Kotoba is teaming up with @KindredVentures for a night on the frontier of Voice AI for East Asia 🎙️We'll cover our datacenter + on-device speech AI, our speech-to-speech simultaneous interpretation model, and the Kotoba app — now with hundreds of thousands of users in East Asia — plus where voice agents go 4pm next Wed, June 24 · San Francisco RSVP 👇 #VoiceAI# #GenAI# #Speech# #OnDevice#

chinese startup built an AI collar that translates barks and meows into full sentences. 95% accuracy. cost $118. 10k people have already pre-ordered it. It uses mics, motion sensors, and AI to read body language and vocalizations.

I'm fascinated by efforts to make animals (or parts thereof) photosynthetic. For the latest attempt, published last week, researchers took thylakoids from plant chloroplasts (the little membranes that carry photosystem I and II proteins) and inserted them into the eyes of mice. Specifically, they gave the animals eye drops containing the thylakoids twice a day for five days. The thylakoids went into corneal cells (apparently they are small enough that the cells endocytose them?) and did photosynthesis, making NADPH and ATP from light. This isn't a gene therapy, though; the eye cells cannot make more of these plant enzymes, and so the photosynthesis only happens for about 8 hours before the enzymes are destroyed. Why do this in the eye? One reason is that light doesn't penetrate tissue deeply; maybe a millimeter. Therefore, the eye is one of the few parts of the body that actually gets light exposure. It is also -- maybe equally important -- immune privileged, meaning these plant proteins don't trigger an inflammatory reaction (which would likely happen in other tissues). I'm not sure this paper has any real utility, at least not clinically. The authors say that it does (to help treat corneal injuries, for example) but I think it's just expected for authors to make up claims like this to get published in CNS journals. The more interesting reason to read this paper, I think, is just that it shows light can be used as a direct "metabolic input" in mammalian cells. You can use light to make energy molecules and NADPH, which can then be used by the cell's normal pathways. This isn't the first paper to do stuff like this, either. There is a rich history of (temporary) photosynthetic animals! In 2011, Christina Agapakis & co. injected living cyanobacteria into zebrafish embryos, and it worked. (No developmental impact on the fish.) In 2024, a Japanese team put chloroplasts from red algae into Chinese hamster ovary (CHO) cells, and the chloroplasts apparently survived and did photosynthesis for two days. Biotechnologists have a great ability to harness GENES taken from nature to build useful tools and therapies. We can sequence the natural world, collect genes in databases, and use tools like AlphaFold to figure out what they code for. But our ability to harness entire organelles -- structures crafted over millions of years of evolution, which perform functions that cannot be matched by genes alone -- is severely limited. Animal photosynthesis, and pursuits thereof, might be a useful way to start closing this gap.

We are releasing Carbon: a crazy fast DNA model Carbon is 275x faster than the next best model. So fast you can process the whole human genome on a single GPU in <2 days. Here are the tricks we used: When modelling DNA sequences a lot of the performance comes down to tokenizing the sequences in a smart way. BPE tokenizer struggle because there are no whitespaces and character (called base in DNA) level tokenizers waste a lot of compute on too many tokens. Carbon is built with a unique tokenizer: we split sequences in chunks of 6 bases, but during both training and inference we can work with single base resolution. That's similar to having word tokens but resolving them at the character level. All possible thanks to the DNA tokens unique structure. The architecture combined with the tokenizer makes the model 275x faster than the previous SoTA (Evo2) at this size. We built an interactive demo so you can explore how the model can generate DNA sequences, investigate the structure of genes, predict the effect of mutations, generate and fold proteins and even reconstruct parts of the tree of life.