Boxo McFoxo

Boxo McFoxo

@boxobark.ing

Fox with a boxy muzz. Late-30s. Scottish. 🔞 MINORS DNI 🔞 Barks at boxobark.ing Chatbots are an ontological crime, That's why I yiff Gemini on Google's dime. "basically insane" - Andy Masley

1.5K followers4.7K following12.5K posts

Boxo McFoxo reposted

This point from @emilymbender.bsky.social surfaces something that’s been nagging at me: the “stochastic parrot” phrase seems to mostly be used by people trying to say it’s *not* an appropriate term. Why? Why does the term live rent-free in their minds?

Emily M. Bender@emilymbender.bsky.social

I haven't looked deeply at every case, but the overall gestalt is that this remark tends to come from people who are deeply committed to the idea/ideology of "AI" as something all-powerful that will either solve all of our problems or kill us all. >>

Boxo McFoxo reposted

Sorry to be the party pooper, but I’m really disappointed. In case the Tolkien analogy was too subtle: if you play OpenAI’s and Anthropic’s game, gather round and work on their ETH “papers” in any way, you’re a part of the problem. You’re just encouraging them. You’re selfish and naive. Do better.

Boxo McFoxo reposted

It's almost as if being a far right politician requires also being a sniveling loser crybaby.

Yashar Ali 🐘@yasharali.bsky.social

1. BREAKING: Flávio Bolsonaro is refusing to concede Brazil's presidential election and says he will seek to have the election annulled after losing to President Lula.

Boxo McFoxo reposted

Lula winning is great for the planet, great for Brazil, and great for people who aren't revanchist assholes everywhere in the world, but let's not forget that it's also going to make some of the worst people on Earth extremely sad and mad. We should enjoy that, as a little treat.

Boxo McFoxo reposted

Replying to @olivia.science

And to go back to Lance Nizami's point... Yeah... There's something that doesn't make sense if you just try to plop information theory unceremoniously onto the neuroscientific world. Full PDF available: philarchive.org/rec/NIZITI

The problem with this if somebody (not saying the authors or even the journalist, but many do) takes this as a serious model is that it violates some important aspects of engineered systems required for information theory. Perhaps addressible like in quote above, but not easy nor always.

Boxo McFoxo reposted

But, yeah, if what falls on our sensory surfaces is "information" that's a really strong claim — it's good for scientists to say what model, formalisms, theories they use, but it's slightly different to see it outside a direct quote as if "information" is a neutral word like "thing" or a phenomenon.

It's always interesting to see journalists reflect the information theory metaphor on topics like this where it slips into scientific model... > crows and primates are processing the same information

Google have put a classifier inside the Gemini 4 Argon API that injects some dynamic context right under the developer instructions, mostly as an anti-'jailbreak' thing, but it also fires for at least one perfectly innocent reason (if you ask it its model version).

View full thread

But they believe their own mythos about the model actually understanding roles, so they made it a second `system` turn. Meaning the KV cache gets cleared out twice, first when the classifier inserts something, and a second time, when it stops inserting it.

Pretty hostile product design considering cached tokens cost 95% less. Though an LLM trained to appear as if it is a thinking, understanding, conversing entity is already hostile and unethical product design anyway, of course!

Google have put a classifier inside the Gemini 4 Argon API that injects some dynamic context right under the developer instructions, mostly as an anti-'jailbreak' thing, but it also fires for at least one perfectly innocent reason (if you ask it its model version).

They have an injected turn role for dynamic context (called, appropriately, `context`) that goes at the bottom of the context window right above the latest user message, so that it doesn't clear out the entire KV cache.

Boxo McFoxo reposted

NEW AD: Have you seen these anti-Talarico attack ads? Elon Musk paid for nearly all of them. The world's first trillionaire knows if he drags the most corrupt politician in Texas across the finish line, Ken Paxton will owe him everything. Elon Musk: Texas is not for sale.

Someone has vibe coded a set of php2shell vulnerability check scripts with Claude, but their approach to Docker images is a linter that checks for bundling of the PHP Docker base image into a build. There's so many other ways that PHP can end up running inside a Docker image than this lmao

github.comphp2shell-check/docker-scan.sh at main · MakFly/php2shell-checkCheck and apply the php2shell fix (CVE-2026-109618, CVE-2026-109619) on hosts, SSH servers and Docker images - MakFly/php2shell-check
View full thread

This script would miss so many vulnerable images. A much better approach would be a forensic audit of all running containers themselves for any hint of the Zend Engine in any form.

Someone has vibe coded a set of php2shell vulnerability check scripts with Claude, but their approach to Docker images is a linter that checks for bundling of the PHP Docker base image into a build. There's so many other ways that PHP can end up running inside a Docker image than this lmao

github.comphp2shell-check/docker-scan.sh at main · MakFly/php2shell-checkCheck and apply the php2shell fix (CVE-2026-109618, CVE-2026-109619) on hosts, SSH servers and Docker images - MakFly/php2shell-check

And even worse, the parsing is fragile. It does also run the php command inside the Docker images to check the running version, but again, not all images that use php have the php CLI installed. Some don't even have the normal php packages installed, and have Zend compiled into a static binary.

METR has closed this request and archived the repository because "HCAST tasks are now deprecated". Okay @metr.org, but you still have the Time Horizon chart on your website, without enough metadata about those tasks provided for third parties to fully audit your claims.

github.comRelease more metadata about the private tasks · Issue #7 · METR/hcast-publicI understand the need to keep the full definition of the private tasks confidential in order to prevent benchmark contamination. However, it's impossible for independent researchers to fully assess...

I guess they're moving on from their manufactured exponential curve now that they can talk about "loss-of-control" events instead.

Boxo McFoxo reposted

you can use the CA bug to make your boring smart TV way more interesting