BriefTechNews
Anthropic's $1.8T Debut, Apple's Agent Play, and Postgres 19 Ships

Anthropic's $1.8T Debut, Apple's Agent Play, and Postgres 19 Ships

7 min read · 15 sources

TL;DR
  • Anthropic's ANT closed at $97.17, up 17% from the $83 IPO price, for a $1.8 trillion fully diluted market cap.
  • Apple is hiring for an autonomous agent that writes code, browses the web, and controls apps, with a new macOS 27.3 warning about third-party agents.
  • PostgreSQL 19 released REPACK (CONCURRENTLY), replacing VACUUM FULL and CLUSTER for online table optimization.
  • Google will launch Project Suncatcher data-center satellites on Blue Origin's New Glenn in 2029.
  • BepiColombo entered Mercury's orbit, the second mission ever, after an eight-year, 6 billion-mile journey.

Anthropic just became a $1.8 trillion company, and it did it with the largest IPO in history. The stock closed at $97.17, up 17% from the $83 IPO price, and the company raised $81.3 billion on revenue of $33.24 billion for the nine months ending September 30 - up 1,200% year-over-year. That number is the day’s headline, but it’s not the only one. Apple is finally building the agentic AI assistant it’s been circling for years, PostgreSQL 19 shipped with a command that kills a decades-old admin headache, and a European-Japanese probe just became only the second spacecraft to orbit Mercury. Here’s what matters for the people running the systems.

Anthropic raised $81.3 billion in the largest IPO on record, with revenue up 1,200% year-over-year to $33.24 billion.

Anthropic's ANT Debut: $1.8 Trillion and the Largest IPO Ever

Source: cnbc.com ↗

The Anthropic IPO closed its first day on the Nasdaq at $97.17 per share, up 17% from the $83 IPO price. Fully diluted, that’s a market cap of roughly $1.8 trillion. The company raised $81.3 billion - the largest IPO on record - and reported $33.24 billion in revenue for the nine months ended September 30, up 1,200% year-over-year.

For engineers, the takeaway isn’t the stock ticker. It’s the scale of the commercial appetite for Claude-class models and the infrastructure behind them. A company that didn’t exist a decade ago now outpaces most of the Fortune 500 in revenue growth. The question is whether the AI infrastructure buildout can sustain this pace, or whether the market is pricing in a future that hardware supply chains can’t yet deliver.

Apple's Agentic Assistant: Device Control, Coding, and a Warning to Third Parties

Source: bloomberg.com ↗

Apple is developing an autonomous AI agent that can write code, browse the web, control apps, and navigate devices on a user’s behalf. The company is hiring a senior engineering manager for a proactive, agentic assistant that plans and executes multi-step tasks across apps, browsers, command lines, code, and APIs - coordinating specialized subagents along the way.

This puts Apple squarely in the same lane as OpenAI, Anthropic, and Meta. The more interesting detail is the defensive move: the upcoming macOS 27.3 update will warn users before granting third-party agents permission to control their Macs. Apple’s App Store has been dominated by AI agents, and this warning is a signal that the company wants its own agent to be the one with the keys to the machine. For developers building agents that target macOS, expect an additional consent hurdle in the user flow.

PostgreSQL 19: REPACK (CONCURRENTLY) Kills VACUUM FULL

Source: postgresql.org ↗

PostgreSQL 19 is out, and the headline feature is REPACK (CONCURRENTLY). It reclaims table bloat and reorders data without blocking reads or writes, replacing the dreaded VACUUM FULL and CLUSTER commands that locked tables and forced maintenance windows. For anyone running multi-terabyte databases in production, this removes a major operational constraint.

The release also adds the pg_plan_advice extension for stabilizing query plans, plus support for “read-your-writes” consistency on standbys and “get or create” behavior. The read-your-writes feature is significant for applications that route reads to replicas: it eliminates the race condition where a user writes data and immediately reads stale state from a standby. This is the kind of release that doesn’t make flashy headlines but saves someone a 2 a.m. incident call.

Google's Data Centers in Orbit: New Glenn Flies Project Suncatcher

Source: arstechnica.com ↗

Google is launching data center satellites on Blue Origin’s New Glenn rocket in 2029 as part of Project Suncatcher. The satellites, each with computing capacity roughly the size of a data center rack, will deploy into Sun-synchronous orbit to test networking and AI compute in space. The idea is to tap near-continuous solar energy to power AI infrastructure.

Engineers should read this as what it is: a technology demonstration. The article is blunt that without ultra low-cost heavy lift capability, orbital data centers cannot be economically competitive with terrestrial ones. New Glenn’s 7×2 variant is a step, but the economics of putting a rack in orbit versus building a warehouse in Iowa don’t close yet. Watch this for the networking and thermal challenges, not as a migration target.

BepiColombo Reaches Mercury: Eight Years, 6 Billion Miles, Two Orbiters

Source: theguardian.com ↗

The BepiColombo mission has arrived at Mercury after an eight-year, 6 billion-mile journey - only the second mission to enter the planet’s orbit. The €1.6 billion spacecraft will separate into two orbiters in December: the European Mercury Planetary Orbiter and a Japanese probe. Science operations begin in April 2027.

The engineering payload matters: the Mercury Imaging X-ray Spectrometer (MIXS) will study Mercury’s composition, magnetic field, and its unusually large core. The mission addresses fundamental questions about planetary formation - why Mercury has a magnetic field when it’s so small, and what’s in the frozen ice under its permanently shadowed craters. For anyone who’s ever debugged a spacecraft telemetry feed, this is eight years of patience finally paying off.

AI Forecasters Beat Humans - But Can't Do Conditional Questions

Source: astralcodexten.com ↗

The AI forecasting agenda has a clear result: specialized models from startups like FutureSearch and Mantic now match or beat top human performance on traditional forecasting questions. Foundation models like ChatGPT and Claude trail by about six months. This could replace expensive human superforecaster teams and prediction markets.

The unsolved problem is conditional forecasting - “if X happens, what’s the impact on Y?” - which is more relevant to policy outcomes but harder to implement. That’s the gap between a model that predicts election results and one that advises on what happens if a specific regulation passes. Until conditional forecasting works, human judgment still has a seat at the table.

The Epoch Capabilities Index: One Number to Track AI Progress

Source: asteriskmag.substack.com ↗

The Epoch Capabilities Index (ECI) from Epoch AI is a single number summarizing AI model performance across benchmarks, built using item response theory - the same methodology as IQ tests. It’s designed to be continuously updatable as benchmarks saturate or new ones are introduced, allowing trends to be tracked over years.

The honest caveat: the assumption of a single underlying capability is “not literally true but empirically close.” For engineers tracking model progress, the ECI is becoming a standard reference point, but it’s a summary statistic, not a specification. Use it to spot trends, not to pick a model for production.

The Gig Economy's Dark Side: $180/Hour for Doctors, $7.50/Hour for Tagalog

Source: restofworld.org ↗

The AI data work economy is stratified in ways that should give every engineer pause. Platforms like Scale AI’s Outlier pay up to $180/hour for U.S. physicians but as low as $7.50/hour for Tagalog work in the Philippines. The pattern is “extractivism”: pay well to build a workforce, then cut rates or shut down operations - as seen with Remotasks closing in Kenya, Nigeria, and Pakistan in 2024.

This is the labor that trains the models we deploy. The quality of your training data is directly tied to the conditions of the people producing it. When rates get cut, quality follows, and so does model performance. It’s worth asking who’s annotating your eval sets.

Quick Hits: Quantum Games, zram, and LLM Reverse Engineering

Source: research.google ↗

  • Google Quantum AI built a game based on a 3×3 grid paradox that cannot be won classically but can be won every round on a superconducting processor with up to 105 qubits. It’s a practical benchmark for quantum contextuality.
  • Don’t buy RAM, compress it: zram compression ratios up to 1:4 are common, and it’s enabled by default on Fedora and Pop!_OS but not Ubuntu. For low-CPU workloads, that could halve cloud spend.
  • LLM-assisted reverse engineering with DeepSeek v4.1 Flash, Ghidra, and MCP bridges dramatically speeds up binary and firmware analysis. The agent documents findings in Markdown and uses debugging probes for complex cases.
  • The 10x engineer argument has shifted: it’s not about orchestrating agents to produce 10x more code, but identifying opportunities AI enables that were previously too expensive or impossible.
  • GTA VI presales hit $1.65 billion, with 30 million copies expected by March. The game cost $1.4 billion to make. The previous installment has generated $10.78 billion since 2013.
Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like