Tech for Good This Week: AI Diagnosis, Cleaner Grids and Scam Busts

For the first time in any quarter, grid-scale batteries supplied more of Australia’s evening peak than gas: 8.0% of generation against 5.1%, according to the Australian Energy Market Operator’s Q3 report. In the same week a personalised mRNA therapy passed its first phase 3 test in melanoma, and police and prosecutors on three continents put numbers on their fight against online fraud.

This is TSN’s weekly round-up of tech for good stories from Saturday 17 to Friday 23 October 2026. Each line links to our full report, and every figure comes from that report. Where a result rests only on a company’s own testing, we say so, because a good share of this week’s news is still at that stage.

Medicine and health: a melanoma result and a diagnostic AI under peer review

• Melanoma. Adding the investigational mRNA therapy intismeran to Keytruda cut the risk of recurrence or death by 42% in a 1,137-patient trial of people whose high-risk melanoma had been removed, Merck and Moderna said. At two years, 82.2% on the combination were alive and recurrence-free, against 70.0% on Keytruda alone. It is not an approval. Read our report.
• Diagnosis. Microsoft’s MAI-DxO reached the right diagnosis in 83.2% of complex teaching cases, against 36.1% for generalist doctors using reference tools, a paper in Science reports. These were past cases on a benchmark Microsoft built, and no product has been announced. Microsoft’s diagnosis study.
• Ovarian cancer. A King’s College London model, HRDPath, spots a DNA-repair defect from the routine slides hospitals already make, with an AUC of 0.846, its authors say. They are clear it is not ready for clinical use. KCL’s HRDPath.
• Preterm labour. Stanford’s SIMCO “digital twins”, built from 218 million immune cells, picked aspirin plus lansoprazole as a pair to test. The evidence so far is simulations, lab work on donated blood and, Stanford says, an animal experiment. Stanford’s immune digital twins.
• Pathology. Sophont says its Holon-S1 model predicts which genes are active in a tumour from ordinary stained slides. It is for research use only, and the results are the company’s own retrospective work. Sophont’s Holon-S1.
• Protein misfolding. A Johns Hopkins-led team won up to $52.1m over four years from ARPA-H to build a spinal-fluid test for Parkinson’s, selective binders and an open model. It is a research plan, with no results yet. The Hopkins award.
• Clinical trials. Oracle and AstraZeneca signed a non-binding agreement to explore using AI and real-world data to find trial patients. No money, timetable or health systems were disclosed. Oracle and AstraZeneca.
• Drug delivery. Metis TechBio says its deal with Eli Lilly to design tissue-targeting fat particles with AI could be worth up to about US$1.1bn in milestones. Only low double-digit millions are guaranteed up front. Metis and Lilly.
• Hospital drones. Matternet and Apian drones now carry urgent children’s pathology samples from Great Ormond Street Hospital in London to laboratories. The companies claim delivery times fall by up to 90% but have given no flight counts. Great Ormond Street drones.
• Hantavirus. Scientists mapped in atomic detail how the Andes virus spike grips its human receptor, PCDH1, in a Nature paper. There is still no approved vaccine or treatment. The hantavirus structure.
• Eye exams. In a Johns Hopkins-led study of 397 young people with diabetes, 99.5% of White and 98% of non-White participants completed an autonomous AI eye exam offered at a routine clinic visit, against prior exam rates of 79.8% and 61.8%. The study ran at Baltimore sites only. The AI eye exam study.

AI for discovery: protein design went open, but most results are self-reported

• Antivenom research. Anthropic says Claude produced 318,484 protein designs in about 64 hours; of 211 tested in the lab, 17 met its criteria for binding scorpion neurotoxins. The company says it has not produced an antivenom. Claude’s toxin binders.
• Open protein design. The University of Washington and NVIDIA released RFD4-Proteina under an open BSD-3 licence. The team claims five to ten times higher design success in computational tests, with no peer-reviewed paper yet. RFD4-Proteina.
• Proteins from text. Westlake University’s 16bn-parameter Pinal model turns plain-language descriptions into proteins that worked in the lab, according to a Nature paper. Its weights are public. Westlake’s Pinal.
• DNA in a chat. Hugging Face’s open 14.8bn-parameter Carbon-X reads raw DNA and answers in English on a single 80 GB GPU. Its benchmark scores are Hugging Face’s own, and the model is not clinically validated. Carbon-X.
• Gene targets. In Xaira’s liver-cell test, genes picked by its X-Scientist agents hit 7.1% of the time, against 6.9% for human picks and 0.9% for random ones, the company says. The agents were faster, not more accurate. Xaira’s X-Scientist.
• Phage therapy. Radical Numerics claims phages designed with its Omnii-1 model were given to one patient with a drug-resistant E. coli infection. It has not said how the patient is doing, and its report is not peer reviewed. The Radical Numerics claim.
• Robot labs. Google Research says AI agents running liquid-handling robots completed a full aptamer design-and-test cycle in 7 to 10 days, with hit rates rising to 42.1% for one target. The reports are self-published. Google’s wet-lab agents.

Climate and Earth observation: Australia’s batteries passed a real-world test

• Batteries over gas. Battery charging or discharging set Australia’s wholesale power price in 48% of dispatch intervals in Q3, up from 18% a year earlier, AEMO says. Batteries beat gas.
• Flexible solar. A Soochow University-led team reported flexible perovskite modules at a certified 20.2% efficiency on 750 cm², in Science. It is a lab and pilot result, not a product. Flexible perovskite modules.
• Seafloor geothermal. Endurance Energy says its test plant on an underwater volcano produced more power than it used for 8.7 seconds. That output came from a natural vent, not its new well. Endurance’s claim.
• Wildfires. ICEYE and OroraTech will combine radar and thermal satellite data to track fires from first hotspot to damaged buildings. No customers, prices or launch date have been given. ICEYE and OroraTech.
• Lightning. Vaisala Xweather says NVIDIA’s AI models will stretch lightning warnings from one hour to three. It has published no accuracy figures. Vaisala and NVIDIA.
• Satellite maps. Ai2 released OlmoEarth v1.3, which sums up a year of imagery as 128 numbers for every 10-metre square of land, with free global maps for 2023 to 2025. OlmoEarth.

Fraud and cyber defence: arrests and frozen funds, with refunds still unclear

• Africa. Police in 21 countries arrested 1,462 suspects and recovered more than $105m in Interpol’s Serengeti III operation, which also disrupted 10,909 malicious networks, Interpol said. Serengeti III.
• Scam compounds. The US Scam Center Strike Force says it has restrained $938m in crypto from Southeast Asian scam operations. Restrained money is held back, not yet repaid. The strike force figure.
• NovaTech. New York’s attorney general secured nearly $6m from four NovaTech promoters, who are banned for life from the securities, commodities and crypto industries. More than 11,000 New Yorkers lost money. The NovaTech settlement.
• Italy. Italian financial police arrested 30 people over an alleged mafia “hidden bank” that used crypto in €68m of laundering, ANSA reported. The suspects are presumed innocent. The Italy case.
• Safer code. Anthropic says Claude proved that 239 entry points in 123 open-source libraries cannot crash on any input up to 64 MB, finding hundreds of bugs on the way. The proofs hold only under stated conditions. Claude’s proofs.

Access, connectivity, child safety and jobs: Greece set a date for its under-15 ban

• Local languages. GoTo released a 27bn-parameter open-weight model for Indonesian, Javanese, Sundanese, Batak Toba and Balinese. Running it at full precision needs about 54 to 60 GB of graphics memory. GoTo’s model.
• Phones from space. The US FCC authorised Globalstar’s C-3 system of 48 low Earth orbit satellites to connect directly to smartphones. Amazon’s purchase of Globalstar is still pending. The FCC approval.
• Children online. Greece’s parliament barred under-15s from social media from 1 January 2027 and made platforms responsible for checking ages, the Greek ministries said. Fines can reach 6% of global turnover. Greece’s ban.
• Child-harms list. In its first statutory report on content harmful to children, Ofcom recommended adding body stigma and depression content to the Online Safety Act list. It is advice to the Secretary of State, not law. The Ofcom report.
• AI skills. New York plans an AI Jobs Institute in Queens, with $10m of public money, aiming to train up to 40,000 people by 2032. It has not opened yet. The AI Jobs Institute.

The strongest results this week came with independent checks

Line the week up by how each result was tested and a pattern appears. The items readers can lean on hardest had someone outside the company check them. AEMO’s battery figures are a grid operator’s own count of what happened. The melanoma result comes from a 1,137-patient phase 3 trial whose data go to a medical congress and a journal. KCL’s ovarian cancer model, the hantavirus structure, Pinal and the perovskite modules all went through peer review. Microsoft’s diagnostic study did too, though its test set was Microsoft’s own design.

Much of the AI-for-science news sits a step lower. The binder counts, the open-model benchmarks, Xaira’s hit rates, Radical Numerics’ phage case and Endurance’s 8.7 seconds are companies reporting on their own work. Some may hold up well. For now they are claims, and several of the companies say so themselves.

The fraud stories need a different caution. Arrests, restraints and settlements are real legal steps, but they are not refunds. Interpol recovered $105m against losses of more than $385m, and the US strike force has not said how much of its $938m has reached victims. For anyone working in anti-money-laundering, the useful number is what gets back to victims, and that is the number usually missing.

The test for each of these stories is the same: when does someone other than the maker check the result?

Several of those checks are close. Merck and Moderna present their full melanoma data at the ESMO Congress in Madrid on Saturday 24 October. Platforms in Greece have until 1 January 2027 to show they can check ages.

Sources

This round-up draws only on TSN’s own reporting, linked above. Each post lists its full primary sources. Primary sources quoted directly here:

• AEMO: Quarterly Energy Dynamics Q3 2026 (PDF)
• Merck: Merck and Moderna will present results from phase 3 INTerpath-001 trial at ESMO 2026
• Science: Sequential diagnosis with language models
• Interpol: Cybercrime crackdown in Africa nets 1,462 arrests, recovers over USD 105 million
• US Attorney’s Office, District of Columbia: Scam Center Strike Force
• New York Attorney General: Attorney General James secures nearly $6 million from crypto scammers
• Greek Ministries of Health and of Digital Governance: joint release on the under-15 social media bill

Related stories

• Tech for Good hub
• Personalised mRNA Therapy Cuts Melanoma Recurrence or Death Risk 42%
• Microsoft Diagnostic AI Scored 83% on Past Cases, Science Paper Says
• Batteries Beat Gas at Australia’s Evening Peak, AEMO Says
• Interpol-Led Africa Crackdown Nets 1,462 Arrests and Recovers $105m
• Greece Passes Under-15 Social Media Ban, With Age Checks From 2027