@NeverSayClever @BenShindel @SkyeSharkie You are not eligible for MAID in canada if you have only psychological conditions, it *requires by law* a physical condition.
English
Jan
771 posts
Jan
@AnInsanityCheck
CS-PhD. Inquiring non-expert. The world is too complicated to be certain.
Tham gia Haziran 2019
306 Đang theo dõi53 Người theo dõi
@BenShindel @SkyeSharkie I don't think this is within the bounds of Noah's statement, although I guess you could read it that way. He's not saying that someone who would otherwise be a euthanasia candidate is invalidated by depression; he's saying not to treat depression with death. Right?
English
@FakePsyho True. 100% is almost impossible, and basically unachievable in Kaggle compute limits. I'd argue the previous grand prizes were also unachievable, which makes marketing with the prize pool a bit dishonest. Still a nice benchmark though.
English
At this point, I've played through ALL of the games there. I might do a longer write-up describing all of the flaws. Sadly, this would probably get lost in a flood of "AI is 100x worse than humans!!!111" posts from clueless influencers.
While it's clear this whole thing is designed in a way that the models will artificially get as low score as possible, I also suspect that game designers / tester team were just not experienced enough to design something fair. I'd expect that the private test games have an order of magnitude more terrible ideas (considering they are expected to be harder).
English
AI (or any human) will never get 100% in ARC-AGI-3
Let me introduce you to the worst game mechanic you can find in a puzzle game: fog of war
At the start, if you go right instead of bottom, you're wasting many moves. Your score on this level literally depends on a conflip!
English
@FakePsyho Except the humans get to retry after they figure it out x.com/i/status/20373…
Rishi Mehta@rishicomplex
@fchollet according to your paper: "Participants were limited to a single attempt per environment and could not revisit previously completed levels. However, they were allowed to reset the current level at any time. In some cases, participants reset levels after reaching a solution in order to improve efficiency, though this typically increased total interaction time." So humans could play around with the task a bunch, and then just reset the game when they figured it out to get the optimal trajectory? Is AI allowed to do this?
English
ARC-AGI-3 is an important benchmark. However, I have a major issue with the “Human score 100%” statement. How many humans have tested all 1000 puzzles? How were people selected? This was not published for previous ARCs either. In one case, the human score was based on I think 2 people.
This is really an unscientific way, as it assumes all humans are the same or that previous exposure to puzzles or video games, for example, is not considered. What education level and background did these humans have? I am sure humans will still score highly, but it would be very surprising if this was 100%.
Without this data and scientific measurement, this appears a biased test that assumes solving 100% of the puzzles is purely intrinsic intelligence common to all humans.
ARC Prize@arcprize
Announcing ARC-AGI-3 The only unsaturated agentic intelligence benchmark in the world Humans score 100%, AI <1% This human-AI gap demonstrates we do not yet have AGI Most benchmarks test what models already know, ARC-AGI-3 tests how they learn
English
@loveofdoing I am 99% certain your code has a bug. For example, in sampling mode you do not compare proposed solution to correct one. Always check predictions and always check counterexamples. This is just AI slop, and everyone falling for it needs better epistemic heuristics.
English
316 ARC-AGI tasks solved with zero learning. No neural net, no training, no DSL — just 19th-century projective geometry.
Encode grid cell relationships as Plücker lines in P³, find transversals via Schubert calculus, score candidates by geometric incidence.
95% solve rate on the eval set (of non-timeout tasks).
Single C file, runs in seconds.
Beff (e/acc)@beffjezos
The masculine urge to try to hack a new solution to ARC-AGI benchmarks
English
I have lived in Europe for more than 10 years and I don’t kid you when I say all the people I have met on welfare were citizens. Not even once have I met an immigrant on welfare. Not to say they don’t exist, they do. But I have never met one and my circle is full of immigrants.
@levelsio@levelsio
Current state of Western Europe
English
Jan đã retweet
@max_spero_ @pangramlabs Depending on model cost and storage cost you might consider briefly caching results for specific websites (i.e. social media). Probably makes no sense rechecking viral tweets thousands of times.
English
Roadmap for @pangramlabs Chrome Extension
1) Train pangram-small
2) Proactive AI checks on every surface of the internet. AdBlock for AI-generated content
@jason@Jason
X Product request: a toggle to turn off all AI generated content. Thank you for your attention to this matter!
English
@pangramlabs @tenobrus @iruletheworldmo Ohno, the AI checker is showing signs of malicious misalignment!
English
@tenobrus @iruletheworldmo We believe that this document is fully human-written
Disclaimer: For text under 75 words, results may be less accurate.
pangram.com/history/5b745a…
Pangram Labs@pangramlabs
@tenobrus @iruletheworldmo We are confident that this document is primarily human-written, with some AI-generated content detected pangram.com/history/4cecf1…
English
@goblinodds That's nice and all, but excuse me for being skeptical when kristi noem lies publicly about the case and then doubles down later on: youtu.be/T-UL_rEH1WQ
YouTube
English
as for people assuming the US officials lied, i mean, possibly, but i went to a temporary chat w chatgpt to sanity check myself
English
i sure hope everyone who was insisting that Good's car completely missed the officer right up until this news broke is updating their beliefs rn
CBS News@CBSNews
BREAKING: The ICE agent who fatally shot Renee Good on Jan. 7 in Minneapolis, Jonathan Ross, suffered internal bleeding to the torso following the incident, according to two U.S. officials briefed on his medical condition.
English
@arithmoquine Skill issue. Simply always think that you should follow your souls call rn.
English
3 years later and everyone i know who's tried this to any extent has discovered entirely new ways of being fucked up
collapse the bountiful multiplicity of the human soul with one easy trick
shaurya@shauseth
you have access to your optimal policy at all times. you just choose not to follow it. you can literally access it by asking “what should i be doing rn?”
English
there is absolutely zero chance of this
Guillermo Rauch@rauchg
There's definitely a non-zero chance we get "generate your own GTA6 in a few minutes" before GTA6
English
@celestepoasts The NoPE architecture also works, it just doesn't do positional encoding at all. IIrc it does exactly what you think: Uses causal mask to infert position implicitly.
Also means that these approaches will not work on e.g. Masked diffusion LLMs or diffusion models.
English
oh ok wait yeah causal masking already kind of does this like atleast the permutation symmetry stuff and absolute number of tokens if im understanding this right
still cursed as fuck this works
English
ml is fundamentally unserious man what do you mean this works
hardmaru@hardmaru
One of my favorite findings: Positional embeddings are just training wheels. They help convergence but hurt long-context generalization. We found that if you simply delete them after pretraining and recalibrate for < 1% of the original budget, you unlock massive context windows.
English
well a bit on W_o matrix and why it's so critical for transformers? btw, this was in my drafts from 3 weeks back!
well without the W_o projection, multi-head attention would be basically useless!
here's the intuition behind it:
now, so in multi-head attention, we split our embedding into multiple heads right where
head_dim = embedding_dim / num_heads and each head does its own separate attention calculation:
head₁ → output₁
head₂ → output₂
...
head_n → output_n
and to think about it, if we just concatenated these outputs together, the heads would never share information!
think about it - if head₁ learns syntax and head₂ learns entities, they'd never combine that knowledge, so concatenated heads → no information sharing → model can't combine different patterns!
it's like there are separate juices sitting next to each other in the same glass but not actually mixing and the concatenation just puts them in the same container
but W_o is like the blender that combines all these different "flavor patterns" (what each head learned) into a rich mixture where they can interact and enhance each other!
now as we want the model to blend these specialized representations, we need that extra step:
output = W_o × concat(output₁, output₂, ..., output_n)
when we apply W_o (which is a d_model × d_model matrix), it literally mixes the outputs from all heads! it's like taking different juices and blending them together!
and that's where the magic kicks in:
without W_o: heads are independent silos
with W_o: each output position gets info from every head
so now the model can actually combine syntax understanding with this entity recognition or maybe any kind of logical relations!
it's like saying for transformer models, we actually want all these different "attention head juices" to blend into one rich understanding, i mean without this matrix, it's just multi head is useless!
English
@Agrippa84837 @OriginalGoalie @JimSTruthBTold @SocDoneLeft You claimed that the shooter was not a white nationalist. His manifesto shows otherwise. You ignore it. Why would they continue arguing as if you were a good faith interlocutor?
English
@OriginalGoalie @AnInsanityCheck @JimSTruthBTold @SocDoneLeft Now you're threatening me with violence?
I can't say I'm surprised. That always happens when you leftists run out of arguments.
English
Right-wing perpetrators committed 93% of political killings since 2000.
Even if we include anti-police terrorist killings by non-leftists, "the left" did just 6%.
Exclude those? Just 5 killings, 1.4%.
The US right is far, far more violent than the US left.
Skscartoon@skscartoon
The modern left has firmly established themselves as the ideology of violence and terrorism. They don't believe in conversation, reason, or civility; and unfortunately, I don't think they can be dealt with such.
English
@Agrippa84837 @OriginalGoalie @JimSTruthBTold @SocDoneLeft You seriously want to tell me the guy who wrote this is left wing? Read the first page hoplofobia.info/wp-content/upl…
Literally the first sentence: "If there’s one thing I want you to get from these writings, it’s that White birth rates must
change."
Absolute joke. Brainworms
English
@OriginalGoalie @JimSTruthBTold @SocDoneLeft I don't know the other three. But the Buffalo shooter wrote in his manifesto that he identifies as a communist. The Club Q shooter identifies as non-binary.
Typical white nationalists.🤡
English
@littmath @davidbessis @dean_weez Wait, I don't understand. Afaict, the expected ratio should be 1 (same number of boys and girls in expectation), but the expectation of the ratio is bigger than 1:1, is that what this is about?
English
No. It depends on the formalism. If the population consists of N families who have kids until they have a girl and then stop, the expected ratio will depend on N but be greater than 1/2. It will tend to 1/2 as N-> \infty.
If you model things differently (e.g. kids have more kids) you get a different answer.
English
Call me dumb, but I don't even understand how this could be a problem about conditional probabilities. If the chances of having a girl are 50%, and if all draws are independent, then—come on!—what else could happen than 50/50 long term trend?
Am I this bad at probabilities?
Lyman Stone 石來民 🦬🦬🦬@lymanstoneky
reasonably confident this is a yale demog prof publicly getting a fairly common quant demo problem wrong very publicly i could be wrong, but the probability nerds in the comments and the people like me doing simulations all seem to agree that it's 50/50
English
@lymanstoneky It's just an interpretation difference. The number of boys and girls will be the same in expectation, but the expected ratio can behave weirdly. I.e. 1/1, 1/2, 1/3 and so on. Calculating the expectation over these ratios gives a different result (as it is a different question)
English
reasonably confident this is a yale demog prof publicly getting a fairly common quant demo problem wrong very publicly
i could be wrong, but the probability nerds in the comments and the people like me doing simulations all seem to agree that it's 50/50
Emma Zang@DrEmmaZang
Interesting problem. It looks obvious until you actually think about it. Many people gave the wrong answers because they assumed: E[G/(G+B)] = E[G] / (E[G] + E[B]), which is not true. Ratios are sneaky. The denominator moves, and suddenly all our nice intuition falls apart. The quick intuition: once family sizes vary, each family contributes differently to the population ratio, so you can’t just take expectations on the top and bottom and call it a day. And if you try to “just simulate it,” the result depends on how many families you draw. With small N you might get something that looks close to 0.5, but it’s always a bit above 0.5. Simulation will happily mislead you if you let it. If I ever teach PhD formal demography, I’m definitely putting this in the problem set. It’s too good not to. 😄
English
@willdepue This is apparently calculated including benefits, which is why more children increases marginal tax rate, as people lose these benefits at some income level.
Dishonest/Misleading way of displaying it.
English
