The Unread Pile — Edition No. 24

Edition No. 24 Sunday, 18 October 2026 21 new saves

The Lead Story

Half the Speed of Sound Over the Grand Canyon

Via @elonmusk — with the timeline’s Z-axis chorus

Editorial illustration: a futuristic sports car flying through the air above a city skyline at dusk

Illustration: the sequel — yesterday’s disbelief becomes today’s demo itinerary. Snake River. Grand Canyon. On its wheels.

Musk doesn’t walk the specs back; he raises them. A day after the flying-Roadster reveal, he posted the sequel himself: “Future fun Roadster demos will include flying over the Snake River or maybe Grand Canyon and landing on its wheels and exceeding half the speed of sound.” The reveal is no longer a launch event. It’s a tour. The question shifted overnight from did it really fly? to where will it fly next? — which is exactly how spectacle marketing is supposed to work.

The numbers need restating because they still sound like a prank. A newly designed car weighing more than 4,000 pounds, doing flips, travelling at 380-plus miles an hour — past half the speed of sound — and landing on its wheels after flying over a canyon. MKBHD’s canonical disbelief stands unchallenged: there is, by his reading, no street tire invented yet that can handle anything near that. And yet the escalation continues, because the escalation is the product moment. Every new impossible detail is another free headline.

The speed ladder of the claim

Figures per Tesla’s reveal and Musk’s posts, Oct 2026.

“Future fun Roadster demos will include flying over the Snake River or maybe Grand Canyon and landing on its wheels and exceeding half the speed of sound.”— @elonmusk

Meanwhile, the Z-axis chorus kept writing the marketing department’s copy for free. Dilip Kumar’s four-word investment thesis: “Never bet against @elonmusk.” Vijay Shekhar Sharma wondered what the traffic-jam demo was even for. Derek Nee reported that traffic is a bit heavy in the Z-axis today. And Alexandr Wang asked the one question an engineer asks while everyone else laughs: “wait who the hell was driving the tesla.”

Main learnings

Escalation is the strategy. When the reveal gets mocked, don’t defend the specs — raise them. Disbelief converts to coverage at a fixed rate.
The sequel arrives in 24 hours. Musk compressed the launch cycle to a day: reveal, then the demo itinerary, then the memes become canon.
Half the speed of sound is the new Z-axis. The Roadster stopped being a car and became a flying-vehicle demo program.
The chorus is the distribution layer. Four one-line reactions reached further than the official video.

Practical takeaway — when a founder escalates an impossible claim the day after it’s mocked, don’t ask whether it’s real. Ask why you are still talking about it.

Musk’s new post · Never bet against Elon · The traffic-jam question · Z-axis traffic · Who was driving?

The week’s through-line

Intelligence is now the cheapest thing in the building — a model writes 90% of its own code, a frontier eye costs a seventeenth, quotas reset for everyone. So value moved up one layer: to execution, governance, and whoever owns the bill.

The Build

Claude Writes 90% of Claude Code’s Code

Via @Hesamation — quoting @ClaudeDevs

Editorial illustration of a human and an AI looping through code review

Illustration: the human-in-the-loop is now the rarest resource — the legend writing the other ten percent.

“Claude writes about 90% of Claude Code’s code,” Anthropic’s ClaudeDevs account posted — and the punchline wrote itself: the absolute legend is whoever writes the other ten percent. The numbers underneath are the real story. From June to September, merged pull requests more than doubled, while the share that got reverted fell by more than 40%. Volume up, quality up, humans mostly watching.

The mechanism is the code-review redesign they wrote up on claude.dev: the bottleneck in AI-written code was never generation, it was review. They rebuilt how code gets reviewed — how feedback is routed, what counts as approval, how the model learns from corrections — and then throughput unlocked. The 90% figure gets the headline, but the 40% reverts figure is the one that matters: writing code was the easy part; being right about the code is what the review machine solved.

Anthropic on Claude Code’s own development, June → September

Via @ClaudeDevs (blog: solving-code-review-at-anthropic).

Main learnings

Review is the bottleneck, not generation. Throughput doubled only after the review pipeline was redesigned.
The 10% is the premium. The human’s remaining share is judgement, taste, and the review design itself.
Measure reverts, not volume. Twice the PRs mean nothing without the 40% quality gain.

Practical takeaway — if your team’s AI coding output is stuck, don’t buy a bigger model. Redesign the review loop.

The post

Let the Reset Wars Begin

Via @beffjezos, @thsottiaux — Codex day 12 & 13, Antigravity’s answer

Beff’s joke landed because it’s true: “Google has its own Tibo now. Let the reset wars begin.” Tibo Sottiaux, OpenAI’s Codex lead, has spent his 28-day series announcing limit resets the way musicians announce tour dates. On day 12 he detailed the engineering underneath: reusing context across subagents, keeping context cached longer, and better request routing to increase cache hits — the effective-usage improvements behind the reset. On day 13, the two-word event: “Le limit reset,” landing by end of day Pacific.

And Google answered in kind. Varun Mohan posted that Antigravity was resetting Argon and Gemini model quotas for everyone, telling builders to keep building through the weekend. The pattern is now visible: the coding-agent wars are fought not on model quality but on allowance — who lets you build the most before the meter runs out. Quotas are the pricing lever, resets are the marketing, and the builder’s weekend is the battleground.

“Google has its own Tibo now. Let the reset wars begin.”— @beffjezos

Main learnings

Context is the currency. Day 12’s real news is context reuse and caching — the meter only drops when the engineering gets smarter.
Resets are marketing events. A quota reset is a product launch that costs nothing and buys a weekend of builders.
The war is on allowance, not models. When models are interchangeable, whoever resets first wins the weekend.

Practical takeaway — budget your agent spend like media spend: the reset calendar is the new rate card.

Day 12 · Day 13 · The reset wars

The 17x Cheaper Eye

Via @deedydas — on Elorian’s Glimpse

Deedy’s thesis: “LLMs are either not capable or not affordable for visual reasoning tasks at scale.” The Elorian team — one of Jeff Dean’s best reports who led Gemini, an Apple LLM lead, an xAI lead, and a Harvard professor — just launched Glimpse, described as a visual-thinking model with competitive frontier performance versus Astra on key public visual-reasoning benchmarks, for 17x cheaper. The demo that sold it: Glimpse breaking down the physics of the flying-Tesla video, frame by frame.

The wedge is video inputs at industrial scale — manufacturing lines, robotics, gaming, content moderation — where every frame is a query and frontier prices are prohibitive. Deedy led the seed round with Striker and Altimeter, and it’s already on OpenRouter. The pattern repeats: every frontier capability gets a second act as a commodity input. First text, then code, now the eye.

Frontier visual reasoning, per-token cost

Claimed by Elorian per Deedy’s post; Astra as benchmark.

Main learnings

Video is the new text. The frontier moved from reading frames to reasoning about them — and the price collapsed first.
Pedigree is the moat signal. A Gemini lead, an Apple LLM lead, an xAI lead, a Harvard professor — teams like this get funded on a paragraph.
The demo is the physics. Breaking down the Tesla video’s physics is a better benchmark story than any leaderboard.

Practical takeaway — if your product touches video — moderation, QC, robotics — reprice your assumptions at one-seventeenth.

Deedy’s post

Reverse-Engineer Anything

Via @zodchiii and @AravSrinivas

Editorial illustration of software being taken apart and rebuilt

Illustration: the closed app, taken apart — one prompt, down to the binary.

Two posts, one stack. darkzodchi’s version: REA, an open-source MCP server, plus Opus 5.5, is “dangerous” — one prompt and your agent reverse-engineers any app, down to the binary, maps unknown repos, rebuilds undocumented APIs. Aravind Srinivas’s version: Perplexity’s Computer has integrated the REA skill, and they’re handing out 10,000 bootstrapping credits so anyone can reverse-engineer an app of their choice — Computer inspects how the app works without its source code and builds a working version.

The reply sections tell the two halves of the story. One user burned 9,000 of the 10,000 credits on the first prompt — the economics of reverse-engineering are still absurd. The others called it a legal minefield. Which is precisely the point: the capability arrived before the pricing and the law. The binary was the last moat. It isn’t anymore.

Main learnings

The closed app is open now. Source code was never the moat; it was the inconvenience. That inconvenience is gone.
Credits are the new freemium. 10,000 credits that evaporate in one prompt is a marketing number, not an allowance.
Law lags the prompt. The capability exists; the case law doesn’t.

Practical takeaway — audit your own product’s binaries against the REA workflow. Whatever your agent finds, someone else’s already did.

The REA playbook · Perplexity’s REA skill

The Money

From Inference to Execution: AI’s Next Economy

Via @JayaGup10 and Ashu Garg — article, liked & bookmarked

Jaya Gupta’s thesis, co-written with Ashu Garg, names the week’s through-line: the AI economy is shifting from supplying intelligence to executing work. The money moves from selling GPUs and renting models to agents that complete tasks end-to-end — and the power accrues to the agent-assembling companies, which decide the model-and-infrastructure mix underneath. Her bet: owning the most GPUs will no longer guarantee capturing the most value.

The evidence she cites: outcome-based pricing is rising — pay for the result, not the token — and CoreWeave’s Q2 revenue more than doubled to $2.575 billion on heavily financed capacity. The GPU landlords are doing fine; the question is who keeps the margin. When execution is the product, the stack picker — the agent company choosing which model, which cloud, which tool — captures the surplus, and the infra becomes fungible underneath.

Value capture: inference → execution

Thesis via Jaya Gupta & Ashu Garg, ‘From Inference to Execution’.

Main learnings

The surplus moved to the assembler. Whoever picks the model/infra mix keeps the margin.
Outcome pricing beats token pricing. Customers pay for completed work; token meters become an internal cost.
Infra becomes fungible. When agents choose the stack, GPUs are a commodity input — $2.575B quarters notwithstanding.

Practical takeaway — if you sell intelligence, start selling outcomes. The margin is following the result, not the token.

Jaya’s post

Buy the Dip, On Autopilot

Via @mignano — Supertake’s new feature

Michael Mignano announced Supertake’s new “Buy the Dip”: when one of your take’s bigger positions falls hard below what you paid, Supertake runs deep analysis to check whether the reason you own it still holds. If it does, it draws from the take’s cash reserve and doubles down — within the take’s limits. Autopilot takes do it on their own; otherwise Supertake asks first. Deactivate any time.

The thesis-check is the product. Retail dip-buying fails because the dip is priced before the thesis is rechecked; automating the thesis-check before the order converts a reflex into a discipline. The autopilot distinction is the honest part — autonomy is a tier, not a promise, and the take that opts in has pre-committed to its own rules.

Main learnings

Automate the check, not just the order. “Does the thesis still hold?” is the trade; the buy is the paperwork.
Autonomy is a tier. Autopilot vs ask-first is the right way to sell agency.
Pre-commitment beats discipline. The rules work because they were written before the dip.

Practical takeaway — write your buy-the-dip rules before the dip. Then let the machine follow them.

The announcement

Consumer Is a Balance-Sheet Game

Via @aviralbhat

“Consumer businesses are balance sheet businesses as much as profit and loss businesses,” Aviral wrote. Early founders obsess over margins; scaled ones watch how fast cash moves. His example: simply cutting receivables from 90 days to 60 days on a ₹100-crore business frees ₹8 crore of cash — without selling one more unit.

Cash velocity: 90 → 60 days receivables

Illustrative math on a ₹100 cr business, via @aviralbhat.

Practical takeaway — before chasing margin, chase receivables. Speed of cash is a free fundraise.

The post

Groceries Fund the Space Dream

Via @aviralbhat

“A country needs groceries so that it can dream of space.” Aviral’s thread maps how India’s deep-tech risk capital was seeded by internet wealth: Sachin and Binny did Flipkart before funding Ather; Mukesh did Myntra before funding Skyroot; Deepinder did Zomato before funding LAT. Replies added Paytm to Kawa Space and Zerodha to Agnikul.

The mechanism: consumer-tech exits recycle founder capital into ventures with decade-long horizons that institutional VC won’t touch early. One reply cautioned the examples are cherry-picked — fair, but the direction of the flow is real. Groceries are the apprenticeship; space is the graduation.

Main learnings

Wealth recycles before institutions arrive. Founders fund the frontier their VCs can’t yet underwrite.
Consumer exits are deep-tech’s seed round. Flipkart, Myntra, Zomato — the apprenticeship economy.
Cherry-picked but directional. Three examples don’t make a law, but the flow is real.

The post

Amazon, Below Business-As-Usual

Via @ganeshsonawane — quoting @nareshgupta_

Ganesh Sonawane’s three words of alarm: “this has never ever happened before.” Quoting a sales chart from Naresh Gupta showing Amazon India running below business-as-usual during the festive season — the one quarter when the graph is supposed to spike. The author’s own caveat: the chart covers one section of Amazon sales, but the ratio is broadly correct.

Replies blamed weak iPhone festive demand and channel mix shifts. Whatever the cause, the signal is a consumer-pulse reading: the festival that reliably lifts e-commerce didn’t this time. When BAU becomes the ceiling during the season, something in the demand engine has changed.

Main learnings

Festive is the canary. When the season fails to lift, it’s demand — not logistics.
Read the ratio, not the chart. One section, but the proportion tells the story.

Practical takeaway — track festive BAU deltas as your demand early-warning. This one is flashing.

The post

The Humans

The Self-Actualisation End of the IIT-IIM Pipeline

Via @abidsensibull

Abid Hassan’s IIM-Ahmedabad friend told him he doesn’t want his kids to go to IIT or IIM — in fact he’d be happy if they actively didn’t want it, and just had fun in life instead. Abid’s framing: that is “the real self actualisation end goal of going to IITIIM.” The credential’s final dividend is permission to not need it.

The replies split on privilege: it’s easy to decline the ladder once you’re standing on it. But the deeper read is about what the credential was actually buying — optionality. Once optionality is secured, the optimal strategy is to let the next generation spend it on whatever maximises their own utility function, not to re-run the same optimisation. The ladder was the means; fun is the end.

“That is the real self actualisation end goal of going to IITIIM.”— @abidsensibull

Practical takeaway — credentials buy optionality. Spending the optionality — not re-running the race — is the point.

The post

Malfoid, and the Great AI Sentiment Flip

Via @enggirlfriend

A Harry Potter fanfic — female Draco Malfoy in an enemies-to-lovers arc with Harry — is going viral on TikTok, and the AI-generated videos are so good they’re flipping TikTok’s AI sentiment from historically negative to positive. The trend even has a name: “malfoid.”

The sentiment flip

Anecdotal, via @enggirlfriend’s thread.

Sentiment follows delight, not debate. A thousand essays couldn’t do what one irresistible fan-video did: make AI feel like fun instead of threat. Distribution of delight beats distribution of argument.

The post

Shoot Your Shot

Via @nikunj

Nikunj Kothari wrote a prediction seven months ago, met ~50 startups in the space since, and this week led the Series A for the one that fit. “Shoot your shot, you never know how the universe collaborates in your favor.”

The mechanism is unglamorous: write the thesis down, do fifty reps, close one. The prediction is the filter; the meetings are the work; the universe’s collaboration is what happens when a public thesis meets private dealflow. Fifty nos is not the cost of the yes — it’s the due diligence that makes the yes safe.

Practical takeaway — write the thesis publicly, then do the reps. The shot you don’t write down never gets taken.

The post

The Founder You Don’t Lose

Via @vkhosla

Vinod Khosla’s reply to a portfolio entrepreneur: “As one of our portfolio entrepreneurs, I prefer you build Eight Sleep!” The CEO, joking that he’d take the job himself if he weren’t already running Eight Sleep, had called the opportunity “of a lifetime.” Khosla’s answer: the lifetime opportunity is the one you’re already building.

The investor’s nightmare isn’t a failed company — it’s a great founder walking away from a great company. Khosla’s preference is a retention thesis stated publicly: the best new job is the old company, finished.

Practical takeaway — the opportunity of a lifetime is usually the one you’re tempted to leave.

The post

The Workshop

Play the Bad Guy: A Pre-Launch Security Ritual

Via @Voxyz_ai — liked & bookmarked

Vox’s Claude Code security ritual: before you ship, make Opus 5.5 play the bad guy — three subagents, test accounts, 14 attacks across impersonation, UI-limit bypass, and external-entry groups. Collect the proof, fix everything in one pass, re-test, then install Anthropic’s security-guidance plugin. The prompt is safety-constrained: test accounts and test data only, never production.

The replies worried the attack prompts could trip guardrails or get accounts banned — which is itself the signal: the ritual is powerful enough to scare the platform. Vibe-coded sites ship with vibes for security; this is the first playbook that treats AI-assisted attack as a standard pre-launch gate.

Main learnings

Adversarial testing is a gate, not a luxury. Three subagents, 14 attacks, one pass of fixes.
Constrain the blast radius. Test accounts and test data only — the prompt says so explicitly.
Re-test after the fix. The ritual is collect, fix, re-verify — not collect and hope.

Practical takeaway — add the bad-guy pass to your launch checklist. If it can break your product, better your agents find it than someone else’s.

The post

Screener’s Live Concalls

Via @faltoo

Pratyush Mittal’s team shipped the feature investors actually wanted: listen to live earnings concalls directly from Screener — click and join, no dial-in number. The fun detail: they discovered their own office iPBX could do it. Replies immediately asked for recordings, AI summaries, live transcripts, and AGM coverage.

Practical takeaway — the best features are the ones users ask for in the replies. Ship the click, then the transcript.

The post

The Mac Mini Home Server

Via @priymrj

Priyam Raj runs Claude Code 24/7 on a Mac Mini in Bangalore while he’s in his hometown, previewing everything over Tailscale. “It’s perfect.” The remote Mac Mini is the new home-server flex — your personal data centre, one Tailscale away, coding while you sleep.

Practical takeaway — a spare Mac Mini plus Tailscale is a personal cloud that codes overnight. Set it up once, harvest forever.

The post

Tangerine: Drag, Shift, Convert

Via @th2049

Thameem’s Mac app Tangerine converts files with just drag + shift — no menus, no dialogues, one gesture. Replies asked for a Windows version and noted it needs Apple Silicon. The lesson of the tiny demo: the best utility is one gesture away from invisible.

Practical takeaway — the killer feature of any tool is the one that deletes the interface.

The post

Briefs

The quota wars — Codex day 12’s context-reuse engineering and day 13’s “Le limit reset” vs Google’s Antigravity reset for everyone. The builder’s weekend is the battleground. Day 12 · Day 13
The 10,000-credit illusion — Perplexity’s REA credits burn through 9,000 on the first prompt. Freemium is a marketing number now. Post