@alexdaubois Curious about sharing the variable via WeakReference wrapper, then we can ensure that variable only available in our given scope, otherwise original object will be GC-ed completely.
Introducing SubQ - a major breakthrough in LLM intelligence.
It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA),
And the first frontier model with a 12 million token context window which is:
- 52x faster than FlashAttention at 1MM tokens
- Less than 5% the cost of Opus
Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention).
Only a small fraction actually matter.
@subquadratic finds and focuses only on the ones that do.
That's nearly 1,000x less compute and a new way for LLMs to scale.
I’m so happy that I’ve managed to do a lot around my open-source projects with #ai last months before all major players decided to reduce their limits and request more money for the same level. Nowadays even with subscription I’m hitting limits constantly and this is annoying…
@alexdaubois With enabled opcache and disabled file system checks, PHP takes class definitions from the shared memory in the form of opcodes and loads class faster, bypassing all expensive filesystem operations.
As a bonus, property interception now works for properties, declared inside traits as well, which wasn’t possible at all with previous version of AOP engine.
Just merged completely new implementation of property hooks for the #goaop framework that operates on native PHP Property Hooks! No more magic with unset of properties and __get/__set magic methods https://t.co/eFXwGMI0Yc
Together with native PHP8 attributes, plugin for the IDEA and #phpstan generics it works so natively! Can’t wait to finish all planned works around version of 4.x of framework.
@pronskiy@phpstorm@OndrejMirtes@return@DaniilGentili Also, can be a task for the PHP Foundation group - to address this as standard first and then as implementation in #php engine itself. Or for enthusiasts with vibecoding and enough credits/tokens ))
Pro-tip: if you want to adopt #agentic development for your PHP project, after initialization, request first your agent to achieve #phpstan level 10 step-by-step, committing each level separately. This will set a foundation layer for agents to make their work better.
@alexdaubois And special case with WeakReference https://t.co/xKztvL3tQb which doesn’t increase the recount of underlying Zval to allow GC to collect object earlier.
@alexdaubois The only exception with GC - are Fibers, there are gotchas how GC works for Fibers during destruction, check the Changelog https://t.co/9gngMpFqIb
@BenjDicken@blackcat_dev The most powerful combination is when you have a key, UUIDv8, which is primary key in the table (clustered index), partition key and sharding key at the same time. Then no matter how much data service will have - lookup time/insert time will be pretty stable.
@BenjDicken@blackcat_dev The only right decision here is to have, UUIDv7, automated partitions and partition pruning working effectively from the very beginning. We operating MySQL (AWS Aurora) databases, each with ~50Tb of data. Each partition size is almost the same - thus performance remains stable.
@TheCelavi Reason - is to have the ability to stop it. To prevent the humankind from the T-1000. Once it will work 24/7 - this evil will start to grow )
@alexdaubois Would definitely add here the ultimate thing, that I have implemented years ago with @goaop - Aspect-Oriented Programming. Still works, still the best thing that I discovered in PHP