The Week Five Channels Opened

Named for Rosalind Franklin — the British crystallographer whose X-ray diffraction image, Photo 51, revealed the double helix structure of DNA and who was excluded from the 1962 Nobel Prize awarded to Watson, Crick, and Wilkins because she had died in 1958 and the prize is not awarded posthumously — OpenAI's GPT-Rosalind was released this week as a restricted research preview with launch partners at Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific. On BixBench, the standard bioinformatics benchmark, the model scored 0.751, the highest published score among models with results on record. On LABBench2, a suite of 1,892 life sciences tasks, it outperformed GPT-5.4 on six of eleven categories, with the strongest gains in CloningQA — the end-to-end design of molecular cloning protocols. At Dyno Therapeutics, on unpublished RNA prediction tasks, the model's best ten submissions ranked above the 95th percentile of human experts. Joy Jiao, OpenAI's Life Sciences Research Lead, said: "We do think there's a real opportunity to help researchers move faster through some of the most complex and time-intensive parts of the scientific process." Drug development currently takes ten to fifteen years from target discovery to regulatory approval in the United States, and nine out of ten drugs entering clinical trials fail. The model does not cross the finish line. It shortens the approach.

In a separate experiment with no institutional affiliation and no peer-reviewed publication — a blog post at writetobrain.com, four researchers, and a custom ultrasound transducer array — four distinct smells were induced in two human subjects by focusing 300-kilohertz ultrasound at specific coordinates 39 millimeters below the forehead, inside the olfactory bulb. Each coordinate produced a different smell: fresh air, garbage, ozone, campfire. The subject wore noise-canceling earphones during the blind trial; the researcher selected coordinates the subject could not hear. The team — Lev Chizhov, Albert Yan-Huang of Caltech, Thomas Ribeiro, and Aayush Gupta — notes that, as far as they can verify, this kind of stimulation had never been attempted before, not in humans, not in animals. No chemical cartridges. No implants. No electrodes. The smell is a parameter. The olfactory bulb has an address, and it can be reached from outside the body.

Reading thought without surgery is a problem that has resisted neuroscience for as long as neuroscience has existed. The skull attenuates electrical signals from the brain severely enough that conventional EEG — even laboratory systems with 256 electrodes — cannot reliably extract the fine-grained patterns of internal speech, the words a person constructs mentally without vocalizing. Sabi, a California startup backed by Khosla Ventures, Accel, and Initialized Capital, has woven between 70,000 and 100,000 miniature EEG sensors into a fabric hat and trained a proprietary foundation model on 100,000 hours of neural recordings from 100 volunteers. The target output is 30 words per minute, with no daily calibration, from a device that looks like a hat. Vinod Khosla said: "The biggest and baddest application of BCI is if you can talk to your computer by thinking about it." The device has not shipped. The dataset exists. The model is trained. The hat is real.

Gašper Beguš of UC Berkeley, working with colleagues Maksymilian Dąbkowski, Ronald Sprouse, David Gruber, and Shane Gero, published this week in Proceedings of the Royal Society B the result of analyzing 3,948 sperm whale codas — rhythmic click sequences recorded from 15 individuals off the coast of Dominica over 13 years. Stripping the silence between clicks and analyzing each click's frequency structure with the same acoustic framework used to classify human vowels, the team identified two discrete categories: one matching the formant pattern of the vowel in "father," one matching the vowel in "feet." These categories interact with coda rhythm in ways that parallel phonological rules in human language — specific vowel qualities align with specific timing patterns; two lengths of one vowel category exist, as in Finnish or Latin; edge clicks carry spectral traces of adjacent codas, as sounds in human speech carry the influence of their neighbors. Beguš said: "On the surface, their clicks sound nothing like our vowels. But that's because their clicks are slow and our vowels are fast." David Gruber said: "I think it's another humbling moment that we're not the only species with rich, communicative, communal and cultural lives." What the vowels mean has not been established. That they exist has.

Jongpil Kim and colleagues at Dongguk University in Seoul published this week in Cell a gene switch that activates when exposed to a magnetic field and turns off when the field is removed. The mechanism — identified through a genome-wide CRISPR loss-of-function screen — centers on a protein called Cytochrome b5 type B, or Cyb5b, anchored to the outer mitochondrial membrane. When stimulated by electromagnetic fields in the range of 0 to 10 millitesla, Cyb5b triggers a specific pattern of oscillatory calcium release inside the cell. Not generic calcium influx — the specific oscillatory signature, which other biological processes that elevate calcium do not replicate. Only this signature activates the Lgr4 promoter downstream, which drives expression of whatever gene has been placed behind it. In live mice, the team demonstrated partial cellular reprogramming via Oct4-Sox2-Klf4 expression triggered by the field, conditional Alzheimer's pathology modeling, and restoration of serotonin synthesis in a depression model. The switch requires no drug, no implant, penetrates tissue freely, and turns off when the field is removed. The body has a wireless on-button. It is not in the body.