@karpathy.bsky.social posted this on the other site (link in alt) and I wanted to talk about it for a second. There's no doubt that Nano Banana is impressive. But there's errors in its solutions. Worse, the errors are subtle. The final output is right, but steps aren't. Making it wrong!

Did Kohler really see Smart Pipe and go "What a great idea!" For $600, and a monthly payment of $7, you too can send Kohler pictures of your poop! www.kohlerhealth.com/dekoda/

I'm not sure I'll ever learn how to type correctly on an iPhone... The phrase I want to type is "shoot themselves in the foot" Apple please: 1) don't correct multiple words at once. It creates more work 2) censor words or phrases. You're not protecting anyone, you're creating more harm

Great! I always wanted to spy on my neighbors through the walls! Seriously, I can't see how the potential for abuse doesn't significantly outweigh the utility. Abuse not only by autocrats (turnkey tyranny), but also by hackers and abusive partners. Not cool... news.ycombinator.com/item?id=4442...

My favorite #Apple thing is: Opening a new window and it moving my workspace to that window despite my cursor being somewhere else. e.g. Logging into a website (literally Google), getting a new popup window for my security key, popup on wrong monitor and shifts focus away just so I can move back.

Got the #CVPR outstanding reviewer award! There's a lot of noise in the reviewing process, so write the review you wish you received. Finding flaws is easy, but review is about the progress and making the work the best it can be. Remember, we're all on the same team.

If you're at #CVPR, stop by Hall D from 10am-11am. I have two posters (166 & 167) I'm presenting with the eLVM Workshop. One work on mitigating the downsides of restricted attention mechanisms, creating efficient high performance GANs. The second on the distillation of Normalizing Flows.

You can now call me Doctor. You don't have to, but you /can/

People often talk about how theory and research generate little to no economic value. I'm confused by this What's the economic value of... calculus? Certainly an ABSURD number... I get it. We need to make things, but you can't walk or run without the ground you stand on. Don't take it for granted.

Let the record show, on May 27th, 2025 GPT-o3, a model for advanced reasoning, still cannot figure out the trivial boat crossing problem without some extra help. This came up because a discussion of how more parameters does not mean better generalization. It's the nuances that matter, that's what

I want this (NeurIPS?), and easy navigation from the pdf to the abstract. Google always shows the pdf but never the abstract, which has the bibtex. Make it easy, like arxiv changing /pdf/ to /abs/ I even keep a list of urls because it's easier than regoogling github.com/stevenwalton...

Rereading the excellent talyarkoni.org/blog/2018/10... and thinking that perhaps what has changed is the need to win. To beat the incentives you have to be willing to not get the grant, the paper acceptance, the tenure

These conversations always put people into two camps: - Those who love coding for coding's sake - Those who love the result and code is a means to an ends Everyone cares about the result! So, I don't buy it. What is the divergence then? news.ycombinator.com/item?id=4377...

Ezra doesn't tell the full story. It's worse. We are watching Turnkey Tyranny unfold in real time I try not to be political on social media but were witnessing blatant violation of US law, the Constitution being undermined, and violation of international law.

How do I know when my cat is fully charged?

In case anyone was wondering, yes, you can transcode 8 videos to AV1 WHILE playing Elden Ring at maximum settings on a NVIDIA 4080S. I would suggest keeping it at 7 though. Little extra wiggle room.

When someone says: "let's start from first principles" you can be confident the next thing they say isn't going to start from first principles. Why is this a thing? Why is it so common in CS?

Filling out applications to #NVIDIA and I found this in the skills section. I didn't know they had such an interest in camping! Rest assured NVIDIA, I have the skills to survive the robot apocalypse and live in the woods! I hope it doesn't come to that, but I'm ready!

Today's OPSEC lesson is this: all of the encryption in the world is not going to save you if you add the editor of The Atlantic to your war-planning Signal group chat. www.theatlantic.com/politics/arc...

I'm petty disappointed with a lot of the conversation on this Atlantic piece, especially the CNN coverage I saw. Stop making it about using @signal.org, show how dumb they are. Let me explain. www.theguardian.com/us-news/2025...

When you ask Siri: "What month is it?" This is a pretty big fail. IME a lot of Apple ML is not highly robust. Needs more augmentation. Hopefully this isn't because lack of data harvesting (I believe it can be solved without that) (I tried doing video but it cuts audio when talking to Siri)

The algorithms aren't alright How do you serve a leech who has fewer views over the official source? The problem is this seems to be the norm, not a one off. Anyways, some cool stuff with GTC. Watch here www.youtube.com/live/_waPvOw...

@andrewgwils.bsky.social, you used "we" on a single author paper. You know what you have to do... en.wikipedia.org/wiki/F._D._C... This is a great paper and you should read it even if there is no cat coauthor.

Every work has flaws Every work has reason to be rejected The purpose of review is to: - sanity check papers, ensuring claims are reasonably evidenced - works aren't plagiarized (i.e. "not novel"). We publish to communicate. We review to improve communication

Crows are extremely intelligent animals and can learn complex behaviors. Crows were the first to invent calculus, beating Newton by 200 years! Crows also like food

Love Signal. Recently got my mom on it as well.

There's something deeply confusing about modern tech and I need someone to explain Watch this video and turn on closed captions. 2 minutes will do, no need to watch the whole thing (the video is fun, but I'm not sharing it for any other reason than it is in German) www.youtube.com/watch?v=4VrL...

@thomwolf.bsky.social hits on one of my biggest gripes in academia. Being a good scientist REQUIRES one to be a bit anti-authoritarian. To really embrace the "verify" part of "trust but verify." You cannot shift paradigms by maintaining the current one. It just is incompatible.

There's a crossing in Eugene where on a daily basis I see people almost get hit. What happens is: - Person in one lane stops to let pedestrians cross (or stops because pedestrians are actively crossing) - Other car swerves around, slams on breaks in the middle of the intersection

I think this is representative of a problem in academia. Success is not at odds with wellbeing. It is required you struggle to grow. It is good to gain resilience. But that's different. In fact, humans LOVE to struggle and challenge themselves. That's how we got here in the first place!

If you're a fan of @3blue1brown.com videos, let me introduce you to jHan's channel. He doesn't publish videos as frequently but each of them are masterpieces. His latest video is on the history of solving Fermat's Last Theorem. www.youtube.com/watch?v=GY4y...

It's no secret that I have a lot of unpopular or dissenting opinions, but I'd argue that this is _not_ because I'm _not_ well read, rather because I am. There's a mistake I think many make when reading. Statement, claims, proofs are all read as a matter of fact. But I believe this is fundamentally

Is Big Brother criticizing out of legitimate concern or jealousy? I'll commend the action but I'm hesitant regarding the motivation given the UK's own fascist tendencies. An optimist pointing out a fire will still pick your pocket while you're distracted. www.theverge.com/news/617273/...

The actual answer can be found here (A6): maa.org/putnam/ I found the right answer in ~3 minutes. Am I becoming super human? I'm seriously concerned about the inability to do the basics: CHECK YOUR OUTPUTS Grok didn't solve this. Even with internet access. (Pic from @littmath.bsky.social)

Academia: Forged by those who go against the grain Built to change the world But structured with the belief that paradigm shifts occur by following the current paradigm How many dark horses, how many idols, do we need before we stop making the shade that made them dark in the first place?

@wikipedia.org could you please set the "automatic" option as the default option? I fail to see the utility of its existence if this is not set as the default. I believe this can be resolved by replacing 'skin-theme-clientpref-day' with 'skin-theme-clientpref-os' for the default

I'm often baffled at how often we create extra unnecessary work. My department (CS) just sent me a "survey" to "apply" for TA positions. Most of the questions are like "what courses have you previously TA'd for" and were filled out on the previous survey... last term. Or "what academic year are you

I think a lot of these math benchmarks are absurd. It seems people are chasing the score without thinking about what we're actually trying to do. Not getting 100% on 3 digit times a 3 digit but getting a 6 digit x 4 digit should make us question everything. Something fundamental is wrong.

Reading rebuttals: Reviewer (not me): work is incremental Authors: thanks for the complement I gotta hand it to them. Respect

The Bitter Lesson of DeepSeek is that scaling isn't all you need Of course people have been screaming this for years and it's demonstrable, but now it's clear that no one can win the AI race by scale alone I've seen too many works get rejected with "just tune a large model, why train from scratch"

I'm not sure if I'll ever get used to the #Apple ecosystem and I'm not sure what people mean by "it just works." e.g. Sitting in front of my macbook listening to music from my iPhone on my AirPods. Why am I getting "Steven's AirPod Pros connected" on my mac several times an hour?

We need more #vim posts. Here's a fun and common pattern I use when making classes. Grab the lines you want to define and use substitute. Just use regex capture groups to capture the leading whitespace and then our target word. Substitute is really underappreciated but extremely powerful.

Hey, do you make websites? - Please add "Dark Mode" - Make the default "automatic" not "light" The tyranny of light mode users MUST end. This is low hanging fruit with high utility

Question: For #CVPR what constitutes "highly irresponsible"? Would it be "highly irresponsible" for a reviewer to list under "Major Weaknesses": "symbols are not defined, e.g. ∑" I am having a hard time processing this. This can't be a serious complaint. My browser must be rendering it wrong...

Review process in a nutshell: Creates sliced bread Reviewer 1: Why would anyone want this, I can slice my bread at home. Reject Reviewer 2: There's no novelty, I already slice my own bread. Reject Reviewer 3: This might be useful for some but I don't think it is substantial enough. Weak reject

Repost from Twitter from 2020, before ViTs: I trained a CNN w/o pooling on MNIST and produced a colored visualization of the feature maps, translating a digit on an input canvas. This illustrates the equivariance property of convolutions. More on why: chriswolfvision.medium.com/what-is-tran...

What is the point of having "advanced search" if you aren't going to follow the search parameters? I understand this search is overloaded but that's explicitly what negations are for! This kind of problem is quite prevalent and honestly, at times I feel like I don't even know how to search.

My favorite feature of BlueSky? The feed doesn't update every time I leave the app for 10 seconds. So many times I want you read a post but can't at the moment or there's an interruption. On Twitter I just was constantly sharing those with my browser...

I'm worried an AI Winter has begun. Winters happen when you make promises you can keep: "OVER-hype". It happens when we put all our eggs in one basket. While LLMs were a great leap forward, to the general public they have already become lackluster. We've seen too www.theverge.com/2025/1/6/243...

We get too caught up in what is the "right" research to do. History has shown us time and time again that big breakthroughs frequently come from unexpected or even rejected ideas! I think it's probably better we stop worrying about what's "best" and pursue what is the most interesting. We don't