claytonius

Part of what’s interesting here is that there haven’t been any robust featurizers for DNA in the same sense that we have robust featurizers for proteins (like ProtBERT et al.) that just work out of the box for amino acid sequences. Since many genes have >100k bp vs. proteins that are often less than 1k AAs, the much longer context window is needed. ProtBERT gets ~200k downloads per month on huggingface and DNABERT gets ~10s.

Oocytes are one of the most difficult cells to engineer as well... The mitochondria in Oocytes are turned off from birth because otherwise reactive oxygen species will damage the mtDNA. One would have to use CRISPR to re-build the mtDNA, use telomerase / crispr to get the telomeres to exactly the right length, fix any mutations within the genome, etc. There are many many technical challenges to overcome, requiring further development of multiple immature technology stacks, and the bar is much higher than it is for mice since you would want a baby with the cellular metabolism profile of a 30y/old.

It may help to have the set of keywords that will yield better search results. Geneticists generally assume that every trait is determined by a combination of Genes and Environment (GxE), and the differences between fraternal and identical twins are used to determine what component is heritable/genetic and what component is environmental.

I don’t think it’s straightforward to do a head to head comparison.

from: https://www.youtube.com/watch?v=yso2S2Svdlg

@ 25:14

James Wang: "If a model doesn’t fit into a GPU’s HBM, is it smaller when it’s laid out in the Cerebras way relative to your 18 gigabytes?"

Andrew Feldman: "It is — it’s smaller in that we hold different things in memory than they do. One can imagine a model that has more parameters than we can hold — one can posit one, but remember our memory is doing different things. Our memory is basically holding parameters. That’s not what their memory is doing. Their memory is holding the shape of the model, their model is holding the results of the batches. We use memory rather differently. We haven’t found models that we can’t place and train on a chip. We expect them to emerge, that’s why we support clustering of chips and systems, that’s why we do that in whats called a “model parallel” way, where If you put two chips together you get twice the memory capacity. That’s not what you get when you put multiple GPUs together. When you put multiple GPUs together you get two versions of the same amount of memory, you actually don’t get twice the memory. I see you smiling here because you know that’s a problem… …With us if we support 4 billion parameters and you add a second wafer scale engine, now you support 8 billion parameters ,and if you add a third you can support 12 billion. That’s not the way it works with GPUs. With GPUs you just support two chips, each with a few million - tens of millions of parameters."

Congratulations to Regina, it's well deserved. She's clearly doing something right over there since members and former members of her lab have been consistently ahead of the rest of the pack on a variety of tasks by a wide margin.

Something like a CycleGAN and/ or weight sharing would do the trick

I think you're right that most DeepFakes use Variational AutoEncoders (As described by the article) where the latent space is jointly distributed between the two domains, allowing for translation. For additional context: AutoEncoders are also frequently adversarially regularized so you're not wrong about GANs being involved. My understanding is that the most accurate implementations also have some manual feature engineering for more tractable problems like facial orientation.

One of my favorite related papers: https://arxiv.org/pdf/1703.00848.pdf

The prices were very low -- if I remember correctly the zero was £4.50 and the zero W was £5.00. I went in looking for a motor so I didn't catch the price of the Pi B+ etc. Sorry!

I had a minute to check it out over lunch - most of the floorspace is dedicated to demonstrating what the raspberry pi can do at a high level. They had stations for coding, gaming, sensors, etc. but only ~1/4th of the space was devoted to inventory. While they have a decent selection of Pis, sensor kits, and accessories, it's definitely not a replacement for Maplin. Additionally, Already__Taken seems to have been correct about the level of knowledge among the staff - not everyone working there was technical. This is definitely aimed at the general public.

Not quite... https://www.history.com/this-day-in-history/fdr-takes-united...

Unfortunately even with just collagen we can see an immunogenic response that is annoyingly exacerbated by UV cross-linking of the individual collagen proteins. That's why many of the best manufactured ECMs are cross-linked using a mixture of UV exposure and EDC-NHS. Without photolithography i'm not sure you'd be able to get a resolution high enough to mimic native ECM structures.

You're right that we could already re-seed decellularized pig ECMs with human stem cells, but we have also already been able to derive stem cells from the stomach. I think the main claim that they're making is that by deriving both the ECM and the Stem cells from the patient, the resulting tissue engineering scaffold is less immunogenic than other methods.

While this is somewhat interesting, they get into trouble by going on to claim that they've been able to generate "functional cardiac, cortical, spinal cord, and adipogenic tissue implants". From what I can see in the paper they saw increased expression of a few markers associated with those cell types and some minor electrical activity, etc.

The idea that differentiating stem cells into vaguely tissue-like phenotypes is equivalent to being able to manufacture organs brings back memories of the shenanigans that went down at UCL a few years ago [1].

1. https://forbetterscience.com/2016/12/21/birchalls-trachea-tr...

It seems to me that the main advancement here is that the native extracellular matrix is being re-seeded with differentiated stem cells, as opposed to ordinary tissue engineering which uses a manufactured ECM or tissue transplants, which do not involve making stem cells.

There’s still a long way to go before organs can be engineered, even with this tech. We will need much finer control over the delivery of differentiating signal factors for pluripotent stem cells as well as ECM structure.

I'd be interested in knowing where you've been getting your information from. Teslas don't have engines but the AC induction motors they use are substantially more advanced than anything that existed before [1]. Tesla is also fairly famous for having a $0 marketing budget.

[1] https://cleantechnica.com/2018/03/11/tesla-model-3-motor-in-...

It's also important to consider that photosynthesis only captures photons in a very narrow spectrum relative to the total spectral output of the sun.

Looks useful -- Out of curiosity, how do you determine where to place holes in the sidewalls or sockets for the board to screw into?

From what I understand, artificial ECMs can be more difficult to build in some ways, but if you can figure out how to incorporate the right signaling factors into an artificial collagen matrix, you can avoid a lot of challenging biocompatibility/immune reaponse problems. Also, I believe most cells are differentiated before being seeded onto the scaffold.

First sentence... “Germany’s Bayer (BAYGn.DE) will wrap up the $63 billion takeover of Monsanto (MON.N) on Thursday”

I'm currently sitting next to Paolo Bombelli... He laughed and said that your version of events is fairly accurate, although they used liquid nitrogen as an "anesthetic".