There doesn't seem to be a good in-depth academic resource for advanced compiler optimization. I've searched a lot and all the courses I found were introductory, the actual interesting techniques require diving deep into the source code of popular OSS compilers. I found that quite surprising.

It seems to me that most academic courses overfit massively on parsing, while only teaching the rest superficially.

As if he wasn't playing with the prototypes daily before the launch.

Maybe he should step down instead. How about that? Nah, the money and prestige is too good. Better compose a PR piece to control the damage.

It is possible that he did, but that his brain did not compel him to give it politically spicy prompts. However, I think that should be expected from a general audience. I am more curious of who they used as a test group to give this a once-over and a "looks good to me". There is no way it was representative of a wide public consumer base.

With all the spyware there is on a company’s laptop nowadays I wouldn’t even dare to think about querying such questions in an internal-only version of the AI.

These were not politically spicy prompts though. They were normal questions with faulty answers generated.

Screw a test group, a test department, called QA, people who are paid to understand the system and how it works enough to evaluate if it's fully working or not and empowered to stop ship when it's catastrophically not working.

And if you silo all your employees to basically be clones of each other? I suspect it did pass QA with rave reviews!

This might sound harsh, but:

Maybe the wise choice is to change the path? That feeling of resignation, while a bitter pill to swallow, might be a good thing in the end. I used to beat myself over for not working hard and dedicating myself to my ideas, yet a few years down the line it became obvious the startups I wanted to build would have been made obsolete by other tech. I've dodged a few wasted years, a burnout or straight-up insanity right there. Trust your gut.

Why do something you are not enjoying? Life is short, it might end faster than you think. The people close to you can also disappear sooner than you think. Try to make the most of it. It's not all about making it big and being successful/rich.

Something to think about. Thank you.

ChatGPT 4:

Ten elephants would have a total of 32 legs if two of them are legless.

Analysis:

# Calculating the total number of legs for ten elephants, considering two of them are legless

# Number of legs each elephant normally has legs_per_elephant = 4

# Total number of elephants total_elephants = 10

# Number of legless elephants legless_elephants = 2

# Total number of legs # (Total elephants - Legless elephants) * Legs per elephant total_legs = (total_elephants - legless_elephants) * legs_per_elephant total_legs

Was this its raw response to the same query as in the OP? It seems odd it would provide a response using variables named with underscores, rather than using spaces, or more traditional algebraic notation (x/y/z).

ChatGPT (paid version at least) writes a quick python script in cases like these, and then executes it to get the result. For transparency, the script is shown in the output as well. Probably to avoid embarrassments like the ones we saw above.

Example (expand the program by clicking the little link at the end of the response): https://chat.openai.com/share/31e9a077-4c24-4c17-b30d-8d0c4d...

It's cool how rapidly these oft-called "fundamental problems with LLMs" vanish with bigger/better models.

GPT-4 examples elsewhere in the comments suggest otherwise.

That's just you showing that you clearly misunderstand. These aren't examples of fundamental problems, instead they are clear examples that these things are just autocomplete regardless of how many people think they are doing something more complex. It's not that getting the example correct is interesting, it's that getting it wrong is a clear sign of stupidity.

Did you use the same prompt?

Yes, I copy pasted it (How many legs do ten elephants have, if two of them are legless?).

Even the small Mixtral model gets this one right.

A large SAD light therapy light (beurer). Fixes my insomnia. Bonus points for acting as a webcam light.

What are they using to benchmark single-threaded performance? I don't think it's good to state that without clarifying what is the task being run. You could have a benchmark that increments a number in register. Since add/increment runs in a single clock, a high frequency CPU (Pentium 4 at 3.8Ghz?) would win over a wide core with great memory bandwidth (M2 at 3.7Ghz). Technically, that's a greater single thread performance, but not really applicable to the real world.

If it's over a variety of tasks and/or in a task that's similar to real world computation, kudos to them!

Just geekbench, not much else.

Because the temporal factor is the hardest one to game. Thus a strong predictor of quality/spaminness. Spammers used to do churn and burn and rank in a month or so. This has been killed mid 2013.

I've seen it work 'til 2019 at least

Use distilled water. I bought a distilling machine just for this and my pm2.5 sensor barely moves. Used to peak up to 120 on tap water. Now it’s consistently under 2 and the air is humid.

Distilling machine is a good idea since distilled water seems hard to find sometimes.

Only if they are poorly predicted.

New Intel/AMD CPU's do a register based popcount in a single clock.

Used to be three cycles.

Unfortunately, the original AMD64 back in 200x lacked a popcount, so most software built for PCs even today lacks any instances of the instruction. Means to get the instruction generated are finicky, non-portable, and often result unexpectedly in a function call, instead. E.g., without a "-m" option, Gcc and Clang will turn "__builtin_popcount()" into a function call. Likewise, "std::popcount()" and "std::bitset<>::count()". Always use at least "-mavx".