I guess we can start a chain of comment listing their favorite ones, I haven't tried many but "wobble stack" is pretty cool and simple, "Starfall Defense" is a much more complex game but I love tower defense.
Maybe a year ago I’d agree but the gap has grown. I also pay for Cursor which is based on Kimi and there is no comparison for complex code gen vs Fable. It mostly succeeds well at small rapid fire stuff which is the only reason I pay for it (plus the IDE DX). But any heavy feature planning and prototyping I use Claude.
I predict they will all be mostly the same in 5+ yrs but coding is serious work and companies aren’t going to pay for almost good.
I know nothing about this field, but I imagine the actual problem is how do you deliver the Cas12a2 protein to each individual cancer cell compare to a viral gene therapy?
There are two major problems, delivery is one of them. Collateral damage of mass cell destruction leading to systemic inflammation is the other.
The approach I'm reviewing now uses lipid nanoparticles (LNPs) for delivery. It isn't great for targeting my bone marrow condition but its workable. The team hasn't optimized it at all, either. There are also viral delivery mechanisms that I haven't studied yet.
The collateral damage problem is the backpressure on the delivery problem. If you get really good at delivery, you can destroy A LOT of cells very quickly. The human body (usually) responds to these events by releasing a lot of pro-inflammatory cytokines. This can lead to cytokine storms or worse.
As you "get good" at killing the target cells, the net effect can turn bad. It will probably be a balancing act.
Lipid nanoparticles are quite old as-is. How do you target cells specifically?
> If you get really good at delivery, you can destroy A LOT of cells very quickly.
You can destroy cells quickly. Ok. So the question is: how do you detect specifically only cancer cells via lipid nanoparticles? That was already a problem years ago with Herceptin. The rationale that is always used is that "we need to do something" for certain aggressive cancers. It has never been a super-effective technique, despite all the promo of how monoclonal antibodies are so accurate.
> As you "get good" at killing the target cells, the net effect can turn bad. It will probably be a balancing act.
That's already the status quo in the whole cancer field. I don't think that more than sloppy accuracy is acceptable for any gene therapy - and the off-target cleaving of CRISPR has always been the number #1 problem here.
> I don't think that more than sloppy accuracy is acceptable for any gene therapy
Valid critiques of Cas12a2 must acknowledge the mechanistic differences between Cas9 and Cas12a2. There is no research to suggest Cas12a2 is "sloppy" and significant research that demonstrates it is not "sloppy."
I appreciate the skepticism but I would encourage you to study the actual mechanism discussed in the paper.
> So the question is: how do you detect specifically only cancer cells via lipid nanoparticles?
You don't. Healthy cells will also get these nanoparticles, but without the triggering DNA sequence, the mRNA payload will remain inert and eventually will be degraded.
> Healthy cells will also get these nanoparticles, but without the triggering DNA sequence, the mRNA payload will remain inert and eventually will be degraded.
Naively, I would deal with this by deciding how many cells I want to kill each day and then figure out a dosing schedule that achieves that. Or maybe it's better to do one dose every few days. But yeah either way.
> I would deal with this by deciding how many cells I want to kill each day
Yes, this is an approach. This starts to exceed my understanding of the problems the teams are facing - but there are concerns about the efficacy of Cas12a2-based approaches fading. Not because the cells adapt, but because your immune system starts acting funny in the presence of the payload.
I don't recall the specifics but there seems to be a window of opportunity for these things.
Well with a standard autoregressive model you can generate for example 256 tokens at once if you have 256 users, with this approach you can generate 256 tokens for a single user but you need several forward steps.
So the diffusion process takes more GFLOPs, if you have enough users you can already balance memory and compute.
What's novel here is the extremely small KV cache memory usage per long context windows, like 0.77GB with 512K, a 90% memory usage reduction compare to the already really small KV cache memory usage of Deepseek V4 Flash.
reply