Hacker News: abel_

New comment by abel_ in "Small models also found the vulnerabilities that Mythos found"

abel_ — Sat, 11 Apr 2026 18:53:42 +0000

This misses the broader ongoing trend. For a few million dollars, of course you can create a startup that builds tools it can use to more efficiently find code vulnerabilities. And of course you can do this with weaker models with scaffolds that incorporate lots of human understanding. The difference now is that you don't need an expensive team, nor a bunch of human heuristics, nor a million dollars. The requisite cost and skill are falling rapidly.

New comment by abel_ in "Imagen, a text-to-image diffusion model"

abel_ — Tue, 24 May 2022 10:01:02 +0000

On the contrary -- the opposite will happen. There's a decent body of research showing that just by training foundation models on their outputs, you amplify their capabilities.

Less common opinion: this is also how you end up with models that understand the concept of themselves, which has high economic value.

Even less common opinion: that's really dangerous.

New comment by abel_ in "Gato – A Generalist Agent"

abel_ — Wed, 18 May 2022 00:28:40 +0000

This was done so that the IRL robot manipulation tasks could be done fast enough. In the future, we may always need small models mixed with large models for some tasks (e.g., for slow long term planning and fast short term planning), though compute does have a tendency to improve exponentially...

New comment by abel_ in "Gato – A Generalist Agent"

abel_ — Wed, 18 May 2022 00:23:19 +0000

There's a neat argument against these models doing interpolation: the manifold of the data is so sparse that it's infinitesimally unlikely for a good predictor to be doing interpolation between existing points on the manifold.

https://arxiv.org/abs/2110.09485

New comment by abel_ in "Bits of advice I wish I had known"

abel_ — Fri, 29 Apr 2022 02:19:25 +0000

I wonder what the net effect of such pieces of writing is. The problem is that these abstract and contextless statements make sense only if they cause the reader to reflect on some experience, and thus only mildly reinforce currently held beliefs. Otherwise, I can't see how the statements would stick for most people (not even as cached memory).

What would add significantly to this is a bunch of Gwern-style links embedded within each of these quips. The author is clearly speaking from a vantage point not many others have attained, and he'd be able to provide a story or other context to each.

New comment by abel_ in "You’re muted – or are you? Videoconferencing apps may listen when mic is off"

abel_ — Thu, 14 Apr 2022 00:21:27 +0000

The problem with software-controlled permissions is that nation-state actors (who have unbounded resources) can snoop on your private matters with significantly greater ease.

At least with a hardware switch, someone would have to physically intercept the air waves in the room you're in. In software, the surface for OS-level vulnerabilities is massive, and state sponsored mass surveillance just gets easier.

Sadly, this is a trade-off we have made as a society for "ergonomics".

New comment by abel_ in "It’s Like GPT-3 but for Code – Fun, Fast, and Full of Flaws"

abel_ — Sun, 20 Mar 2022 02:29:17 +0000

While others here have touched on the idea that Codex has changed their coding habits, what I find interesting is that Codex has changed how I write code altogether. For example, I had to connect a database to an API a little while ago. Obviously I had the option to use an ORM as one would normally. But instead, I just wrote out all the SQL commands and wrapper functions in one big file. Since it was all tedious and predictable, Codex helped me to write it in just a few minutes, and I didn't need to muck around with some complex ORM. These are the tradeoffs I'm personally excited about.

New comment by abel_ in "It’s Like GPT-3 but for Code – Fun, Fast, and Full of Flaws"

abel_ — Sun, 20 Mar 2022 02:16:30 +0000

As someone else already mentioned, the scaling laws paint a different story empirically: we haven't hit diminishing returns at all, and there's no end in sight.

But more anecdotally, the first applied neural network paper in 1989 by LeCunn has pretty much the same format as the GPT paper: a large neural network trained on a large dataset (all relative to the era). https://karpathy.github.io/2022/03/14/lecun1989/

It really just seems that there are a certain number of flops you need before certain capabilities can emerge.

New comment by abel_ in "A DARPA Perspective on Artificial Intelligence [pdf]"

abel_ — Sun, 20 Mar 2022 00:30:54 +0000

The most pressing dangers of AI most researchers see:

- error rate too high

- you can trick a classifier with noise

- it's racist sometimes

Actual dangers of AI:

- stop problem

- infeasibility of sandboxing

- difficulty of aligning black boxes with human values

New comment by abel_ in "Newly published 9/11 footage [video]"

abel_ — Sat, 26 Feb 2022 13:54:26 +0000

While progress has been made in computer vision, that progress has been relatively narrow up until now, and I think the activation energy required to produce this level of quality would be more than it's worth. As others have mentioned, new footage comes out all the time.

However, I agree with the sentiment. Someday, we will have a massive foundation model capable of producing any video with a little conditioning on text. But we don't currently have such a model. In some sense, we're still in the era of easily verifiable video, and this era might end someday soon.

New comment by abel_ in "Get better sleep – Anecdata and sleep tech"

abel_ — Tue, 22 Feb 2022 06:48:35 +0000

Interesting! I hadn't heard this perspective before. Sounds like Goodharting [0] applies to sleep too

[0] https://en.wikipedia.org/wiki/Goodhart%27s_law

New comment by abel_ in "Don’t point out something wrong immediately"

abel_ — Fri, 18 Feb 2022 08:54:29 +0000

Completely disagree. I agree that the _delivery_ is important, but the interval between when discovering a problem and announcing it should be minimal.

There's a huge difference between maintaining the social etiquette of allowing your conversationalist to explain themselves fully, and waiting a little while before announcing a problem. Announcing right away in a respectful manner also let's you get a correction right away.

New comment by abel_ in "Show HN: Clone your voice and speak a foreign language"

abel_ — Tue, 04 Jan 2022 03:20:46 +0000

An interesting reflection is how quickly research around TTS/STT has progressed. I remember reading [0] thinking we were a long ways away. And things will get way better with multi-task learning and multi-modal learning in the coming years (or months really).

In fact, just a year after this post was written, CoquiAI started their open source projects [1].

[0] https://news.ycombinator.com/item?id=22869365 (https://thegradient.pub/towards-an-imagenet-moment-for-speec...)

[1] https://star-history.com/#coqui-ai/TTS&coqui-ai/STT

New comment by abel_ in "Who won the Amstel Gold Race? Human error in photo-finishes"

abel_ — Thu, 29 Apr 2021 06:09:32 +0000

I think the distortion of the video-finish camera should be considered when making these estimates. There's quite a lot of it happening in the shot. Narrowing down the specific camera intrinsic parameters may be a challenge. The EXIF data of the original footage might be fruitful.

It might also be possible to undistort using some assumptions on colinear points in the video (i.e., the finish line and signs should have straight lines).

New comment by abel_ in "Thanks for the Bonus, I Quit"

abel_ — Sun, 18 Apr 2021 19:58:06 +0000

While I see the value in gratitude as described by the author, the lesson learned still rings with a certain amount of complacence with the situation at hand. Had there been more transparency in the bonus program, there's a chance it would have been reformulated as to not fuel internal wars and allow for the long term success of the company. While it's important to always exercise gratitude, it's also important to demand good incentive structures.