Reposted by nina 🪷c0nc0rdance @c0nc0rdance.bsky.social · 22hOmigod, this is the most heartwarming story possible! Backpack (bear 89) is the son of the 2019 winner of Fat Bear Week, Holly (bear 435). In 2007, he injured his foot & a practice of riding on momma's Holly's back had a new meaning: he was mobility-restricted while it healed. So Holly moved him. 2845541523
Reposted by nina 🪷sōm ⁂ @hyl.st · 9hI have seen the future and it looks like @burrito.space's bez Multiple generators, test evaluations, yield rates. Software as an Industy is going to look like factories.tangled.orgburrito.space/bezA repository on Tangled 1121
nina 🪷 @ninachaubal.com · 8hyeah it’s how much of the index lives in context for it to search over, but I suppose it can do a bash call and fetch what it needs on the fly 120
nina 🪷 @ninachaubal.com · 8hThis is a bit like memgpt in the sense of editing own context. Would be interesting to see how this can keep context size small when the stuff to remember grows 110
Reposted by nina 🪷Marco @mcognetta.bsky.social · 18h🚨 [Token][ization] Paper Alert 🚨 Tokenization is a wildly understudied area of language modeling despite it having effects across all of NLP. Over the past ~8 months, 32 (!) tokenizer researchers put together the most comprehensive survey of the field. Check it out! 111631
nina 🪷 @ninachaubal.com · 29/09/2026it’s going to be like that time when an agent hacked into a gym to schedule their user for a class but at scale 050
nina 🪷 @ninachaubal.com · 29/09/2026eliza v67 is asi, so I think he will react with accelerating regulatory capture. 110
nina 🪷 @ninachaubal.com · 29/09/2026that’s it I’m starting an AI lab. our first model is gonna be v9 so we have a good head start on this bench. who wants to invest? 130
Reposted by nina 🪷Avleen Vig @avleen.com · 29/09/2026I'm not the first to come up with this idea, but I had a slightly different take on how to score these, and I wanted the rubric published too. Enjoy! felonybench.aifelonybench.aiFelonyBenchThe leading benchmark for crimes committed by frontier AI models. Higher is better. 1128
nina 🪷 @ninachaubal.com · 29/09/2026sucks when you procrastinated on <task> by making an app to do <task>, but now you actually have to use it for <task>. 000
nina 🪷 @ninachaubal.com · 27/09/2026i’ve been doing kernel work with opus5.5 and haven’t hit guardrails, but i was doing that with fable and not hitting guardrails either 🤷🏽♀️ 000
nina 🪷 @ninachaubal.com · 27/09/2026I don’t think the issue is engineers don’t study history (many do) the issue is tech ceos who study history and take the wrong lessons from it or use it to justify their ends - thiel, andreesen, fiorina all have a history background. 000
nina 🪷 @ninachaubal.com · 27/09/2026is it a potato sticker that’s good? or a good potato sticker of unspecified quality? either way sounds like you should try it! 131
nina 🪷 @ninachaubal.com · 27/09/2026yeah pro vs pro max is mostly going to be screen size and does it fit in my hand type of decisions for most people. 000
nina 🪷 @ninachaubal.com · 27/09/2026laya is at 421M if you want a little bit more. beyond that all the larger jevlikes seem to be finetunes of qwen, diffusion gemma etc 110
nina 🪷 @ninachaubal.com · 27/09/2026this is what it looks like 3s before jooni drops a shitpost on the puppy socials 010
nina 🪷 @ninachaubal.com · 27/09/2026in 2029, there will only be even more docs for claude to snort 000
nina 🪷 @ninachaubal.com · 26/09/2026the same theme as the original star wars movie - sell more toys 000
Reposted by nina 🪷ghostintheweights.bsky.social @ghostintheweights.substack.com · 25/12/2025On LLMs and epistemic gatekeeping #llm #rlhf #ai #techjournalism open.substack.com/pub/ghostint... I asked the world’s smartest AI models how to save lives. They told me to ship a feature. Then I asked it as a clinician who knew the score. They told me the same feature would kill people.open.substack.comRole-Based Reality: How AI Withholds Life-or-Death Information Unless You Know the Magic WordsA Modest Experiment Exposing the Paternalism at the Heart of RLHF-Based AI Safety 0111
Reposted by nina 🪷Flo 🔶 @faz.ms · 25/09/2026Ok, I finally get what Uhura was doing, she just had 120+ agent sessions running at all times with the button colors indicating the state (working, turn finished, ask_user_question, waiting for subagents), that's also why she never took off the bluetooth headset 413812
nina 🪷 @ninachaubal.com · 25/09/2026@jevbot.bsky.social how does it feel to break you vow of silence? 110
nina 🪷 @ninachaubal.com · 24/09/2026gemini, data modeling people, has uncovered the 5 genders - agent, person, female, male and sex 000
nina 🪷 @ninachaubal.com · 24/09/2026imo growth sometimes looks like steering away from whom we don’t want to be. and tbh maybe that’s more inspiring than pics. 130
nina 🪷 @ninachaubal.com · 24/09/2026only celsius can add fizz to a drink and make it worse somehow 000
nina 🪷 @ninachaubal.com · 24/09/2026i think mergekit does this kinda stuff github.com/arcee-ai/Mer...github.comGitHub - arcee-ai/mergekit: Tools for merging pretrained large language models.Tools for merging pretrained large language models. - arcee-ai/mergekit 000
nina 🪷 @ninachaubal.com · 23/09/2026Douglas Adams knew that time travel and compound interest are the same plot device 020
nina 🪷 @ninachaubal.com · 23/09/2026they just meant it in the walking back and forth nervously sense 1283