I haven't read it yet to be certain, but my impression is that rather than using lossless compression of entire chats, this uses lossy summarization to get the gist of chats. There will be tradeoffs between the two methods, hopefully this paper covers that.
Not being snarky, at the simplest level by knowing x ahead of time you can change your code from y = n/2 into y = n * 0.5. Not wildly important by itself, but in a loop - maybe… I suspect compilers optimize for that already.
I still don't get it. How do you know that 1/x = 0.abc... without performing the division? I mean in general case, not in something matching binary tricks. Such as 1/3 for example. Unless you mean you somehow know the value of 1/x ahead of time. But where does it come from?
If you know the divisor x ahead of time you can pre-compute 1/x at compile time, so that your actual compiled code never does the division--only your compiler does. Your actual compiled code just does the multiplication by a pre-computed constant (and the compiled code doesn't have to know that that constant was pre-computed as the reciprocal of x).
Often, you can amortize a division or reciprocal by calculating it once and then reusing it. Frequently the divisor is dynamic, but reused locally.
For example, if you want to normalize a 3D vector you could do:
mag = sqrt(x*x + y*y + z*z)
x /= mag
y /= mag
z /= mag
That's three divisions with the same divisor. But you could instead do:
invMag = 1.0 / sqrt(x*x + y*y + z*z)
x *= invMag
y *= invMag
z *= invMag
There's still a single division (or reciprocal) done here. But you've eliminated at least the other two. (And it's even better if you have an rsqrt function.)
In some areas it does a good job of giving me a better answer than me checking 5-10 random stack overflow and blog posts with different answers over 20 years...
But, it is completely incapable of saying "I don't know the answer to that" and will spew the wrong answer to anything. For fun I have arguments with ChatGPT that it is just making things up and it should just tell me "I don't know" and it will happily say "I'm sorry, you are right. Here is the right answer <insert made up fact>" and it will do so generating random garbage indefinitely.
The problem is that I can't trust it to provide facts any better than asking my bull-shitting co-worker who is convinced he is right about everything and when called out says "yeah, that's what I was saying, <insert contradictory thing>"
I believe the name is hydrogen fuel cell. It takes hydrogen as a fuel and generates electricity like a battery, instead of combusting it like a typical gas powered motor
The output of a hydrogen fuel cell is water, so it is a fairly sustainable loop if you can capture the water and generate hydrogen from it. But you need an efficient process to convert water into hydrogen
That fuel cell is more like an engine than a battery. You have to get O2 into it, get waste heat out (harder than the engine because the temperature is lower), keep it from drowning in the water it makes, etc.
It is hard to see the fuel cell EV competing with a battery EV.
Ok - it is a true shame that the YouTube copyright claim system is so broken as to enable those shady practices, and that politicians still haven't upped their knowledge of the internet beyond a 'series of tubes'.
But surely the answer should be to fix the broken YT system and to educate politicians to abstain from baseless threats, not to make AI researchers pay for it?
Just to be clear, because it's in the title, the reporter was threatened with a lawsuit for looking at source code. I cannot find anyone acually sued for it. BTW, here's an article saying said reporter wasn't sued: https://www.theregister.com/AMP/2022/02/15/missouri_html_hac...
Anyone with a mouth can run it and threaten a lawsuit. If fact, I threaten to sue you for misinformation right now unless you correct your post. Fat lot of good my threat will do because no judge in their right mind would entertain said lawsuit because it's baseless.
I’ve switched back to IC. But my tendency to manage things has pushed me into a hybrid tech lead role that feels like a blend of IC and management. But the things I’m managing are cost savings, migrations, rearchitecting… so I’m very happy with this role. I don’t know what the jobs and roles are where you are, but I’m at a young company that is still growing and have been able to forge my own path. If you can use your management connections, find a technical area that is lacking and that you have the unique technical skills to solve, that could lead to some exciting work.
I’d worry that working on just any other team that you don’t feel passionate about would make you start questioning the move
But. Careers are not just about maximum growth. It sounds like you need to take some time to take your foot off the gas. Switching to an IC role could let you focus on what YOU want to do for a while and then pick something new when the time comes
I’m not sure what you mean by “bumping the version in source after a release”. If it’s after a release then you released it with the old version? Or do you mean prep the repo for the next version by incrementing it optimistically, in which case it probably depends on your branding model. But I don’t understand the “immediately” part of your question.
Where I work we don’t increment until we know what branch/version we are working on, and determine the semantic version, and then we still use tagging because we tag the version + build id which is computed at build time
My understanding is that initially they have version 1.4.2 and release 1.4.2. Then in the next commit, immediately after release, they bump the version to 1.4.3.alpha0 or something of that nature.
I personally think that's messy because it requires two changes to the version instead of one and there's no way of knowing that the next version will be 1.4.3. If a developer makes a breaking change, will they update the version in the code to 2.0.0.alpha0 or do they fix the version when they release?
I suppose if you have nightly releases then it could make sense but then I think using the date or the commit hash as the version number would make more sense.
