Breakthrough Tracker
Every discovery we follow after the headline fades, with the date it last moved. When nothing changed, that is stated rather than left blank.
LLM-assisted reanalysis of unsolved rare disease genomes
Eighteen families got an explanation for their child's condition after a language model took a second look at their genome data.
- Field
- Medicine and health
- AI contribution
- Central
- Status
- Retrospective multicohort study complete: 18 new diagnoses in 376 unsolved cases
- Last status change
- June 25, 2026
- Next milestone
- Prospective multicenter evaluation with predefined endpoints and comparator baselines
- Primary source
- NEJM AI
The study was retrospective, and the authors describe the absolute number of new diagnoses as modest. It did not systematically address structural variants, repeat expansions, deep intronic changes or mosaicism on a large scale, although the authors report finding them in a 15 case pilot set, and it did not measure time to diagnosis, cost, interrater agreement or changes in patient management. Four of the authors list OpenAI as their affiliation, and the paper says prompt development and model testing were supported by Boston Children's Innovation and Digital Health Accelerator and OpenAI.
AI diabetic eye screening in Tanzania
People with diabetes in Tanzania were nearly twice as likely to reach an eye specialist when AI read their eye photos and they got results on the spot.
- Field
- Medicine and health
- AI contribution
- Significant
- Status
- Randomized trial complete (ISRCTN18317152); Ministry committed to integrating AI diagnostics into screening in multiple regions
- Last status change
- October 12, 2026
- Next milestone
- Rollout of AI-supported diabetic eye screening in additional Tanzanian regions
- Primary source
- The Lancet Global Health
Even with the AI pathway, most people who needed referral did not reach the clinic within eight weeks (157 of 437 did), and the trial measured clinic attendance, not whether vision was saved. The gain reflects the instant result and the face-to-face counseling together, so the software's own share cannot be separated. The trial was funded by the British Council for the Prevention of Blindness, CBM (Christian Blind Mission) and the Sir Halley Stewart Trust; one author reports equity in Big Picture Medical and consulting fees from several eye-care companies, and the declaration of interests does not mention EyRIS, the AI maker named in the protocol.
BRAVO brain-computer interface trial
People who cannot speak could one day talk with words and gestures together, through a single brain implant.
- Field
- Medicine and health
- AI contribution
- Central
- Status
- Listed as recruiting; trial began November 2018, data for this study collected September 2022 to November 2025
- Last status change
- September 14, 2026
- Next milestone
- Decoding tested in more participants and with larger vocabularies of words and gestures
- Primary source
- Nature Neuroscience
This is three people, and the simultaneous speech-and-gesture work was done in only two of them, using vocabularies of discrete phrases and gestures rather than open-ended language. The paper states the research is not intended to inform conclusions about the primary outcomes of the clinical trial, and the avatar is a laboratory system, not an approved device. In the paper's competing-interests statement, two authors, including Edward Chang, are inventors on pending UCSF patent applications relevant to the decoding methods, and Chang is a cofounder of a neurotechnology company.
WeatherNext Cyclones in operational hurricane forecasting
People in a hurricane's path could get about a day more warning, from an AI model hurricane forecasters have already begun to use.
- Field
- Climate and planet
- AI contribution
- Central
- Status
- Running live since June 2025; experimental guidance used by NHC forecasters; peer reviewed in Nature (Aug 2026)
- Last status change
- August 6, 2026
- Next milestone
- Season verification for 2026 and extension to rainfall, storm surge and gusts (named by the authors as future work)
- Primary source
- Nature
The model still depends on ECMWF's analysis of current conditions as its starting point, and it does not yet forecast rainfall, storm surge or gusts, which cause much of the damage. The authors note that traditional models still add real value for intensity, and most of the authors are Google employees.
IBM verifiable quantum advantage (doped Clifford sampling)
If quantum computers are ever going to be trusted with answers people rely on, claims about them have to survive outside scrutiny.
- Field
- Quantum computing
- AI contribution
- Central
- Status
- Original claim classically simulated (Aug 2026); revised periodic-boundary experiment posted Sept 2, 2026 as a preprint
- Last status change
- July 28, 2026
- Next milestone
- Peer review and any classical challenge to the revised 64-qubit circuit
- Primary source
- arXiv
Both papers are preprints and have not been peer reviewed. The revised 64-qubit experiment has not yet faced a published classical challenge, and the task (sampling from specially designed circuits) is a benchmark, not a useful calculation.
Mobilio navigation app
A smartphone app could one day help blind and low-vision people walk outside more safely and on their own.
- Field
- Engineering and materials
- AI contribution
- Significant
- Status
- Peer-reviewed study with 14 participants; not yet publicly released; Harvard is pursuing commercialization
- Last status change
- August 24, 2026
- Next milestone
- Wider real-world testing around Boston and work toward a product
- Primary source
- Nature Biomedical Engineering
This was a small study (14 participants) on set routes in the Boston area, compared with Google Maps plus a cane rather than with guide dogs or other specialized aids. The app is not yet available to download, and how well it works across many cities, weather and lighting conditions is still to be tested. Harvard's Office of Technology Development says it has protected the innovations from this research and is actively pursuing commercialization opportunities; the paper itself declares no competing interests.
Vesuvius Challenge: reading the Herculaneum scrolls
Readers and historians have the first words from a Herculaneum scroll too fragile to unroll, still sealed since Vesuvius buried it in AD 79.
- Field
- History and archaeology
- AI contribution
- Central
- Status
- Surviving text of PHerc. 1667 read end to end (announced June 2026, preprint); first letters found in PHerc. 343 (October 2026)
- Last status change
- October 8, 2026
- Next milestone
- $1M Grand Prize deadline, June 25, 2027
- Primary source
- Vesuvius Challenge
This is a reading of a few letters from one small segment, checked by the challenge's papyrologists but not yet published in a peer-reviewed edition. Who wrote PHerc. 343, and what it is about, are still unknown.
DESI Strong Lens Foundry
Fifty five confirmed cosmic lenses give scientists new ways to study dark matter and how fast the universe is expanding.
- Field
- Science and discovery
- AI contribution
- Central
- Status
- 55 of 76 network-selected candidates fully confirmed with VLT/MUSE spectra, 15 awaiting more observation, 6 ruled out; published in The Astrophysical Journal Supplement Series, August 2026
- Last status change
- August 13, 2026
- Next milestone
- Further spectroscopy and Hubble imaging of remaining candidates from the neural network search
- Primary source
- The Astrophysical Journal Supplement Series 286, 9
Only 55 of the 76 candidates are fully confirmed with both lens and source redshifts. Fifteen have a lens redshift only and the authors say they need more observations, and six were not lenses. The campaign covered 76 of the thousands of candidates the network searches produced.
Quantum gravity-aided navigation at sea
Ships that lose or distrust satellite signals could steer by gravity instead.
- Field
- Quantum computing
- AI contribution
- Central
- Status
- Six-hour GNSS-free sea trial reported in preprint (Aug 2026)
- Last status change
- August 26, 2026
- Next milestone
- Multi-week GNSS-free navigation and peer-reviewed results
- Primary source
- arXiv
This is a preprint by company scientists, not yet peer reviewed, and the GNSS-free navigation run lasted only about six hours on one coastal route. The authors say multi-week navigation, sharper gravity maps and a smaller system are still needed.
EAGLE esophageal cancer screening
The chest CT scans many people already get for other reasons could one day double as a check for early esophageal cancer.
- Field
- Medicine and health
- AI contribution
- Central
- Status
- Published September 22, 2026; validated across 12 centers and 80,612 patients, including a prospective hospital cohort
- Last status change
- September 22, 2026
- Next milestone
- Broader international validation, especially in populations with more esophageal adenocarcinoma
- Primary source
- Nature Medicine
Sensitivity for precancerous lesions was 52.5 percent in external testing, so EAGLE misses about half of them at the high-specificity setting. The authors state that the diversity of training and validation populations requires further expansion and that broader international validation remains essential, especially where distal cancers and adenocarcinoma are more common, and they report sex-based differences in sensitivity that may reflect imbalance in the training data. Nothing here shows that using EAGLE reduces deaths.
Machine learning child malnutrition forecasts in Kenya’s drought early warning system
Kenya's drought agency now uses machine learning forecasts of child malnutrition, aiming to get help to families before hunger peaks.
- Field
- Climate and planet
- AI contribution
- Central
- Status
- Integrated into NDMA early warning (April 2026); NDMA using retrained model to trigger anticipatory cash transfers to mothers (Cornell, July 2026)
- Last status change
- July 28, 2026
- Next milestone
- Rigorous evaluation of the anticipatory cash transfers; county pilots in Isiolo, Tana River and Marsabit
- Primary source
- PNAS
The forecasts miss crises and raise false alarms: according to Cornell, roughly half of one-month alerts and a third of six-month alerts matched real crises in testing, and the paper shows accuracy collapsed for about two years after sentinel sites moved in 2016. Whether the anticipatory cash transfers actually protect children has not yet been evaluated. Two of the paper's six authors, Clinton Ouma and Nelson Mutanda, work for Kenya's NDMA, which collected the data and now runs the tool; the paper declares this as a competing interest.
Google Quantum AI: error-corrected Willow processor
Quantum computers that could one day help design medicines and materials will need to run for days without stopping, and this is a step toward that.
- Field
- Quantum computing
- AI contribution
- Central
- Status
- Learning-based control of error correction demonstrated in a memory experiment (Nature, July 2026)
- Last status change
- July 8, 2026
- Next milestone
- Faster agent-to-processor feedback loop and more advanced learning methods, per Google's July 22, 2026 blog post
- Primary source
- Nature
The 3.5-fold stability gain was measured against drift the researchers injected on purpose, not against months of natural drift during a real algorithm. The processor was storing information (a memory experiment), not running a useful calculation, and the large-code results are simulations.
Retina4IRD clinician decision support
People with inherited retinal disease could reach the gene behind their failing sight sooner, with an AI assistant guiding their specialist.
- Field
- Medicine and health
- AI contribution
- Central
- Status
- Randomized trial published July 24, 2026; primary outcome met in 295 analyzed patients; the registry lists the trial as completed (February to July 2025)
- Last status change
- July 24, 2026
- Next milestone
- Validation in more populations and imaging setups, and coverage beyond 17 genotype categories
- Primary source
- Nature Medicine
The system predicts only 17 genotype categories, and even assisted clinicians put the right gene first only 37.8 percent of the time. The PAP science service report on the study notes that most of the development data came from Asian populations and that a final diagnosis still requires genetic confirmation. The paper itself is paywalled beyond the abstract, so the figures here come from the abstract, the author list and the trial registry. In the paper's competing-interests statement, senior author Tien Yin Wong discloses consultancies with several companies and patent interests, and one coauthor is affiliated with a commercial retinal-imaging company.
Conversational epilepsy device platform
People whose seizures medication does not control could one day have a device that tells them what it sees and learns from their answers.
- Field
- Medicine and health
- AI contribution
- Central
- Status
- Inpatient prototype validated in 13 patients, reported October 13, 2026
- Last status change
- October 13, 2026
- Next milestone
- Prospective testing beyond the monitoring unit
- Primary source
- Nature Medicine
This is 13 inpatients in a hospital monitoring unit, only 2 of them with intracranial recordings, and the detector improvement came from a pseudoprospective simulation rather than live closed-loop use. Nothing here shows better seizure control or fewer injuries. In the paper's competing-interests statement, three authors, including the first and senior authors, are named inventors on a pending US provisional patent filed by the University of Pennsylvania.
ART (array associated reverse transcriptases)
The kind of repeated DNA that led scientists to CRISPR has turned up beside enzymes in viruses that attack bacteria, and nobody yet knows what it does.
- Field
- Science and discovery
- AI contribution
- Central
- Status
- Preprint posted October 11, 2026; arrays shown to be expressed during phage infection, function unknown, not peer reviewed, priority disputed by a University of Copenhagen researcher
- Last status change
- October 11, 2026
- Next milestone
- Experiments showing what the enzyme does to the array RNAs, and independent replication outside the company
- Primary source
- bioRxiv (DOI 10.64898/2026.10.05.755809)
The function of the ART system is unknown: the authors have not shown that the enzyme is active, that the repeat arrays act as templates, or that anything here edits DNA. The work is a preprint that has not been peer reviewed, all authors are Anthropic employees, some may hold equity, and a provisional patent has been filed. Priority is disputed: Mario Rodriguez Mestre, a PhD candidate at the University of Copenhagen, says his group has studied these enzymes since 2022 without yet publishing and that he shared his dissertation and manuscript drafts with Claude; Anthropic denies its models were trained on user transcripts or that its biology team had access to them, and no evidence has established that his material was used.