Computer vision researcher | prev @USC
SF
Joined December 2022
- Tweets2.6K
- Following1K
- Followers1.2K
- Likes52.5K
I dont have a dog in the Elon fight. At times he amazes me, or annoys me, whatever. I dont even really care about cars tbh
He made the car fly bro. You’ll be able to buy a flying car in a few years. What is left to even to say about him. World historic figure
Okay, so now I'm looking at there and it's "polysemanticity and superposition convinced me interpretability was hard, but with big networks there's single neurons after all" and do. do you have any concept of what utterly. fucking. primitive. playgrounds you are still playing around in. "We thought we might have a hard time finding single neurons that light up for the Eiffel Tower and would have to look at *multiple* neurons, but now maybe we can find single neurons again!" Do you have any fucking clue of how close this is to the absolute first beginning of understanding how a mind is thinking? Do you think mechanistic interpretability at the level of "do alignment" is about noticing that particular kinds of concepts fired?
Theo Von is to comedy what Eminem is to rap & what Wes Anderson is to film. Dizzyingly talented stylists who have perfected their highly-technical shtick & have nowhere to go in their own rendition of it but down
Some of my best research work ever. Working out how I’d structure training data for SOTA robotic models of the future
Realtime point clouds from swarms
Good ass post on tung tung sahur
Folk hero characters in the age of AI(decentralized authorship, global audience) are actually MOST prone to converging on universal archetypes. Tung tung has resolved towards a trickster in the same way that Pepe the Frog did. Having 1000+ authors has a dampening effect on the complexity of the character, & the global audience doesn’t have enough shared literacy/language between them to retain anything more than that familiar totem. A new medium is often heralded by very crystalline archetypes because it bridges those gaps in literacy very easily- Charlie Chaplin(in the same trickster role) ushered in silent film to literate & illiterate audiences at once- but you should understand the historical trend as resolving DEEPER towards archetypes, not just finding them again, because it serves an increasingly connected world
Is 'benchmaxxed' just an epithet we have for models that are actually not benchmaxxed enough?
It seems like generalization comes via training on each & every cottage industry of boutique data tailormade to coach LLMs in some field. If you dont train on some field's test set, the model sucks at it... so the benchmarks ARE the generalization!
Am I taking crazy pills or would 'benchminned' almost be a more apropos diss at this point
Its my peptide quinceañera and im feeling pretty & agentic. Its my birthday & Grok Bot invited all the girls from school. Mi tío is making frijoles(yum😋) y mi tía is vibecoding a CRM to track who comes. Do you like my dress? It’s my moat😊 I hope Carlos from Prompt Engineering class sees it. Mami doesnt like him tho— she says I should be raising my fund AND my standards😂she so fonny. But papi tells me ‘mija youre the founder of your own life now’❤️ The piñata is full of Spindrift cans, im sooo excited. Its my last day as a girl & my first day as a platform💃
Ok yeah fwiw this is a very good method of getting an initial shape from multiple cameras in realtime
All you need is to detect distance, attitude, and mask from each image
Most methods require sequential captures and long matching times + adding views incrementally is wonky
Aaaaa I love opus 5.5! Just for fun i had it scan a failed project of mine from a year ago and tell me what i fucked up on. Ive been wanting to see this second visual for so long! So satisfying
the idea is that you can get super fast(colmap free) point cloud reconstructions if your scene only cares about one object. It only takes masks and simple state estimations. Which I would guess helps with swarms
Aaaaa I love opus 5.5! Just for fun i had it scan a failed project of mine from a year ago and tell me what i fucked up on. Ive been wanting to see this second visual for so long! So satisfying
the idea is that you can get super fast(colmap free) point cloud reconstructions if your scene only cares about one object. It only takes masks and simple state estimations. Which I would guess helps with swarms