🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

My brother’s family is touring London and I’ve been turning their vacation photos into AI clips. The kids love it.

anyone from the Sora team happen to be following me? 😬

The only true AI innovation in the "thinking" space has been LLMs. All other embodiments, modifications, extra training runs, bolt ons, tools, CoT etc are products and optimisations, not the science of intelligence. Behind all the smoke and mirrors and "thinking", there's just

we don't talk enough about external memetics vs internal memetics when it comes to LLMs, but this becomes more important when models take on roles due to external input (esp with reasoning models that have web access) vs something that comes from the model itself

you're not a real founder if people don't think you're actually just unemployed

i used to think multitasking is hard, but what's really hard is switching between a mindset that's zoomed-in and a mindset that's zoomed-out

🚨The UK AISI identified four methodological flaws in AI "scheming" studies (deceptive alignment) conducted by Anthropic, MTER, Apollo Research, and others: "We call researchers studying AI 'scheming' to minimise their reliance on anecdotes, design research with appropriate

I've been yapping for months about bad evaluation setups and how results/AI behaviors are reported, and this new AI Security Institute paper does so much more clearly. In short: There's a massive difference between showing a model can do something sketchy versus showing it tends to

seeing ai writing everywhere is turning me slightly nuts. right now it always pulls me out of the reading experience w uncanny valley effect I know it’s not going anywhere so I just hope the models get better It’s not the emdashes that do it for me it the verbal patterns.

fixing AI slop turns out to be as painful as fixing your own slop

SceneScript treats 3D reconstruction as a language problem rather than a geometry one. The model watches a video of a room and just learns to write a script for it. It autoregressively spits out text commands like make_wall(...) or make_bbox(...) that define the scene.

if researchers labeled their own data we would have AGI by now no one looks

is opus 3 down for anyone else in the app? seeing this error for both opus 3 and haiku 3.5 other models work fine

for a long time i’ve suspected that “long context” models don’t process every input token uniformly. thanks to kelly’s work in chroma’s latest tech report, today we have proof. input length, distractors, and many other factors influence performance on many simple tasks.

spot on those who could use AI to do real harm are actively choosing not to, and are instead alerting labs to issues through responsible disclosure paradigms vast majority of "ai safety" research is getting the model to say bad stuff and freaking out when it follows directions

"Neon" vec3 c,p,v;for(float i,z,d;i++<5e1;o.rgb+=-c*d*d/z/z+(cos(p.y+vec3(6,1,2))+1.1)/(length(tan(p.y/.3)/cos(p.xz/.1))+d*d/.01)/z)c=normalize(FC.rgb*2.-r.xxy),p=z*c,p.xz-=t,z+=d=.4*max(dot(cos(v=p-sin(p).xxz).xz,sin(v.zx/.6))+.6,v.y+3.);o=tanh(.1*o);

we're hurtling through the void on top of a ball of lava. the lava has a very thin crust and a tiny bit of ice. the ice is being liquidified by a ball of plasma 94 million miles away as we spin around it at 15 miles a second. it's all so weird but you gotta feel lucky somehow

every time i click on this it's either "Multifaceted Therapeutic Realignment: A trauma-informed framework for evaluating attachment styles in LLMs" or "We've pleased to announce we're partnering with the Pentagon to put Claude on adrenochrome-seeking kamikaze drones"