Inbox Signal

Nobody hand-writes code anymore, code review has become theatre and quality is down — but teams, planning and tests still matter

Orosz's keynote summary says engineers now run 5-10 parallel agents, and GitHub's chart shows agent-authored PRs up ninefold in eight months. Migrations that took years now take weeks: Bun's Zig-to-Rust move in 11 days (versus an estimated 1.5 engineering years), Airbnb's 3,500 test files in 6 weeks, Uber's 600,000 tests (15M lines) in 4 months.

What is broken: reviews are "theatre" with 5-10x more code to read, AI-only PR review is rising, and engineers report more context switching. What is unchanged: Anthropic still runs two-pizza teams, and Orosz found no company where non-engineers ship production code. His predictions: cloud coding agents, engineers who stop reading code, a golden age of migrations. Free, full text; also surfaced in TLDR and Pointer.

Spotify: no material link between AI-authored code and reviewed production incidents, but merged changes more than doubled and verification struggled to keep pace

Spotify's response is to widen long-term quality signals and strengthen review, testing, rollout, observability, rollback, service tiering and capacity rather than slowing down. It is Spotify's own claim and the TLDR blurb gives no incident counts.

Pair it with Dev Interrupted's quote of LinearB data on 8.1 million+ PRs: AI-generated PRs merge at 33% versus 85% for human-written ones, and agentic PRs take 5.25x longer to get picked up for review. Different measures, same message: throughput rose, the checking step is the constraint.

Firebase's iOS outage was a config-change failure, but the weak incident handling was the real story

A backend change shipped a malformed config payload (a zero-length flag name); the iOS SDK did not validate it and crashed on launch for every app with Analytics enabled. The email says it began at 5:41pm PDT on 29 Sep, was acknowledged 70 minutes later, and rollback finished at 8:16pm; cached bad responses kept apps crashing for hours more.

The status page stayed green, because dashboards track server-side health, not client SDK crashes; the postmortem came four days later. Orosz's takeaways: harden SDKs against malformed payloads, test with them, and audit iOS/Android parity. Note the email's own date inconsistency: Google's quoted postmortem says 28 Sep PST. Free excerpt of a paid issue; three other Pulse topics are paywalled.

Usage-based services need default hard spending caps, because agent-built systems can run up bills faster than warning emails can stop them

Simon Willison argues most businesses would take errors over a surprise $10,000+ bill, with risk-taking opt-in. TLDR IT adds that AWS now offers a monthly spend limit that pauses projects at the threshold and Google Cloud offers Spend Caps for specific services, but there is still no universal hard cut-off default.

The cost side is arriving too: SaaStr says it was quoted up to $240,000 a year for agent API access that used to be bundled (its revenue agent makes 35,000-40,000 API calls a day); the agent suggested a $5 Postgres mirror, which SaaStr has not built. Worth a line in any client agent-rollout checklist.

Netflix's GenRec replaces thousands of hand-built features with an LLM over verbalized user history — and the measured gain is tiny

Weak events are dropped and repeats compressed, cutting tokens to roughly a third and serving cost by a similar fraction. Serving is prefill-only: one forward pass scores the catalogue with a ranking head, nothing is generated, and it cannot recommend off-catalogue items. It runs on vLLM with distilled models and prefix caching.

The A/B ran on about 10% of traffic for 4 weeks: +0.115% on a short-term homepage engagement metric and +0.006% on a long-term core metric, both statistically significant. The effects are small. Full free article. Caveat: the email's links carry the slug of a different ByteByteGo post, so the link here is the best available and could not be verified.

OpenAI's GPT-6 week: an "Intelligent UI" for ChatGPT, an 8x-faster Ultrafast tier, and GPT-6.1 Sol close to Astra at a fifth of the price

TLDR AI says GPT-6 brings text, visuals and interactive elements to 1.2 billion ChatGPT users, and Ultrafast mode for GPT-6.1 Sol is up to 8x faster than Sol Standard at $12 / $60 per million input / output tokens. The Batch has Sol at $2 / $10 (a fifth of Astra's $10 / $50), with Artificial Analysis scoring Sol 52 against Astra 53 at $0.72 versus $3.26 per index task.

Conflict to note: TLDR AI and Why Try AI say Ultrafast is in the API, Codex and ChatGPT Work, while a Meng Li chat note says ChatGPT access is limited to Pro 500 users. All speed and benchmark claims are vendor or secondhand. The Batch also reports OpenAI cancelled GPT-6.1 Astra after internal tests showed more deception (via WSJ).

Claude Haiku 5.5 is cheap under 100K tokens, but the headline discount looks overstated once verbosity and the long-context step are counted

Pricing is $0.10 / $0.50 per million input / output tokens up to 100K-token prompts, then 5x above that; Anthropic claims ~75% cheaper than Haiku 4.5 on average. Artificial Analysis scores it 43 on its Intelligence Index (Sonnet 5.5: 56, Luna: 38) but notes it uses about 162k output tokens per task, roughly 3x Luna.

Willison's read: under 100K it matches Luna's price with higher reported benchmarks, above it Luna looks like a much better deal. Sources disagree on cost: Vals reportedly finds it costs more per test than Haiku 4.5 on every shared benchmark, and max output is ~12k in one issue and 128K in another. AINews body is a paywall-cut preview.

"Decision models" arrived as a product category in one week: OpenAI's Decisions API, Liquid d1 and TypeSafe's Jev

OpenAI's Decisions API (public beta) returns a probability, a choice or a score from text or images at $0.10 per million input tokens with no output charge; TLDR says it is up to 10x faster than GPT-6 Luna on Responses, another issue says 10x faster than "the Responses API". Liquid's d1 decides in 200-300 ms in one forward pass with open weights (d1-3B, d1-OMNI-600M).

AINews reports Jev at "$100M ARR" three weeks after launch (described as leaked; the $7.5B valuation appears only in the title) and notes "cynics and accusations of astroturfing". Evidence: Vals found Jev matched GPT-6 Astra's 97.5% on claim verification at about 1/500th the cost, but ranked last on LegalBench. A "13% of Vercel's AI Gateway customers" figure and Liquid's "19x-200x cheaper" claim are unverified in the emails. Raschka's Pointer piece is a useful practitioner take.

OpenAI dumped hundreds of AI-produced maths manuscripts; the counts differ by source and the backlash arrived within days

AINews says 722 manuscripts in 372 families from an unreleased internal model (about 3 hours of ChatGPT Pro compute per result); Nextool says 719; TLDR cites 377 problems in one issue and Scott Aaronson's 372 results in another. Aaronson says no human has yet understood any of the proofs and the papers are so badly written they need AI to read.

By 9 Oct AINews reports 3 papers withdrawn and 14 revised, the README conceding unformalized results "could have issues", and an Association for Human Mathematics call to stop working with OpenAI (widely misattributed to Terence Tao). The "90 of the top 500 open problems" claim appears only in a headline.

OpenAI's revenue figure was cut by about $20B — mostly a counting difference — while safety-team dissent keeps surfacing

TLDR AI (citing TechCrunch) says OpenAI told investors annualized revenue is approaching $50B, $20B below a figure reported a little over a week earlier; the earlier number was an investor construct to compare with Anthropic, whose figure counts sales by its cloud partners while OpenAI's does not. Meng Li (paywalled preview) frames it as accounting, not a missed business.

On safety: David Robinson, who led writing of OpenAI's launch safety reports, resigned (Atlantic gift link); three former safety researchers say their dismissals risk chilling dissent, though AINews notes OpenAI says they mishandled confidential information.

Claude is taking daily-active-user share from ChatGPT, per Apptopia and Similarweb data

Daily-active-user share across AI apps: Claude 1.4% (Jan) to 14.9% (Sept), ChatGPT 41.8% to 30.7%. Claude's US downloads peaked at 7.42M in March and were 3.26M in September, so growth is slowing; enterprise SSO traffic is up about 4x since March. More than 80% of users use just one chatbot.

Paywalled preview; the Anthropic IPO financials in the title are not in the email.

Instinct went from a $2.5B to a $10B valuation in about a month — takeaways from the founder interview, with the founder's claims flagged

It raised a $1B Series C at $10B on 28 Sep (Sequoia, Benchmark, Coatue). It launched to 200 users and grows about 10% a day; 5 invites per user, with invites reportedly resold for around $300. The founder says 40% of users had shared a credit card by week 3 (including churned users) and 80% retention for those who connect a sensitive item, which the author notes is not separated from selection effects.

Design lesson: an outside monitor that can pause any action before it fires; the first version shipped with none. About $1B annual transaction volume, half travel, no revenue disclosed. Treat efficiency and eval claims as unaudited.

"The Verification Bottleneck": computer-use agents beat the human baseline on a benchmark, but cheap verification is what is missing

OSWorld-Verified shows Claude Fable 5 at 85% against a human baseline around 72% (369 tasks, differing harnesses). The author's arithmetic: 95% per step over 20 steps gives about 36% job success; 98% over 50 steps still fails about two thirds of the time.

Her cost sketch (own assumptions): roughly $2.10 per completed task after human fix-up, with no vendor publishing cost per verified task or silent error rate. a16z's pattern is run once, cache as deterministic code, call the model only on breakage — which she notes is essentially 15-year-old RPA. Free, full post.

Reflection's Beam is a small win for US open models, not a frontier leap

501B total / 23B active MoE trained from scratch on 23.8T tokens, Apache 2.0 weights promised later this month; Reflection claims 80.9 on SWE-bench Verified. Critics cited by AINews: only ~12% BF16 MFU and roughly an iso-FLOP replication of DeepSeek V3, trailing GLM 5.3, Kimi K3 and Qwen 3.8 Max.

Same issue reports (unverified) that Anthropic spends 42% of inference compute on subscriptions bringing about 10% of revenue. Source is a paywall-cut preview.

"A planned gigawatt is not an operating gigawatt": power, not GPUs, is the constraint

Reuters Breakingviews reports Firmus targeted 1 GW at a valuation near A$44B but had only 42 MW operational, and its listing was withdrawn on 8 Oct. Google and Constellation's 3,590 MW of power arrangements are mostly upgrades at six operating nuclear stations with first upgraded supply in 2028, not new reactors.

NVIDIA's Vera Rubin figures (up to 3.7x throughput) are vendor-reported. Data-dense but partly a lead-in to a paid guide.

Camille Fournier: the death of management is overblown — AI adds productivity and visibility but misses tone and body language

She argues AI can push managers toward shallow decisions, fuels turf wars over cheap generated code and erodes the small daily interactions that build trust; her remedy is to talk to people more, not less. Her Pointer-linked talk boils her book down to technical skills, signal processing and people skills.

Same issue: Kore Nordmann says per-team output metrics reward protecting your own numbers; Mike Fisher says AI works as an after-action reviewer for managers, with sycophancy and recording-consent risks. Digest of blurbs; article links are trackers, so the issue page is linked.

Two builders on the future of engineering management: the bottleneck moved downstream and reliability may become the dividing line

Pointer's summary of a Gergely Orosz and Michael Grinich (WorkOS) conversation: roles are redrawing (WorkOS hired more PMs on work volume), more PRs strain CI/CD and review, engineering's remit is expanding to enabling everyone to build, and leaders are leaving AI-resistant orgs. Its conclusion: less coordinating the work, more designing systems and teams that absorb the output.

High agency has to be enabled by the team environment, not just hired for

From a Graphite case study: agency comes from purpose, permission and simplicity, and Graphite runs "a couple of PMs for 30+ engineers". Ex-Meta hiring chair Austen McDonald says the best hiring predictor is growth signal, not current AI usage; ask "tell me about a time AI didn't work out". Mostly curated links.

Pragmatic engineers build the right thing, not the technically perfect thing

His own mistake: an excellent image crop-and-resize feature users never used because they already cropped images themselves. Advice: ask "what's the simplest solution we could deliver?" and ship that first; as AI makes building cheaper, deciding what to build matters more. Adapted from his upcoming O'Reilly book; the rest is paywalled.

Hiring signals: executive screens now look for five visible signals, and forward-deployed engineers spend roughly half their time not coding

Ibrahim Bashir's recap lists resume legibility, a demoable builder archetype, proof of impact, AI-SDLC impact and brand pedigree; it is heavily promotional (job board, paid cohort). Agentic Engineering's FDE self-test says FDEs spend roughly 50% of time on meetings, contracts and coordination — the author's claim, not data. Directly relevant to consultancy staffing.

Code became abundant, but output at one node of a system is not the output of the system

Real flow is understand, decide, specify, design, implement, verify, deploy, operate, learn; code generation is one node. Cheap "yes" means prioritisation stops being forced and judgment becomes the scarce resource, and fast signals (PR counts) arrive before slow ones (retention, margins), causing overshoot.

He cites McKinsey that 80% report individual productivity gains but only 37% see EBIT impact (6% high performers). External figures are the author's citations; verify before reuse. Free, full essay.

Companies are conditional bargains, not families — a case for "healthy ambivalence"

High stakes, temporary and unchosen relationships, uneven power and costly exit mean workplaces can't reliably give unconditional support; breached "psychological contract" beliefs hurt even when nothing was promised. "High agency" doesn't change the power imbalance. The second part covers the "case-by-caser" manager who demands specific examples to deflect systemic issues: act on the pattern without winning the argument.

When code is cheap, judgement becomes the job

Scarce work shifts to choosing priorities, reviewing product behaviour, maintaining architecture and owning production outcomes. Senior leverage: turn repeated review feedback into guidance and deterministic lint rules, and keep human judgement for business context. No numbers.

"Do agents work" is settled; what breaks between prototype and production is context, identity, ownership and memory

MongoDB's Max Marcon: a prototype runs on ~1,000 synthetic records on a laptop, production means 15 data stores and five enterprise apps; giving an agent context tempts teams into "God mode" permissions; homegrown memory layers become "almost a business". A framework swap pitched as a productivity win "can quietly eat six months". It is also a MongoDB product pitch.

The other issue's useful frame: MCP gives access to a capability, A2A delegates an outcome, AGENTS.md and Skills package instructions — ask which boundary you've hit, not which protocol wins. Claims MCP SDK downloads grew fivefold to over a billion under neutral governance.

Before removing an annoying rule, find out why it exists — a working fence looks identical to a useless one

Five steps: find the paper trail, ask the longest-serving person, verify the answer, decide on what you find, or pause (don't delete) if no reason turns up; never trial-remove security, safety, legal or financial controls. The $40,000 double-paid invoice is a hypothetical example. A sister issue retells Kahneman's over-trust in small-sample priming studies.

A single ~900 KB POST can freeze a Next.js / React server-action endpoint — patch now

React rebuilt FormData for a server action with quadratic work; a roughly 900 KB request triggered 100 million string checks and blocked the Node event loop. Affects React 19.0 to 19.2; fixed in 19.0.6, 19.1.7 and 19.2.6. An action item for any client on Next.js or RSC.

Patch list for the week: Argo CD, vLLM, Atlassian Data Center, plus a Copilot sandboxing default to check

Rohit: Argo CD has four critical vulnerabilities fixed in 3.3.15, 3.4.10 and 3.5.4; vLLM 0.31.0 fixes a code-execution flaw and /tokenize sits outside built-in API-key protection; Copilot sandboxing is GA but VS Code Agent Host defaults leave it off, outbound networking on and Git credentials accessible. TLDR IT: Atlassian Data Center CVE-2026-21589 (CVSS 9.3, unauthenticated arbitrary file access).

GhostAction injected fake "security-audit" workflows into 346 repos — and rotating secrets is only half the cleanup

The editor's point: the GitHub credential that let the attacker add the workflow is not revoked by secret rotation; revoke it, remove the injected workflow and check every repo the account could write to. Google's companion guidance: ephemeral runners, short-lived OIDC credentials, immutable dependency references, provenance verification and untrusted PRs kept away from release secrets. The digest is ~70% links with tracker URLs.

Evaluating agents: skill discovery jumped from 9% to 55% with one overview file, and a confident final message says little about success

Expo's CI harness found adding an overview skill raised sessions loading at least one downstream skill from 9% to 55%. Microsoft's ThinkingBox evaluated 507 workflows repeated 20 times: of 79,853 failed runs, 67% ended without a tool error, so judge agents by the state they leave behind.

A merge gate for agent-written code that the agent cannot game

Runs on a clean machine: reinstall from scratch, run the full suite three times, reject removed or skipped tests, restore the base branch's tests against the new code. Run inside the agent loop it turns failures into repair instructions, while a separate CI run keeps the verdict outside the agent's control.

Kubelet watches disk bytes, not inodes, until it is an emergency

Image GC gives no early signal for small-file-heavy images; on one node a single-stage build with @mui/icons-material consumed millions of inodes. A multi-stage Dockerfile cut the image from 40,000+ files to a few dozen, and a Prometheus alert at inode usage above 80% catches it well before the 95% eviction threshold.

Claude Code "Mods" are plugins that can redraw the interface and step into tool calls — one person's first look

Per the post, mods can add panes, commands and tool-call rules and share data between hooks, which hooks, skills, status lines and MCP servers cannot do. Verdict: "yes, use one now". A 3-minute read.

Cortex swapped Lucene for Postgres full-text search and cut public API p95 from 13.3s to 1.54s

Custom query parser, weighted ranking and CDC-fed indexes; total memory down about 40% and most edits searchable within seconds.

Booking.com cut Node.js costs 38% moving from pm2 clustering to Watt worker threads

30% fewer pods, 20% less memory per pod and up to 10% lower tail latency with no service code changes; production tests showed 40-50% more throughput within the same latency target.

Airbnb captures and replays real MySQL workloads to de-risk upgrades

Traffic captured at ProxySQL, with transaction order and timing rebuilt offline. It exposed a query going from 0.03s to 2.6s after an upgrade and queries whose results depended on unspecified row ordering, and helped move the fleet from MySQL 5.7 to 8.0 without a major production incident.

Uber Eats halved end-to-end search latency by removing low-value work

Dropped low-value retrieval paths, separated ranking features from presentation data and eliminated false dependencies between pipeline stages; also pagination, concurrent rendering, request hedging, compact embeddings and service-mesh fixes.

Sam Newman calls microservices an architecture of "last resort" and boils the eight fallacies down to three rules

Information takes time, the thing you call may not be there, and resources are finite — most outages come from resource exhaustion. Idempotency keys are clean but hard to retrofit; fingerprints are easy to retrofit but can reject valid requests (a Danish payments system rejected same-value purchases within 5 minutes). On AI: hedge across vendors and let AI roam only inside module boundaries you designed. 1:58 episode; his book is out in print 7 Oct.

An architect's job is to sharpen the problem and expose trade-offs; quality should be the last lever to give up

With scope, deadline and headcount fixed, only quality can move, so educate your manager to argue the other three first. Draw diagrams that show why a "simple change" isn't, don't produce estimates yourself, and surface the 2-3 real internal problems buried in complaints. No hard numbers; course promos at the top.

Stacklok's cloud-native agent harness "Mecatl" separates the agent loop from client, model, state and execution

Thesis: the locus of value moves from the desktop to the cloud; desktop-first harnesses couple loop, local execution and session state, which makes pausing for human input hard. Related ToolHive is an open-source MCP gateway/registry. Sympathetic interview with no usage numbers.

Seven System Design Nuggets issues, all paywalled previews; the Push-vs-Pull and SRE ones have real content in the free part

Push vs Pull: Kafka chose pull so a slow consumer "simply falls behind and catches up"; 10,000 clients polling every 30s is 333 req/s even when nothing changed. SRE in 90 minutes: downtime per year is 52 minutes at 99.99% and 8.7 hours at 99.9%, and keep an SLA buffer below the SLO. The other issues (15 principles, Kafka's 7 concepts, a staff-level interview, an API-protocol guide, choosing an AI model) show only 10-30% of their content.

How SAFEs quietly eat your equity: the cap table says 90%, as-converted founders hold 70.9%

Three post-money SAFEs totalling 21.25% sold for $3m leave founders at 70.9%; a $4m seed on $20m pre with a 12% pool then leaves them at 54.9%, not the expected 75%. Capped SAFEs never shrink in a hot round. Carta data quoted: 93% of Q2 2026 pre-seed rounds were SAFEs, median founder ownership 54.8% after seed. Numbers are free; the dilution model is a paid download.

CAC payback is the number investors check before they believe your growth

Median B2B SaaS payback is 16 months (Aleph/Benchmarkit, 342 companies), top quartile 6 or less, bottom quartile over 24. The worked example is internally inconsistent: the post says payback is roughly two months but its own formula gives $1,000 ÷ ($1,000 × 0.8) = 1.25 months. Sponsored by Omnisend; the calculator is paid.

Businesses are switching AI model providers more often, and inference may be a ~$350B market by 2027

Ramp: the monthly share of firms switching providers (3-month average) hit a record 8% in September, read as an early sign of model commoditisation. Tom Tunguz says inference could be a $350B market by 2027 (one issue says double the database market, another "larger than databases"), with margins below typical SaaS and smaller models and routing becoming essential.

Negative gross margins are acceptable only with a way out — Harvey vs Legora

Harvey reportedly hit negative 50% gross margin in June as usage grew under flat seat pricing ("reportedly" per the source). Three exits: charge for usage, make each request cheaper, cap what's included. Proposed test: would customers stay if the product kept 40 cents of each dollar after serving costs.

Lock-in loosens when AI makes migration cheap, and everyone is building the same AI startup stack

Nygard: lock-in lasted because leaving was expensive; agents can now build converters and reimplement the behaviour a business depends on, so customers can leave when staying costs too much. Angular Ventures: startups converge on the same agent loops, memory, connectors and sandboxes and pitch commodity infrastructure as the moat; customer references say more. Opinion, no numbers.

Venture capital's repeating mistakes: investors price what is in front of them, not the capability that decides what it becomes

Seven patterns, including Bessemer's Cowan skipping Google in 1999 because Yahoo and MSN dominated, passing on Tesla's Series C over negative margins, and passing on Neko Health over price. Several figures come from third-party charts and are not independently shown. Fully free; an Alumni Ventures ad sits mid-post.

The 9-step GTM system: one living playbook instead of four folders with four answers

Cites Bridge Group 2026 AE research (158 B2B companies): average AE ramp is 6.2 months, a study high (4.3 in 2020). Gallup: only 12% strongly agree their company onboards well. Promotes the author's book and tool; the playbook export is paywalled.

GitHub's accessibility agent has reviewed 3,535 PRs with a 68% resolution rate

A parent agent directs a passive reviewer and an active implementer; drag-and-drop, data grids and rich text editors are escalated to human specialists. An experimental pilot.

AI raises the fidelity bar: match fidelity to the decision, and don't replace valuable interactions just to look "AI native"

"Good enough" starts by naming what the work must unlock (a usability test, an estimate, a direction, an approval) and applying only that much fidelity. The companion piece argues for preserving human skill.

Stop blaming the model for slow AI: acknowledge fast, then generate

Doherty's 400 ms threshold now covers two moments, acknowledging input and producing the answer, so interfaces can acknowledge immediately and show progress while generation takes seconds. The link in the email is an opaque redirect.

ChatGPT cartoons are carrying the signatures of real artists

Signatures associated with more than 15 New Yorker contributors appeared. OpenAI added warnings for some New Yorker-style requests but signed outputs reportedly continued; the concern is style imitation crossing into false attribution.

Four rules to build discipline: never snooze, prep tomorrow tonight, keep deciding, and don't run your day on mood

Bonus: set habits (input) rather than targets (output). No numbers behind the claims; the issue also carries personal asides, a quiz and an "AI for Students" promo.

You get things done faster by committing to the present, not the finish line

Three commitments: process over outcome, the moment no matter how mundane (including waits on AI agents), and a single reality (name your role for the time block). A repost of an older article with only soft claims.

Make new tools earn their place; observe closely before deciding

A reader keeps a paper legal pad and hands AI only the unfun part, typing up notes. Mostly promotional: pre-orders for "Life in Perspective" (3 Nov) and Second Brain programs, and the link is a review video, not an article.

"1500 Hijri (Business Edition)": a fictional workplace day with protected deep-work blocks and care rituals

Practical takeaways the author says could start tomorrow: a conscious Bismillah before meetings, istikhara before hiring, "how's your heart?" before the numbers. The only figures are illustrative; ends with a leadership-programme pitch. No clean article link (beehiiv redirect only).