Anthropic Cuts Claude Distillation by Up to 99%
ALSO: Meta’s Muse profiles everyone, OpenAI’s 700+ papers
Welcome to Guru99 AI Report!
Top Story: This week Anthropic claims it can cut what copycats harvest from Claude by up to 99%, and a researcher finds Meta’s Muse keeps a profile on everyone you know. Plus, 700+ AI-written math papers. Let’s dig in.
🛡️ Anthropic Says New Shields Cut Claude Copying Up to 99%
Brief Buzz:
Anthropic says some AI labs are simply copying rivals’ models at a fraction of the cost. The technique is distillation: take a large model, bombard it with questions, then train a smaller model on the answers it gives back. Anthropic now says its new anti-distillation defenses cut what copycats can harvest from Claude by 90–99%.
- Reasoning is the prize. In Anthropic’s test, a copy trained on full reasoning hit 80.4% on a coding task versus 71.8% on a summarization task.
- Preserved thinking blocks API accounts created on or after August 31, 2026 from editing Claude’s earlier context, making repeated extraction harder.
- A two-stage classifier (an AI filter) modeled on Anthropic’s bioweapons safeguards flags extraction attempts while wrongly blocking under 0.1% of requests.
- Tens of thousands of accounts were banned, often within two hours of starting a distillation campaign.
- Zhipu’s GLM-5.3 — which Anthropic claims distills US models’ hacking prowess — was rated by NIST the most cyber-capable open model, though still behind the US leaders.
💡 Why Should You Care?
Anthropic warns that copied models could ship without strong guardrails, putting hacking and surveillance power within reach of far more actors. Developers face trade-offs — integration updates, rare false blocks, charges for blocked requests — and Anthropic expects the copycats to adapt.
🕵️ Meta’s Muse Keeps a Page on Everyone You Know
Brief Buzz:
Muse (Meta’s AI agent) wants to know your inner circle. A researcher examined its internal instructions and found it builds a profile page for each person in your life. Those pages refresh every hour — and can be created for people who never signed up at all.
- Independent researcher Karan Joshi got the files simply by asking Muse to copy them through its normal chat.
- Pages span family, partners, friends, colleagues, collaborators — and even people you merely follow.
- Each page can log where someone lives, their job, their birthday, your shared history, and tips on improving the relationship.
- Meta says Muse draws on public information plus what you share, all stored in each user’s own dedicated virtual machine.
- One YouTuber claims Muse — set to “Allow Always” — gave out his home address to a Marketplace buyer.
💡 Why Should You Care?
Your friends and family never opted in. Before linking your email or contacts to any AI agent, check what it stores about other people — and whether that data can be deleted.
🎒 Google’s AI Push Into Schools Sparks a Backlash
Brief Buzz:
AI was pitched as a personal tutor for every student, but reality is messier. A Wall Street Journal investigation into Google’s classroom AI push found students leaning on the chatbots, then struggling to think for themselves once the AI was taken away.
- In the OECD’s PISA 2025 results, students who used AI to summarize, draft, or research scored about 20 points lower in science — roughly a full year of schooling.
- Reading and math scores hit record lows across the OECD average, though the OECD stresses the AI link is an association, not proof of causation.
- A Brookings Institution study of students, parents, educators, and tech experts found AI’s current risks to learners outweigh its benefits.
- ChatGPT-assisted essay writers showed the weakest brain connectivity in an MIT Media Lab preprint — though the study ran with just 54 people.
- There’s no off switch: Google Search’s AI Overviews and AI Mode can’t be disabled through the admin controls Google gives school districts, per the WSJ.
💡 Why Should You Care?
The point isn’t to ban AI — PISA shows students who use it as a study aid and learn to question it match or beat peers who skip it. The red flag is letting AI do the thinking.
🧬 Google, Meta, and US Government Join $1.8B Virtual Cell Push
Brief Buzz:
Biohub — the Zuckerberg-backed nonprofit — just landed a slate of major new partners. Google DeepMind, Meta, and the US government are all joining the effort to build open biology data for AI. The goal: a “virtual cell” that lets scientists test ideas on a computer first.
- $1.8 billion total, including $500M of Biohub’s own funds from April — a mix of cash plus existing data, compute, and lab infrastructure.
- $300M comes from Google DeepMind, Meta, and drug-discovery startup Isomorphic Labs.
- The DoE has committed $500M+ over five years, and NIH is adding datasets themselves built with $500M+ in prior federal funding.
- Corporate funders get early access before the data goes public, Biohub’s science head told Reuters; government-funded data carries no such embargo.
- The first dataset is due in about a year, with Biohub targeting accurate predictive models within five years.
💡 Why Should You Care?
Drug discovery runs on years of lab work and trial and error. Virtual cells could let scientists test ideas digitally first, speeding the path to a cure. The catch: corporate funders see the “open” data before anyone else.
🔩 This $250M Startup Turns Code Into Tiny Working Machines
Brief Buzz:
Atomic Machines just came out of stealth. The California startup spent six years under wraps and has raised $250 million. The pitch is simple: write code, and its AI-run factory produces a small working machine. Change the code, and you get a different machine.
- Jeff Holden founded Amazon Prime and was Uber’s first chief product officer, where he set up its self-driving unit.
- The Matter Compiler needs no molds or retooling; the company says it verifies each step and learns from every build.
- The goal is machines with micron-scale features — tiny motors, gears, and pumps — that chip fabs were never designed to make.
- First up is PrimeSwitch, a fingertip-sized switch rated for 150 amperes and built for AI data centers; Atomic claims it’s about 1,000 times faster than ordinary contactors.
- It’s not fully hands-off yet — engineers still shuttle parts between the 13 machines and help assemble the final device.
💡 Why Should You Care?
Vibe coding turns plain-English ideas into apps. Atomic wants the hardware version — prompt-to-product manufacturing. If it works, new gadgets could ship in weeks instead of years. But the claims haven’t been independently tested yet.
🧮 OpenAI’s Unreleased AI Drops 700+ Math Papers at Once
Brief Buzz:
OpenAI just posted hundreds of math papers written by an AI model it hasn’t released yet. They tackle unsolved problems in math, computer science, and physics — none peer reviewed, and a couple already pulled for errors.
- 722 manuscripts from 372 families were uploaded to OpenAI’s public math repository on October 6; after the withdrawals, the count now sits at 719.
- OpenAI put roughly 4,000 problems to the model, with each run taking about three hours of ChatGPT Pro compute time.
- About 42% of the main results carry Lean proofs (code a computer steps through to verify); the rest have no machine verification.
- Three papers were pulled a day later over a single sign error, and 14 more were later revised.
- OpenAI says it consulted the IAS independent AI/math advisory group and will fund the related workshops.
💡 Why Should You Care?
AI may generate research faster than humans can check it. The real test isn’t how much you can produce — it’s how many of those proofs survive independent review.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
