tintor

3 ms·

31.
by tintor 1mo ago
None of the videos on website say how much are they speed up (or not), and is robot teleoperated or running autonomously.
32.
by tintor 1mo ago
The humans at OpenAI assumed that secure sandboxes are secure against their models, without safety guardrails. The two big questions are: 1) Why did they resume training without rolling model back to state before the first sandbox compromis
33.
by tintor 1mo ago
It is from the Gates video in the article.
34.
by tintor 1mo ago
Is Switzerland exposed to similar risk, given the amount of glaciers?
35.
by tintor 1mo ago
I can visualize algorithms, graphs, 3d objects, 3d environments, but no colors.
36.
by tintor 1mo ago
AlphaZero works in simple and fully observable environments like board games. RLVR for LLMs works for coding and math domains which can be verified by code. AlphaZero, RLVR and other RL methods don't work for learning dextrous manipula
37.
by tintor 1mo ago
"... AI will avoid climate change." xAI's Colossus 1 and 2 data center mobile gas turbines beg to differ. AI data centers are driving electricity consumption in 100s of GW, and most of that energy is not clean renewable nor n
38.
by tintor 1mo ago
Adults want time limit they can set in the app themselves.
39.
by tintor 1mo ago
So nice of them to let kids lose their minds 2 hours per FB and 2 hours per IG each day.
40.
by tintor 1mo ago
HN is social media only in strictly technical sense of the term. HN isn't optimized daily for engagement like FB and IG are. There isn't a single image, or video on website. Ads (for YC companies) sometimes show up and are rare. C
41.
by tintor 1mo ago
People who don't use social media to catch up with distant friends and family just ... phone or video call. No addiction, no infinite scroll, no exhaustion, no casino-like mechanics, no endorphins, no political rage, no AI companions,
42.
by tintor 1mo ago
It is not sufficient to secure VM the agent has CLI permissions on. We must also secure GPU and CPU nodes on API side which generate LLM tokens.
43.
by tintor 2mo ago
They could etch the model architecture, without the weights into the chip. This way newly post-trained model can be loaded and served the same day.
44.
by tintor 2mo ago
second hand smoke
45.
by tintor 2mo ago
Not sharing my private git repos with Elon Musk owned company.
46.
by tintor 2mo ago
Because these new data centers were built near people's homes.
47.
by tintor 2mo ago
"your deposits are FDIC insured" claim was a lie, for anyone building on top of Synapse
48.
by tintor 2mo ago
This Sokoban solver works on tiny and simple Sokoban levels, lagging behind the several SOTA Sokoban solvers that are available. http://www.sokobano.de/wiki/index.php?title=Solver_Statistic...
49.
by tintor 2mo ago
Natural selection. Agents that survive will be resilient.
50.
by tintor 2mo ago
Maybe because frontier labs buy the same RL tasks from task producer companies.
51.
by tintor 2mo ago
They are not. They can't do dexterous manipulation by controlling a humanoid robot.
52.
by tintor 2mo ago
People who have nothing to hide get ambushed by police at gunpoint, because surveillance system misidentifies them.
53.
by tintor 2mo ago
PrimeIntelect is not on official ARC-AGI-3 leaderboard: https://arcprize.org/leaderboard
54.
by tintor 2mo ago
It is easier to raise >$100M than few million.
55.
by tintor 2mo ago
> The only reason we don't have $6 gas right now ... I see you are not in California.
56.
by tintor 2mo ago
I am amazed how in US closing elevator doors switch to opening if someone touches the door edge. This results in abusive behavior of people intentionally jamming their arms in the path of closing doors, and delaying everyone else inside the
57.
by tintor 2mo ago
How do you tell from VR if hallway is cramped?
58.
by tintor 2mo ago
What is the price for this upgrade?
59.
by tintor 3mo ago
> AI managed to escape using standard and well documented script kiddie methods > AI broke in using standard script kiddie methods. Go ahead and show us how easy it is to break into HuggingFace (and OpenAI) networks.
60.
by tintor 3mo ago
OpenAI's router chooses between models of different sizes, which are still trained on roughly the same data. Its purpose is to reduce infra cost for OpenAI for simpler queries. No need to pay for GPT 5.6 Sol inference for "Hello&q

More ›