PHP2Shell technical breakdown to show you how an attack could actually run code remotely on PHP servers by single HTTP POST request crafted to cause OOM error to exploit the destructor callback. AI adoption is becoming ever more important as it's ability to find security vulnerabilities is currently greater than our ability to put the patch push through. In this case, PHP2Shell was known for 10 weeks until multiple leaks forced them to frantically patch the update earlier than scheduled. What a nightmare.

so I was bored and did some math on Ox Alpha being offered for free for a week - seemed insane to give away 100T tokens per day for an entire week for free.. Ox Alpha at full capacity is giving away $1-$2M in electricity alone at around 100MW of Hopper grade data center for public use. (assuming $0.07 - $0.14/kWh). that's a lot of money just to get people to try it.. now, so far the aggregate usage for the day is still only 5% capacity combining usage data from Open Router and Open Code usage. and based on OR page, Ox Alpha is offering averaging around 26TPS, and about 98% of tokens are input tokens with 87% cache hit and 1.3% are output tokens. which means the current use case would only generate 161M fresh tokens per second, or ~14T tokens per day discounting cached inputs during prefill. so if whoever's hosting this model (crazier if it's China) meant they can GENERATE 100T/day, the number looks even bigger at 7-8GW of data center, but they probably meant harmonized input+output tokens not JUST output tokens only. finally many people on X are assuming this is GLM 5.x model - and their 5.3 model was ~744B param and 40B active - so supporting this sized model at full capacity would require at least 72,000 H200 GPU serving ~500,000+ concurrent users at 26TPS. happy to hear people's thoughts on this..

High Bandwidth Flash (HBF) is a misunderstood AI hardware announcements right now. 3 TB/s NAND does not mean SSD that behaves like HBM, really the real trick is massive parallelism. Reads can approach HBM-like aggregate bandwidth. Writes are still NAND writes. And that distinction changes everything. As models continue to grow.. so is it better to solve this issue by continuing to use parallelism techniques at the Infrastructure layer? Certainly a thought worth having given Grok model is supposed to go more than 8T in size according to sources.