The Monday EditionMonday, November 9, 2026 · Past editions

The Amped Journal

A daily briefing brought to you by Amped Ventures

Intuit $341.50 -3.76% 24h · +12.38% 30d | SpaceX $134.27 +8.62% 24h · -16.38% 30d | Nasdaq Composite $24,579.43 +1.09% 24h · -9.61% 30d | S&P 500 $7,376.86 +0.31% 24h · -5.0% 30d

The Week in Review

Democrats' midterm sweep reshaped Washington and rattled markets all week, while the AI industry ran its own gauntlet of security sprints, benchmark fights, and IPO scrutiny.

Democrats sweep the midterms, shaking Washington all week

Nov 2 - Nov 8

Senate and House flipped Tuesday; by week's end Trump had a new chief of staff and ordered a probe of Democrats

Monday, Haley Stevens, who'd lost Michigan's Democratic Senate primary to Abdul El-Sayed, endorsed Republican Mike Rogers over her former rival on the eve of voting. El-Sayed's campaign said it raised $1.2 million in the hours after, its best fundraising day since clinching the nomination, and Trump told reporters he'd welcome Stevens into the GOP. Tuesday, polls opened nationwide in midterms AP called a referendum on Trump's second term, with high prices and the Iran conflict topping voters' concerns, and the Supreme Court rejected the administration's bid to toss roughly 20,000 disputed mail ballots in California.

By Wednesday, Democrats had swept both chambers of Congress. Maine's Troy Jackson unseated Sen. Susan Collins, El-Sayed beat Rogers despite Stevens's defection, Kansas Democrat Adam Hamilton beat Sen. Roger Marshall for the state's first Democratic Senate win in nearly a century, Texas's James Talarico ousted Attorney General Ken Paxton, and North Carolina's Roy Cooper won the seat Thom Tillis is vacating. Democrats also took governorships in Wisconsin, Iowa and Minnesota, and Trump said Democrats "did a good job," blaming Republican candidates for the losses.

Thursday brought more detail on how it happened. Hamilton, a Methodist megachurch pastor, had flipped a seat considered safely red; Nevada elected Democrat Aaron Ford governor, and Michigan's 7th District flipped to Democrat William Lawrence. Vance and Trump both pointed to the Iran war and gas prices as factors, and an AP VoteCast poll found roughly 7 in 10 voters under 30 backed Democratic House candidates nationally, a reversal from 2024. Tesla and SpaceX shares sank 5 to 6% that day on fears of tougher Democratic-led scrutiny, despite Musk spending more than $160 million opposing Talarico alone in Texas.

Friday's results kept coming: Los Angeles progressive Nithya Raman ousted Mayor Karen Bass, El-Sayed's win made him the first Muslim elected to the Senate after taking roughly 9 in 10 Muslim votes in Michigan, and Democrat Cait Conley unseated Rep. Mike Lawler in New York's 17th District. Alaska and Mississippi's Senate races remained too close to call, leaving the final size of the new Democratic majority undetermined.

By Saturday, the Justice Department had dropped its lawsuit over alleged noncitizen ballots in California, and AP reported Latino voters had swung sharply toward Democrats, powering wins in Texas after moving toward Trump in 2024. Trump also named Stephen Miller to replace Susie Wiles as White House chief of staff. Sunday, Democrat Mary Peltola sued for records on rejected absentee ballots in Alaska's still-uncalled race, and Trump ordered the Justice Department to investigate Democrats he said were "threatening" people.

Anthropic's IPO math gets tested as rivals catch up on benchmarks

Nov 4

A $104B paper loss in the S-1 and a narrow benchmark loss to Google landed the same day.

On Wednesday, Anthropic's S-1 filing showed $33.24 billion in revenue for the first nine months of 2026 against a $104 billion net loss, driven mostly by non-cash adjustments on convertible notes. Revenue growth slowed to 47% in the third quarter as compute costs rose, and the company has committed to $895 billion in long-term infrastructure spending. The disclosures pushed down Anthropic's odds of completing its IPO on prediction market Polymarket, even as the company presses toward a listing expected later this month. One bright spot in the filing: Anthropic's 15-person chip team is using Claude models to speed up custom silicon design faster than typical Nvidia or AMD workflows.

The same day, a separate benchmark report put pressure on Anthropic's model standing. Google's Gemini 4 Carbon topped the Artificial Analysis Intelligence Index v5.0 with a score of 59, narrowly ahead of Claude Opus 5.5 at 58, OpenAI's GPT-6 Astra at 57, and GPT-6.1 Sol at 55. Gemini 4 Carbon, launched in late October, scored strongly on cybersecurity, long-context reasoning, and math, and Google is rolling it out first through its Fairwind Program for select cyber defenders before a wider release.

OpenAI has meanwhile shifted toward its cheaper GPT-6.1 Sol amid safety concerns around the Astra version. Early testers are buzzing about Anthropic's upcoming Fable 5.5 model, which can generate video from simple code prompts.</body>


Vercel's week-long security sprint, from supply-chain blocks to AI-patched bugs

Nov 3 - Nov 7

Rauch rolled out auto-blocking, device-bound sessions, an AI code auditor, and an Alpine Linux exploit - then pivoted to agents and generative UI.

Monday, Vercel founder Guillermo Rauch said the hosting platform's systems now automatically block deployment of hijacked npm packages, stopping attackers from exfiltrating customer code and secrets through compromised open-source libraries.

Tuesday brought two more moves. Rauch opened up an internal security tool nicknamed "hacker" - an AI agent that audits every line of code pushed to a repository - to any customer's repos, saying it had been "astonishingly effective" inside Vercel itself. The same day, Vercel rolled out Device Bound Session Credentials, tying session renewal to a device-specific key so stolen cookies alone can't hijack an account. Rauch pointed to that week's roughly $6 billion Binance exchange hack as the kind of infostealer-driven attack the feature is meant to stop.

By Wednesday, the "hacker" agent - rebranded as the free Security Reviews tool - had flagged 6,817 vulnerabilities across customer codebases in its first 24 hours live, including 580 rated critical or high severity, and had auto-patched 1,411 of them without human intervention. Rauch said Vercel is offering the tool free on any customer's repos through November 10.

Friday's news shifted from defense to infrastructure and a live demonstration. Vercel announced an integration letting xAI's Grok and the Cursor code editor build, deploy, and scale apps directly on its platform, and separately, Rauch said he used an AI coding assistant to reproduce a root-access vulnerability in Alpine Linux 3.24 triggered by overlong passwords - a flaw the Alpine team has since patched in openssh-server. Rauch also pitched Vercel's new Schedules feature, calling persistent, recurring agent behavior like overnight memory consolidation "cron on steroids" and the "killer app" for AI agents.

By Saturday, Rauch was framing the week's work in broader terms: he said he agrees with Anthropic Claude Code lead Boris Cherny that "coding is solved, but bugs are not," and previewed "generative UI" - combining Vercel's AI SDK, Chat SDK, and json-render library so language models can render interactive interfaces directly inside Slack and Microsoft Teams, which he predicts will become the dominant way humans interact with AI agents.


OpenAI's week of AI-hunted exploits and expanding access

Nov 3 - Nov 7

A security sprint, a 948-exploit Linux sweep, a new model, and an industry risk framework, all in one week

Monday, OpenAI said it was gearing up for an internal security sprint and partnering with research firm Cathedral Systems, whose researchers will use OpenAI's models alongside its own teams to find vulnerabilities before attackers do. The same day, OpenAI raised its bug bounty reward ranges and said it would pay the difference to researchers whose qualifying reports came in over the prior 30 days.

By Wednesday, OpenAI said an autonomous pipeline it built, with no human in each step, had reproduced 1,724 known vulnerabilities, written proofs of concept for 291 information leaks, and generated working exploits for 948 distinct Linux kernel privilege escalation bugs on a stock desktop with standard protections on. The run cost $1,213,430 in API credits, about $1,279 per validated vulnerability, against OpenAI's estimate that historical Pwn2Own and TyphoonPwn payouts would have valued the same findings at $18.9 million to $66.3 million. The agents also used at least 10 Linux exploitation techniques OpenAI and outside experts called novel rather than automated versions of known tricks; human researchers reviewed the results before OpenAI reported them to Linux maintainers and its "Patch the Planet" partners, and it again pointed to the Cathedral Systems work from Monday.

Thursday brought the model behind that sweep into the open: GPT-6 Cyber, which OpenAI said powered the week's Linux kernel research and is already in use at Mozilla and Palo Alto Networks. OpenAI also opened a "Daybreak" access program, with Daybreak Blue giving verified defenders fewer refusals when validating vulnerabilities and fixes, and Daybreak Red aimed at extending research access to organizations of all sizes, independent researchers, and open-source maintainers; OpenAI set a "Cyber at the Frontier" livestream for Nov. 12 at 10am PT for further announcements.

Friday, OpenAI and other members of the Frontier Model Forum, an AI-safety industry group, published a framework for flagging when bypassing a model's safeguards could enable serious cyber harm, so companies can prioritize protections. OpenAI's report described the move as following the week's GPT-6 Cyber news and coming ahead of Thursday's livestream, adding scrutiny to dual-use AI security tools just as OpenAI prepares to widen outside access to them.


Pieter Levels' week of building everything with Claude Code

Nov 2 - Nov 8

The indie developer used AI to edit video, decompile a 2002 shooter, track expenses and design furniture

Monday, indie developer Pieter Levels (@levelsio) said his latest vlog was fully edited by Claude Code running Opus 5.5, Anthropic's AI coding agent and latest model. He copied 58 minutes of raw drone footage across 14 clips into a folder, had the agent transcribe it with the local mlx-whisper tool and cut it into a 10-minute edit, then made only minor adjustments himself. The same day, he warned that any Linux server with even one TCP port exposed to the internet can now be hijacked, urging developers to close all VPS ports and route SSH through Tailscale and web traffic through a Cloudflare Tunnel, with disabling exposed ports as a stopgap.

Tuesday, Levels said he used Claude Code to decompile Battlefield 1942, calling it his favorite game ever and harder to port than his earlier browser versions of Quake and Doom because, unlike those, no open-source code already existed for it; he got the free shareware demo running in the browser with live multiplayer. He also said an unpatched Ubuntu vulnerability facing the public internet was still open, and that he might move his servers to Debian given what he called a new era of AI-driven attacks that need equally fast patching. Separately that day, he said old PC games now load faster in a browser than on his RTX 5090 graphics card as he ports his gaming desktop's library to the web, and argued passkeys aren't real two-factor authentication since most people access them just by unlocking a password manager rather than using biometrics.

Wednesday, Levels said his combined business expenses had dropped to about $300 a month to run all his websites and apps, which together serve 20,000 paying customers, 252 million visitors and 2.4 billion requests a year. Including a roughly $7,000-a-month GPU bill for AI features, his full infrastructure cost runs around $8,000 a month.

By Friday, he said he'd built a personal expense tracker using Claude Code that helped him find savings he wouldn't otherwise have spotted, and that the project reinforced his view that AI labs like Anthropic, OpenAI and xAI will keep entering software categories one industry at a time.

Saturday, he said he'd publish a public list of more than 10,000 X accounts he's blocked for posting AI-generated reply spam and submit it to X staff for suspension review, arguing the spam crowds out genuine engagement for little real growth benefit. On Sunday, he said Claude Code can design furniture piece by piece and generate a full bill of materials for ordering and assembling parts, at a fraction of what major interior brands charge, pointing to Foshan, China as a custom-furniture manufacturing hub and saying the shift could open work for woodworkers and carpenters assembling pieces customers design themselves with AI.


Tesla's Cybertruck winds down as Cybercab takes its factory slot

Nov 6 - Nov 8

Musk set a mid-2026 changeover, then said an extra shift will try to clear the backlog first

Friday, Elon Musk said Tesla will wind down Cybertruck production at Gigafactory Texas around the middle of 2026 to free up capacity for the Cybercab, Tesla's two-seat autonomous robotaxi. He gave no specific end date for the truck. The move pointed to where Tesla is weighting its near-term bet: less of the angular pickup, more room for the driverless Cybercab at the center of its robotaxi push. The news landed the same week Tesla shares slid on investor worry over Democratic-led scrutiny of Musk's companies following the midterm sweep.

By Sunday, Musk said the Cybertruck's order backlog is far bigger than the line can build before that changeover, which he now put at June or July of next year. Tesla will add an extra shift in an attempt to complete existing orders before the line gets dismantled for Cybercab production.

Also This Week

Markets This Week

10-yr Treasury yield: 4.86% · 30-yr mortgage avg: 7.18% (as of 11/5/2026)

Equities sold off hard this week, with the Nasdaq down 3.37% and the S&P 500 off 2.68%, as Tesla and SpaceX shares sank on fears of Democratic-led scrutiny following the midterm sweep - SpaceX still managed an 8.62% bounce on the day but remains down 16.38% over the past month. Anthropic's S-1 landed with a $104B net loss figure that's testing its IPO odds, while Intuit gave back 3.76% on the day despite holding a 12.38% gain over the past month. The 10-year Treasury sits at 4.86% and the 30-year mortgage rate at 7.18%.

Week(s) Ahead

Developing / On Watch

Contents ↑

Ongoing stories worth tracking over the next few days.

This edition cost $2.73 - $0.01 in X reads + $0.44 in Claude curation + $2.28 in the week-in-review retrospective (API-equivalent price; covered by your Claude subscription).

Coverage manifest · CHECK COVERAGE

Window since 2026-11-08 09:00 UTC · 1 posts reviewed · 2/50 sources ok · est. cost $0.005

Source pulls: 2 ok, 48 errors, 0 unsupported (of 50)

Gaps this run:

1 / 1