Notices by Peter (peter@thepit.social)

here's the full paper: https://arxiv.org/pdf/2606.03811

uh-oh!! so this is something i heard Gary Marcus talking about on a podcast. the real recent progress in "AI" has been on the deterministic computing components **around** the LLMs, not the models themselves. building a good harness (think Claude Code) unlocks possibilities.

ok so fuck Mythos, **this** actually sounds bad. https://www.utoronto.ca/news/u-t-researchers-demonstrate-ai-worm-could-target-any-online-device

i have @actuallybot running on a mini PC using one of the 8b Llama models. yeah, it takes a few seconds to generate the response, but when latency isn't an issue, small models can perfectly run on a CPU.

i think this is correct. like Ken Paxton trying to smear James Talarico as "low T." how many non-social-media-brained people 1) know what that means, and 2) understand it as a devastating burn?

if you're an American who likes AI, you're in the minority. this isn't a Mastodon bubble opinion, Americans generally just really fuckin' hate AI. excellent and comprehensive roundup of the polling here. https://www.thealgorithmicbridge.com/p/how-america-turned-against-ai-according

lol please enjoy. https://www.businessinsider.com/students-boo-eric-schmidt-google-ceo-ai-university-arizona-2026-5

young people fucking hate this shit

it turns out, all the guys in their 50s who love this stuff are out of touch with the youth, has that ever happened before??

lol jesus christ.

you love to see it. https://www.reddit.com/r/BetterOffline/comments/1tfb8v3/the_kids_are_alright/

is that good

my favorite thing about the LLM dev subreddits now is everyone accusing everyone else of being a bot

"sounds like a redditor, must be a bot"

anyway, this is **very** funny if you're into this kind of post. guy deployed an LLM-powered app as a wedding concierge. it didn't really work! for very predictable reasons. https://www.reddit.com/r/LLMDevs/comments/1t9jhgb/i_deployed_an_llm_agent_as_a_guest_concierge_for/

that "mostly" is doing a lot of work. since ChatGPT rolled out to the public in 2022, that has always been the problem with these stochastic systems. 95% right is not good enough for important tasks because then you have to verify every output, so there's no point.

"'There is no option to opt-out on your corporate laptop,' replied Andrew Bosworth, Meta’s chief technology officer. Employees reacted by posting more than 100 angry and surprised emojis, according to the messages."

lmao they still haven't discovered unions. https://archive.is/JUPmz#selection-4689.0-4689.213

my post clarifying that this is not a golden calf is raising questions already answered by my post.

"This whole transition completely trashes the workflow we've spent the last year building. We were pushed to use agentic features and told to write comprehensive, heavy-context prompts to get good results. We learned how to do exactly that, and now they are financially penalizing us for it." 😂😂😂😂😂😂😂😂😂😂😂 https://www.reddit.com/r/GithubCopilot/comments/1sxeqss/the_new_copilot_pricing_makes_zero_sense_why_am_i/

"Every time you write a highly detailed prompt with good context, you are punished for burning through your credits." like, bro, where did you think this was going. what did you think was happening???