Daily AI Frontier Briefing: 2026-11-09
Hacker News front page for 2026-11-08 (UTC), 88 stories. AI coding tools collided head-on: one developer publicly quit Claude Code for Pi, an open-source harness (635 pts/406 comments), while Goldman Sachs pulled Claude Code from traders and bankers, keeping it for engineers only (402 pts/249 comments). Two serious local-deployment tools: llama-modchip squeezes a 284B DeepSeek-V4-Flash into 64GB of plain RAM (305 pts), and Maple prunes Gemma 4 26B to sub-2-BPW for 45 TPS on a 6GB GPU. Model research: Q Labs' new scaling law claims compute-optimal models could be 7,000× bigger than Chinchilla prescribes (64 pts/58 comments), and LoreKit's author proposes an OpenTelemetry-style standard for agent memory (158 pts). Industry: OpenAI closed $11B at a $1.25T valuation, below the $1.4T it first sought (174 pts/195 comments); Nvidia's board opened a probe into Jensen Huang with Q3 earnings Nov 17 and a closed-door briefing request Nov 18 (154 pts); Anthropic took over 1.7GW of gas-generator orders for $3.4B (37 pts); scavenged Boeing jet engines are entering data centers (15 pts); Tim Cook is pressing Apple's software chief to turn Siri into a ChatGPT-style agent (33 pts/55 comments); drugmaker CEOs rejected the 'AI cures 90% of disease in seven years' claim (74 pts/80 comments); and a 10-year ML veteran declared he's done using LLMs for coding (98 pts/211 comments, most contested of the day). Developer tools: GitHub's new PR overview page in public preview (175 pts), clj.rs (Clojure in Rust, 152 pts), PacketHound, a Little Snitch for Windows (60 pts), Mizu for zero-copy R/Python over shared memory (68 pts), a 15-year-old's LLM weight-editing tool (13 pts), and RAM-Altman, a script that emails Sam Altman your unused RAM daily (134 pts). Also: Outlook silently bouncing mail from domains without DMARC (106 pts), ad SDKs as a police tracking tool (74 pts), ISponsorBlockTV for every YouTube TV client (116 pts), an anti-'AI brain fry' ticket workflow (82 pts), and a professor's case for writing as thinking (226 pts). 33 items.
The HN front page for 2026-11-08 (UTC) had 88 stories, 4 of them with more comments than upvotes. AI coding tools collided head-on: one developer publicly left Claude Code for an open-source harness, while Goldman did the opposite and took coding agents away from traders. Local deployment produced two serious tools: a 284B model on plain RAM, and a 26B model on a 6GB GPU. On the industry side, OpenAI raised at a discount, Nvidia’s board opened a probe into its CEO, and Anthropic started buying power equipment. This US-shift update adds 9 more: a $10 homelab, Apple’s spelling grey list, a Rust SketchUp alternative, the Dux Wayland desktop suite, November’s project thread, a GitHub outage, Tesla’s product drought, the UK phone-scanning proposal, and Windows 11 losing PC migration. 33 items.
1. One developer ditched Claude Code for Pi, an open-source harness
244 releases a year will do that to a person?
Caleb (caleb.software) wrote about leaving Claude Code for Pi, an open-source coding harness, 635 points, 406 comments, rank 2 for the day. His reasons were an accumulation of small frustrations: 244 Claude Code releases in 2026, over three-quarters of them changing prompts or tool descriptions, with the Bash tool’s description alone nearing 3,000 tokens; AGENTS.md and CLAUDE.md overlapping in responsibility; flaky VS Code integration and GUI extensions. What Pi offers, he says, is stability, minimalism, and deep customizability, he wrote an extension in about 20 minutes that syncs IDE text selection across tabs. If your harness feels bloated, this is the case for a smaller tool you can bend yourself. Post · HN discussion
2. Goldman pulls Claude Code from traders and bankers, engineers only
Six months after praising vibe coding?
The State of AI reports that Goldman’s October internal guidance restricts Claude Code and GitHub’s coding agents to engineers, about 12,000 of its roughly 47,000 employees keep access, while traders and bankers lost it. The stated concerns: unreviewed, agent-written code running on trading desks staffed by people without engineering training; information-barrier requirements between banking and trading; and per-token cost. Strats, the hybrid trading/quant/engineering roles, kept access. The irony: Goldman and Morgan Stanley are lead underwriters on Anthropic’s IPO, marketing shares in a company whose product most of their own bankers can no longer use. Barclays is running the opposite play, expanding Claude Code to half its developers while cutting 5,000 jobs. 402 points, 249 comments. If you write enterprise AI policy, here’s your case study and its counter-case in one story. Report · HN discussion
3. A 284B DeepSeek model running in 64GB of RAM, zero GPUs
GPUs are expensive. RAM is just sitting there?
The GitHub project llama-modchip (Apache-2.0) ships kernel patches, quantization transcoders, and runtime tools that run large MoE models entirely in system RAM via llama.cpp, 305 points, 116 comments. The core is a custom “Fractal-64” quantization format (GGML type 40) that compresses DeepSeek-V4-Flash 284B from the standard 76.87 GiB build down to 56.56 GiB (1.71 bpw), so the weights plus KV cache fit in 62.5 GiB of usable RAM with zero NVMe swap. All 6 routed experts stay enabled by default, bit-exact, logit cosine similarity 1.0; AVX-512 multi-level tiling lifts prefill by 77%, and with an n-gram self-speculative drafter a 16-core Zen 5 machine with no GPU hits 44.4 tok/s prefill and up to 14.7 tok/s decode. The catch: stock llama.cpp, Ollama, and LM Studio reject the custom tensor type, so you must use the project’s patched binaries. If you have a big-RAM server and no interest in GPUs, this is currently the way to run a 284B MoE. Repo · HN discussion
4. OpenAI closed $11B at a $1.25T valuation, below its asking price
Raising more while the anchor drops?
247 Wall St reports OpenAI raised $11 billion at a $1.25 trillion valuation, below the $1.4 trillion it first sought, 174 points, 195 comments. Media reports put Anthropic’s IPO valuation expectation in the same window, trimmed from $1.4 trillion toward $1.2 trillion. Both frontier labs are still raising, but the valuation anchors are moving down. If you track the AI capital cycle, the discount is a harder signal than the round size. Report · HN discussion
5. Nvidia’s board opens a probe as Jensen Huang admits an affair
Earnings on the 17th, a closed-door briefing on the 18th?
Nvidia’s board issued a statement on November 8: a committee of independent directors, assisted by outside counsel, will investigate the intelligence reports targeting CEO Jensen Huang, 154 points, 117 comments, with follow-ups from Bloomberg, the FT, and the BBC. Huang told employees the same day that he had been unfaithful to his wife, apologized, and committed to cooperating; the board says it has no reason to believe U.S. national security or Nvidia secrets were compromised. The allegation: a U.S. intelligence assessment concluded Huang had a romantic relationship with Alice Huang (Huang Le), an executive at Singapore cloud provider Megaspeed, and assessed she was recruited by China’s Ministry of State Security to collect semiconductor intelligence and lobby for chip exports to China. Three House committee chairs have asked Huang to attend a closed-door briefing on November 18, one day after Nvidia’s Q3 earnings on November 17. China’s embassy calls the claims smears; Alice Huang has not been charged, and Megaspeed says it hired outside counsel and is cooperating. If you hold NVDA or work the China compute chain, circle both dates. Statement · HN discussion
6. GitHub ships a new pull-request overview page, in public preview
One page for everything reviewers keep tab-hopping for?
GitHub’s November 5 changelog announces a new pull-request overview page in public preview, 175 points, 122 comments. The view aggregates a PR’s description, review status, and check results in one place instead of hopping between tabs. It’s a rare big change to the PR workflow. If your team reviews on GitHub daily, you can flip on the preview now and send feedback. Changelog · HN discussion
7. Write it down or you haven’t thought it, argues one professor
You understand it until the blank page disagrees?
Distributed-systems researcher Murat Demirbas makes the old case that writing is thinking, 226 points, 72 comments. His ammunition: Manuel Blum’s line that without writing you’re a finite automaton, while writing gives you the Turing machine’s tape; Lamport’s blunter version, “If you think without writing, you only think you’re thinking.” His own calibration: he thinks he understands a paper until writing the review exposes that he got about 60%. A 2018 freewriting post on Paxos vs. Nakamoto consensus later grew into a research paper. For the LLM era he adds a twist: models know the literature but not your half-formed intuitions, and since they can produce plausible six-page memos, Bezos-style memo culture is getting diluted, “writing is now a comparative advantage.” His process: draft in bulk with no self-editing, then switch to ruthless editor mode, and write plain. If you do research or run a team, adding a “write one page first” step beats talking it through in your head. Post · HN discussion
8. clj.rs brings Clojure to Rust, aimed at agent runtimes
No JVM, but keep the macros?
clj.rs is a Clojure dialect implemented in Rust, 152 points, 58 comments, published as the crate cljrs on crates.io. The stated goals: persistent data structures, a hosted macro system, and native interop, usable compiled or interpreted; the landing page explicitly positions it for agentic coding runtimes and ships a browser-accessible REPL. The author calls it early stage with rough edges and publishes no benchmarks. If you want Lisp for your agent toolchain without leaving the Rust ecosystem, this is the only active option. Site · HN discussion
9. Agent memory needs its OpenTelemetry moment, says LoreKit’s author
Four tools, four incompatible memories, fine with you?
Mads Thines, who builds the open-source memory tool LoreKit and works with OpenTelemetry at Dash0, argues agent memory needs an OpenTelemetry moment, 158 points, 102 comments. He counts what a typical developer maintains today: AGENTS.md/CLAUDE.md files, Cursor rules, Anthropic’s memory tool, Copilot’s hosted facts, plus assorted MCP servers. The problem isn’t a missing tool; it’s a missing agreement. His proposal mirrors OpenTelemetry’s four layers: semantic conventions (each record declares address, key, kind, lifecycle, expiry), an MCP-based read/write contract, and conformance tests, with no reference implementation. He distinguishes export (carrying your memories to another vendor) from live addressing (multiple tools sharing one memory location at once), and concedes the xkcd-927 risk, rival proposals already exist in MacPaw’s Portable Memory, the Open Memory Protocol, and Cognee’s COGX. If you build agent frameworks or MCP servers, the shape of this interface is worth a look now. Post · HN discussion
10. 98 upvotes, 211 comments: a veteran ML engineer quits LLM coding
Efficient, sure. But whose code is it?
Kristian Polso announced he will stop using LLMs for coding, 98 points against 211 comments, a 2.15 comments-to-points ratio, the most contested story of the day. He describes roughly 10 years of production ML work, from Keras and TensorFlow onward, and gives six reasons: research (including Microsoft’s) on “cognitive debt” and reduced critical thinking, with his own experience of feeling disoriented after long LLM sessions; generated code that looks impressive at first glance but is often “complete bullshit”, he cites Claude overcomplicating a fix a few lines of migration would solve; agent-written code nobody ends up understanding, which he frames as outsourcing what your code means to OpenAI; the fact that he simply enjoys writing code; data-center power use; and CLOUD Act exposure making U.S. providers a GDPR problem for Europeans. He answers “AI is here to stay” with Esther Weidauer’s retort: so is heroin, and she wouldn’t recommend that either. If you’re pushing AI coding in your team, these 211 comments are the objection list you’ll receive. Post · HN discussion
11. Outlook.com is quietly bouncing mail from domains without DMARC
No announcement, no warning, on brand?
Since November 5, Outlook.com, Hotmail, Live, and MSN have been silently rejecting some mail from domains with no DMARC record, 106 points, 91 comments, with no Microsoft announcement. Unlike the May 2025 rule for bulk senders (5,000+ messages a day), this hits ordinary low-volume senders, and only some messages per domain bounce, so senders can wrongly assume everything is fine. The bounce reads 550 5.7.1 (Message lacks SPF/DKIM/DMARC Auth Records, AS9031519); check with dig +short TXT _dmarc.example.com, no output means exposed. The fix is one TXT record: v=DMARC1; p=none. If your domain sends verification codes, invoices, or newsletters, check today. Post · HN discussion
12. Skip YouTube ads on every TV client with this open-source tool
Premium priced out, ads unbearable, so, door number three?
ISponsorBlockTV is a SponsorBlock client for TV platforms, 116 points, 50 comments, covering Apple TV, Android TV/Google TV, Samsung Tizen, LG webOS, and the rest of YouTube’s TV clients. It calls SponsorBlock’s crowdsourced segment database to auto-skip sponsored segments during playback; it’s open source in Python at github.com/dmunozv04/isponsorblocktv. If your living-room TV runs YouTube and you don’t want Premium, this is the cleanest option. Repo · HN discussion
13. A script that emails Sam Altman your PC’s unused RAM, daily
So which is the real bottleneck, VRAM or RAM?
RAM-Altman is a joke project, 134 points, 46 comments, a script that checks your computer’s free memory every day and emails a report to Sam Altman, riffing on the local-model community’s old line that models fail because “you don’t have enough RAM.” The project lives at ram-altman.lol with open-source scripts. If you run local models, installing it gives you a RAM monitor and a vote in the memory-versus-VRAM argument. Site · HN discussion
14. Compute-optimal models could be 7,000× bigger than Chinchilla says
Out of data, so just train longer and ensemble?
A four-person team at Q Labs Research (Akshay Vegesna, Ethan Baron, Anirudh Satheesh, Samip Dahal) proposes a new scaling law expressing loss in four variables, unique tokens U, parameters N, epochs R, and ensemble members M, 64 points, 58 comments. The core claim: when data is limited you shouldn’t cap model size at Chinchilla; instead let size, epochs, weight decay, and ensemble members all grow with compute, and extrapolating to 5T tokens at 10²⁹ FLOPs the compute-optimal parameter count lands about 7,000× above Chinchilla’s prescription. Supporting findings: optimal weight decay rises steeply with epochs and scale, contradicting prior heuristics; snapshot ensembling from cyclic learning rates adds members for free during multi-epoch runs, then deep ensembles take over, with the optimal member count growing roughly linearly in compute. The motivation number: by the paper’s math, humans learn from about 1:500,000 as much data, versus a 20:1 tokens-to-parameters ratio for a 1.5T-parameter model on 30T tokens, a gap of roughly ten million×. If you work on pretraining recipes, the “more epochs plus ensembles” route deserves a small-scale replication before you believe it. Research · HN discussion
15. Drugmaker CEOs aren’t buying the AI-cures-everything story
The shovel sellers promise cures; the surgeons shrug?
The Wall Street Journal reported on November 4, 74 points, 80 comments, that pharma CEOs are publicly rejecting Silicon Valley’s AI-cures-disease narrative. Vertex executive chairman Jeff Leiden was bluntest, calling the 90%-in-seven-years claim “complete BS”: Vertex used roughly 50 AI companies to analyze 10,000 compounds, and none produced a working drug. AstraZeneca CEO Pascal Soriot said disease and biology are complicated, and calling AI a cancer cure “is maybe a little bit of a jump”; Regeneron’s George Yancopoulos said AI will assist in the lab but won’t “miraculously come in and tell us how to do things.” The Silicon Valley side for contrast: Dario Amodei said in September that curing most human disease in five to ten years may be possible, and OpenAI bought an AI drug-discovery biotech in October. The hard metric: no AI-designed drug has won FDA clearance, and McKinsey’s Alex Devereson notes about 25% of mid-to-late-stage failures come from off-target safety issues current models can’t predict. If you trade AI-healthcare stocks, FDA approvals are a more real progress bar than model leaderboards. Report · HN discussion
16. One developer’s workflow against AI brain fry
Every thread one step from done, and your brain is soup?
Sergey Kuhne wrote up his fight against “AI brain fry”, 82 points, 51 comments, his term for the collapse that follows hopping between 20 simultaneous AI coding threads. His setup: tickets are markdown files sitting next to the code and committed to git, so “nothing important lives in a chat window.” Each agent cycle does four things: research tickets, build the ones a human tagged safe, run tests plus a browser check, and verify deployed work. Sensitive changes go on a branch and merge only after human approval; the agent never deploys to production. Anything needing human judgment lands in an “owner inbox” generated by a script, prioritized. The reading UI is a Claude artifact rebuilt from git each cycle, so context loads in about a minute rather than twenty. If you run multiple agents at once, this one-thing-at-a-time pipeline is directly copyable. Post · HN discussion
17. Tim Cook is pressing Apple’s software chief for a bolder AI strategy
Years behind, and finally in a hurry?
Bloomberg’s November 8 Power On newsletter: Apple CEO Tim Cook is pressing software chief Craig Federighi to pursue a more ambitious AI strategy, 33 points, 55 comments. The same report says Apple is exploring turning Siri’s AI into a ChatGPT-like chatbot/agent, with an iOS 28 shake-up re-organizing the software teams behind it. Apple’s generative-AI lag is public record; this is evidence the fix is starting at the management layer. If you build AI products on iOS, the day Siri becomes a system-level agent, Apple’s own permissions become a moat nobody else gets. Report · HN discussion
18. Anthropic bought 1.7GW of gas generators for $3.4B
The grid can’t deliver, so buy the turbines?
The Financial Times reported on November 8, 37 points, 25 comments, that Anthropic is taking over gas-generator orders San Antonio-based equipment supplier OnePWR had placed with Caterpillar, about 1.7GW in total, for roughly $3.4 billion. The structure is an equipment-order takeover, not an acquisition: no employees transferred, but Anthropic picks up the installation engineering capability too, it initially planned to just buy power from OnePWR and ended up taking the orders. By MLQ’s math that’s about $2,000 per kW, which the FT calls the largest known single purchase of power-generation equipment by an AI lab. Context: Anthropic disclosed roughly $895 billion in compute and infrastructure commitments through 2045 in its amended IPO filing, and peers are making similar moves, xAI’s gas turbines in Memphis, Alphabet’s $4.75B acquisition of Intersect. If you model AI power demand, labs bypassing utilities to lock in equipment is becoming the norm. Report · HN discussion
19. Scavenged jet engines are now powering AI data centers
New turbines take 8 years; a used engine ships in 24?
The New York Times reports, 15 points, that retired Boeing 737 and Airbus A320 engines are being dragged in as on-site data-center generators. The examples: xAI reportedly installed eight converted aircraft turbines in Mississippi to power its Colossus 2 supercomputer, per state permits and satellite imagery; FTAI Aviation converts CFM56 engines into 23–25MW generator sets and took a $1.465 billion order from an unnamed cloud company in July 2026; ProEnergy turns 747 CF6-80C2 engine cores into 48MW turbines; and the U.S. EIA estimates the roughly 4,000 retired military aircraft at Davis-Monthan’s boneyard could theoretically represent up to 40GW. The tradeoff is worse efficiency and more pollution than purpose-built turbines, but delivery runs about 24 months versus 3-to-8-year waits for new ones. If you follow the AI power chain, equipment lead time has replaced electricity price as the first variable. Report · HN discussion
20. A 26B model at 45 tokens per second on a 6GB GPU
Out of VRAM, so spend bits instead?
Show HN: Maple uses MoE pruning to compress Gemma 4 26B below 2 bits per weight and runs it at 45 TPS on a 6GB GPU, 19 points, repo at github.com/AmSach/maple. If your card has 6–8GB of VRAM, this is a ready-made route to running a 26B model; how much quality the sub-2-BPW quantization costs is a test you can run on your own prompts. Repo · HN discussion
21. An open-source Little Snitch for Windows, no kernel driver
You really don’t know what’s phoning home?
PacketHound is billed as a Little Snitch for Windows, 60 points, 28 comments, free, open source under GPL v3, prompting the first time any app on Windows 10 or 11 makes a connection, with rules down to a single domain. It builds entirely on Windows Firewall with no kernel driver; Windows components auto-allow but get verified by path, ownership, and Microsoft signature, while abusable built-ins like PowerShell, curl, certutil, mshta, and rundll32 still prompt. Current release 1.2.0, a 68.2MB installer, with 5/15/60-minute pause options. If you want to see what connects out on Windows, this is far lighter than enterprise EDR. Site · HN discussion
22. Ad SDKs in popular apps are becoming a police surveillance tool
The free app was never free?
A Bloomberg investigation, 74 points, 15 comments, reports that ad SDKs embedded in popular apps are giving police a powerful location tool. Device and location data harvested through the ad-tech supply chain lets law enforcement locate phone users without going through carriers, and any app running a mainstream ad SDK is a potential source. If you ship apps, your SDK supply chain’s data flows are now a legal exposure; if you use a phone, the off switch isn’t yours. Report · HN discussion
23. Zero-copy R and Python over shared memory, with Mizu
Serialize, deserialize, was any of that necessary?
Mizu lets R and Python run in parallel over shared memory with zero copies, 68 points, 10 comments, repo at github.com/shikokuchuo/pymizu. Each side runs its own process and big data objects live in one shared memory segment, skipping the serialize-and-copy round trip. For data teams mixing R and Python, this is a shorter path than shuttling data frames through disk. Repo · HN discussion
24. A 15-year-old built a tool that edits an LLM’s output matrix
Fine-tune the whole model, or just rotate the last layer?
Show HN: Laplace’s Scissors is a Rust/Triton tool for editing an open LLM’s lm_head output matrix, 13 points, by a GitHub user, Miyamoto-Renya, who says they’re 15. The method, “protected readout rotations,” edits the final output matrix to fix wrong answers while constraints preserve the ones the model already gets right, with a default protected-output tolerance of 0.00002. The workflow has three commands: probe (finds weak output directions via Jeffreys divergence), diagnose (solves for a rotation), and export (writes an FP32 checkpoint); a Qwen3-0.6B reference test is included. Code is MIT-licensed; model checkpoints keep their own licenses. If you work on model editing, touching only the output layer is far cheaper than a full fine-tune, and small deployments are the right place to try it. Repo · HN discussion
25. A $10 4G dongle running ten services on 512MB of RAM
Your smartwatch has more compute than this. So what?
A blogger at bosscher.ch bought an OpenStick, a hackable 4G LTE modem dongle, for $10 including shipping, 799 points, 136 comments, rank 2 for the day. The specs are a quad-core Cortex-A53 at 1GHz, 512MB of RAM, and 4GB of eMMC, yet flashed with Debian and Podman it runs ten services: AdGuard Home in 40MB blocking 60% of DNS queries, an AI agent called Zeroclaw idle at 3MB of RSS, plus Cloudflare Tunnel, Beszel monitoring, Gatus, an RSS reader, and a bookmark archiver at 10 to 20MB each. His previous box ran OpenClaw, which ate over 1GB of memory and broke with every update. If you want to start self-hosting for the price of a meal, this parts list is copy-paste ready. Post · HN discussion
26. Apple’s grey list: words your iPhone refuses to help you spell
The keyboard already decided which words you shouldn’t write?
The Guardian reports, 541 points, 360 comments: iPhone keyboards carry a non-public “grey list” of phrases that get no spell-check or swipe-typing help. HN users verified it hands-on: “propaganda” alone succeeds 10 out of 10 times, while “left wing propaganda” succeeds 0 out of 10 and keeps coming out as “left wring propaganda”; “united ireland” gets rewritten as “united iceland”. The list is invisible, uneditable, and Apple has published no criteria; a former Apple employee says the company always picks zero risk over suggesting offensive content. If you type on an iPhone, the existence of the list is itself worth knowing. Report · HN discussion
27. KetchUp, an open-source SketchUp alternative in Rust and WebGPU
SketchUp has been paid software for years. Where’s the rival?
KetchUp, from the BenchUp3D group, is an open-source 3D modeler, 583 points, 172 comments, written in Rust and rendered with WebGPU (or WebGL2), running in the browser with desktop builds for Windows, macOS, and Linux, MIT-licensed at github.com/BenchUp3D/editor. It covers SketchUp’s core moves: draw-to-face, push/pull solids, typed dimensions, groups and components; it imports .skp geometry and exports OBJ without materials, and textures and scenes are unsupported. Work saves only to the local browser, no account, no cloud. If you occasionally need to sketch a shelf or a cabinet without paying for SketchUp, the web page is all you need. Site · HN discussion
28. Dux, a full open-source Wayland desktop, MIT-licensed
Linux desktops ship in pieces. This one ships as a set?
Dux is a suite of open-source Wayland desktop components, 307 points, 114 comments, MIT-licensed at github.com/duxlabs-dev/dux-shell, currently at DMS 2.1. The suite includes a dynamically themed shell, Dux Files (a keyboard-first file manager), Dux GOP system and process monitoring, a greeter, file search, and a calendar supporting Google, Microsoft, CalDAV, and iCloud. Its theme engine derives a 16-color palette with an in-house “dank16” algorithm and syncs GTK3/GTK4, Qt, and terminal colors automatically; the shell targets niri, Hyprland, Sway, and other compositors, and the project is funded by 115+ Ko-fi supporters. If you run a Wayland desktop and are tired of gluing pieces together, this is a working whole. Site · HN discussion
29. 814 comments on November’s “what are you working on?”
Everyone else’s side project is how far along?
HN’s monthly thread drew 232 points and 814 comments, 3.5 comments per upvote. This month’s concrete entries: kube-dc, an enterprise private-cloud platform on the Kubernetes API with 46 custom controllers deployed across 12 datacenters, aiming for one DevOps per 300 people including AI agents; OneBusAway, open-source transit software said to serve millions of riders daily across New York, Seattle, and San Diego, with UCSD using it for 35,000+ students; and a from-scratch C browser built with no libraries, TLS and PNG/JPEG decoding included, at roughly 68,000 lines and 4,822 commits, planned to go open source early next year. If you want side-project ideas or just a look at unusual builds, the thread is worth climbing once a month. HN discussion
30. No successful Tesla product launch since the CTO walked out in 2019
Missing the product, or missing the person who said no?
A CleanTechnica commentary from November 7, 200 points, 269 comments: Tesla has had no successful new product launch since former CTO JB Straubel stepped down in July 2019. The core case is the Cybertruck, announced in November 2019 and delivered more than four years late; social-media figures quoted by the piece (and flagged by its author as unverified) put the line at 125k units a year of capacity, about 70k sold over three years, and a $1.9B net loss, against an original target of 250k a year. The hit products all shipped on Straubel’s watch (Model S, Model 3, Model Y, and the Megapack), and the affordable car never arrived. It’s commentary, not reporting, but “who counterweights the CEO” applies to every tech company. Article · HN discussion
31. GitHub went down again while the status page stayed green
Degraded performance, you say?
The HN thread “GitHub Down Again?” drew 121 points, 131 comments: widespread 504 and 503 errors, logged-out users hit hardest, with release downloads, the GitHub Actions API, and repo pages all failing; one user reported CI and Terraform dead in the water, and a European user clocked the outage at roughly two hours. The status page showed all green at first, then described the incident as “investigating degraded performance”; the incident page is githubstatus.com/incidents/pyynvrm1vskw. In the comments, one user confessed to having an agent commit to a private repo hundreds of times a day, and another claimed platform-wide commits grew about 5× over a year. If your CI is pinned to GitHub, mirrors and spare runners belong on the schedule. HN discussion
32. UK Lords to propose mandatory CSAM-scanning software on smartphones
This time the justification is hard to argue with, right?
The Observer reports, 86 points, 88 comments: a cross-party group of House of Lords members will propose requiring smartphone makers to preinstall CSAM-detection software, with a cited poll claiming about 90% of voters support it. The scheme could scan all photos and videos on a device, with older phones covered via forced updates. Opposition centers on false positives (Google locked a father’s account in 2022 over medical photos of his son’s groin) and on precedent: Apple abandoned a similar scanning plan in 2021 under pressure, and commenters cite the UK Post Office Horizon scandal as what happens when buggy software feeds law enforcement. The proposal is not law yet. If you build end-to-end encryption, cloud photo storage, or sell into the UK, watch where this goes. Report · HN discussion
33. Windows 11 drops PC-to-PC migration and points you to OneDrive
Another local thing that quietly became a cloud thing?
ghacks reported on November 3, 21 points, 9 comments: Microsoft removed the PC-to-PC migration feature from Windows 11 and points users to the OneDrive-based Windows Backup instead. The thread’s dominant reaction is resentment: direct machine-to-machine copying is gone, and the replacement path runs through a Microsoft account and cloud storage. If you regularly set up Windows machines for family or friends, stage a network share or a third-party migration tool before the day arrives. Report · HN discussion