I asked super intelligence to determine if a series converges or diverges yesterday. It said cos(1 + n * pi) is always positive for all n… I must be wrong, not the super intelligence.
This is a view I strongly agree with. While the “worked example effect,” which is apparently a key part of the platform, does seem to accelerate learning compared with discovery based learning [0], I do not think the learning achieved by the regurgitation of another person’s ratiocination is nearly as useful or rich as learning acquired by discovery. In my view, the best way to fully and completely understand something is to come to it in your own mind.
If you want to do this, there are at least two ways:
A) use the AoPS books, which present exercises before explanations, and expect students to attempt them first, or
B) Use Math Academy the way my son does: skip the explanation and worked example and immediately attempt the first problem. Only if stuck go back to the explanation and example.
(B) isn't what Math Academy recommended, presumably because you might be able to solve the problems in the current lesson, but still miss out on something else the lesson is supposed to teach you.
One of the most egregious cases of this, in my opinion, is only publishing metrics that cover part of the confusion matrix. “The false-negative rate? That could not possibly matter for a variant effect prediction model; why would we include that in the paper?” Example: AlphaMissense.
I really want to like Julia. It has a nice type system, a good ecosystem, reasonable syntax, and it’s far faster than Python. But there are several issues with its DX that prevent me from using it: the most severe of which being the complete lack of a cache for the JIT (or JIT like system), inducing multi second compile times for scripts that run in <100ms. There are some external packages that try to solve this issue, but they are far from first-party quality, and, in my experience, are quite buggy.
(I last tried Julia a few years ago; perhaps this has been improved since?)
pkgimages are keyed on package (and dependency) version so if you frequently add or update packages they don't help as much as they should. this new cache is at the function granularity and caches the llvm optimization and native codegen steps.
I imagine this will only become more important as we have more agents working on codebases in parallel; so a clean solution would be extremely valuable.
On the flip side, the rise of agentic coding means that the library/batteries mismatch against python will largely stop mattering. It should be very easy to have an agent implement large libraries -- especially if they can be just a translation from one language to another (esp. one with with better primitives!)
The benefit is probably from the removal of fat, not a direct antibacterial/antiviral effect. Fat plays a complex immunoregulatory role in human physiology: it down-regulates some pathways, while up-regulating others (notoriously, the production of IL6 is carried out, in part, by adipocytes). The overall effect of fat on the immune system, however, is negative: it tends to increase the chances of rheumatological disorders, cancers, and many other diseases. Alternatively, the effect may be due to some sociological factor that their analysis failed to account for.
Nobody's denying this, but the interesting story about GLP-1s is that after you control for fat loss (for instance, by taking cohorts of patients that aren't losing weight) you still get evidence of these off-target effects.
I agree that purchases on e-commerce sites are virtually never this fast; but I have used some extremely fast physical payment terminals (e.g., Square): sometimes it seems like the transaction is sub 300ms, which I find very impressive given the complexity of payment processing.
reply