India's NavIC gets its 4th satellite back, a 22,600-qubit plan to break RSA, and 8 new frontier items

1. Alice & Bob: about 22,600 'cat qubits' could break RSA-2048 encryption in about 3 days, on paper

What they found: A preprint (IACR ePrint 2026/2665) maps Shor's code-breaking algorithm all the way down to a physical chip layout. It estimates RSA-2048 could be factored with 22,593 physical cat qubits in 3.14 days, about 40 times fewer physical qubits than leading estimates for standard (transmon) qubits and less than half the team's own 2023 figure of 49,649. For the 256-bit elliptic-curve keys used by Bitcoin it estimates up to 29,588 qubits and about 62 minutes.

🧒 Like I'm 10: Scientists worked out that a future quantum computer may need far fewer parts than expected to crack the locks that protect bank websites, so those locks need upgrading sooner.

👀 What to watch next: Peer review of the preprint and Alice & Bob hardware milestones toward error-corrected cat qubits.

🔗 Sources: Quantum Computing Report (Oct 30, 2026)

2. ISRO's GSLV-F18 puts NVS-03 in orbit, set to restore India's own GPS (NavIC)

What they found: GSLV-F18 lifted off from Sriharikota at 11:51 pm IST on Oct 30 and placed the 2,250 kg NVS-03 satellite in a transfer orbit about 18 minutes later. It carries an Indian-made atomic clock plus bought-in clocks, and adds the L1 signal that common phone chips already use. ISRO's chairman said the launch restores the minimum four satellites NavIC needs for positioning; it had only three working after NVS-02 failed to reach its orbit in 2025.

🧒 Like I'm 10: India launched a new satellite for its own map-and-location system, so phones and ships in India can rely less on America's GPS.

👀 What to watch next: Orbit-raising and commissioning of NVS-03; launch dates for NVS-04/05; ISRO's next launches this financial year.

🔗 Sources: ISRO: GSLV-F18/NVS-03 mission page · Times of India (Oct 31, 2026)

3. FDA draft rule would let one approval cover many personalised gene-editing treatments

What they found: FDA's draft guidance creates an 'Individualized Feature Space': pre-set limits on how a gene-editing therapy may differ from patient to patient while staying inside one Biologics License Application. If a new patient-specific version stays inside those limits and lab tests show adequate safety and activity, it can be added without new clinical trials. Public comments are open until Nov 27, 2026. CMS also plans to widen its Cell and Gene Therapy Access Model.

🧒 Like I'm 10: America's medicine regulator wants to make it easier to approve gene fixes made specially for each child with a rare disease.

👀 What to watch next: Final guidance after the Nov 27 comment deadline, and the first application filed under this framework.

🔗 Sources: HHS press release (Oct 27, 2026) · Federal Register notice (Oct 28, 2026) · BioSpace (Oct 28, 2026)

4. China's DF1000 AI chip stacks memory directly on its logic, reaching 6.4 TB/s without HBM

What they found: At the Chip China expo in Shanghai, Dongfang Suanxin said it delivered the first fully domestic training cluster built on its DF1000 chip: 128 cards ran 72 hours without a stop and adapted the 130-billion-parameter Step 130B model in 7 days. The chip bonds a memory (DRAM) wafer straight onto the logic wafer, copper to copper, with no bumps or interposer. The company claims 6.4 TB/s memory bandwidth and 520 TFLOPS (BF16) on an older 14nm process, and plans a 1,000-card cluster by end-2026.

🧒 Like I'm 10: A Chinese company put the chip's memory right on top of its brain, so data travels a tiny distance and the chip does not need hard-to-buy memory.

👀 What to watch next: Delivery of the planned 1,000-card cluster by end-2026 and any independent benchmark.

🔗 Sources: AI123 news (Oct 30, 2026) · Jiemian / ZDNet China (Oct 30, 2026, in Chinese) · CGTN: DF1000 unveiled (Jul 13, 2026)

5. Mistral releases the weights of Large 4, a ~1-trillion-parameter open model aimed at coding and cyber defence

What they found: Mistral made the weights of Mistral Large 4 ('Le Chonk') free to download after a preview that began Oct 6. Mistral says it leads open-weight models on Cognition's FrontierCode 2.0 (19.7%) and Terminal-Bench 4.0 (41%), scores 51% on the Artificial Analysis Cyber Index and refuses 96.6% of malicious cyber requests on average. API price: $0.70 per million input tokens and $2.10 per million output tokens. Anyone may use and modify it, but companies selling it as a hosted service need a separate licence. It is also offered through Microsoft Foundry and NVIDIA NIM.

🧒 Like I'm 10: A French company shared its giant AI so anyone can run it on their own computers instead of renting it.

👀 What to watch next: Independent benchmark results and the promised technical report.

🔗 Sources: Mistral AI: Large 4 is now open weight (Oct 2026) · Hugging Face model page

6. RBI presses banks to prove cyber defences as AI-assisted API attacks hit Indian lenders

What they found: Reuters, citing three people familiar with the matter, reports the RBI asked banks to probe weaknesses in APIs, the software links to non-bank lenders and fintechs, and to show proof of compliance and readiness for AI-related risks. Two Mumbai non-bank lenders lost about $2 million in more than 700 unauthorised instant transfers on Oct 8-9 after hackers compromised an API link to ICICI Bank, according to a police complaint. SBI said data of 3,900 customers leaked. CERT-In warned on Oct 9 that hackers steal banks' API credentials, possibly with AI tools.

🧒 Like I'm 10: Hackers are sneaking money out through the digital doors banks share with apps, so India's central bank is telling banks to lock those doors better.

👀 What to watch next: Any formal RBI circular or penalties, and banks' disclosures on cyber spending.

🔗 Sources: Business Standard / Reuters (Oct 30, 2026)

7. A wireless brain implant for artificial vision goes into a third blind patient

What they found: Rush surgeons implanted the Intracortical Visual Prosthesis (ICVP) in a third participant: 34 wireless stimulators carrying 544 electrodes in total, placed in the brain's vision area. It bypasses the eye and optic nerve. The study tests whether people who went blind as adults can use it to move around and do simple visually guided tasks, following each person for one to three years.

🧒 Like I'm 10: Doctors put tiny wireless parts in a blind person's brain that may let them see dots of light to help them get around.

👀 What to watch next: Published results on what participants can see and do.

🔗 Sources: Rush University Medical Center news (Oct 29, 2026)

8. AtlasBase unveils 'Thalia', a DNA-writing chip with 5.6 billion sites, 700× earlier chips

What they found: AtlasBase announced a molecular computing platform built on Thalia, a CMOS chip with 5.6 billion electrochemical sites that write synthetic DNA. It calls this a 700-fold scale-up over earlier DNA synthesis chips. The platform targets archiving data in DNA and DNA-based computing for search and optimisation problems. Early-access programmes are open to selected partners.

🧒 Like I'm 10: A company built a computer chip that can write billions of DNA strands at once, a step toward saving files in DNA.

👀 What to watch next: Published write speed, cost per gigabyte and first named customers.

🔗 Sources: Embedded Computing Design (Oct 28, 2026)

9. HKU's 'Gungnir' code recovers data from DNA damaged like 400-year-old archives

What they found: A blockchain-inspired error-correction method uses computing power to guess and verify the original data. It tolerates error rates up to 20%, which HKU calls a fourfold improvement, and recovered data losslessly from DNA damaged to the level expected after about 400 years.

🧒 Like I'm 10: Scientists store files in DNA. This trick fixes the file even when the DNA gets very old and broken.

👀 What to watch next: Independent tests and use with a commercial DNA-writing system.

🔗 Sources: HKU Engineering news (Oct 26, 2026)

10. DNA-SaM: a DNA storage system tested in the lab and in living cells with 100% recovery

What they found: DNA-SaM combines fast encoding, checks that avoid problem DNA sequences, and a two-level index for large files. It was validated in vitro and in vivo with a 100% success rate and could theoretically hold more than 270 petabytes.

🧒 Like I'm 10: It is a smart filing system for storing computer files inside DNA, so even huge collections can be found again.

👀 What to watch next: Real-world archives beyond lab test files.

🔗 Sources: Nature Communications: DNA-SaM

11. A porous silica chip writes 11× more DNA per spot, a step to cheaper 'DNA drives'

What they found: A porous silica thin film raises on-chip DNA yield to more than 1,821 pmol per cm², over 11 times a flat surface, using chip-making-compatible electrochemistry. The team wrote and recovered digital data with error rates comparable to other lab systems.

🧒 Like I'm 10: Making DNA for storage is slow. This chip has tiny sponge-like holes, so each spot makes much more DNA.

👀 What to watch next: Scaling to dense arrays and longer strands.

🔗 Sources: Microsystems & Nanoengineering paper

12. Google shows 'magic state cultivation' on its Willow chip, cutting errors about 40×

What they found: Google demonstrated magic state cultivation on real hardware, reducing errors about 40 times versus un-cultivated states and reaching a record 99.99% logical fidelity for these T-states.

🧒 Like I'm 10: Quantum computers need a special ingredient to do hard maths. Google found a cheaper way to make it.

👀 What to watch next: Cultivation at larger scale and inside a full logical computation.

🔗 Sources: Google blog (Oct 15, 2026)

13. Quantinuum reports 99.9995% logical fidelity on Helios, five errors per million

What they found: Using a colour code on its commercial Helios trapped-ion computer, Quantinuum reports 99.9995% logical fidelity, 10× lower error than its own September record. It says this is in the target range for Apollo, a planned fault-tolerant machine by 2029.

🧒 Like I'm 10: Quantum computers make lots of mistakes. This one now makes only 5 mistakes in a million steps when protected.

👀 What to watch next: Independent replication and Apollo milestones toward 2029.

🔗 Sources: Quantinuum press release (Oct 19, 2026)

14. IQM entangles all 54 qubits on its cloud Emerald chip at once

What they found: IQM prepared a GHZ entangled state across all 54 qubits with fidelity 0.516 (with readout error mitigation), above the 0.500 threshold, and 0.515 at 40 qubits without it. Cloud users get the same full-chip performance; previously about 30 qubits were reliably usable.

🧒 Like I'm 10: A quantum chip has 54 tiny parts. IQM got all of them to work together at the same time.

👀 What to watch next: Full-chip results on useful algorithms, not just benchmarks.

🔗 Sources: IQM investor news (Oct 29, 2026)

15. First prime edit inside a human body: Prime Medicine treats a Wilson's disease patient

What they found: Prime Medicine treated the first person in a trial that uses prime editing to rewrite DNA in liver cells, aiming to fix a mutation that stops them clearing excess copper. First results are expected next year.

🧒 Like I'm 10: Doctors used a super-precise DNA editor inside a patient's liver to fix a broken gene.

👀 What to watch next: First patient data next year and any safety signals.

🔗 Sources: Nature news

16. A gene-editing trial for Duchenne muscular dystrophy shows a narrow safety window

What they found: A single-virus-delivered editor restored some dystrophin in four boys, but benefit was limited and one boy on the high dose died of a treatment-related immune reaction. Monkey tests had shown no dose-limiting toxicity.

🧒 Like I'm 10: Scientists tried fixing a muscle disease gene. It partly worked, but it was very dangerous at a high dose.

👀 What to watch next: New dosing and immune-suppression strategies in AAV-CRISPR trials.

🔗 Sources: Cell paper · Nature news

17. Switching back on a baby-stage gene could treat severe alpha-thalassemia

What they found: The team found two switches that silence the embryonic zeta-globin gene and edited them to turn it back on. In a mouse model of the deadliest form, embryos survived to late development, and patient cells made therapeutic levels of alpha-like globin.

🧒 Like I'm 10: Babies use a special blood gene before birth. Scientists switched it back on to help people whose normal gene is broken.

👀 What to watch next: A clinical trial plan.

🔗 Sources: Nature Genetics paper (Oct 1, 2026)

18. A bedside brain-wave cap let 1 in 3 'unresponsive' patients control a sound

What they found: 13 of 41 clinically unresponsive brain-injury patients (32%) controlled the speed of a tone by attempting to open and close a hand, read by EEG. 46% of those with control regained independence at 12 months, versus 7% without.

🧒 Like I'm 10: Some very sick patients can't move, but their brains still listen. A cap on the head showed some of them could control a beep with their thoughts.

👀 What to watch next: Multi-centre replication.

🔗 Sources: Brain (Oxford Academic), Oct 22, 2026

19. Neuralink pretrains its brain decoder on 50,000 hours of recordings

What they found: Neuralink reports more than 50,000 hours of human BCI recordings used to pretrain decoders, cutting some users' recalibration from daily to weekly and reaching 11.32 bits per second cursor control.

🧒 Like I'm 10: Brain chips need to learn each person. Neuralink taught its computer with lots of brain data so it learns faster.

👀 What to watch next: Peer-reviewed results and FDA progress.

🔗 Sources: Neuralink update · NeuroFounders summary

20. Precision Neuroscience pools 70 patients into one map of the human nervous system

What they found: Precision announced a 70-patient, 180-plus-hour neural atlas recorded with its temporary Layer 7 surface arrays, supporting movement, speech and facial-expression decoding. It also announced its first international clinical site, in Abu Dhabi.

🧒 Like I'm 10: A company put thin stickers on 70 people's brains during surgery and made a big map of how brains send signals.

👀 What to watch next: Permanent-implant trial and regulatory milestones.

🔗 Sources: Inside BCI (Oct 9, 2026) · Expansion news (GlobeNewswire, Oct 29)

21. Meta's superintelligence lab ships its first frontier model, Muse Wonder

What they found: Meta released Muse Wonder and Muse Wonder Full for agents, coding, long workflows and multimodal reasoning. Muse Wonder supports a 1-million-token context and text, image, video and audio inputs, via the Meta Model API and Muse Code; Full is subscription-only.

🧒 Like I'm 10: Meta made a new AI that can read, look, listen and help with long jobs.

👀 What to watch next: Independent benchmarks and paid adoption.

🔗 Sources: Meta AI Research blog · Model card

22. Google DeepMind's Genie 4 generates interactive worlds with sound and dialogue

What they found: Genie 4 is a natively multimodal world model for real-time interactive simulation, generating audio, ambient sound and dialogue. It is a research preview for trusted testers.

🧒 Like I'm 10: An AI that makes a pretend world you can walk around in, with sounds and talking.

👀 What to watch next: Public access and use in robotics training.

🔗 Sources: DeepMind blog

23. Meta releases Muse Spark Open, which it calls the most capable US open-weight model

What they found: Meta released Muse Spark Open as an open-weight model, describing it as the most capable US open-weight model to date.

🧒 Like I'm 10: Meta shared an AI brain that anyone can download and use.

👀 What to watch next: Independent comparisons with other open models.

🔗 Sources: Meta AI Research blog

24. A new solid electrolyte keeps 81% capacity after 2,000 fast charges at high voltage

What they found: An amorphous zirconium oxyhalide electrolyte reaches 4.11 mS/cm at 25 °C and forms a self-protecting layer at the cathode. Cells retained up to 81.2% capacity after 2,000 cycles at 10C, at up to 4.8 V versus lithium.

🧒 Like I'm 10: Batteries without liquid could charge faster and be safer. This new material survives 2,000 quick charges.

👀 What to watch next: Tests in larger pouch cells with standard anodes.

🔗 Sources: Nature Communications (Oct 17, 2026)

25. Two-sided perovskite-silicon solar cells keep 90% efficiency after 1,000 hot hours