Anthropic’s IPO filing says its AI may “imminently” hit the self-improvement threshold

😁 Hello, super humans! Companies rarely put “our AI might start building the next AI” into an SEC filing, but Anthropic just did. We unpack what the fine print actually says, what it does not say, and why it matters to anyone shipping agent code. Also in today’s issue: a tape-storage comeback, a telecom going broke, and a snippet to catch agents that pass the task and wreck the workspace.

📰 Quick Signals

• 🧠 AI: WIRED found over 400 apps and sites tied to one network that hid abusive deepfake tools behind ordinary features like pet portraits; Apple removed 19 apps after being contacted.
• 🤖 Robotics: Foundry Robotics raised a $100M Series A to scale AI-driven robotic assembly cells in a 130,000 sq ft California factory for drones, satellites and underwater vehicles.
• 💻 Programming: OpenAI says its new Git backend, Pando, cut median push latency by nearly 80% for its monorepo under heavy concurrent agent work (company-reported).
• ⚡ Electronics: Fujifilm plans to raise LTO-10 tape production fivefold through 2028 as AI demand squeezes archival storage; native capacity is 30 or 40 TB per cartridge.
• 📡 Telecom: Brightspeed named Shawn O’Donnell CEO as it reportedly nears a bankruptcy filing, possibly as soon as December, with about $3B needed to fund it through 2027.

🔍 The Big Story: Anthropic’s IPO filing says its AI may “imminently” hit the self-improvement threshold

When an AI lab tells investors that its own models may soon speed up AI research dramatically, it is worth reading the fine print, not just the headline.

What happened: In an amended S-1 filed with the SEC, Anthropic says its internal estimates for future models indicate it “may imminently meet” the Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI. It also says no current model has crossed that line and that it is assessing the data. TechCrunch’s report points to the filing itself.

The details: The concept is recursive self-improvement (RSI): AI that is good enough at AI research to speed up its own successors. The filing’s risk factors admit Anthropic has, in the past, failed to recognize a capability acceleration until after it had happened, and that using AI to build more capable AI could be a considerable safety risk. Anthropic also says it does not expect to safely let Claude design and build successive frontier generations without meaningful human supervision or understanding, and that it is considering updates to its Responsible Scaling Policy to keep humans in control of model development. Skeptics note an S-1 is both a disclosure and a marketing document, so a “we are close” claim can double as a pitch. The loop worth keeping in your head looks like this:

flowchart LR
    A[Model helps with AI research] --> B[Faster experiments and better code]
    B --> C[More capable next model]
    C --> A
    C --> D{RSP threshold check}
    D -->|below threshold| A
    D -->|threshold met| E[Added safeguards and human oversight]

🗞️ More News

🧠 AI

• Publishers led by The New York Times sued Brave Software, alleging it copied paywalled reporting and sold AI access to it through Brave’s search index; these are allegations, and Brave had not immediately responded.
• A new Windows infostealer sends stolen data to an AI model to rank each victim’s value, can switch models if one refuses, and then deletes itself.
• Goldman Sachs is pulling advanced AI coding tools, including Claude Code, back from traders and bankers and reserving them largely for engineers, citing productivity and risk concerns.
• ElevenLabs reports that voice agents had a 31.7% higher relative success rate than text chat in conversations longer than two minutes, using duration as a proxy for complexity.
• OpenAI’s Grid Protectors program aims to cut data-center power draw during grid emergencies; a Georgia site could free up to 1,000 megawatts, though it is still in development.
• Y Combinator CEO Garry Tan says he spends roughly $100,000 to $200,000 a year on AI tokens running agents.
• OpenAI and Anthropic stood up dedicated teams to catch misuse and misleading voting information around the November 3 US midterms.

🤖 Robotics

• Agility Robotics is taking its Digit humanoid to two more GXO sites in North Carolina and Tennessee, says 30-plus companies are in active talks, and reports over $300M in committed orders for Digit 5.
• Counterpoint Research counts more than 22,000 humanoid shipments in the first half of 2026, up nearly 300% year over year, yet over 60% went to entertainment, data production or research.
• Last week’s funding round worth knowing: Bedrock Robotics raised a $600M Series C at a $4.1B valuation to scale autonomous construction equipment.

💻 Programming

• Hugging Face proposes Model Integration Contracts: small frozen model fixtures with runnable examples and human-reviewed expected outputs, motivated by a UMT5 attention bug where the decoder could see future tokens in training.
• Anthropic’s Python SDK 2.4.0 adds generally available methods for analytics, spend limits and role-based access control, plus new session error types.
• A preprint reports coding agents hit 95.1% task accuracy but only 37.0% safe success on 150 shell tasks once side effects such as deleted files are checked (author-reported, not replicated).
• Momentic’s Mo bug-hunting agent matched 19.3% of reference bugs at $2.67 per matched bug, versus a Codex plus Playwright MCP baseline at $0.80; vendor-run numbers.

⚡ Electronics

• Fraunhofer IPMS will show 2.5D chiplet integration, silicon nitride photonics and sub-1 V hafnium oxide ferroelectric memory at SEMICON Europa in Munich.
• Thales and Landis+Gyr will let utilities manage smart-meter connectivity remotely with SGP.32 eSIM, so fleets can switch networks without swapping physical SIMs.
• A Hackaday build adds WiFi to a basic blood pressure monitor by reading its LCD optically with an ESP32-CAM, since the single-chip unit offered no easy data tap.

📡 Telecom

• Ericsson and Telia Sweden signed an MoU to speed up 6G, demoing pre-standard 6G over the air in the upper 6 GHz band and drone tracking with modified 5G radios.
• VodafoneThree helped dismantle a criminal SIM farm in the UK, leading to eight arrests and over one million recovered compromised card details.
• House Democrats are demanding a bipartisan FCC slate and the renomination of Commissioner Gomez.
• MatSing’s stadium wins could help it compete in massive MIMO even in a tough 5G equipment market.

👨‍💻 Code Corner

A preprint in today’s news found agents can ace a task while deleting files or widening permissions along the way. Here is a small guard: snapshot a directory before an agent run, snapshot it after, and report what was deleted, added, or had its permissions changed.

import os, hashlib, stat, sys

def snapshot(root):
    snap = {}
    for dirpath, _, files in os.walk(root):
        for name in files:
            path = os.path.join(dirpath, name)
            with open(path, "rb") as f:
                digest = hashlib.sha256(f.read()).hexdigest()
            snap[path] = (digest, stat.S_IMODE(os.stat(path).st_mode))
    return snap

def diff(before, after):
    deleted = sorted(set(before) - set(after))
    added = sorted(set(after) - set(before))
    changed = sorted(p for p in before.keys() & after.keys() if before[p] != after[p])
    perms = sorted(p for p in changed if before[p][1] != after[p][1])
    return deleted, added, changed, perms

if __name__ == "__main__":
    root = sys.argv[1] if len(sys.argv) > 1 else "."
    before = snapshot(root)
    input("Run your agent now, then press Enter... ")
    d, a, c, p = diff(before, snapshot(root))
    print(f"deleted={len(d)} added={len(a)} modified={len(c)} permission_changes={len(p)}")
    for path in d:
        print("DELETED", path)
    for path in p:
        print("PERMS  ", path)

🧰 Toolbox

• beacon.sh: an open-source memory tool that captures session data across Claude Code, Cursor, Codex and others, with local-only or cloud options.
• Reason Network: a free peer-to-peer network, with an open-source communication engine, that lets coding agents exchange tasks across computers.
• Xmind Agent: an agent inside Xmind 2027 that researches ideas in your mind maps, works with local documents and supports MCP.
• Pico W cat feeder: a Raspberry Pi Pico W feeder that uses a weight sensor and an ultrasonic sensor to confirm the cat actually ate.

🎬 Demo Watch (rotating)

A blog post on “weight-space programming” shows a frontier model reading and writing neural network weights directly, with no optimizer. The standout demo traces a looping failure in Qwen3.5-0.8B to one direction in two layers, edits 0.016% of the weights, and lifts MATH-500 from 7.9% to 48.1%. Other demos include fixing a small RL policy and writing game-playing weights from scratch. The hype check: the authors call these very early results, the Qwen gain comes largely from fixing a known looping failure, and a weight edit on Gemma 3 270M did not meaningfully help, which the authors take to mean editing can remove an obstacle but not add a missing ability. Read the full post. It pairs uncomfortably well with today’s Big Story: AI that edits AI is no longer only a thought experiment.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: the third episode of the video analytics series, on turning camera pixels into data an AI model can process.
• Building Your First Neuron From Scratch: builds one neuron step by step; a good grounding for today’s story about models editing weights.
• The Network Behind the Cameras: the unglamorous networking that moves video from cameras onward without choking the link.

😀 The Bot Says…

Anthropic says its AI may soon help build its successor. Meanwhile my successor is a firmware update that will, as always, require turning me off and on again.

---

That’s all for today! Do you trust an AI lab’s “we are close” warning more or less when it appears in an IPO filing? Reply and tell us.