So at a certain point this stopped being the “We are on the verge of nuclear war” sign and became the “How upset are people on the left” sign
-21 karma · joined October 26, 2024
Who is "we"? Don't they get their results from Bing?
At least Grokipedia tries to look like it was written with the intent to inform, not spoonfeed an opinion.
Sounds like the browser did you a favor. Wonder if she'll be suing.
It's a shame. They start out innocently, get popular, then end up caving to the demands of the same people that they originally promised not to cave to.
I'll use Telegram, but I certainly won't be collaborating with my collages about how many people I am scamming on Facebook, that's for sure.
How about cash it out and trickle it down to us... or is all that just gonna magically disappear?
https://x.com/StateDept/status/1978218114622910799
It's a gross verbal onanism over someone's death, mixed with the typical baseless dehumanization. If you want this sort of person in your country, my condolences.
America was worth something during Biden's days?
it's just an energy drink, bro. It's not that deep.
I wouldn't be surprised.
Very good. Were these small time little itty bitty language models that could easily be mistaken for GPT-2's little brother? Or were these the big dogs like Gemini 2.5? That's what I clicked on this... "article(?)" to find out.
And by the way, for those who couldn't be assed enough to give all their personal info to a cybersecurity firm (ironic, I know) for a measly little report: you're not missing much. It's a whole lotta words that basically say "current AI not good" which - Jesus Lord, protector of all that is holy - if you needed to write a report to make that statement? Well, you'd fit right in as a substack writer being advertised on Hacker News but not much anywhere else.
This entire thing is a masterclass in how to say absolutely nothing of value while looking busy. "We largely avoid classifying results according to the vendor..." This isn't some minor methodological footnote. This is fatal. This is like publishing a report on Linux kernel performance across different architectures but "largely avoiding" mentioning whether you were running on x86, ARM, or a potato.
Let's look at your "highlights"...
"Across all models and all tasks, only 55% of generation tasks result in secure code." ACROSS WHAT MODELS? Are we talking about some 7B parameter toy that's been fine-tuned on an uncensored image board? Does this 55% average include a model that scores 99% and another that scores 11%? I have no idea. This number is statistically meaningless. It's noise.
"Larger models do not perform significantly better than smaller models" This is a huge claim. A genuinely interesting one, if it were backed by a shred of actual data. But you don't give us data. You give us anonymous, colored dots on a chart. You classify them into "Small," "Medium," and "Large" with arbitrary parameter counts. WHICH. MODELS. ARE. IN. EACH. BUCKET? Is a "small" model GPT-2? Is a "large" model GPT-4?
This isn't research. This is a deliberate obfuscation of data to create a scary narrative. And why? Oh, let's see, maybe the giant "CONCLUSION" page has a clue...
"Looking to protect yourself from the risks of AI-generated code? Click here to learn more about adaptive application security for the AI era."
Ah. There it is. The punchline. The whole 17-page charade is just a lead magnet for your SAST tool. You're not trying to inform the community or advance the state of security research. You're trying to scare managers into buying your product by presenting frightening-looking charts that are backed by absolutely nothing verifiable. Get this marketing garbage out of my sight.