1,064 karma · joined December 15, 2015
Nowadays, you can just write in python with no care for the hundreds of thousands of cycles and megabytes of memory you are wasting.
All indications today point towards the same happening with the cost of intelligence.
Scale matters a _lot_.
This is temporary. There is a future where AI builds tools made by AI for AI, evolving at speeds that we may no longer be able to follow.
Even with our current nascent technology, we can already observe this happening. See the recent laments on the AI Bun rewrite that "it's a million lines of code that nobody has ever reviewed." Or Cursor building a new source-control system for LLMs, because git is designed for human collaboration speed, not hundreds of changes per second.
That's just 4 years in. How much of humanity's code will remain hand-written vs LLM-written 2, 5, 10 years from now?
Indeed, if you remember before AI took the world by storm, HN used to be chock-full of articles about how the hiring process is broken for both employers and candidates, where you can never tell if what you see is what you get.
When I run a local LLM I get none of that. I hit the intelligence walls or buggy behaviour, but it doesn't matter if it's 8am or 8pm, the model behaves exactly the same. If something doesn't work as I wished, I can retry as many times as I wanted without the model getting angry at me.
UB supersedes volatile, once the compiler hits UB then all bets are off. Compilers can and do optimize out UB branches, which is almost never what you want... yet here we are.
There is a list of open items here, it's looking pretty good tbh: https://github.com/orgs/uutils/projects/1
In fact, the gap is so large that it can be really hard for a person on one side of it to understand how people on the other side think.
It didn't bench well against the other benchmaxxed models, and it was too expensive to run, but it was a glimpse of the future where more capable hardware will lead to appreciably smarter models.
The real question here is how would _we_ be able to recognize that? And would we even have the intellectual honesty to be able to recognize that, when at large we seem to be inclined to discard everything non-human as self-evidently non-intelligent and incapable of feeling emotion?
Let's take emotions as a thought experiment. We know that plants are able to transmit chemical and electrical signals in response to various stimuli and environmental conditions, triggering effects in themselves and other plants. Can we therefore say that plants feel emotions, just in a way that is unique to them and not necessarily identical to a human embodiment?
The answer to that question depends on one's worldview, rather than any objective definition of the concept of emotion. One could say plants cannot feel emotions because emotions are a human (or at least animal) construct; or one could say that plants can feel emotions, just not exactly identical to human emotions.
Now substitute plants with LLMs and try the thought experiment again.
In the end, where one draws the line between `human | animal | plant | computer` minds and emotions is primarily a subjective philosophical opinion rather than rooted in any sort of objective evidence. Not too long ago, Descartes was arguing that animals do not possess a mind and cannot feel emotions, they are merely mimicry machines.[1] More recently, doctors were saying similar things about babies and adults, leading to horrifying medical malpractice.[2][3]
Because in the most abstract sense, what is an emotion if not a set of electrochemical stimuli linking a certain input to a certain output? And how can we tell what does and what does not possess a mind if we are so undeniably bad at recognize those attributes even within our own species?
[1] https://en.wikipedia.org/wiki/Animal_machine
https://gcc.gnu.org/legacy-ml/gcc/2016-02/msg00381.html
"The fact is, undefined compiler behavior is never a good idea. Not for serious projects."
"THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE."
To use the software you have to accept the license, which means you explicitly confirm that they are not your supplier. Pretty clear cut, no?
Edit: EULA-loving companies don't want to accept the license terms for the _free_ software they themselves are using - the hypocrisy is nothing short of staggering.
Good riddance for Lightning.
Good job.
Our reinvented wheel using posgresql, rabbitmq and EC2 runners has ~10x better throughput and scales linearly with the number of pending tasks, whereas the airflow falls apart and fails to keep the runners fully occupied the moment you out any real load on it.
On the plus side, it does have offer communication advantages if you have entirely independent BE and FE teams, and it can help minimize network traffic for network-constrained scenarios such as mobile apps.
Personally, I have regretted using GraphQL every time.
The improvement loop is likely not fully autonomous yet - it is currently more efficient to have a human-in-the-loop - but there is certainly a lot of LLMs improving LLMs going on today.