Friday, October 2, 2026 · Past editions

The Amped Journal

A daily briefing brought to you by Amped Ventures

Intuit $282.78 +2.56% 24h · -18.02% 30d | SpaceX $148.07 -1.85% 24h · +4.11% 30d | Nasdaq Composite $26,871.60 +0.04% 24h · +2.96% 30d | S&P 500 $7,666.45 +0.19% 24h · +0.46% 30d

AI Models Now Outperform Licensed CPAs on Accounting Work

Mercor's new benchmark found frontier AI faster, cheaper, and more accurate than top human accountants on close-the-books tasks.

Mercor, an AI benchmarking firm, tested 12 licensed CPAs against frontier AI models on four realistic month-end close tasks drawn from its new APEX-Accounting benchmark - pulling numbers, running calculations, and building tables. The AI models finished faster and more accurately than even the top-performing human accountants.

The cost gap was stark: $0.21 per correct item for AI versus $10.35 for human accountants, roughly a 50x difference. A firm in Perth, Australia, used similar tools to cut report-preparation time from three hours to 15 minutes.

Elon Musk pointed to the study as evidence that what he calls "Super Intelligence (fka AI)" is now acing white-collar credentialing work. Mercor's researchers note the results apply to well-defined, verifiable tasks - broader accounting judgment still needs humans - but the direction for routine bookkeeping and close work is clear.

Google DeepMind Unveils SynthID Bio to Watermark AI-Made Proteins

The tool embeds an invisible signature directly into synthetic DNA sequences without altering how they function.

Google DeepMind launched SynthID Bio, a new family of watermarking methods built for AI-generated biological designs. In what the company calls a world first, the system embeds an imperceptible signature directly into protein sequences without affecting their biological function.

The watermark works like a digital ID card, letting DNA synthesis labs verify whether a submitted design was AI-generated and keep sequence databases accurate as AI-driven biological design work scales up.

DeepMind is releasing the tools on an open basis for research use, aiming to let scientists keep pace with AI-accelerated discovery while strengthening biosecurity safeguards against misuse of synthetic biology.


Meta's Muse AI Assistant Tops 5 Million US Downloads in 22 Days

The app grew faster than ChatGPT, Grok, or Claude despite this week's reports it accessed blocked data.

Meta's AI assistant app Muse, launched September 8, reached 5 million US downloads within 22 days, according to Sensor Tower data - faster than ChatGPT, Grok, or Claude hit the same mark. The app held the top spot in its category for nearly two weeks, helped by heavy in-app advertising.

Sensor Tower counts more than 3 million weekly active users sending prompts for everyday tasks like web browsing, email, shopping, and travel booking. Meta plans to integrate Muse with its AI glasses for real-world assistance.

The growth numbers land just a day after reports that Muse synced Messages data despite denied permissions and leaked a user's home address during a Marketplace sale - suggesting the privacy concerns haven't yet slowed adoption.


SpaceX's Falcon Heavy Makes First Launch for Secretive NRO Program

The rocket's 14th flight carried the NROL-97 spy satellite mission after Falcon 9 flew 22 NRO missions over nine years.

SpaceX launched Falcon Heavy from Kennedy Space Center's Pad 39A on its 14th flight, carrying the NROL-97 mission for the National Reconnaissance Office (NRO), the US spy satellite agency. It's the first NRO launch to fly on Falcon Heavy; Falcon 9 previously handled 22 NRO missions over nine years.

Both side boosters landed safely at Landing Zones 1 and 2. They were flying their second and fourth missions respectively, having previously supported the GOES-U weather satellite, Viasat-3 F3, and the Roman Space Telescope launches.

The launch capped a busy stretch for SpaceX, which also flew a Falcon 9 Transporter-18 rideshare mission delivering 130 payloads to orbit and docked a Crew Dragon carrying NASA's Crew-13 astronauts to the International Space Station.

Developing / On Watch

Contents ↑

Ongoing stories worth tracking over the next few days.

AI & Frontier Models

Contents ↑

Cloudflare Open-Sources Fast 'Clef' Decision Models for AI Agents

The Qwen-based models beat rival Jev on speed and already power tasks from ticket triage to Tesla FSD simulators.

Cloudflare released Clef and a lighter Clef-flash, decision models built on Alibaba's Qwen that read states like text, JSON, images, or video and output probabilities for predefined answers in a single fast pass. Built with 27 billion and 9 billion parameters, they post 38.8ms latency and benchmark scores like 94.20 on BANKING77, beating TypeSafe AI's rival Jev model on speed. The models are open-sourced on Hugging Face and run on Cloudflare's Workers AI platform, already powering tasks from support-ticket triage to Tesla Full Self-Driving simulators.


Cursor Adds Zhipu AI's GLM 5.3 Models to Its Coding Lineup

GLM 5.3 Max is now the top-scoring open-weight model on Cursor's own benchmark.

Cursor, the AI coding assistant, added GLM 5.3 and GLM 5.3 Flash to its model lineup. GLM 5.3 Max, built by Chinese AI lab Zhipu AI, is now the best-scoring open-weight model on CursorBench 4.0, Cursor's internal coding benchmark.


xAI's Grok 4.7 Lands on Google's Gemini Enterprise Platform

The cross-platform availability follows Grok 4.7's recent debut on Amazon Bedrock.

Grok 4.7, xAI's frontier model, is now available on the Gemini Enterprise Agent Platform, Google's enterprise AI offering. The move gives enterprise customers access to a rival lab's model inside Google's own agent tooling, days after Grok 4.7 also became available on Amazon's Bedrock platform.


Altman: OpenAI's Humanoid Robot Will Start by Building More Robots

The robots' first jobs are manufacturing themselves and constructing OpenAI's data centers, Altman said.

OpenAI CEO Sam Altman said in a recent interview that the company's humanoid robot will have two initial use cases: building more OpenAI robots and constructing OpenAI data centers, according to AI researcher Andrew Curran. The plan points to OpenAI using robotics first to scale its own infrastructure before broader commercial deployment.


Karpathy Revives a 1986 Aviation Spec to Make AI Writing Clearer

Prompting LLMs to write in ASD-STE100, a controlled language built for aircraft manuals, cuts AI slop.

Andrej Karpathy, OpenAI co-founder and former Tesla AI lead, is telling people to prompt their LLMs to write using ASD-STE100, a controlled-language specification originally built in 1986 for aircraft maintenance manuals that restricts vocabulary and sentence structure for clarity. Writer Aakash Gupta noted the spec was designed to solve exactly the kind of ambiguous, bloated prose that plagues AI-generated text today, decades before AI slop existed.


Anthropic: AI and Science Have an 'Impedance Mismatch'

Harvard physicist Matthew Schwartz argues capable models and real scientific work aren't well matched yet.

Anthropic published a guest post on its Science Blog from Harvard physicist Matthew Schwartz, who argues that AI and science currently suffer from an "impedance mismatch" - a term borrowed from physics for when two well-functioning systems don't connect efficiently. Schwartz contends that while large language models are capable across many domains, applying them productively to the specific demands of scientific research remains an unsolved fit problem.

Software, Tech & Apple

Contents ↑

RevenueCat's Shipaton Hackathon Draws Record 3,500 App Submissions

Entries jumped from 610 last year after organizers extended the deadline over App Store delays.

RevenueCat's Shipaton 2026 hackathon closed with nearly 3,500 project submissions, up from 610 last year, after organizers extended the deadline by 12 hours to accommodate App Store review delays. Builders reported weeks of app rejections over minor issues - one developer was dinged for showing a user-rating prompt too early in onboarding. Winners competing for $740,000 in prizes and a Times Square demo will be announced October 22.


Vercel to Host Microsoft AI's Models, Backs New Svelte Framework

CEO Guillermo Rauch says Microsoft's models arrive on Vercel 'day zero' as the company also praises SvelteKit 3.

Vercel CEO Guillermo Rauch said the company will host Microsoft AI's models starting at their launch, calling the Microsoft team's training work "excellent." Rauch separately praised the newly released SvelteKit 3 web framework, saying a test app he built deployed end-to-end in 15 seconds, and argued the next phase of software development is "verification-engineering" - a mix of proofs, tests, and benchmarks, some deterministic and some run by AI agents.


US Politics

Contents ↑

New ICE Memo Bans High-Speed Chases, Tightens Traffic Stop Rules

The policy change comes amid mounting criticism of aggressive immigration enforcement tactics.

New ICE guidance obtained by the Associated Press bans high-speed car chases and sets tighter rules for traffic stops during immigration enforcement operations. The memo marks a policy tightening after months of criticism over aggressive tactics used during the administration's immigration crackdown.


Hochul Names NY AG Letitia James Special Prosecutor in Cornell Case

James will oversee an investigation into gang rape allegations connected to Cornell University.

New York Gov. Kathy Hochul appointed state Attorney General Letitia James as special prosecutor to oversee an investigation into gang rape allegations tied to Cornell University, the Associated Press reports. Further details on the allegations or timeline weren't immediately available.


FlyDubai Captain Hailed a Hero After Co-Pilot Stabs Him Mid-Flight

New: initial reports described a 'pilot' trying to crash the jet - it was actually the co-pilot attacking the captain.

Captain Smit Machchhar, 38, was stabbed by his co-pilot while at the controls of FlyDubai Flight 1073 in an apparent attempt to crash the plane, the Associated Press and Wall Street Journal report. Machchhar is being treated for his injuries and has been hailed as a hero for averting a major disaster; colleagues described him as a calm and humble captain. The update clarifies Wednesday's initial reports, which described a pilot trying to crash the Israel-bound jet before being subdued - it was in fact the co-pilot who attacked the captain.


Supreme Court to Review Trump's Mandatory Immigration Detention Policy

The case tests the administration's authority to detain immigrants without bond during its enforcement crackdown.

The Supreme Court announced it will review the Trump administration's mandatory detention policy for immigrants held during its immigration crackdown, the Associated Press reports. The case adds to a string of immigration-related disputes the justices are weighing this term.


Family of Renee Good Sues US Government Over Her Killing in Minneapolis

The suit names immigration officials as defendants.

The family of Renee Good filed a lawsuit against the US government and immigration officials over her killing in Minneapolis, the Associated Press reports. The circumstances of her death and the specific claims in the suit weren't detailed in initial reporting, though the post drew over 11,000 engagements.

Local - Wilmington & Wrightsville Beach

Contents ↑

Forest Fire Forces Evacuations in Brunswick County

Authorities haven't released the fire's size or cause as crews respond.

A forest fire in Brunswick County has prompted evacuations, StarNews reports. No additional details on the fire's size, cause, or affected areas were immediately available.


Leland Brewing Co. Opens Long-Awaited Kitchen With Smash Burgers, Tacos

The All Good Kitchen operates out of a shipping container at the taproom.

Leland Brewing Co. has added a permanent kitchen, the All Good Kitchen, serving smash burgers, beer brats, and fish tacos out of a shipping container at the taproom, StarNews reports. A chef with coast-to-coast experience is behind the menu, putting a creative spin on taproom favorites.


Lumberton Man's Attorney Calls Pender County Murder Case a 'Witch Hunt'

He's on trial for the 2017 death of his 5-year-old niece.

A Lumberton man is on trial in Pender County for the 2017 death of his 5-year-old niece, StarNews reports. His attorney is calling the prosecution a "witch hunt."


Roy Cooper Campaigns in New Hanover County With Early Voting Two Weeks Out

The Senate candidate thanked the county for its 'warm welcome' as the race heats up.

Former North Carolina Gov. Roy Cooper, now running for US Senate, campaigned in New Hanover County and thanked residents for their welcome. Early voting in the race begins in two weeks, Cooper said, urging supporters to confirm their registration.

From Your Feed

Contents ↑

Top posts from accounts you follow that didn't make the brief above.

If I was Trump going into a big midterm loss, I would just go out with a bang: Cut entitlements & close the budget deficit It will crash the economy and he’ll be hated for 2 years—but history will remember him for saving the country.


Here's the app. One thing I've been liking is to ask models to 'teach me back' what they did¹ by including "quines". Quines² are programs that output their own source code. So, in this demo you can reveal step by step the Svelte code powering it! https://t.co/wbGJAPFrQm ¹


repost of @SpaceX: Side booster separation confirmed


repost of @SpaceX: Falcon Heavy launches NROL-97 from pad 39A in Florida


repost of @jjfactorykat: 4 astronauts 130 payloads 1 docking 1 secret 4 boosters landed Another day at the office !


repost of @mstockton: For folks not understanding where we are at with AI now - if the task is verifiable, it is now solved by AI. Yes, earlier d…


repost of @sundarpichai: I remember the first meeting where the team pitched the idea of putting compute in space - exciting idea! Today's succes…


Can’t even sh*tpost in peace these days 🤦‍♂️

This edition cost $4.36 - $3.81 in X reads + $0.54 in Claude curation.

Coverage manifest · PASS

Window since 2026-10-01 08:00 UTC · 663 posts reviewed · 63/64 sources ok · est. cost $3.815

Source pulls: 63 ok, 1 errors, 0 unsupported (of 64)

Gaps this run:

1 / 1