Progress with RL is very interesting, but it's still too inefficient. Current models do OK on simple boring linear code. But they output complete nonsense when presented with some compact but mildly complex code, e.g. a NumPyro model with some nesting and einsums.
For this reason, to be truly useful, model outputs need to be verifiable. Formal verification with languages like Dafny , F*, or Isabelle might offer some solutions [1]. Otherwise, a gigantic software artifact such as a compiler is going to have a critical correctness bugs with far-fetched consequences if deployed in production.
Right now, I think treating a LLM like something different than a very useful information retrieval system with excellent semantic capabilities is not something I am comfortable with.
I really like everything Uri Alon (last author) publishes, but these types of studies have a history of inflating genetic contributions to phenotypes. Decoupling genetics from environment is not easy as they are both highly correlated.
In fact, the article discussion states: "Limitations of this study include reliance on assumptions of the twin design, such as the equal environment assumption". My take on this is that the main result of the article is probably true, but the 50% figure is likely to be inflated.
I hit the jackpot with the ultrasound technician who spoke passionately about what she believed about lifestyle risk for cardiovascular conditions and she believed quite strongly that heart disease runs in families more because lifestyle runs in families than because of genetics. She's not at the top of the medical totem pole but I can say she inspired me to take responsibility for my health than the specialist who I talked to about the results.
If the environment was significantly more varied in health impact between twin comparisons than expected, then the correlations they found under estimate the genetic component.
Some randomness is part of the signal being studied, and some is undesired measurement noise to be controlled for. And it is only the latter that is beneficial to be carefully removed or otherwise controlled for.
I prefer X11 as well, but it has some security issues. Notably, all applications can read your input at any time. It's really hard to sandbox.
Wayland brought some irritations, including increased latency, and an architecture that requires rethinking all window managers. A rewrite is not enough. Very annoying.
I will never understand why "the computer can tell what input it is receiving" has turned into an accepted threat model.
I understand that we have built a computer where our primary interface depends on running untrusted code from random remote locations, but it is absolutely incredible to me that the response to that is to fundamentally cripple basic functionality instead of fixing the actual problem.
We have chosen to live in a world where the software we run cannot be trusted to run on our computers, and we'd rather break our computers than make another choice. Absolutely baffling state of affairs.
Defense in depth. One compromised application may do a lot of harm if it has access to your keyboard inputs. Supply chain attacks are not that uncommon. While you can trust software developers, you cannot completely trust their builds.
I agree. I think fixing the keylogging issue should be possible without dumping the entire architecture. Perhaps the new X11 fork https://x11libre.net will achieve that? At least, it's encouraging to hear it's getting maintained.
Regarding (recent) supply chain attacks, Linux needs to take supply integrity and sandboxing more seriously. The tools to do so are there (e.g. Nix and firejail/bwrap) and, unlike Wayland, they play well with existing software.
And when someone violates that trust, do you then tear the house down and build one with only external doors, requiring inhabitants to circle in the yard to move between rooms? The point of the Wayland security model is that the inhabitants of the house do not trust each other, and the architecture of the house must change to accommodate that.
I'm not impressed with the analogy. I am not confused about the goals of Wayland's security model. I am dismayed at the poor judgment elsewhere in computing that has led to its necessity.
Deep SSMs, including the entire S4 to Mamba saga, are a very interesting alternative to transformers. In some of my genomics use cases, Mamba has been easier to train and scale over large context windows, compared to transformers.
I doubt it, I feel like it might improve shops that already care and are already creating with rigor. I don't think it'll raise the bar for the avg shop. However, perhaps that's just be being cynical. By real time embedded is the same do you mean the same in the sense that they are just as poor in quality?
> [...] the same in the sense that they are just as poor in quality?
I mean some real-time software for critical embedded systems has an incredible level of rigor, making heavy use of static analysis, model checking, and theorem proving.
reply