting skepticism

TLDR says Anthropic told investors Q3 revenue rose about 45% to $16.8B and that an IPO could value it near $2T, with trading possibly starting next month; AINews cites Bloomberg for a listing before Thanksgiving. Meng Li's skeptical read: secondary-market quotes below $1.7T, buy-side institutions wanting a listing price under $1.5T, and alleged window-dressing through gross margin that excludes training costs and gross-basis revenue.

The accounting claims are the author's allegations; the preview does not give the detail. Preview only.

Per-token prices mislead: a 33% price cut produced a 4% higher bill, and five frontier models sit within 5 points at a 3.8x cost spread

Tom Tunguz reports five frontier models within 5 points on a benchmark but a 3.8x cost spread, with Gemini 4 Argon cheapest at $1.99 per task; billed reasoning tokens at max effort mean you should test your own requests. A separate test found a 33% Sonnet token-price cut produced a 4% higher bill across 191 constructed requests (thinking off; tokenizer change and longer answers). That is a small sample.

Ramp's AI Index: 1% of firms account for 80% of OpenAI and Anthropic spend, and September spend per employee was 3.1% below July's peak while token volume rose.

CoreWeave's debt is the week's AI-financing stress test; the two sources give different figures

TLDR cites CNBC for CoreWeave carrying more than $51B of debt and obligations, with its five-year credit default swap spread soaring; nobody doubts demand, but cash flow has to justify capex. Meng Li, citing disclosures as of 30 June 2026, has total debt of about $35.6B and a revenue backlog of about $104B, notes Nvidia is both major shareholder and key supplier with a contractual duty to buy part of unsold capacity, and says Jensen Huang denied on 15 October that CoreWeave needs a backstop.

The figures are not the same measure: $51B is debt plus obligations, $35.6B is debt only. TLDR (19 Oct) adds that hyperscalers are paying modestly more interest and some data centers are delayed over permitting.

Safe agent autonomy depends on where controls sit, not how many restrictions you add

Two Sigma's John Feminella attacks 'approval fatigue': approving everything blocks work and adds risk once attention lapses, so use composable controls, with classifiers for ambiguous calls and deterministic rules for clear ones ('Why do cars have brakes? So that you can go fast and stop safely'). Bloomberg's Andrey Rybka argues for deterministic controls outside the model: scoped credentials, tool-call controls, outbound traffic limits, audit trails, budgets and kill switches.

Anthropic's Keon Shahab describes a trust layer of identity, sandbox and approval before action; OpenAI's Joseph Kim predicts finance agents doing idea generation and trading in 2027. These are vendor presentations, and there are no figures.

A small energy-based model solved 96.2% of Sudoku puzzles where a panel of frontier LLMs managed 2%

In Logical Intelligence's public experiment, a panel of frontier LLMs solved 2% of puzzles at about $11,000 of API calls and the best single LLM solved 6.9%. Kona, an energy-based reasoning model described as about 1000x smaller, solved 96.2% for about $4 of GPU time, at 313 ms per puzzle against up to 90 seconds. The argument is that token-by-token generation is a structural limit on constraint problems, so this approach suits machine-checkable tasks (Sudoku, Lean proofs, chip layout), not open-ended ones.

All results are company-reported and come from a guest interview, so treat them as self-reported.

Gemini 4 Argon is generally available with a price cliff on 30 November and mixed independent evidence

Introductory API pricing is $2 / $10 per million tokens, doubling to $4 / $20 after 30 November (Artificial Analysis via AINews). It matches GPT-6 Astra on that index at about 60% of Astra's cost, $1.99 per task, with a 15% hallucination rate on AA-Omniscience; Cursor reports 52.8% on CursorBench, on par with Opus 5.5. Google's internal claims (a six-million-line Chromium folder split in weeks, 35% more agentic code submissions) are flagged by AINews as deployment claims, not controlled measurements.

The New Stack ran each model once: Argon matched Opus 5.5 on 14 planted bugs at lower cost but was worse on billing and long-context dependencies, where one wrong API call caused most billing failures.

Measure agent ROI per goal, not per session or token: Ramp's data shows session-level views misread cost

Ramp's agent Inspect wrote 75% of accepted code changes by August 2026. One cash-planning bug fix opened 8 sessions across 3 codebases, used about 5.6M tokens, cost $93 and shipped 2 fixes. Grouping about 200,000 runs over 3 weeks into work items showed only 17% of runs pursued more than one goal, but they took 58% of spend. Ramp has not published an ROI number.

The piece cites METR 2025 (16 programmers took 19% longer while believing they were 20% faster; a small sample), Gartner expecting over 40% of agentic projects cancelled by end 2027, and a board-sentence template: what we spent, the accepted-result cost against the old way, and which metric moved. Free in full.

Three proposals for what replaces line-by-line human review of agent PRs

'Human PR review is next' argues review shifts to requirements, acceptance testing and release ownership, with independent review models and guardrails; no numbers. 'Gittifacts' proposes humans approve intent while machines prove conformance, routing changes to auto-merge, quick or full review by test evidence and policy. Evidence generation took 13-17 minutes per change, and auto-merge still relies on the submitting agent's own test environment, which is the gap to watch.

A third piece says to measure agent work by the human cost of getting an accepted result (briefing, supervising, reviewing, correcting, verifying, plus agent spend), keep failed attempts and manual rescues in the data, and track later defects. All three are opinion pieces summarised by TLDR blurbs.

Put a name on every project: one hands-on IC as the directly responsible individual, with a weekly written update

James Stanier says flatter orgs and AI-aided solo ownership make named ownership the simplest fix: one IC per project, not the team or the manager. Precedents are Apple's DRI, Amazon's single-threaded leader and Stripe's 'who wants to DRI that?'. At Nordhealth about 30 projects each have a DRI who posts a weekly Linear update; leadership reads all of them in 30-60 minutes, and TLDR Founders adds that this surfaces 4-6 needing attention. There is a Wednesday office hour with leadership.

He pairs it with blameless post-mortems so DRIs keep making bold calls. Free, full text.

LinearB data: AI PRs merge at roughly a third the rate of human ones and wait 5.25x longer for review

LinearB's 2026 Benchmarks, as quoted, have agentic AI PRs waiting 5.25x longer for pickup (17.6 hours against 3.4 at p75, which works out nearer 5.2x than the stated 5.25x) and AI-generated PRs merging at 32.7% against 84.4% for human-written ones; the other episode rounds this to 33% against 85% across 8.1M+ PRs. LinearB's own rule is a spec for every story above three points: a project estimated at six months without AI shipped in about four weeks, at least two of them spent on specs.

Both episodes are heavily promotional for LinearB and TypeSafe products and a workshop, so the benchmark numbers are the substance.

Code is no longer scarce; build an 'AI harness' around scarce human attention

Guest Adrian Gonzalez Sanchez argues the constraint is human attention for specifying, verifying and deciding, and proposes five parts: durable driving artifacts (specs, executable tests), machine-checkable guardrails, engineering memory (decision records), attention gates (a named reviewer when work starts, a WIP limit on review, confidence thresholds so agents emit nothing rather than plausible-wrong output) and delegation rules (reversible, contained and clear means the agent decides).

Start by measuring time to first review, review duration and rework rate. It cites about 90% daily AI use (DORA 2025) and METR's 19% slower result. The email opens with a webinar promo and the author's book and course are plugged at the end.

Companies are building 'internal Lovables' for non-engineers: Ramp's Glass and Stripe's Harbor

Ramp's four-person team built Glass, which is part of new-hire onboarding and is now launched externally; Stripe's Harbor was built by one engineer in about six weeks. Reasons to build rather than buy: a tiny investment, safe exposure of internal data, an enforced internal design library, high non-engineer adoption, no vendor wait and in-house AI infrastructure skills. TLDR Dev adds higher adoption than vendor products.

Free excerpt of one of four Pulse topics; the rest of the issue is headline-only.

Unhealthy execution is a system problem: too much WIP, constant priority changes and deadlines nobody believes

LT#356's lead pick says the fix is to reduce WIP, control how work enters the queue, make priorities explicit and give estimates that mean something. Others in the two issues: Jason Cohen on OKR-style frameworks being unfalsifiable because every failure is blamed on execution (at WP Engine, letting teams pick one metric they cared about worked better); Maarten Dalmijn on refinement meetings mixing four abstraction levels; a PayPal C++ class that grew past 5,000 methods and took two years to untangle.

Blurbs only, and the newsletter's links are subscriber-tracking redirects, so none is linked here.

The senior-to-architect gap is deciding across team boundaries; an architect's job is to push scope, timeline and people before quality

The author scores a typical senior developer 1-5: architecture 1, data management 1, modeling 1, communication 1, facilitation 1, against DevOps/observability 4 and coding 4. Those scores are his opinion. The gap is moving from decisions within one team to decisions across teams; practise by tracing one design choice across team boundaries.

A week earlier the same author says that when scope, deadline and headcount are fixed only quality moves, so the architect should educate managers on trade-offs, have the engineers who write the code make the estimate, and surface the two or three important problems buried in complaints.

Knowledge capture compounds, and good engineering is an exercise in human relationships

Teams lose decisions, the 'why' behind systems; capture is now cheap because meeting notes, ADRs and product decisions can be summarised and reused by agents, turning documentation from a tax into an asset. The second half is a two-year-old Kent Beck interview: XP practices only work if people communicate and give feedback, and waterfall appeals partly because it replaces grassroots communication with a plan passed down a hierarchy.

Luca Rossi's readings are linked only through tracking redirects, so they are not linked here.

AI has killed Amazon's writing culture: a six-pager read for the first time two minutes into the meeting

The anecdote: an L7 principal PM presented a six-pager he clearly read for the first time two minutes in, and nobody spoke up. Bezos said a six-page narrative guarantees clear thinking; with AI that guarantee is gone, and a senior engineer now tags documents 'AI;DR'. Proposed rule: use AI for two-way-door (reversible) decisions, have humans write and read for one-way doors, and own every word; some documents stay human-written.

The author left Amazon in 2024, so the Amazon claims are secondhand by his own admission.

The feature factory is a spectrum, not a binary, and outcome language can be a proxy too

Cutler is 'tired of outcomes vs outputs, projects vs products' binaries. 'Projects to products' led ZIRP-era companies to over-hire PMs and force premature convergence; features are a proxy teams use under low trust or many dependencies, and OKRs and metrics are proxies as well ('performative outcome-centricity'). Sometimes shipping a feature list is right.

What holds teams back is access to customers and data, autonomy, training, promotion structure and money; AI amplifies both good and bad habits. No quantitative claims. Full essay.

Two Sean Goedecke essays: the human-plus-AI 'centaur' phase may last decades, and engineers don't get to be Luddites

Software's centaur age may last decades: as in chess, human plus AI beats either alone, so there is no need to panic or jump ship early (Refactoring's Luca Rossi doubts the timeline but likes the advice). The Luddite essay says engineers helped transform or destroy other jobs, so it is only fair their own is transformed.

Both are blurb-level in TLDR; the arguments are opinion without data.

Don't read AI-written code, constrain it with tests and QA, but liability law fits badly when nobody wrote the code

The essay says releasing AI-generated software without human review is feasible only with extreme constraints and automated testing (a self-driving analogy), and that liability law, which targets human developers, fits badly when no one wrote the code. TLDR AI attributes it to Robert C. Martin while the link is to blog.ploeh.dk, so the author is not confirmed here.

Andrew Ng: LLMs are far stronger on verifiable tasks than on judgement tasks, and people overgeneralise

Reinforcement-learning gyms and automatic reward checks drive progress on verifiable work like code and math; judgement tasks (customer complaints, product features, strategy) have better and worse answers but no ground truth, so labels and rewards are harder. He expects humans to keep a context advantage on judgement for a long time and recalls overgeneralising from Deep Blue, Watson and AlphaGo. No numbers.

The email has no clean web link (tracking redirects only).

Two long reads on why change gets expensive: incompatible local models, and domains where feedback is scarce

'Decoherence' (48 minutes) defines it as the accumulation of locally valid but mutually incompatible models, so the cost of change becomes dominated by translating between them and grows with the rate of change. 'The expensive-loop problem' says high-value human work is in expensive-loop domains (real-world constraints and costs) that need theory building, proxy management and diagnosis under scarce feedback, which cheap-loop training does not teach.

TLDR blurbs only; neither is checked against the full text.

Knowing the shape a program should have before you write it is what AI can't do for you

Judging clean from messy code is becoming tool-assisted; deciding what shape a program should have first is not. The suggestion is that the best way to improve at AI coding may be more thinking about problems and less about prompts, and that the people who benefit most already knew what they wanted. Blurb only.

Agentic verification: quality comes from loops the agent can run itself

Verification, not code-writing, is the bottleneck. Osmani prefers end-to-end tests of a small set of high-value flows as ground truth, property-based tests (fast-check, Hypothesis, proptest) for what must never happen, and differential tests that use the old system as an oracle for migrations. Tests must be fast and deterministic or the agent learns to retry rather than fix; isolate flaky ones.

Write the check before the code, and have humans review evidence, not just diffs. No hard numbers. Free.

Are harnesses worth the cost? One experiment says a harness lifted a cheaper model but did not help the stronger one

In ten runs on one Java API service, a behavioral harness raised GPT-6 Luna from 3 to 10 successes, while GPT-6 Sol passed 10 of 10 with or without it; harnessed Luna was about 33% cheaper per run than unharnessed Sol. The author expected the harness to help every model and says the results contradicted that.

Limits: harness maintenance cost is excluded and an unusually detailed brief limits how far it generalises. Also in Pointer #760.

Design an MCP server around jobs, not endpoints

Auto-wrapping every REST endpoint as a tool makes agents burn tokens chaining calls. Notion's first MCP mapped endpoints one-to-one and Notion itself said it created suboptimal experiences for agents; Block's first Linear MCP had 30+ tools and a simple question took four to six calls. Tool names and descriptions are runtime prompts, errors should tell the agent what to try next, and the checklist is under 7 workflows, actionable errors and testing on three hosts.

It is partnered with Manufact, which is disclosed, with a plug for its MCP cloud near the end.

Spec-driven development: spend the time on the spec, because fast code from unclear requirements is waste

Two workflows: OpenAI Codex tech lead Michael Bolin writes the spec in Notion, gets a colleague review and builds with a simple prompt (a higher-tier model writes the spec, a lower-tier one builds); Larridin's CTO uses seven steps from a spike to a smallest unambiguous spec, then test plan, implementation plan, implement, verify.

Paywalled after step 3, and it is an adapted chapter of his paid book.

Datadog claims its Scout SRE model matches frontier models at 3-8x lower estimated inference cost

Scout 1.0 is trained with reinforcement learning on simulations from historical incidents and is said to match GPT-6.1 Sol and Opus 5.5 at 3-8x lower estimated inference cost. Datadog's service-criticality post says an LLM analysed several hundred SEV-1/2 postmortems for about $32 to derive four signals, and the resulting tiers drive deploy and auto-merge decisions.

Both are vendor-authored.

TanStack Start hit 1.0, with three companies reporting big gains after leaving Next.js

Reported figures: Inngest local page loads from 10-12 seconds to 2-3, Railway builds from over 10 minutes to under 2, and Lovable 49% lower median TTFB, and Lovable made it its default framework. All are self-reported company figures, and the URLs were decoded from Fireship's tracking links.

Deno joins Cloudflare: runtime development ends after a year of fixes, Deno Deploy shuts in six months, JSR continues

Per TLDR Dev, runtime development ends after one year of fixes and Deno Deploy shuts down in six months, while the JSR registry continues. Fireship ran a eulogy in a joking tone. If you have Deno runtime or Deploy dependencies, that is the planning window.

Self-managed GitLab 19.2-19.4 has an unauthenticated code-execution flaw; OpenCode disclosed a breach

A crafted CI/CD regular expression allows unauthenticated code execution on self-managed GitLab CE/EE 19.2, 19.3 and 19.4, fixed in 19.2.8, 19.3.5 and 19.4.2 (advisory 12 October; GitLab.com and Dedicated already patched). OpenCode disclosed a breach affecting 1.4% of its Console users, those with external providers connected, with an update on 17 October.

Also noted: LeaderWorkerSet 1.0 changes the DisaggregatedSet default to one unavailable and no surge replica, so pin rollout settings first.

WebMCP: letting sites expose JavaScript tools to agents beat browser automation in a Shopify checkout test

In a Shopify checkout experiment WebMCP was 2.7x faster and 58% cheaper, with 60 of 60 successes against 56 for browser automation. The idea is explicit tool schemas on the page instead of agents parsing the DOM or screenshots; imperative tools are needed alongside declarative HTML for multi-step tasks. One experiment, from a blog summary.

Uber's Genie answers on-call questions with RAG over internal docs instead of fine-tuning

Uber's platform teams field about 45K support questions a month in Slack, with answers scattered across Engwiki and an internal Stack Overflow. Genie chunks that content, embeds it with an OpenAI embedding model, serves from the Sia vector database and must cite source URLs; buttons offer follow-up, resolved or escalate, and it mutes once a human takes over. After a year it covered 154 Slack channels and answered more than 70K questions.

The article reports a 48.9% helpfulness rate and about 13,000 engineering hours saved, and equates helpfulness with resolution, which is ambiguous. Full article.

Why HTTP/2 still stalls on TCP packet loss, and why HTTP/3 and QUIC exist

Part 1 covers HTTP/1.x and connection setup: browsers typically open 6 parallel connections, TCP slow start begins at about 10 packets (roughly 14 KB) and doubles each round trip. Part 2 covers HTTP/2: Server Push was disabled in Chrome and removed in Firefox, Early Hints are used on only 2.9% of desktop pages, and TCP head-of-line blocking can make HTTP/2 on one connection sometimes worse than HTTP/1.1 on six on lossy networks.

Part 2 is a paywalled preview that stops at QUIC, so the QUIC, QPACK and connection-migration content is missing. Part 1 is mostly complete and free.

Scaling an AI chat app to 1M users is mostly an inference-cost problem

On illustrative assumptions (8 LLM requests per conversation, 1,424 input and 300 output tokens each, GPT-5.4 at $2.50 / $15 per million), a conversation costs about $0.064, so the bill is about $193 a month at 100 daily users, $193,440 at 100,000 and $1.93M at 1M. Routing 70% of requests to GPT-5.4-mini cuts it about 49% to about $98,635; adding semantic caching (30% hit rate) and prompt trimming reaches about $48,000, a 75% cut.

Paywalled preview: the self-hosting claims are teased only. Guest-authored, and the numbers are the author's assumptions, not measurements.

Ten scaling bottlenecks, walked through a hypothetical task app: missing index, connection exhaustion, CPU

The scenario is invented: Postgres, 3 app servers and Redis under $2,000 a month. An unindexed project_id on an 8-million-row tasks table takes task-list latency from 80 ms to 4 seconds; the fix is a composite index built concurrently. Connection exhaustion (max_connections 200 against 225 requested) is fixed with PgBouncer in transaction mode, 1,000 client connections onto 50, and CPU at 90-95% from Markdown rendering by caching rendered output.

Paywalled preview showing 4 of 10 bottlenecks; the numbers come from a made-up scenario.

Resilient AI-powered software: RPO and RTO of zero are impossible, and failover capacity may not exist when memory is scarce

Disaster-recovery patterns rise in cost from backup-and-restore through pilot light and warm failover to multisite active-active. GPU, CPU and memory shortages mean on-demand failover capacity may not exist, so reserve it, prioritise critical services or run redundant infrastructure. For AI features, plan for the model being down with fallback vendors, evals and dynamic timeouts, and make spend visible per person (James Stanier: a QA engineer burned over 1,000 euros of Claude Code in a week and more than doubled productivity).

The issue includes unsubstantiated claims, flagged in its own text. Book excerpts; some pattern sections are cut off.

Networking fundamentals for software engineers, part one of ten concepts

Figures in the body: a 10-byte payload over TCP/IPv4/Ethernet carries 58 bytes of overhead; a /24 gives 256 addresses (254 usable), a /16 gives 65,536; Linux ephemeral ports run 32768-60999. It cites Facebook's 2021 BGP withdrawal, which left its services unreachable for about six hours, and stresses that NAT is not a firewall. Free; part 2 is to come.

Three Postgres items: zero-downtime migration steps, a UUIDv7 benchmark, and a silent insert ceiling

The migration piece uses expand, backfill, switch, cleanup, short lock timeouts and concurrent indexes. The UUID benchmark (PG18 on a laptop, 10M rows) put UUIDv7 at 32.3 seconds, bigint at 29.7 and UUIDv4 at 60.7. EDB found a silent insert ceiling from relation-extension locks and measured 7.2x throughput from hash partitioning in HammerDB.

The UUID run is a laptop benchmark and the EDB result is a vendor benchmark.

ClickHouse chased a 32 ms tail latency: idle connections reset the TCP congestion window

On high-RTT networks, simple SELECTs after an idle period took 32 ms longer because idle resets TCP's congestion window and unbatched metadata packets exhausted it. Application-level batching halved the TCP segment count and removed the penalty. 14-minute read; vendor engineering blog.

Rippling's outbound engine only pays because of multi-product expansion

Swapping sequencing software for SDRs working the same list lifted conversion from about 1% to 3-7%; Rippling books about 1,300 meetings a month, roughly half arranged by phone, and outbound drives more than half of revenue (per its CRO). The author estimates $35-45M a year all-in SDR cost against about $50M of new ARR, a payback near a year that is only mediocre on its own: it works because the product line (20+ products, a $45K land) expands per customer.

A single-product seller copying it would burn about 18 months of runway on the author's math. Several figures are the author's estimates. Free, full essay.

G2: 91% of B2B buyers find per-seat AI pricing unfair, and credit-based pricing leaves buyers unable to tell what they are buying

G2's 2026 AI Agents Report says 91% of B2B buyers find per-seat pricing unfair, 57% prefer usage-based pricing, 74% are open to switching at renewal and 46% say agent costs came in higher than expected, with buyers wanting spend caps. It is a vendor survey, and Stripe's 'two-thirds of the Forbes AI 50 use usage-based pricing' comes from Stripe's own blog.

Elena Verna's piece adds that Lovable's $25 a month buys 100 credits that buyers cannot translate, and the same outcome can cost very different amounts because agents retry; her fix is to quote in recognisable units such as minutes or pages.

PostHog's 38 lessons from ten AI products: build for the agent first, and context beats the harness

PostHog says five times more people use its MCP server daily than PostHog AI; its install wizard turns two hours of work into about 8 minutes, converts users to paid 5x more often and activates them 2x faster, at $6.67 a run. One AI bet got only 23% 'very disappointed' in the PMF survey. The harness is no moat (PostHog AI went through three in a year); giving agents more context beat better reasoning, subagents or pruning, and a general executeSql tool turned a four-call question into one.

Company content marketing, but the advice is concrete. Also in Pointer #759 and TLDR.

Why lean AI-native startups beat traditional SaaS, per Stripe data cited by the author

The post cites Stripe data that AI-native startups reach $30M annualised revenue in about 20 months against 60+ for SaaS peers, at roughly 6x revenue per employee. The moat shifts from features to accumulated operational context, and a $50K monthly API bill that enables $2M of output is not overhead. Menlo is cited for AI-native startups taking 63% of enterprise AI app-layer spend in 2025, up from 36%.

The proposed 'Unit Economics of Intelligence' metric is explicitly non-standard. Free, full essay.

How VCs actually make money, and why they need huge exits

A fund earns about a 2% management fee and 20% carry: a $300M return on a $100M fund gives the VC $40M. Correlation Ventures' 21,640 US deals had 65% returning less than cost and 4% returning 10x or more; Kauffman's review found 69 of 88 funds failed to beat the stock market by enough to justify the risk and the wait. The founder rule of thumb is to divide fund size by their stake: a $100M fund owning 10% needs a $1B exit.

Ends with a paid spreadsheet upsell, though the article is complete.

Lightsprint raised $6.1M to sell 'agreement', not coding speed: non-technical staff describe a change and engineering reviews the PR

Lightsprint (YC Spring 2026, led by 645 Ventures) lets product, design or business staff describe a change in plain language; agents write the code and it lands as a PR that engineering reviews. The thesis is that agents write code faster than teams can review it. Claimed traction: 3x more code shipped per team, about 80% of merged code written by agents, and 1,000+ non-technical PRs merged.

Paywalled preview, written by an investor in Lightsprint who discloses the conflict.

Higgsfield says it went from $1M to $1B annualised revenue in 18 months; all figures are the founder's claims

From a 20VC interview: ARR is four weeks of live revenue times 13, with multi-year enterprise deals excluded. Finance projects $4.5B in 12 months while the founder says over $10B. He claims over 300% month-12 net revenue retention in the business segment and about 30% of users dropping in month one; the writer wants cohort data before trusting the 300%. Margins are over 80% on open-weight models and 20-30% on closed ones.

Support and legal were not eliminated: legal is over 10 people and customer success over 40. Free in full.

The 2026 fundraising market is concentrated: AI took 82.7% of US venture dollars through September

PitchBook-NVCA has AI at 82.7% of US venture dollars through September; US rounds of $100M+ were about 5% of Q3 deals but 75% of dollars. Carta's benchmarks for 1,000+ software rounds: median seed $4.1M at a $24.3M valuation with 18% dilution, median Series A $14.4M at $80M. Only 211 emerging VC firms closed a new fund in the first nine months of 2026, down from 927 in 2022.

A skeptic's note says AI's share of US dollars fell from 88.4% in Q1 to 65.9% in Q3. Partial paywall: 10 of 105 links are free.

App Store 15% fee versus web checkout: a $9.99 subscription nets $8.49 in-app against about $9.33 on the web

Losing about 9% of buyers erases the gain, and an Adapty example was rolled back after a 9% fall in revenue per paywall viewer. Useful arithmetic for any client weighing web checkout. Blurb only.

NN/g: synthetic respondents reach the same conclusion as humans but converge on far fewer answers

In NN/g's comparison, 120 synthetic respondents and 604 humans reached the same conclusion on UX terminology, but the synthetic ones converged on about a dozen labels against 50+ for humans, a mode-collapse effect. Fine for hypotheses, not for minority views.

NN/g's 'new big ball of mud': non-technical 'vibe architects' grow fragile agentic systems

Participants kept throwaway drafts, bolted on features and let AI patch files until they could not locate or recreate what they had built. The suggested fix is separating global, local and ambient context. 12-minute read.

Low-contrast text on 83.9% of top-million home pages, and a pure-CSS icon system

Clearleft's 'This hue shall pass' fixes WCAG AA/AAA failures by nudging lightness with hue fixed. Lea Verou's 18-minute post builds an icon system from --icon custom properties, @property, style queries and mask, with no HTML changes.

Make room for hard things: the hunger for hard problems comes from a simple life, not a trait

Regular runners (4+ hours a week) are unusually affluent and educated though running is cheap, which Luca reads as Maslow: covered basic needs free mental space for harder challenges. Energy is a jar: sand (chores, fake urgencies, many tabs) crowds out the rocks of stretch work, so spend at least half your time on stretch work.

The same issue has a Rands interview about delivering hard feedback with a personal story instead of 'pull it together'. Author's own essay.

Judge a decision by the quality of its reasoning, not its result

The essay uses Annie Duke's 'resulting' idea: Bezos's regret-minimization framework would have been identical had Amazon failed, and Amazon traded at about $106 in December 1999, under $6 in September 2001 and did not regain the 1999 level until October 2009. Habits: ask whether a door is one-way or two-way, and write down your assumptions on the day.

It notes regret minimisation ignores what you can afford to lose. Complete essay.