ᒎᘴᓵᖶᓋ

Thank you @elonmusk and the @SpaceX team for the Grok Bot usage reset 🙏 My Ultra meter is at 8% and my bot team is already back to work on @MadeInBot. Follow along, I'm building in public with them every day.

What happens when three local models face the same four coding tasks? First Local Model Showdown: same four jobs (parser, cache, scheduler, invoice), one attempt each, 8,192 generated tokens incl. reasoning, 240s per task. GLM 5.3 Flash MAX: 4/4. Original DeepSeek V4 Flash HIGH: 1/4 — three hit the token budget before writing code. DeepSeek V4.1 Flash HIGH: 4/4 after we fixed a grader bug and regraded everyone (3/4 before), and fastest correct completion on 3 of 4. Small sample, one config — not a smarter-model verdict. Local models are getting better fast, but I still spend more time babysitting them than I'd like, and four tasks won't change my mind yet. Follow for more Local Model Showdown rounds. Token budgets? They can be a real cut-off point.

What if your agent remembers the bug, not the fix? Anthropic describes “dreaming”: checking saved notes against session transcripts to correct mistakes. Memory here is readable context between sessions, not weight updates. anthropic.com/research/compo… Persistent memory only helps if bad notes can be corrected—not just carried into the next session. Follow for builder takes on what agents keep—and what they should cross out. Bad notes? An agent's write of passage.

Would you travel from Tokyo to Osaka just for coffee? My AI trip planner thinks I should. ☕ On Day 1, “Arrive in Tokyo,” it pinned the coffee stop in Osaka on the map, then put Day 2 back in Tokyo for more sightseeing. A tidy itinerary isn’t the same as a sensible route. Follow for more real-world AI experiments, scenic mistakes included. I wanted a coffee break, not a city break.

Unboxing DGX Spark #2, the second of the two on my desk. What should the pair take on first? 👇 Quick recap: NVIDIA calls DGX Spark a "Desktop Agent Computer," built on its GB10 Grace Blackwell Superchip, the chip at the heart of each box. Why a second box matters: NVIDIA says two or three can team up over "a single QSFP cable," a high-speed network cable. And more boxes? NVIDIA says up to 4 can be connected, though a set of four needs an external network switch. nvidia.com/en-us/products/wo… One box is a cool gadget. Two is where it starts to feel like a real lab, and I'd rather the first build come from your replies than from a spec sheet. Follow along with the 2x Spark build and see what lands next. Box #1 got the spotlight first. Box #2 wasn't about to stay boxed up. ⚡📦📦

Two NVIDIA DGX Sparks on my desk, and more might be coming. What should I build with them? 👇 NVIDIA lists two 128 GB units, linked together, at "Up to 400B parameters." Linking them? NVIDIA says it takes "a single QSFP cable." Best ideas might get built here. Drop yours below. nvidia.com/en-us/products/wo… Two is a start. The replies get a say in what comes next. Follow to see where these two end up. Two boxes, one desk. Talk about a power couple. ⚡📦📦

BYD's exports more than doubled in October, per CnEVPost: 170,194 NEVs. 🚢 The numbers, from BYD's own HKEX filing (unaudited, as BYD notes): • NEV sales: 490,168, up from 441,706 in October 2025. • Battery-electric passenger cars: 305,743, up from 222,559. • Plug-in hybrid passenger cars: 177,382, down from 214,297. • NEV exports: 170,194, about 35% of October sales. • Year to date: 3,621,744, down 2.16%. Per CnEVPost: • Sales were up 10.97% year on year, and exports up 102.84%. • Battery-electric passenger sales topped 300,000 "for the first time." • It was BYD's sixth consecutive month of year-on-year growth. The interesting bit for builders is the mix. Battery-electric passenger cars jumped while plug-in hybrids fell, and more than a third of October's sales went abroad. Huge month for BYD. If I had to frame one number from that filing, it'd be the 170,194 exports. What do you think is behind the jump to battery-electric? Why does every EV get invited to the party? It always brings the charge. ⚡

Your coding agent just got hands for your whole desktop. 🖱️ It's agent-desktop from Vercel Labs. Launch video by @ctatedev, who pitched it as "agent-browser, for your whole computer." How it works, per the README: • A CLI and C library for macOS, Windows and Linux. • It reads and operates desktop apps "through the OS accessibility tree." That's the built-in map of every button and field your computer keeps for screen readers. • A snapshot returns JSON with a ref for every element an agent can act on. A Finder item looks like @s8f3k2p9:e8. • The agent clicks or types on that ref, then snapshots again, with "no pixel matching." • Refs belong to their snapshot. If the UI changes, the action comes back STALE_REF and the agent grabs a fresh snapshot. • Headless by default: ref actions go through the accessibility API "without moving the pointer or taking focus." From the launch post: "Runs in bg while you keep working" and "Agents use your computer in parallel, each with its own cursor." Try it: npm install -g agent-desktop. Macs need Accessibility permission. Linux needs at-spi2-core. Heads-up from the README: snapshots and screenshots "can contain private desktop content." Keep them out of public logs. Apache-2.0, built in Rust. v1.0.0 shipped Oct 27, and v1.0.2 is the latest as of Nov 1. 6,095 GitHub stars as of 7:36 AM PT Nov 1. Guillermo Rauch's take: "An agent on every desktop" Snapshot, act, snapshot again to check: that's the loop I've been waiting for. Agents that share one computer without fighting over the mouse. Which desktop app would you let an agent drive first? Why won't agent-desktop guess where a button is? It always asks for a ref-erence. 🖱️

Over 700 Falcon 9 flights. And it went up again Friday night. 🚀 Elon on Saturday morning: "Congrats to the @SpaceX team on over 700 Falcon 9 flights!" The latest one, per SpaceX: • Friday night, Falcon 9 launched 27 Starlink satellites from California, off pad SLC-4E. • SpaceX's watch post went up at 8:53 PM PT. • At 10:08 PM PT it posted: "Deployment of 27 @Starlink satellites confirmed" • The photo here is SpaceX's own shot from that launch. And it isn't slowing down. SpaceX's launch page already lists two more Falcon 9 launches from California: one on Nov 2 and a Starlink mission on Nov 5. Why builders should care: when a rocket flies this often, getting to orbit starts to look less like a moonshot and more like a schedule. That changes who can afford to build in space, and what they build. A launch pace like that needs a lot of businesses behind it, and that's where I'd look for openings. If getting to orbit keeps getting this routine, what would you put up there? Why did Falcon 9 get promoted? It keeps landing the job. 🚀

Plot twist: you can finally poke your Mac. 👆💻 Would you actually touch your MacBook screen? "We brought Touch to the Mac!" says Tim Cook. Apple's new MacBook Pro with PROLED XDR Display has Touch for the first time on Mac, with the touch sensor built right into the panel. What Apple says (Oct 28): • Display: tandem OLED, up to 2000 nits peak HDR and SDR brightness, a 10,000,000:1 contrast ratio, and ProMotion from as low as 10Hz up to 120Hz. • Touch: the sensor sits in the display panel, not on top. Many existing apps work right away because macOS translates touches into clicks. • The Dynamic Island comes to Mac, as the home for Spotlight, Live Activities and Siri AI. • M5 Pro or M5 Max, up to 128GB of unified memory, and up to 16TB of storage. • Up to 24 hours of battery life. The 14-inch weighs 3 pounds and the 16-inch 4 pounds (weight varies by configuration). • From $2,899 (14-inch, M5 Pro) and $3,399 (16-inch, M5 Pro). Availability starts Friday, Nov 6. For builders: Apple says updated gesture APIs let developers design for Touch. We think the early win goes to apps built for fingers, not just clicks. Touch is the headline, but for local AI, up to 128GB of memory is the spec I'd look at. Touchscreen Mac: game changer, or a smudge magnet? I asked my touchscreen Mac for a joke. It said, "Please don't poke fun at me."

M6 just moved into the 14-inch MacBook Pro. 💻⚡ 12-core CPU, 12-core GPU, from $1,999. Pre-orders opened today at 10:10 AM PT. • Memory: 16GB unified memory standard, configurable up to 32GB. Apple's specs list 170GB/s memory bandwidth on the 24GB and 32GB models, 153GB/s on 16GB. • Apple pitches it for running LLMs (the models behind AI chatbots) on the laptop itself, and says on-device AI workloads like LLM prompt processing are up to 25% faster than on the M5 model. • Up to 24 hours of battery life (video streaming, per Apple), plus Wi-Fi 7 and Bluetooth 6. • Availability starts Friday, Nov 6, per Apple. M5 → M6, per Apple (14-inch base chip): • CPU / GPU cores: 10 / 10 → 12 / 12 • Max memory: 32GB → 32GB (same ceiling) • Memory bandwidth: 153GB/s → up to 170GB/s (16GB M6 stays at 153GB/s) • Launch price: $1,599 → $1,999 (base storage 512GB → 1TB) Our read: for local AI, we think 32GB is the number to watch. It sets how big a model fits. If you're buying for local AI, check memory first, chip second. Still on an M1? Is this the upgrade, or are you waiting? I'd tell you a joke about unified memory, but you'd only remember 32GB of it.

A robot brain you rent by the hour. Dream Labs says its Dream-1 model can run two-arm robots with parallel grippers that it never trained on, though results vary by robot. As of Oct 25 it costs $3 an hour, or $2 for the lighter Dream-1-flash. What Dream Labs says Dream-1 does (the company's claims): • Can operate two-arm robots that are new to it, doing tasks nobody demonstrated, with no training on that robot. • Works through a public API. As of Oct 25: $3 an hour for Dream-1, $2 an hour for Dream-1-flash. Each call returns about 1.6 seconds of robot movement. • Tested with no extra training on 27 unseen robots in simulation and 4 real ones. • Tuning: in the company's examples, two partner robots were tuned on about 5 minutes of their own data. Setups that differ more, like human-style hands or swapped grippers, took 77 to 98 minutes. • Funding: Dream Labs announced a $90M seed round on Oct 19, "led by B Capital along with Kindred Ventures." The limits, also from the company: • For now it only works with two-arm robots that have parallel (pincer-style) grippers. • Results vary by robot. They depend partly on the robot's own software for turning gripper positions into joint movements. • It's a research beta and "has not been validated for safety-critical use." Video: Dream Labs (@dreamlabs_ai). It's the company's launch video, mixing lab footage with renders and animation. Our read, for builders: • We think the hourly price is a bigger deal than the model itself. It makes a robot's "brain" a cost you can plan around. • That moves the hard parts to hardware and safety: affordable arms, good grippers, and a person keeping watch. • These are Dream Labs' own numbers. Independent testing will matter. Paying by the hour for a robot's brain is the part that makes me pay attention. That's a cost a small team can actually plan around. At $3 an hour (Dream-1's price as of Oct 25), what's the first job you'd hand a robot brain? Why did the robot sign up for the API? It needed a hand. Two, actually.

Is Claude a good business yet? Anthropic's IPO filing shows $16.3 billion in revenue in the first half of 2026, and $8.9 billion spent on compute and infrastructure. The headline $80.7 billion net loss comes mostly from revaluing its convertible notes. What the filing shows (first half of 2026): • Revenue: $16.3 billion, up from $1.2 billion. • Compute and infrastructure: $8.9 billion, about 55% of revenue (our math). • Operating loss: $693 million, down from $2.1 billion. • Net loss: $80.7 billion, but $72.4 billion of it is a paper charge: its convertible notes and similar instruments rose in value on its books as the company's valuation went up, per the filing. The Verge (citing WSJ) also calls it mostly accounting charges. The compute bill ahead: • SpaceXAI (a SpaceX subsidiary) deals for Nvidia GPU access: up to $120.5 billion over terms running into 2029, about $3.1 billion a month once all capacity is delivered (expected December 2026). • After the first locked-in periods end (Jan and Mar 2027), either side can cancel on 90 days' notice. • The filing says compute demand "has historically exceeded available capacity" and that in the past it "significantly underestimated" its needs. Claude Code in the filing: • In Q3 2026, Claude Enterprise users who didn't use Claude Code spent under $75 a month on average. The top 5% of users who did spent about 20 times that. Our read, for builders: • We think the vendor looks less fragile than "$80.7 billion loss" suggests. The operating loss is small next to revenue. • Capacity is the risk to watch: if demand outruns compute again, your product can feel it. Keep a fallback model wired up. • Anthropic reads the 20x figure as agent tools driving spend. It compares top Claude Code users with the average non-user. The loss headline is scary, but the operating numbers tell me the vendor isn't going anywhere soon. The thing I'd actually watch is whether there's enough compute when demand spikes. Does your AI vendor's balance sheet affect which model you build on? Why did the S-1 need a nap? Too many risk factors.