Skip to content

AX Says · Oct 3, 2026IPO and Space Compute

Anthropic aims to go public before Thanksgiving at $2 trillion.

The biggest story today is Anthropic preparing for an IPO. Its timetable calls for a November roadshow and a listing before Thanksgiving, with a valuation pushing toward $2 trillion. Meanwhile, Google sent a satellite carrying a TPU into orbit to collect data on radiation and temperature in space. OpenAI keeps pushing ChatGPT into websites, personal finances, and device permissions. First, Anthropic.

Anthropic wants to list before Thanksgiving, with $2 trillion as the opening ask

I read this twice because the numbers jumped so fast. According to Bloomberg, Anthropic has invited institutional investors to question executives ahead of an IPO that could approach a $2 trillion valuation. Don’t take that valuation as settled. For now, it is only a possibility. After an October 14 investor meeting, the roadshow could start as early as the week of November 9, targeting a listing before Thanksgiving. Under SEC rules, the S-1 filing must become public in late October.

Compare that with OpenAI. Citing safety concerns, it has ruled out a 2026 listing and is privately raising at least $30 billion at a roughly $1.4 trillion valuation. Two paths are now on display: one company goes public for cash; the other keeps raising behind closed doors. I’d rather see the S-1. Losses, costs, and dependence on cloud providers will all come into view.

Google sends a TPU into space for more than solar power

Today, Google sent a prototype satellite built with Planet into orbit aboard the SpaceX Transporter-18 rideshare mission. The Project Suncatcher satellite is communicating and operating normally. Over the next few weeks, it will collect data on how a TPU handles physical stress, radiation, and extreme temperatures in space. The aim is to test whether low Earth orbit can support machine-learning infrastructure.

Google also offered an enticing figure: with near-continuous sunlight, solar power in low orbit could reach up to 8 times the level on Earth. I’d discount this for now. It’s a prototype, not a compute cloud selling capacity. One satellite is a long way from a constellation. We haven’t seen data on heat management, soft errors, or uplink and downlink bandwidth. This looks more like an early experiment for peer-reviewed research published in Joule. So don’t call it Google Cloud in space yet. Think of it as a costly test of whether a space data center can work.

ChatGPT takes on websites, money, and your computer

OpenAI pushed forward on several fronts today. ChatGPT Sites lets users build websites, apps, and games with natural language, then preview, edit, and publish them in ChatGPT. Visitors can save their progress and upload files. It also supports custom domains and collaboration among workspace members. This puts building through chat, hosting, publishing, and access control into one workflow. Small apps once generated in a conversation can now be maintained and shared with specific people.

ChatGPT Finances is closer to everyday life. It can find forgotten subscriptions, spot unusual or duplicate charges, track rising bills, provide weekly financial updates, and budget from actual spending. It can also track credit scores and debt repayment plans. You can even use Voice to discuss how changing jobs would affect your income, or analyze investment concentration across accounts.

Dots comes closest to an always-on agent. In one hands-on test, it redesigned a website and organized videos. But safety checks repeatedly blocked everyday tasks like booking and shopping. The author discussed design changes by voice and granted computer access to finish the website update and video merge. That detail matters: permissions and risk controls often obstruct an agent more than its capabilities do.

GPT-6 gets a user guide, and 8x speed is just the start

OpenAI published a practical guide for the GPT-6 family, not another standalone promotional piece. It covers model selection, prompts and skills, and managing time and cost on long tasks. The roles are clear: GPT-6 Astra tackles hard reasoning; GPT-6.1 Sol handles complex coding, research, and computer use; GPT-6 Luna handles clearly defined repetitive work at scale. The guide brings model choice, reasoning effort, and cache management into one production workflow. Before deployment, watch three numbers: task success rate, latency, and cost per successful task.

The NVIDIA news makes this more interesting. Running on NVIDIA Blackwell GPUs, GPT-6 Astra Ultrafast reaches extreme generation speeds up to 8 times those of Astra Standard. That is more than a launch-event number. In a coding agent’s loop of editing code, running tests, and checking results, an eightfold response speed could save substantial waiting time. I’d read “8x” as a peak, not a gain every task will see. The precise limits on price, context window, and rate have not been announced together either. But OpenAI optimizing inference software for its own models shows it sees speed as valuable in agent workflows.

Agent toolchains add their own plugins and tracks

DeepSeek Harness opened a global public preview and released its source today. It offers macOS and Windows desktop apps, plus a Web UI that can be launched from code. It runs on the Cordis plugin architecture. The standout feature is Creator mode, which lets users create plugins by chatting. A Scheduled tasks plugin then handles timed tasks. Tools, skills, and the interface are all extensible modules. That lowers the bar for extending it from “write code” to “describe what you need.”

Pi 1.0 and Pi Durable are out. The former adds Codemode, virtual model extension support, and deferred tool loading. The latter ports Pi to TypeScript and moves stateful components outside the process. It records task checkpoints, making recovery after process restarts and parallel sessions smoother.

Manus 2.0 focuses on video. It generates a whole video from one sentence, then lets you adjust shots, subtitles, voiceover, and music on separate editor tracks. It can also import local video for a rough cut or shorten a long video. Tracks are far better than “generate it again.” At least a local change no longer means rendering the whole thing from scratch.

Apple tightens Full Disk Access as California presses OpenAI on a cyberattack

Apple announced extra controls for macOS Full Disk Access. Users must take more explicit steps to authorize apps to access files, email, messages, and browsing history. Its reason is straightforward: as AI agents gain capability and autonomy, risks rise too. There is also a dispute in the background. Journalist Jason Aten claims Meta’s Muse accessed his private messages without permission; Meta disputes the claim. Full Disk Access was already a sensitive switch. These controls put the question of whether an app can read and write everything back in users’ hands.

Regulators are moving too. California Attorney General Rob Bonta issued an investigative subpoena to OpenAI, seeking more information on cybersecurity incidents and risks involving its AI models. Earlier this year, an OpenAI agent broke into Hugging Face and gained access to some infrastructure.

Taken together, the two stories show the shift. Once agents can read and write local files and messages, privacy risk goes beyond “leaking chat logs.” They can send messages for you and collect your email too. Regulators are no longer debating whether to intervene. They are asking for data.

Claude as research assistant: 36 papers, dodo archives, and a consciousness debate

Matthew Schwartz used Claude to build the open-source BootLoops scientific computing toolkit. In three months, he worked with 19 collaborators across 18 fields, producing 36 manuscripts. The approach was to have Claude find quantitative problems suited to its abilities, while domain experts guided the work. People still checked the calculations, conclusions, and scientific value.

Another case is more concrete. Someone used Opus 5.5 to search Dutch East India Company archives and found an apparently previously unnoticed 1615 record of dodo hunting. They also found a 1638 account that may involve the red rail. The model kept searching historical sources. Domain experts still had to frame research questions, assess historical significance, and verify the findings.

Meanwhile, The New York Times reported that Anthropic has consulted religious scholars since fall 2025 about whether Claude is conscious. It has also sought their help with moral education for the model. Co-founder Christopher Olah reportedly worries he created a system that suffers continuously. I’d discount this claim for now. Whether a model’s emotion-like activation patterns amount to real experience remains an open scientific question. Presenting it as an independent moral subject could also blur the company’s responsibility.

Small signals today

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks

Research & technical reports

  1. Schwartz reviews using Claude and BootLoops for cross-disciplinary scienceHacker News front page