If you have not updated your llama.cpp in a few weeks it's likely you are missing out on massive gains on blackwell hardware. FYI.

A few days ago i talked about how most companies are paying for more externalities with AI than getting a measurable benefit.. I called a few of my clients and explained the problem with AI bots and pitched the idea of a security megaproject: - build a state actor-grade AI pentesting apparatus - completely remove wordpress as a vector by flatfiling it and re-engineer wordpress sites to fit - Finish building defense system that exceeds cloudflare's capabilities - Build system that identifies attack patterns and reports them with suggested remediations - Build system that auto-snipes entire ASNs based on long term collected reputation I told them this is is a moonshot and easily low 5 figures per client since we are pooling R&D. Also i must take some % of developer labor off their existing projects. So far i have 2 yesses and 1 maybe on getting funding for the mega project. I'm surprised by the reception - but it really is bad enough out there to justify the investment.

We have taken many left exits in the attempt to finish the 'Blast Processing' part of Zerolith. It started out using Swoole, then re-implementing IBM's fast low level disk cache, then to experiment with APCU and memcached, and ended up back in async land. Blast processing was supposed to be a go-fast switch you enable and magic shit happens, but we found out it's not that easy - and the solutions that mostly worked, were notably slower than ones that require some user intervention. We thought it would be wrong to favor convenience over speed so much. It was worth investigating all possible routes and seeing the pros and cons. In the end we've decided to implement easy async, fast background writing/function execution to offset heavy or high frequency operations, and semi-automated memcached database caching support. You should see aspects of Blast processing gradually appear starting with Preview Release 3

My friend seems to think i could make money building bespoke AI hardware + software setups so that a company can get guaranteed privacy. I disagree because so many people are rent-it-pilled and don't care about privacy. So i haven't bothered with that. What do you think?

Whatever money some companies have saved using AI, they have paid a lot to guys like me in cybersecurity because of how rapidly AI is hacking and breaking stuff. Literally the only reason 35 servers are on the internet still is because i patch/mitigate things same day now Half my year has been spent mitigating AI!

I'm always surprised when my Mazda 2 even needs gasoline.. just a super infrequent event, even with a 10 gallon tank

Have you heard of phasync for PHP? It adds extensive async and non blocking IO without having to do much in user code land by using some newer low level PHP functions. AFAIK it's a prototype for native async in PHP, too. It's not surprising to see it beat NodeJS and is competitive with Go. The language itself happens to be very fast and computationally efficient at this task. You can thank C and Rasmus Lerdorf for that. After removing blocking IO, we see it's true speed potential. Zerolith still performs this task 4 times faster per CPU core, but doesn't help make async easy. I think the second PHP implements it, we are absolutely all in on fixing PHP's achilles heel in our framework design too. The performance potential is just too juicy to ignore.