Keep on to blur preview images; turn off to show them clearly
Mind. Blown. I shared BLITZ Hardcore History podcast on slavery with Kimi and it produced this top-notch presentation. Dig the quality. What's truly next level is that the text is EDITABLE! Black magic.
CPO at https://t.co/BNZzlkTfVp. Founder of https://t.co/hOAmca8qLm and https://t.co/dRwgbZCSOw. Coffee-making, parenting, building, exploring: RU → CN → NZ → CL → UK → NZ → PL → UK?
RT @kadmitriev: NEW 🚨 Zelensky’s closest ally Andriy Yermak — head of the Ukrainian presidential office and Ukraine’s #2 — resigns after an…
Root node of the web of threads: https://t.co/ifH80GcLpo
we haven’t scratched the surface on how much the shoggoth can be behaviorally altered even with these post-trained agentic models lots of ppl believe this but even if we stopped model training today, there’s so much value to be extracted from just designing harnesses for valuable Tasks things like prompts and skills are levers that pull the model in behavioral directions, you can sell harnesses and prompts; the models are rlly smart just takes a lot of care to make them good and reliable evals and dogfooding are our friends 🫡
building agents and harnesses, prev @awscloud, phd cs @ temple
RT @SpencerGuard: 100% agree. Watch the video. Israel’s advancements in technology, tactics, doctrine, innovations displayed and validated…
Subscribe https://t.co/Xm1OaUU8jk • seed investing • writing • ॐ •🙏• I use '—' • tweets saved 90 days • 📷
a bit late to the party but i read deepseek math v2 paper. this is the first and only work replicating the success that was seen in closed source i.e. IMO Gold using large language models. there aren't many new things here. if i was to say what stands out it would be: > simply the idea that just rewarding answers isn't a guarantee of correct reasoning process > so we train a verifier over expert graded data from 17.5k > we train a meta verifier to further ensure verifier isn't grading incorrectly it is in fact as simple as it sounds and a part of me was expecting open source IMO gold model to use something fancier. i had some pet theories i wanted to see validated or trashed but i suppose when we move to generalist models with such capabilities, we would get to see that. until then it is a great north star with a fairly simple but compute-heavy and leaning on data curation approach that gets us something which was considered incredibly difficult for models until last year.
making models learn • eXperiments lab • memes and training lores
RT @geoffreywoo: No one gives a fuck about Asian representation. Don’t be a victim @SimuLiu. Produce your own movies, make your own money,…
🏗️ Love to build stuff (@runwayco, @sandboxvr, @postmates, @zynga) people love. 💸 Investor @amplitude_hq, @mercury, @owner, @elevenlabsio, @meetgamma ++