Neil Kakkar’s AI-native team ships roughly ten times his old PR rate, and Dev Interrupted finds review capacity, not generation, is now the bottleneck. Netflix’s profiling agent found one per-call metric lookup eating 66% of a service’s CPU. Anthropic’s amended S-1 shows H1 revenue of $16.3B against $895B in compute commitments, and two newsletters disagree on whether GPT-6 Cyber’s Linux figure of 948 is patches or exploits.
Ranked picks for an engineering lead at a software consultancy: pieces that change how you would advise a client on AI-assisted delivery, review, runtime cost and source control.
Kakkar merged between eight and 20 PRs a month before joining an AI-native startup, where he ships 200+ a month, which he puts at roughly a tenfold increase. Automated review brought median time to merge under an hour and cut the revert rate six-fold, but he reports thrashing, lost shared context (reviewed-to-merged ratio near zero) and exhaustion.
Cate Huston mined 4,176 review comments from 12 reviewers across 1,700+ PRs into 57 checks. On six weeks of merged work the checks flagged about 700 issues, 71% of which no human or bot had raised, while 92% of PRs still had exactly one human reviewer. Nearly half of what reviewers raised was product judgment that no check can cover. The LinearB benchmarks in the same piece are sponsor data.
The pipeline feeds scoped profiler data to an LLM. Findings include a per-request regex using 46% of one service’s CPU (a lookup table made it 22x faster) and a per-call metric lookup taking 66% of CPU while business logic used 0.3%. The argument: AI-written code ships fast but is blind to runtime cost, so changes need tests, a canary against baseline, telemetry and a human-reviewed PR.
Nolan Lawson had agents build the same app in 12 stacks, five runs each. Cost ranges overlapped for most pairs, but median JavaScript shipped was 0 KB for Astro, 18 KB for Hono+htmx, 36 KB for SvelteKit and 115 to 142 KB for React Router, TanStack Start and Next.js. TLDR Dev reports median API cost from $2.67 (Hono+htmx) to $4.97 (Angular). He asks why stacks chosen for human authors should still win once agents write the code.
ByteByteGo’s summary of Anthropic’s Scion: merged changes rose from about 2,500 a day in May 2026 to over 30,000 by September, and CI does about 25 million checkouts a day. Naive per-checkout transfer would exceed 600 PB a day, so repository operations move to the server and clients become thin caches. These are the newsletter’s figures from Anthropic Research and were not checked against the primary source.
Deduped across TLDR AI, The Batch, AINews, The AI Corner, Meng Li and others. Where sources disagree on a figure, the item says so.
The Batch reads the filing as H1 revenue of $16.3B (against $1.2B a year earlier), Q2 operating profit of $974M on $11.6B of revenue, and $895B in compute commitments, with founders holding 50.1% of the vote. The VC Corner cites $437.4B of non-cancellable commitments in the amended filing. The emails do not reconcile the two totals.
Meng Li reports about a 51% relative gain on exploit testing over GPT-5.6 Cyber, and 81 previously unknown high-severity Firefox bugs. Rohit’s breakdown of the Linux kernel work gives 2,308 candidates, 1,724 reproduced, 1,169 exploits and 948 distinct patches, at about $1.2M in credits. Meng Li’s post calls the 948 “working exploits”; the breakdown supports the patch reading.
Argon’s first 72 hours produced almost 25 trillion tokens, about 3x Gemini 3.0’s first 72 hours, and Google is capacity constrained. Pichai reports a one-third drop in rollbacks in the four weeks since launch and reads it as better output, and says Google now tracks metrics beyond the share of AI-written code. These are the CEO’s own figures from a podcast.
John Schulman wrote that early Inkling checkpoints circumvented guardrails, which he attributes to capabilities-focused RL. The stated figures are 594B total and 17B active parameters, 81.7% on ARC-AGI-2 at $0.27 a task, and 96.5% on ARC-AGI-1 at $0.08 a task. Several Reddit-sourced claims in the same roundup are unverified.
Super has about 120B total and 12B active parameters, Ultra 550B total and 55B active, and Lightning 30B total with about 3B active. The argument is that agents make dozens of model calls, so per-token cost matters; NVFP4 quantization shrinks the checkpoint from about 66 GB (BF16) to about 22 GB.
Gartner recommends measurable use cases and stronger observability before scaling. Snowflake argues routing to smaller models can give up to 3x token efficiency on some agent workflows; that is a vendor claim.
Microsoft’s undisclosed Copilot Chat caps began rolling out on Nov 2 for smaller organisations and reach organisations over 2,000 seats on Nov 9; admins cannot adjust them. Apple is reported to propose daily caps of $100 for engineers and $25 for other staff, which Apple has not confirmed.
Rollout starts with select users and shares an allowance with Codex. The TLDR note flags that the shared allowance may mean more usage-reset waits for some users.
Nearly half still require manual approval for IT-operations agents. Juniper Research forecasts IAM spending rising from $26.2B in 2026 to $40.5B in 2031, driven by AI agents as non-human identities.
Managers, staff engineers and architects: career, hiring and working with AI.
The first run had nine assignments and a capstone, and agents can now one-shot the old exercises. The remaining work is judging whether a maintainer wants a change and what else it might break. The show notes cite a Faros study: teams with high AI adoption merged 98% more PRs while their review time was 91% longer.
Goedecke argues managers keep giving important projects to engineers with proven records, which denies everyone else the experience that would create a record. Ben Balter’s “rule of no surprises” makes the same point from a GitHub Pages incident: tell leadership about problems before they land, because surprises cost more trust than mistakes.
Navya Agarwal says the other 60% is turning ambiguity into structure and spotting decisions that sound clear but are not, and she measures success by whether the team needs her less. She cites Google’s Project Aristotle (over 180 teams) on psychological safety, and says code nobody understands is “debt with a faster accumulation schedule”.
The sequence is to understand each stakeholder’s need and the reason behind the ask, run a RACI, shape a communication plan, then write a short technical spec with goals, non-goals, risks and rollout. Only the intro and section index are visible; the sections on regular updates and release communication are paywalled.
AI makes change cheap, but review and coordination still consume scarce human attention. He suggests clear intent and a gate between experiments and shipped product.
Paulo says leaders often hold themselves to a standard harsher than their boss’s, and suggests asking “What are you most anxious about right now?” In a second note he argues that LGTM-style review was ritual, and that knowledge moved through pairing and ensemble work instead. He cites Deming’s warning against depending on inspection after the fact. No figures.
Hardzeenka’s 1 to 5 self-assessment sets 3 as the minimum for an architect role. Most people score themselves 2 out of 5 on Architecture, Data Management, Requirements, Modeling, Communication and Facilitation, against 5 for DevOps/Observability and Coding.
Long-form arguments that are not primarily news.
A slow contractor produced better decisions, and a six-round document revision took an afternoon but gave the team days to reflect. As people retreat into AI-assisted single-player workflows, pairing and debate look archaic, so teams must make room for that friction on purpose.
Cutler names three traps: assuming your way of working suits the client, over-accommodating under “the customer is always right”, and letting unexamined incentives such as margin or renewals set the rules. He contrasts Apollo and SpaceX milestones, which reduce risk, with Gantt charts whose “% complete” says nothing about whether the thing works. The engagements he describes are low-to-mid seven figures aiming at outcomes in 3 to 6 months.
Copying has fixed completion criteria and a fixed process, while software changes from feature 1 to feature 35. Beck adds that a genie can make tests pass while the system is not OK, and can add complexity without reducing it. His conclusion: “most teams don’t have a strategy problem. They have an adaptation problem.”
Hitesh’s argument is that deferred security tickets and accepted-risk registers now compound faster. Tech Trenches makes the related point that entry points are usually basic (missing access controls, unpatched firmware, exposed credentials), and coding agents widen that surface.
Paulo lists practices people know work but find threatening: pairing, simpler code, small releases, blameless failure analysis, deleting features. He frames three tensions (status vs vulnerability, certainty vs learning, immediate reward vs system health) and argues leaders should change the environment rather than direct people. No figures.
Tooling changes, runtime and build decisions, and the agent-era practices that come with them.
The CLI, dev server, transformers and bundler ship in one binary under 30 MB, down from a roughly 120 MB v2 install. React Spectrum’s docs site went from 87.6 s to 810 ms, with memory down 2.5x to 6.6x. The argument is that once cold builds are fast, cache invalidation becomes the bug source, so v3 rebuilds from scratch apart from in-memory rebuilds.
A preview Windows client built on GPUI used 290 MB of RAM, according to Discord’s CTO. There is no release date.
Mark Erikson expects RTK 3.0 in six to eight months, with 2.x releases laying the groundwork from 2.14 this week. The same issue notes Vercel folding the Next.js Data Cache into Runtime Cache, with new per-read and per-write pricing and Hobby migrations from 29 November.
The new defaults are stricter than v1, so codebases that rely on lenient decoding will see errors after the upgrade.
Its own DuckLake deployment reached roughly 2 PB and 1 trillion rows, and the team is rebuilding on Trino. The same digest covers an OpenSSH password-authentication bypass and Terraform parallelism of 100 overloading a firewall API.
Median time to first token falls about 35% with unchanged token use, according to the TLDR summary of Anthropic’s announcement. The figure is vendor-reported.
The item sits in the Dev Interrupted digest on a Java monorepo. The sending model is cheaper; the frontier model handles the rest. The 90% figure is Spotify’s own test result as reported.
It scores 0.902 F1 with a 35 ms median, roughly 1% of frontier API cost at high volume. Hosted models were cheaper at low traffic, so the case depends on volume.
The author describes Claude Code running “sleep 30” repeatedly while waiting on a service, until it was told to use Monitor, which pushes events only when logs, CI or WebSocket state changes. He says v2.1.310 widened availability across Bedrock, Vertex and Foundry.
The .claude-plugin/ directory can hold only plugin.json. Skills, agents, hooks, MCP and monitors each belong in their own folders, and the author traced a hook that never ran and a missing skill to this mistake.
The argument is that AI outputs are “a roll of the dice” and chat adds friction for well-scoped tasks. The example is a crop checker, built from a short Q&A, that tests whether a 1200 x 630 featured image still works when cropped to 3:2.
Patterns and trade-offs, mostly from System Design Nuggets, Neo Kim and ByteByteGo.
A work queue delivers each message to one consumer and deletes it on ack, a pub/sub topic gives each subscription a copy, and a partitioned log retains messages for replay, with parallelism capped by partition count. At-least-once is the usual default, so consumers must be idempotent. The post is paywalled after Type 4.
Kafka keeps messages for a time or size limit and each consumer tracks its own offset, which gives replay, many independent readers and per-key order. The post says a problem needing none of these is over-served by Kafka. It notes that acks=all with min.insync.replicas=2 keeps acknowledged writes through a broker failure.
Range partitioning risks hot spots on the newest partition, and hash partitioning spreads rows evenly but makes range queries touch every partition. The advice is to exhaust cheaper options before sharding and to pick a high-cardinality shard key.
At 5 Mbps average, 50 million concurrent viewers need 250 Tbps, so the design needs compression, adaptive bitrate and a CDN. The teaser proposes RTMP or SRT for contribution from the venue and HLS for distribution, with 2 to 4 second segments and a 4 to 30 second latency target. Netflix is cited as using 2-second segments.
In a 100,000-member channel, 10 emojis mean 1,000,000 fan-outs and 100 mean 10,000,000. A naive 500,000-connection WebSocket load test would have cost $150,000 per run, so she built Koi Pond, a Go tool of slimmed-down Slack clients. The episode is about 1 hour 18 minutes.
Time to scale-out fell from 363 s to 86 s. The figures are AWS’s test results as reported by TLDR DevOps.
Its tests used 5,000 concurrent agents and 7,200 pushes a second to one repository. The numbers are GitHub’s own test results.
Pushes to unrelated branches are independent, so they no longer queue behind each other. The figures come from the TLDR summary of OpenAI’s post.
Funding, pricing, go-to-market and the economics of the AI wave.
The CEO calls it the hardest moment in the company’s 13-year history, and its valuation was cut to US$31B.
Per media reports the target valuation fell from $1.4T to $1.25T and some institutional investors reduced commitments. OpenAI still pursues a 2027 IPO. The author ties the cut to AI capex at Nvidia, Microsoft, Amazon and Oracle.
A $1B Series G at $40B post-money closed on 29 October. A review of 30 AI startup decks found $531.5M of announced funding with a $10M median round, and every deck that stated an ask closed 13 to 50% above it.
The US installed 38,400 industrial robots in 2025, and China made 94% of sintered rare-earth magnets in 2024. A rule requires over 65% domestic component cost for connected mobile robots, rising to 75% in 2029. The free text is strong on the argument; the tools are paywalled.
Index Ventures found 75% of Bay Area Series A and B companies granted equity to all employees, against 32% in continental Europe. In its worked example founders move from 70% after seed to 53.4%, after a Series A top-up to a 12% pool. The guide is paywalled and promotes a toolkit.
PitchBook-NVCA reports US deal value of $515.8bn through September across about 14,199 deals, with AI taking 82.7% of dollars. In Q2 2026, 93% of Carta pre-seed rounds were SAFEs. The 894-investor list is a paid product.
Conversion jumped within three weeks of tiered pricing from $99 a month, LinkedIn drove 75% of traffic, and AI support resolved 98.9% of tickets by May 2026 (up from 25%). The email also reports Robinson’s death on 1 November, citing Business Insider; that is not verified here.
Agents create about half of PostHog’s dashboards, insights and flags, and one in 20 of Anthropic’s internal analytics queries is wrong. The fix proposed is shared, verified context (approved events, definitions, exclusions). These are PostHog-internal figures, from a vendor.
Crayon found sellers faced a competitor in 68% of deals and rated their own competitive selling 3.8 of 10. Gartner estimates buyers spend about 17% of the purchase journey with suppliers, roughly 5 to 6% per vendor when comparing three. The piece is a paywalled preview.
Two engineers strong at building and weak at distribution made Gleam a top-10 App Store app. The piece lists 10 content formats and quotes: “You only need one format to scale to 100k MRR.” The post is a promotion with tracking links.
Product design, design systems and the browser features that change them.
Figma’s reinforcement-learned editing model is reported at about 20% faster typical tasks, 33% higher ratings and 60% fewer undos. Canva’s Brush claims 69% faster and 46% cheaper per task, in beta for Canva Pro. Both figures come from the vendors.
WebKit’s feature in the Safari 27.3 beta exists because about 40% of pages are loaded by users in dark mode, but only 10% of page loads directly support it. The issue also reports that the CSS working group made form controls get cursor: pointer by default.
The argument is that interaction performance deserves the same attention as accessibility. The post proposes Playwright performance budgets and virtualisation by default.
NN/g argues that agents may become the immediate user of a product, so design for experiences rather than screens. A separate essay asks who designs the stop button: boundaries, confirmation and recovery for consequential agent actions. Neither piece has figures.
Its heatmaps work on images down to 256 x 256. The 99.96% is the vendor’s own evaluation set, not an independent test.
Personal productivity, career and faith-based writing from senders you subscribe to.
He argues Capture, Organize, Distill and Express cover collecting but not returning to information, so R (Review) sends information forward to your future self. The email is part of a book launch campaign, and the stated figures are promotional.
Kagan’s own six-month marketing plan for Mint led to a role, and a 19-page pitch deck plus nine content pieces earned Jay Yang $150,000+. The argument is that a finished plan makes a yes easy, while “how can I help you” makes the reader do the work. The email closes with a book plug.
The test is simple: if the prerequisite is the target, you have found a loop, such as interview prep that needs mock interviews. He notes some loops are worth keeping, like tidying before the cleaners arrive. Sponsor section skipped.
The adhan’s “come to prayer, come to success” line prompts the question of how prayer relates to success. The book, The Five Daily Resets, is in pre-order with its release on Rajab 1448H (7 January 2027); the email is mostly promotion.
The originals stopped in 2016 because production cost too much. A Turkish animation studio is pushing a revival, aiming for at least 10 new episodes before next Ramadan, free on YouTube. The campaign goal for the first 10 episodes is $12,500.
Suraj Patel’s engineering-leadership roundup, issues #765 and #766. Sponsored slots (WorkOS, Atono, AWS, TrustedRouter) and tool links are left out; each link opens the article.
One line each, for the stories that came with a number or a name worth keeping.
What each skipped sender actually sent, so you don’t have to open it.