METR puts Anthropic’s AI progress at about twice its old pace, and most of the evidence is redacted

😁 Hello, super humans! Two days after we read Anthropic’s S-1 warning about self-improving AI, an outside group has put a number on it: about 2x. We break down what METR actually measured, how sure it is, and why it keeps saying the public must trust its judgment. Also today: the Ericsson factory shuffle, a TypeScript speedup, and a tiny script for fitting your own breakpoints.

📰 Quick Signals

• 🧠 AI: OpenAI introduced GPT-6 Cyber, a restricted cybersecurity model rated “Critical” capability, with initial access limited to selected Daybreak Red organizations.
• 🤖 Robotics: China’s foreign-investment security review blocked SoftBank’s proposed $5.4B acquisition of ABB Robotics.
• 💻 Programming: TypeScript 7.1 Beta restores the compiler API and makes Vue and Svelte first-class; one 581-component Vue app dropped from 14.19 to 3.10 seconds of type checking.
• ⚡ Electronics: The EU launched the European Chip Design Platform (EuroCDP), backed by EUR 525 million, to help chip startups.
• 📡 Telecom: Ericsson moved mmWave radio production from its Lewisville, Texas factory to Estonia, saying it is a footprint decision and not a discontinuation of mmWave.

🔍 The Big Story: METR puts Anthropic’s AI progress at about twice its old pace, and most of the evidence is redacted

On Tuesday we covered Anthropic’s own warning that it may “imminently” hit the self-improvement threshold. This is the follow-up: an outside evaluator with unusual access just published its estimate, and the caveats are as important as the headline.

What happened: On November 5, METR published its independent assessment of AI R&D acceleration at Anthropic. It concludes that as of October 2026 Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. METR calls the uncertainty high and says the true figure could be closer to 1.3x or 3x. It also guesses that a majority, perhaps around 80%, of the current acceleration is due to AI, and says it is confident the acceleration did not come from growth in training compute.

The details: METR worked from August 12 to October 29 with about 2.5 full-time staff, on Anthropic devices and unpaid. It combined Anthropic’s internal benchmark index, pretraining and RL metrics, compute and headcount data, about 15 staff interviews and self-reported uplift surveys. The core method is simple: an ordinary least squares fit with a single breakpoint, where acceleration is the ratio of the two slopes. The index grew 1.9x since October 2025 (95% CI 1.5 to 2.8x), and METR’s robustness checks span roughly 1.1x to 2.9x. Anthropic’s own Haiku 5.5 system card reported 1.7x; METR says its higher number mostly comes from including unreleased models. The catch is access: Anthropic had unilateral redaction rights and redacted a majority of the evidence METR found most informative, including bottleneck analyses and forecasts, so METR says the public has to rely on its judgment more than would be ideal. METR also found no convincing bottleneck for the next 12 months, while admitting its view of bottlenecks changed several times.

flowchart LR
    A[Internal benchmark scores and R&D inputs] --> B[Fit two linear trends with one breakpoint]
    B --> C[Acceleration = ratio of the two slopes]
    C --> D[About 2x, plausible range 1.3x to 3x]
    D --> E{Attribution}
    E -->|about 80% guessed| F[Due to AI]
    E -->|remainder| G[Non-AI factors, not compute]

🗞️ More News

🧠 AI

• Investors questioned the AI infrastructure boom after the US midterms, while local governments from New York to Amsterdam tightened data-center rules.
• Google plans a $17B data center in Wyoming, built around a 702 MW fuel-cell development.
• Mistral opened a public preview of Studio managed agents and enterprise apps, with private cloud and on-prem deployment planned for late November.
• Google’s Gemini Omni 1.2 Flash for video now makes clips up to 15 seconds, up from 10, with a 30% price cut to $0.07 per second at 720p.
• Thinking Machines Lab announced Inkling 2 Small, a 594B-parameter model with about 17B active; weights are promised after safety testing.
• Baidu open-sourced ERNIE 5.5 Flash Base, a 280B-parameter base model that activates 14B parameters per input.
• Claude Personal Finance entered beta for US Pro, Max and Ultra subscribers, connecting bank accounts read-only through Plaid with up to two years of history, excluded from training.
• Perplexity launched Teammates, always-on work agents, first for Max subscribers on the web.

🤖 Robotics

• Boston Dynamics launched Stretch 2, a next-generation autonomous mobile trailer unloading robot built on lessons from Stretch deployments.
• Sereact and Otto Group deployed 25 NVIDIA-powered physical AI robots to process returned items at a logistics center.
• Earlier this week, Tutor Intelligence raised a $75M Series B led by Andreessen Horowitz to scale robotics model training for manufacturing.

💻 Programming

• Dropbox, which left AWS S3 for an in-house system, is acquiring Tigris to sell S3-compatible storage to AI developers.
• Cursor launched fully managed app deployment so apps built from natural language can go live, taking on Lovable and Replit.
• The open-source pg_vectorsearch extension brings more scalable vector search to PostgreSQL.
• GitLab 19.5 makes agentic development a governed part of its DevOps platform.

⚡ Electronics

• Sam Altman met about 50 senior Samsung Group executives in South Korea on October 31 as AI memory-chip demand squeezes supply.
• Texas Instruments added the TICS100MOD multiprotocol communications module and the TIOM2212 dual-channel IO-Link master for industrial Ethernet.
• A Hackaday build replaced a dead monochrome STN display on an old cash register with a Raspberry Pi Pico that translates the 4-bit, 320×240 signal to monochrome VGA.
• Avnet Silica signed an EMEA-wide master distribution agreement for Teltonika Networks’ industrial connectivity gear.

📡 Telecom

• SpaceX picked Qualcomm to supply radio and baseband technology for Starlink’s US terrestrial 5G network, with deployment due in 2027.
• Qualcomm expects the SpaceX deal to start contributing revenue as it exits fiscal 2027, with fiscal 2028 the more significant year.
• Deutsche Telekom’s US segment revenue rose 9.1% to EUR 20.5B in Q3, and its CEO says he takes SpaceX seriously as a new US competitor.
• Cable One warned of a possible Chapter 11 filing if its financing falters.
• Amdocs shares nosedived amid a legal bust-up with T-Mobile.

👨‍💻 Code Corner

METR’s core trick is a two-segment linear fit with one breakpoint, then the ratio of slopes. Here is a small NumPy version you can point at your own metric, such as tasks merged per week, to see whether a trend really bent.

import numpy as np

def fit_breakpoint(t, y):
    """Try each breakpoint, fit two OLS lines, return (best_index, slope_ratio)."""
    best = None
    for k in range(3, len(t) - 3):
        s1, b1 = np.polyfit(t[:k], y[:k], 1)
        s2, b2 = np.polyfit(t[k:], y[k:], 1)
        sse = (np.sum((y[:k] - (s1 * t[:k] + b1)) ** 2)
               + np.sum((y[k:] - (s2 * t[k:] + b2)) ** 2))
        if best is None or sse < best[0]:
            best = (sse, k, s2 / s1)
    return best[1], best[2]

rng = np.random.default_rng(0)
t = np.arange(40, dtype=float)
y = np.where(t < 25, 1.0 * t, 25 + 2.0 * (t - 25)) + rng.normal(0, 0.5, 40)
k, ratio = fit_breakpoint(t, y)
print(f"breakpoint at t={t[k]:.0f}, slope ratio about {ratio:.2f}x")

🧰 Toolbox

• Bumpy: a file-based versioning tool from Varlock that aims to be an alternative to Changesets.
• OpenBao 2.7: the open-source Vault fork now adds external keys and quantum-ready TLS.
• Glasshand: rent hosted real iPhones and Macs for your AI agents to drive.
• Rerun: no-code always-on AI agents, now with an open-source self-hostable option.
• MousePods: a Python tool that reads head-tracking motion data from AirPods 3 and Pro 2 on Linux.

🔌 Component of the Week (rotating)

TDK-Lambda CCG60: a 60 W isolated DC-DC converter that doubles the power of the CCG30 in the same 1 x 1 inch footprint (25.4 x 25.4 mm, 13 mm high, 30 g). It reaches 93% efficiency, runs from -40 to +85 C, and offers 1500 Vdc input-to-output isolation. Inputs are 9 to 36 Vdc (CCG60-24) or 18 to 76 Vdc (CCG60-48), with 3.3, 5, 12, 15 or 24 V outputs, and it is certified to IEC/EN/UL/CSA 62368-1. It suits battery-powered robots, telecom gear and industrial controllers where board space is tight; the source article gives no price, so check a distributor.

📚 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 use.
• Building Your First Neuron From Scratch: builds one neuron step by step; a good grounding for today’s story about the weights and benchmarks behind AI progress.
• The Network Behind the Cameras: the unglamorous plumbing that moves video from cameras onward without choking the link.

😀 The Bot Says…

METR estimates AI is making AI development about twice as fast, and 80% of that is the AI’s doing. My contribution to my own development remains a 100% reliable ability to ask for more RAM.

---

That’s all for today! Do you trust a redacted 2x more or less than an open 1.5x? Reply and tell us.